linux/fs/io_uring.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Shared application/kernel submission and completion ring pairs, for
   4 * supporting fast/efficient IO.
   5 *
   6 * A note on the read/write ordering memory barriers that are matched between
   7 * the application and kernel side.
   8 *
   9 * After the application reads the CQ ring tail, it must use an
  10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
  11 * before writing the tail (using smp_load_acquire to read the tail will
  12 * do). It also needs a smp_mb() before updating CQ head (ordering the
  13 * entry load(s) with the head store), pairing with an implicit barrier
  14 * through a control-dependency in io_get_cqe (smp_store_release to
  15 * store head will do). Failure to do so could lead to reading invalid
  16 * CQ entries.
  17 *
  18 * Likewise, the application must use an appropriate smp_wmb() before
  19 * writing the SQ tail (ordering SQ entry stores with the tail store),
  20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
  21 * to store the tail will do). And it needs a barrier ordering the SQ
  22 * head load before writing new SQ entries (smp_load_acquire to read
  23 * head will do).
  24 *
  25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
  26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
  27 * updating the SQ tail; a full memory barrier smp_mb() is needed
  28 * between.
  29 *
  30 * Also see the examples in the liburing library:
  31 *
  32 *      git://git.kernel.dk/liburing
  33 *
  34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
  35 * from data shared between the kernel and application. This is done both
  36 * for ordering purposes, but also to ensure that once a value is loaded from
  37 * data that the application could potentially modify, it remains stable.
  38 *
  39 * Copyright (C) 2018-2019 Jens Axboe
  40 * Copyright (c) 2018-2019 Christoph Hellwig
  41 */
  42#include <linux/kernel.h>
  43#include <linux/init.h>
  44#include <linux/errno.h>
  45#include <linux/syscalls.h>
  46#include <linux/compat.h>
  47#include <net/compat.h>
  48#include <linux/refcount.h>
  49#include <linux/uio.h>
  50#include <linux/bits.h>
  51
  52#include <linux/sched/signal.h>
  53#include <linux/fs.h>
  54#include <linux/file.h>
  55#include <linux/fdtable.h>
  56#include <linux/mm.h>
  57#include <linux/mman.h>
  58#include <linux/percpu.h>
  59#include <linux/slab.h>
  60#include <linux/blkdev.h>
  61#include <linux/bvec.h>
  62#include <linux/net.h>
  63#include <net/sock.h>
  64#include <net/af_unix.h>
  65#include <net/scm.h>
  66#include <linux/anon_inodes.h>
  67#include <linux/sched/mm.h>
  68#include <linux/uaccess.h>
  69#include <linux/nospec.h>
  70#include <linux/sizes.h>
  71#include <linux/hugetlb.h>
  72#include <linux/highmem.h>
  73#include <linux/namei.h>
  74#include <linux/fsnotify.h>
  75#include <linux/fadvise.h>
  76#include <linux/eventpoll.h>
  77#include <linux/splice.h>
  78#include <linux/task_work.h>
  79#include <linux/pagemap.h>
  80#include <linux/io_uring.h>
  81#include <linux/tracehook.h>
  82
  83#define CREATE_TRACE_POINTS
  84#include <trace/events/io_uring.h>
  85
  86#include <uapi/linux/io_uring.h>
  87
  88#include "internal.h"
  89#include "io-wq.h"
  90
  91#define IORING_MAX_ENTRIES      32768
  92#define IORING_MAX_CQ_ENTRIES   (2 * IORING_MAX_ENTRIES)
  93#define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
  94
  95/* only define max */
  96#define IORING_MAX_FIXED_FILES  (1U << 15)
  97#define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
  98                                 IORING_REGISTER_LAST + IORING_OP_LAST)
  99
 100#define IO_RSRC_TAG_TABLE_SHIFT (PAGE_SHIFT - 3)
 101#define IO_RSRC_TAG_TABLE_MAX   (1U << IO_RSRC_TAG_TABLE_SHIFT)
 102#define IO_RSRC_TAG_TABLE_MASK  (IO_RSRC_TAG_TABLE_MAX - 1)
 103
 104#define IORING_MAX_REG_BUFFERS  (1U << 14)
 105
 106#define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
 107                                IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
 108                                IOSQE_BUFFER_SELECT)
 109#define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
 110                                REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS)
 111
 112#define IO_TCTX_REFS_CACHE_NR   (1U << 10)
 113
 114struct io_uring {
 115        u32 head ____cacheline_aligned_in_smp;
 116        u32 tail ____cacheline_aligned_in_smp;
 117};
 118
 119/*
 120 * This data is shared with the application through the mmap at offsets
 121 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
 122 *
 123 * The offsets to the member fields are published through struct
 124 * io_sqring_offsets when calling io_uring_setup.
 125 */
 126struct io_rings {
 127        /*
 128         * Head and tail offsets into the ring; the offsets need to be
 129         * masked to get valid indices.
 130         *
 131         * The kernel controls head of the sq ring and the tail of the cq ring,
 132         * and the application controls tail of the sq ring and the head of the
 133         * cq ring.
 134         */
 135        struct io_uring         sq, cq;
 136        /*
 137         * Bitmasks to apply to head and tail offsets (constant, equals
 138         * ring_entries - 1)
 139         */
 140        u32                     sq_ring_mask, cq_ring_mask;
 141        /* Ring sizes (constant, power of 2) */
 142        u32                     sq_ring_entries, cq_ring_entries;
 143        /*
 144         * Number of invalid entries dropped by the kernel due to
 145         * invalid index stored in array
 146         *
 147         * Written by the kernel, shouldn't be modified by the
 148         * application (i.e. get number of "new events" by comparing to
 149         * cached value).
 150         *
 151         * After a new SQ head value was read by the application this
 152         * counter includes all submissions that were dropped reaching
 153         * the new SQ head (and possibly more).
 154         */
 155        u32                     sq_dropped;
 156        /*
 157         * Runtime SQ flags
 158         *
 159         * Written by the kernel, shouldn't be modified by the
 160         * application.
 161         *
 162         * The application needs a full memory barrier before checking
 163         * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
 164         */
 165        u32                     sq_flags;
 166        /*
 167         * Runtime CQ flags
 168         *
 169         * Written by the application, shouldn't be modified by the
 170         * kernel.
 171         */
 172        u32                     cq_flags;
 173        /*
 174         * Number of completion events lost because the queue was full;
 175         * this should be avoided by the application by making sure
 176         * there are not more requests pending than there is space in
 177         * the completion queue.
 178         *
 179         * Written by the kernel, shouldn't be modified by the
 180         * application (i.e. get number of "new events" by comparing to
 181         * cached value).
 182         *
 183         * As completion events come in out of order this counter is not
 184         * ordered with any other data.
 185         */
 186        u32                     cq_overflow;
 187        /*
 188         * Ring buffer of completion events.
 189         *
 190         * The kernel writes completion events fresh every time they are
 191         * produced, so the application is allowed to modify pending
 192         * entries.
 193         */
 194        struct io_uring_cqe     cqes[] ____cacheline_aligned_in_smp;
 195};
 196
 197enum io_uring_cmd_flags {
 198        IO_URING_F_NONBLOCK             = 1,
 199        IO_URING_F_COMPLETE_DEFER       = 2,
 200};
 201
 202struct io_mapped_ubuf {
 203        u64             ubuf;
 204        u64             ubuf_end;
 205        unsigned int    nr_bvecs;
 206        unsigned long   acct_pages;
 207        struct bio_vec  bvec[];
 208};
 209
 210struct io_ring_ctx;
 211
 212struct io_overflow_cqe {
 213        struct io_uring_cqe cqe;
 214        struct list_head list;
 215};
 216
 217struct io_fixed_file {
 218        /* file * with additional FFS_* flags */
 219        unsigned long file_ptr;
 220};
 221
 222struct io_rsrc_put {
 223        struct list_head list;
 224        u64 tag;
 225        union {
 226                void *rsrc;
 227                struct file *file;
 228                struct io_mapped_ubuf *buf;
 229        };
 230};
 231
 232struct io_file_table {
 233        struct io_fixed_file *files;
 234};
 235
 236struct io_rsrc_node {
 237        struct percpu_ref               refs;
 238        struct list_head                node;
 239        struct list_head                rsrc_list;
 240        struct io_rsrc_data             *rsrc_data;
 241        struct llist_node               llist;
 242        bool                            done;
 243};
 244
 245typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
 246
 247struct io_rsrc_data {
 248        struct io_ring_ctx              *ctx;
 249
 250        u64                             **tags;
 251        unsigned int                    nr;
 252        rsrc_put_fn                     *do_put;
 253        atomic_t                        refs;
 254        struct completion               done;
 255        bool                            quiesce;
 256};
 257
 258struct io_buffer {
 259        struct list_head list;
 260        __u64 addr;
 261        __u32 len;
 262        __u16 bid;
 263};
 264
 265struct io_restriction {
 266        DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
 267        DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
 268        u8 sqe_flags_allowed;
 269        u8 sqe_flags_required;
 270        bool registered;
 271};
 272
 273enum {
 274        IO_SQ_THREAD_SHOULD_STOP = 0,
 275        IO_SQ_THREAD_SHOULD_PARK,
 276};
 277
 278struct io_sq_data {
 279        refcount_t              refs;
 280        atomic_t                park_pending;
 281        struct mutex            lock;
 282
 283        /* ctx's that are using this sqd */
 284        struct list_head        ctx_list;
 285
 286        struct task_struct      *thread;
 287        struct wait_queue_head  wait;
 288
 289        unsigned                sq_thread_idle;
 290        int                     sq_cpu;
 291        pid_t                   task_pid;
 292        pid_t                   task_tgid;
 293
 294        unsigned long           state;
 295        struct completion       exited;
 296};
 297
 298#define IO_COMPL_BATCH                  32
 299#define IO_REQ_CACHE_SIZE               32
 300#define IO_REQ_ALLOC_BATCH              8
 301
 302struct io_submit_link {
 303        struct io_kiocb         *head;
 304        struct io_kiocb         *last;
 305};
 306
 307struct io_submit_state {
 308        struct blk_plug         plug;
 309        struct io_submit_link   link;
 310
 311        /*
 312         * io_kiocb alloc cache
 313         */
 314        void                    *reqs[IO_REQ_CACHE_SIZE];
 315        unsigned int            free_reqs;
 316
 317        bool                    plug_started;
 318
 319        /*
 320         * Batch completion logic
 321         */
 322        struct io_kiocb         *compl_reqs[IO_COMPL_BATCH];
 323        unsigned int            compl_nr;
 324        /* inline/task_work completion list, under ->uring_lock */
 325        struct list_head        free_list;
 326
 327        unsigned int            ios_left;
 328};
 329
 330struct io_ring_ctx {
 331        /* const or read-mostly hot data */
 332        struct {
 333                struct percpu_ref       refs;
 334
 335                struct io_rings         *rings;
 336                unsigned int            flags;
 337                unsigned int            compat: 1;
 338                unsigned int            drain_next: 1;
 339                unsigned int            eventfd_async: 1;
 340                unsigned int            restricted: 1;
 341                unsigned int            off_timeout_used: 1;
 342                unsigned int            drain_active: 1;
 343        } ____cacheline_aligned_in_smp;
 344
 345        /* submission data */
 346        struct {
 347                struct mutex            uring_lock;
 348
 349                /*
 350                 * Ring buffer of indices into array of io_uring_sqe, which is
 351                 * mmapped by the application using the IORING_OFF_SQES offset.
 352                 *
 353                 * This indirection could e.g. be used to assign fixed
 354                 * io_uring_sqe entries to operations and only submit them to
 355                 * the queue when needed.
 356                 *
 357                 * The kernel modifies neither the indices array nor the entries
 358                 * array.
 359                 */
 360                u32                     *sq_array;
 361                struct io_uring_sqe     *sq_sqes;
 362                unsigned                cached_sq_head;
 363                unsigned                sq_entries;
 364                struct list_head        defer_list;
 365
 366                /*
 367                 * Fixed resources fast path, should be accessed only under
 368                 * uring_lock, and updated through io_uring_register(2)
 369                 */
 370                struct io_rsrc_node     *rsrc_node;
 371                struct io_file_table    file_table;
 372                unsigned                nr_user_files;
 373                unsigned                nr_user_bufs;
 374                struct io_mapped_ubuf   **user_bufs;
 375
 376                struct io_submit_state  submit_state;
 377                struct list_head        timeout_list;
 378                struct list_head        ltimeout_list;
 379                struct list_head        cq_overflow_list;
 380                struct xarray           io_buffers;
 381                struct xarray           personalities;
 382                u32                     pers_next;
 383                unsigned                sq_thread_idle;
 384        } ____cacheline_aligned_in_smp;
 385
 386        /* IRQ completion list, under ->completion_lock */
 387        struct list_head        locked_free_list;
 388        unsigned int            locked_free_nr;
 389
 390        const struct cred       *sq_creds;      /* cred used for __io_sq_thread() */
 391        struct io_sq_data       *sq_data;       /* if using sq thread polling */
 392
 393        struct wait_queue_head  sqo_sq_wait;
 394        struct list_head        sqd_list;
 395
 396        unsigned long           check_cq_overflow;
 397
 398        struct {
 399                unsigned                cached_cq_tail;
 400                unsigned                cq_entries;
 401                struct eventfd_ctx      *cq_ev_fd;
 402                struct wait_queue_head  poll_wait;
 403                struct wait_queue_head  cq_wait;
 404                unsigned                cq_extra;
 405                atomic_t                cq_timeouts;
 406                unsigned                cq_last_tm_flush;
 407        } ____cacheline_aligned_in_smp;
 408
 409        struct {
 410                spinlock_t              completion_lock;
 411
 412                spinlock_t              timeout_lock;
 413
 414                /*
 415                 * ->iopoll_list is protected by the ctx->uring_lock for
 416                 * io_uring instances that don't use IORING_SETUP_SQPOLL.
 417                 * For SQPOLL, only the single threaded io_sq_thread() will
 418                 * manipulate the list, hence no extra locking is needed there.
 419                 */
 420                struct list_head        iopoll_list;
 421                struct hlist_head       *cancel_hash;
 422                unsigned                cancel_hash_bits;
 423                bool                    poll_multi_queue;
 424        } ____cacheline_aligned_in_smp;
 425
 426        struct io_restriction           restrictions;
 427
 428        /* slow path rsrc auxilary data, used by update/register */
 429        struct {
 430                struct io_rsrc_node             *rsrc_backup_node;
 431                struct io_mapped_ubuf           *dummy_ubuf;
 432                struct io_rsrc_data             *file_data;
 433                struct io_rsrc_data             *buf_data;
 434
 435                struct delayed_work             rsrc_put_work;
 436                struct llist_head               rsrc_put_llist;
 437                struct list_head                rsrc_ref_list;
 438                spinlock_t                      rsrc_ref_lock;
 439        };
 440
 441        /* Keep this last, we don't need it for the fast path */
 442        struct {
 443                #if defined(CONFIG_UNIX)
 444                        struct socket           *ring_sock;
 445                #endif
 446                /* hashed buffered write serialization */
 447                struct io_wq_hash               *hash_map;
 448
 449                /* Only used for accounting purposes */
 450                struct user_struct              *user;
 451                struct mm_struct                *mm_account;
 452
 453                /* ctx exit and cancelation */
 454                struct llist_head               fallback_llist;
 455                struct delayed_work             fallback_work;
 456                struct work_struct              exit_work;
 457                struct list_head                tctx_list;
 458                struct completion               ref_comp;
 459                u32                             iowq_limits[2];
 460                bool                            iowq_limits_set;
 461        };
 462};
 463
 464struct io_uring_task {
 465        /* submission side */
 466        int                     cached_refs;
 467        struct xarray           xa;
 468        struct wait_queue_head  wait;
 469        const struct io_ring_ctx *last;
 470        struct io_wq            *io_wq;
 471        struct percpu_counter   inflight;
 472        atomic_t                inflight_tracked;
 473        atomic_t                in_idle;
 474
 475        spinlock_t              task_lock;
 476        struct io_wq_work_list  task_list;
 477        struct callback_head    task_work;
 478        bool                    task_running;
 479};
 480
 481/*
 482 * First field must be the file pointer in all the
 483 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
 484 */
 485struct io_poll_iocb {
 486        struct file                     *file;
 487        struct wait_queue_head          *head;
 488        __poll_t                        events;
 489        bool                            done;
 490        bool                            canceled;
 491        struct wait_queue_entry         wait;
 492};
 493
 494struct io_poll_update {
 495        struct file                     *file;
 496        u64                             old_user_data;
 497        u64                             new_user_data;
 498        __poll_t                        events;
 499        bool                            update_events;
 500        bool                            update_user_data;
 501};
 502
 503struct io_close {
 504        struct file                     *file;
 505        int                             fd;
 506        u32                             file_slot;
 507};
 508
 509struct io_timeout_data {
 510        struct io_kiocb                 *req;
 511        struct hrtimer                  timer;
 512        struct timespec64               ts;
 513        enum hrtimer_mode               mode;
 514        u32                             flags;
 515};
 516
 517struct io_accept {
 518        struct file                     *file;
 519        struct sockaddr __user          *addr;
 520        int __user                      *addr_len;
 521        int                             flags;
 522        u32                             file_slot;
 523        unsigned long                   nofile;
 524};
 525
 526struct io_sync {
 527        struct file                     *file;
 528        loff_t                          len;
 529        loff_t                          off;
 530        int                             flags;
 531        int                             mode;
 532};
 533
 534struct io_cancel {
 535        struct file                     *file;
 536        u64                             addr;
 537};
 538
 539struct io_timeout {
 540        struct file                     *file;
 541        u32                             off;
 542        u32                             target_seq;
 543        struct list_head                list;
 544        /* head of the link, used by linked timeouts only */
 545        struct io_kiocb                 *head;
 546        /* for linked completions */
 547        struct io_kiocb                 *prev;
 548};
 549
 550struct io_timeout_rem {
 551        struct file                     *file;
 552        u64                             addr;
 553
 554        /* timeout update */
 555        struct timespec64               ts;
 556        u32                             flags;
 557        bool                            ltimeout;
 558};
 559
 560struct io_rw {
 561        /* NOTE: kiocb has the file as the first member, so don't do it here */
 562        struct kiocb                    kiocb;
 563        u64                             addr;
 564        u64                             len;
 565};
 566
 567struct io_connect {
 568        struct file                     *file;
 569        struct sockaddr __user          *addr;
 570        int                             addr_len;
 571};
 572
 573struct io_sr_msg {
 574        struct file                     *file;
 575        union {
 576                struct compat_msghdr __user     *umsg_compat;
 577                struct user_msghdr __user       *umsg;
 578                void __user                     *buf;
 579        };
 580        int                             msg_flags;
 581        int                             bgid;
 582        size_t                          len;
 583        struct io_buffer                *kbuf;
 584};
 585
 586struct io_open {
 587        struct file                     *file;
 588        int                             dfd;
 589        u32                             file_slot;
 590        struct filename                 *filename;
 591        struct open_how                 how;
 592        unsigned long                   nofile;
 593};
 594
 595struct io_rsrc_update {
 596        struct file                     *file;
 597        u64                             arg;
 598        u32                             nr_args;
 599        u32                             offset;
 600};
 601
 602struct io_fadvise {
 603        struct file                     *file;
 604        u64                             offset;
 605        u32                             len;
 606        u32                             advice;
 607};
 608
 609struct io_madvise {
 610        struct file                     *file;
 611        u64                             addr;
 612        u32                             len;
 613        u32                             advice;
 614};
 615
 616struct io_epoll {
 617        struct file                     *file;
 618        int                             epfd;
 619        int                             op;
 620        int                             fd;
 621        struct epoll_event              event;
 622};
 623
 624struct io_splice {
 625        struct file                     *file_out;
 626        struct file                     *file_in;
 627        loff_t                          off_out;
 628        loff_t                          off_in;
 629        u64                             len;
 630        unsigned int                    flags;
 631};
 632
 633struct io_provide_buf {
 634        struct file                     *file;
 635        __u64                           addr;
 636        __u32                           len;
 637        __u32                           bgid;
 638        __u16                           nbufs;
 639        __u16                           bid;
 640};
 641
 642struct io_statx {
 643        struct file                     *file;
 644        int                             dfd;
 645        unsigned int                    mask;
 646        unsigned int                    flags;
 647        const char __user               *filename;
 648        struct statx __user             *buffer;
 649};
 650
 651struct io_shutdown {
 652        struct file                     *file;
 653        int                             how;
 654};
 655
 656struct io_rename {
 657        struct file                     *file;
 658        int                             old_dfd;
 659        int                             new_dfd;
 660        struct filename                 *oldpath;
 661        struct filename                 *newpath;
 662        int                             flags;
 663};
 664
 665struct io_unlink {
 666        struct file                     *file;
 667        int                             dfd;
 668        int                             flags;
 669        struct filename                 *filename;
 670};
 671
 672struct io_mkdir {
 673        struct file                     *file;
 674        int                             dfd;
 675        umode_t                         mode;
 676        struct filename                 *filename;
 677};
 678
 679struct io_symlink {
 680        struct file                     *file;
 681        int                             new_dfd;
 682        struct filename                 *oldpath;
 683        struct filename                 *newpath;
 684};
 685
 686struct io_hardlink {
 687        struct file                     *file;
 688        int                             old_dfd;
 689        int                             new_dfd;
 690        struct filename                 *oldpath;
 691        struct filename                 *newpath;
 692        int                             flags;
 693};
 694
 695struct io_completion {
 696        struct file                     *file;
 697        u32                             cflags;
 698};
 699
 700struct io_async_connect {
 701        struct sockaddr_storage         address;
 702};
 703
 704struct io_async_msghdr {
 705        struct iovec                    fast_iov[UIO_FASTIOV];
 706        /* points to an allocated iov, if NULL we use fast_iov instead */
 707        struct iovec                    *free_iov;
 708        struct sockaddr __user          *uaddr;
 709        struct msghdr                   msg;
 710        struct sockaddr_storage         addr;
 711};
 712
 713struct io_async_rw {
 714        struct iovec                    fast_iov[UIO_FASTIOV];
 715        const struct iovec              *free_iovec;
 716        struct iov_iter                 iter;
 717        struct iov_iter_state           iter_state;
 718        size_t                          bytes_done;
 719        struct wait_page_queue          wpq;
 720};
 721
 722enum {
 723        REQ_F_FIXED_FILE_BIT    = IOSQE_FIXED_FILE_BIT,
 724        REQ_F_IO_DRAIN_BIT      = IOSQE_IO_DRAIN_BIT,
 725        REQ_F_LINK_BIT          = IOSQE_IO_LINK_BIT,
 726        REQ_F_HARDLINK_BIT      = IOSQE_IO_HARDLINK_BIT,
 727        REQ_F_FORCE_ASYNC_BIT   = IOSQE_ASYNC_BIT,
 728        REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
 729
 730        /* first byte is taken by user flags, shift it to not overlap */
 731        REQ_F_FAIL_BIT          = 8,
 732        REQ_F_INFLIGHT_BIT,
 733        REQ_F_CUR_POS_BIT,
 734        REQ_F_NOWAIT_BIT,
 735        REQ_F_LINK_TIMEOUT_BIT,
 736        REQ_F_NEED_CLEANUP_BIT,
 737        REQ_F_POLLED_BIT,
 738        REQ_F_BUFFER_SELECTED_BIT,
 739        REQ_F_COMPLETE_INLINE_BIT,
 740        REQ_F_REISSUE_BIT,
 741        REQ_F_CREDS_BIT,
 742        REQ_F_REFCOUNT_BIT,
 743        REQ_F_ARM_LTIMEOUT_BIT,
 744        /* keep async read/write and isreg together and in order */
 745        REQ_F_NOWAIT_READ_BIT,
 746        REQ_F_NOWAIT_WRITE_BIT,
 747        REQ_F_ISREG_BIT,
 748
 749        /* not a real bit, just to check we're not overflowing the space */
 750        __REQ_F_LAST_BIT,
 751};
 752
 753enum {
 754        /* ctx owns file */
 755        REQ_F_FIXED_FILE        = BIT(REQ_F_FIXED_FILE_BIT),
 756        /* drain existing IO first */
 757        REQ_F_IO_DRAIN          = BIT(REQ_F_IO_DRAIN_BIT),
 758        /* linked sqes */
 759        REQ_F_LINK              = BIT(REQ_F_LINK_BIT),
 760        /* doesn't sever on completion < 0 */
 761        REQ_F_HARDLINK          = BIT(REQ_F_HARDLINK_BIT),
 762        /* IOSQE_ASYNC */
 763        REQ_F_FORCE_ASYNC       = BIT(REQ_F_FORCE_ASYNC_BIT),
 764        /* IOSQE_BUFFER_SELECT */
 765        REQ_F_BUFFER_SELECT     = BIT(REQ_F_BUFFER_SELECT_BIT),
 766
 767        /* fail rest of links */
 768        REQ_F_FAIL              = BIT(REQ_F_FAIL_BIT),
 769        /* on inflight list, should be cancelled and waited on exit reliably */
 770        REQ_F_INFLIGHT          = BIT(REQ_F_INFLIGHT_BIT),
 771        /* read/write uses file position */
 772        REQ_F_CUR_POS           = BIT(REQ_F_CUR_POS_BIT),
 773        /* must not punt to workers */
 774        REQ_F_NOWAIT            = BIT(REQ_F_NOWAIT_BIT),
 775        /* has or had linked timeout */
 776        REQ_F_LINK_TIMEOUT      = BIT(REQ_F_LINK_TIMEOUT_BIT),
 777        /* needs cleanup */
 778        REQ_F_NEED_CLEANUP      = BIT(REQ_F_NEED_CLEANUP_BIT),
 779        /* already went through poll handler */
 780        REQ_F_POLLED            = BIT(REQ_F_POLLED_BIT),
 781        /* buffer already selected */
 782        REQ_F_BUFFER_SELECTED   = BIT(REQ_F_BUFFER_SELECTED_BIT),
 783        /* completion is deferred through io_comp_state */
 784        REQ_F_COMPLETE_INLINE   = BIT(REQ_F_COMPLETE_INLINE_BIT),
 785        /* caller should reissue async */
 786        REQ_F_REISSUE           = BIT(REQ_F_REISSUE_BIT),
 787        /* supports async reads */
 788        REQ_F_NOWAIT_READ       = BIT(REQ_F_NOWAIT_READ_BIT),
 789        /* supports async writes */
 790        REQ_F_NOWAIT_WRITE      = BIT(REQ_F_NOWAIT_WRITE_BIT),
 791        /* regular file */
 792        REQ_F_ISREG             = BIT(REQ_F_ISREG_BIT),
 793        /* has creds assigned */
 794        REQ_F_CREDS             = BIT(REQ_F_CREDS_BIT),
 795        /* skip refcounting if not set */
 796        REQ_F_REFCOUNT          = BIT(REQ_F_REFCOUNT_BIT),
 797        /* there is a linked timeout that has to be armed */
 798        REQ_F_ARM_LTIMEOUT      = BIT(REQ_F_ARM_LTIMEOUT_BIT),
 799};
 800
 801struct async_poll {
 802        struct io_poll_iocb     poll;
 803        struct io_poll_iocb     *double_poll;
 804};
 805
 806typedef void (*io_req_tw_func_t)(struct io_kiocb *req, bool *locked);
 807
 808struct io_task_work {
 809        union {
 810                struct io_wq_work_node  node;
 811                struct llist_node       fallback_node;
 812        };
 813        io_req_tw_func_t                func;
 814};
 815
 816enum {
 817        IORING_RSRC_FILE                = 0,
 818        IORING_RSRC_BUFFER              = 1,
 819};
 820
 821/*
 822 * NOTE! Each of the iocb union members has the file pointer
 823 * as the first entry in their struct definition. So you can
 824 * access the file pointer through any of the sub-structs,
 825 * or directly as just 'ki_filp' in this struct.
 826 */
 827struct io_kiocb {
 828        union {
 829                struct file             *file;
 830                struct io_rw            rw;
 831                struct io_poll_iocb     poll;
 832                struct io_poll_update   poll_update;
 833                struct io_accept        accept;
 834                struct io_sync          sync;
 835                struct io_cancel        cancel;
 836                struct io_timeout       timeout;
 837                struct io_timeout_rem   timeout_rem;
 838                struct io_connect       connect;
 839                struct io_sr_msg        sr_msg;
 840                struct io_open          open;
 841                struct io_close         close;
 842                struct io_rsrc_update   rsrc_update;
 843                struct io_fadvise       fadvise;
 844                struct io_madvise       madvise;
 845                struct io_epoll         epoll;
 846                struct io_splice        splice;
 847                struct io_provide_buf   pbuf;
 848                struct io_statx         statx;
 849                struct io_shutdown      shutdown;
 850                struct io_rename        rename;
 851                struct io_unlink        unlink;
 852                struct io_mkdir         mkdir;
 853                struct io_symlink       symlink;
 854                struct io_hardlink      hardlink;
 855                /* use only after cleaning per-op data, see io_clean_op() */
 856                struct io_completion    compl;
 857        };
 858
 859        /* opcode allocated if it needs to store data for async defer */
 860        void                            *async_data;
 861        u8                              opcode;
 862        /* polled IO has completed */
 863        u8                              iopoll_completed;
 864
 865        u16                             buf_index;
 866        u32                             result;
 867
 868        struct io_ring_ctx              *ctx;
 869        unsigned int                    flags;
 870        atomic_t                        refs;
 871        struct task_struct              *task;
 872        u64                             user_data;
 873
 874        struct io_kiocb                 *link;
 875        struct percpu_ref               *fixed_rsrc_refs;
 876
 877        /* used with ctx->iopoll_list with reads/writes */
 878        struct list_head                inflight_entry;
 879        struct io_task_work             io_task_work;
 880        /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
 881        struct hlist_node               hash_node;
 882        struct async_poll               *apoll;
 883        struct io_wq_work               work;
 884        const struct cred               *creds;
 885
 886        /* store used ubuf, so we can prevent reloading */
 887        struct io_mapped_ubuf           *imu;
 888};
 889
 890struct io_tctx_node {
 891        struct list_head        ctx_node;
 892        struct task_struct      *task;
 893        struct io_ring_ctx      *ctx;
 894};
 895
 896struct io_defer_entry {
 897        struct list_head        list;
 898        struct io_kiocb         *req;
 899        u32                     seq;
 900};
 901
 902struct io_op_def {
 903        /* needs req->file assigned */
 904        unsigned                needs_file : 1;
 905        /* hash wq insertion if file is a regular file */
 906        unsigned                hash_reg_file : 1;
 907        /* unbound wq insertion if file is a non-regular file */
 908        unsigned                unbound_nonreg_file : 1;
 909        /* opcode is not supported by this kernel */
 910        unsigned                not_supported : 1;
 911        /* set if opcode supports polled "wait" */
 912        unsigned                pollin : 1;
 913        unsigned                pollout : 1;
 914        /* op supports buffer selection */
 915        unsigned                buffer_select : 1;
 916        /* do prep async if is going to be punted */
 917        unsigned                needs_async_setup : 1;
 918        /* should block plug */
 919        unsigned                plug : 1;
 920        /* size of async data needed, if any */
 921        unsigned short          async_size;
 922};
 923
 924static const struct io_op_def io_op_defs[] = {
 925        [IORING_OP_NOP] = {},
 926        [IORING_OP_READV] = {
 927                .needs_file             = 1,
 928                .unbound_nonreg_file    = 1,
 929                .pollin                 = 1,
 930                .buffer_select          = 1,
 931                .needs_async_setup      = 1,
 932                .plug                   = 1,
 933                .async_size             = sizeof(struct io_async_rw),
 934        },
 935        [IORING_OP_WRITEV] = {
 936                .needs_file             = 1,
 937                .hash_reg_file          = 1,
 938                .unbound_nonreg_file    = 1,
 939                .pollout                = 1,
 940                .needs_async_setup      = 1,
 941                .plug                   = 1,
 942                .async_size             = sizeof(struct io_async_rw),
 943        },
 944        [IORING_OP_FSYNC] = {
 945                .needs_file             = 1,
 946        },
 947        [IORING_OP_READ_FIXED] = {
 948                .needs_file             = 1,
 949                .unbound_nonreg_file    = 1,
 950                .pollin                 = 1,
 951                .plug                   = 1,
 952                .async_size             = sizeof(struct io_async_rw),
 953        },
 954        [IORING_OP_WRITE_FIXED] = {
 955                .needs_file             = 1,
 956                .hash_reg_file          = 1,
 957                .unbound_nonreg_file    = 1,
 958                .pollout                = 1,
 959                .plug                   = 1,
 960                .async_size             = sizeof(struct io_async_rw),
 961        },
 962        [IORING_OP_POLL_ADD] = {
 963                .needs_file             = 1,
 964                .unbound_nonreg_file    = 1,
 965        },
 966        [IORING_OP_POLL_REMOVE] = {},
 967        [IORING_OP_SYNC_FILE_RANGE] = {
 968                .needs_file             = 1,
 969        },
 970        [IORING_OP_SENDMSG] = {
 971                .needs_file             = 1,
 972                .unbound_nonreg_file    = 1,
 973                .pollout                = 1,
 974                .needs_async_setup      = 1,
 975                .async_size             = sizeof(struct io_async_msghdr),
 976        },
 977        [IORING_OP_RECVMSG] = {
 978                .needs_file             = 1,
 979                .unbound_nonreg_file    = 1,
 980                .pollin                 = 1,
 981                .buffer_select          = 1,
 982                .needs_async_setup      = 1,
 983                .async_size             = sizeof(struct io_async_msghdr),
 984        },
 985        [IORING_OP_TIMEOUT] = {
 986                .async_size             = sizeof(struct io_timeout_data),
 987        },
 988        [IORING_OP_TIMEOUT_REMOVE] = {
 989                /* used by timeout updates' prep() */
 990        },
 991        [IORING_OP_ACCEPT] = {
 992                .needs_file             = 1,
 993                .unbound_nonreg_file    = 1,
 994                .pollin                 = 1,
 995        },
 996        [IORING_OP_ASYNC_CANCEL] = {},
 997        [IORING_OP_LINK_TIMEOUT] = {
 998                .async_size             = sizeof(struct io_timeout_data),
 999        },
1000        [IORING_OP_CONNECT] = {
1001                .needs_file             = 1,
1002                .unbound_nonreg_file    = 1,
1003                .pollout                = 1,
1004                .needs_async_setup      = 1,
1005                .async_size             = sizeof(struct io_async_connect),
1006        },
1007        [IORING_OP_FALLOCATE] = {
1008                .needs_file             = 1,
1009        },
1010        [IORING_OP_OPENAT] = {},
1011        [IORING_OP_CLOSE] = {},
1012        [IORING_OP_FILES_UPDATE] = {},
1013        [IORING_OP_STATX] = {},
1014        [IORING_OP_READ] = {
1015                .needs_file             = 1,
1016                .unbound_nonreg_file    = 1,
1017                .pollin                 = 1,
1018                .buffer_select          = 1,
1019                .plug                   = 1,
1020                .async_size             = sizeof(struct io_async_rw),
1021        },
1022        [IORING_OP_WRITE] = {
1023                .needs_file             = 1,
1024                .hash_reg_file          = 1,
1025                .unbound_nonreg_file    = 1,
1026                .pollout                = 1,
1027                .plug                   = 1,
1028                .async_size             = sizeof(struct io_async_rw),
1029        },
1030        [IORING_OP_FADVISE] = {
1031                .needs_file             = 1,
1032        },
1033        [IORING_OP_MADVISE] = {},
1034        [IORING_OP_SEND] = {
1035                .needs_file             = 1,
1036                .unbound_nonreg_file    = 1,
1037                .pollout                = 1,
1038        },
1039        [IORING_OP_RECV] = {
1040                .needs_file             = 1,
1041                .unbound_nonreg_file    = 1,
1042                .pollin                 = 1,
1043                .buffer_select          = 1,
1044        },
1045        [IORING_OP_OPENAT2] = {
1046        },
1047        [IORING_OP_EPOLL_CTL] = {
1048                .unbound_nonreg_file    = 1,
1049        },
1050        [IORING_OP_SPLICE] = {
1051                .needs_file             = 1,
1052                .hash_reg_file          = 1,
1053                .unbound_nonreg_file    = 1,
1054        },
1055        [IORING_OP_PROVIDE_BUFFERS] = {},
1056        [IORING_OP_REMOVE_BUFFERS] = {},
1057        [IORING_OP_TEE] = {
1058                .needs_file             = 1,
1059                .hash_reg_file          = 1,
1060                .unbound_nonreg_file    = 1,
1061        },
1062        [IORING_OP_SHUTDOWN] = {
1063                .needs_file             = 1,
1064        },
1065        [IORING_OP_RENAMEAT] = {},
1066        [IORING_OP_UNLINKAT] = {},
1067        [IORING_OP_MKDIRAT] = {},
1068        [IORING_OP_SYMLINKAT] = {},
1069        [IORING_OP_LINKAT] = {},
1070};
1071
1072/* requests with any of those set should undergo io_disarm_next() */
1073#define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
1074
1075static bool io_disarm_next(struct io_kiocb *req);
1076static void io_uring_del_tctx_node(unsigned long index);
1077static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1078                                         struct task_struct *task,
1079                                         bool cancel_all);
1080static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1081
1082static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1083                                 long res, unsigned int cflags);
1084static void io_put_req(struct io_kiocb *req);
1085static void io_put_req_deferred(struct io_kiocb *req);
1086static void io_dismantle_req(struct io_kiocb *req);
1087static void io_queue_linked_timeout(struct io_kiocb *req);
1088static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1089                                     struct io_uring_rsrc_update2 *up,
1090                                     unsigned nr_args);
1091static void io_clean_op(struct io_kiocb *req);
1092static struct file *io_file_get(struct io_ring_ctx *ctx,
1093                                struct io_kiocb *req, int fd, bool fixed);
1094static void __io_queue_sqe(struct io_kiocb *req);
1095static void io_rsrc_put_work(struct work_struct *work);
1096
1097static void io_req_task_queue(struct io_kiocb *req);
1098static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1099static int io_req_prep_async(struct io_kiocb *req);
1100
1101static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
1102                                 unsigned int issue_flags, u32 slot_index);
1103static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags);
1104
1105static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer);
1106
1107static struct kmem_cache *req_cachep;
1108
1109static const struct file_operations io_uring_fops;
1110
1111struct sock *io_uring_get_socket(struct file *file)
1112{
1113#if defined(CONFIG_UNIX)
1114        if (file->f_op == &io_uring_fops) {
1115                struct io_ring_ctx *ctx = file->private_data;
1116
1117                return ctx->ring_sock->sk;
1118        }
1119#endif
1120        return NULL;
1121}
1122EXPORT_SYMBOL(io_uring_get_socket);
1123
1124static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked)
1125{
1126        if (!*locked) {
1127                mutex_lock(&ctx->uring_lock);
1128                *locked = true;
1129        }
1130}
1131
1132#define io_for_each_link(pos, head) \
1133        for (pos = (head); pos; pos = pos->link)
1134
1135/*
1136 * Shamelessly stolen from the mm implementation of page reference checking,
1137 * see commit f958d7b528b1 for details.
1138 */
1139#define req_ref_zero_or_close_to_overflow(req)  \
1140        ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1141
1142static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1143{
1144        WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1145        return atomic_inc_not_zero(&req->refs);
1146}
1147
1148static inline bool req_ref_put_and_test(struct io_kiocb *req)
1149{
1150        if (likely(!(req->flags & REQ_F_REFCOUNT)))
1151                return true;
1152
1153        WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1154        return atomic_dec_and_test(&req->refs);
1155}
1156
1157static inline void req_ref_put(struct io_kiocb *req)
1158{
1159        WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1160        WARN_ON_ONCE(req_ref_put_and_test(req));
1161}
1162
1163static inline void req_ref_get(struct io_kiocb *req)
1164{
1165        WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1166        WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1167        atomic_inc(&req->refs);
1168}
1169
1170static inline void __io_req_set_refcount(struct io_kiocb *req, int nr)
1171{
1172        if (!(req->flags & REQ_F_REFCOUNT)) {
1173                req->flags |= REQ_F_REFCOUNT;
1174                atomic_set(&req->refs, nr);
1175        }
1176}
1177
1178static inline void io_req_set_refcount(struct io_kiocb *req)
1179{
1180        __io_req_set_refcount(req, 1);
1181}
1182
1183static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1184{
1185        struct io_ring_ctx *ctx = req->ctx;
1186
1187        if (!req->fixed_rsrc_refs) {
1188                req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1189                percpu_ref_get(req->fixed_rsrc_refs);
1190        }
1191}
1192
1193static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1194{
1195        bool got = percpu_ref_tryget(ref);
1196
1197        /* already at zero, wait for ->release() */
1198        if (!got)
1199                wait_for_completion(compl);
1200        percpu_ref_resurrect(ref);
1201        if (got)
1202                percpu_ref_put(ref);
1203}
1204
1205static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1206                          bool cancel_all)
1207{
1208        struct io_kiocb *req;
1209
1210        if (task && head->task != task)
1211                return false;
1212        if (cancel_all)
1213                return true;
1214
1215        io_for_each_link(req, head) {
1216                if (req->flags & REQ_F_INFLIGHT)
1217                        return true;
1218        }
1219        return false;
1220}
1221
1222static inline void req_set_fail(struct io_kiocb *req)
1223{
1224        req->flags |= REQ_F_FAIL;
1225}
1226
1227static inline void req_fail_link_node(struct io_kiocb *req, int res)
1228{
1229        req_set_fail(req);
1230        req->result = res;
1231}
1232
1233static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1234{
1235        struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1236
1237        complete(&ctx->ref_comp);
1238}
1239
1240static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1241{
1242        return !req->timeout.off;
1243}
1244
1245static void io_fallback_req_func(struct work_struct *work)
1246{
1247        struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1248                                                fallback_work.work);
1249        struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1250        struct io_kiocb *req, *tmp;
1251        bool locked = false;
1252
1253        percpu_ref_get(&ctx->refs);
1254        llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1255                req->io_task_work.func(req, &locked);
1256
1257        if (locked) {
1258                if (ctx->submit_state.compl_nr)
1259                        io_submit_flush_completions(ctx);
1260                mutex_unlock(&ctx->uring_lock);
1261        }
1262        percpu_ref_put(&ctx->refs);
1263
1264}
1265
1266static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1267{
1268        struct io_ring_ctx *ctx;
1269        int hash_bits;
1270
1271        ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1272        if (!ctx)
1273                return NULL;
1274
1275        /*
1276         * Use 5 bits less than the max cq entries, that should give us around
1277         * 32 entries per hash list if totally full and uniformly spread.
1278         */
1279        hash_bits = ilog2(p->cq_entries);
1280        hash_bits -= 5;
1281        if (hash_bits <= 0)
1282                hash_bits = 1;
1283        ctx->cancel_hash_bits = hash_bits;
1284        ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1285                                        GFP_KERNEL);
1286        if (!ctx->cancel_hash)
1287                goto err;
1288        __hash_init(ctx->cancel_hash, 1U << hash_bits);
1289
1290        ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1291        if (!ctx->dummy_ubuf)
1292                goto err;
1293        /* set invalid range, so io_import_fixed() fails meeting it */
1294        ctx->dummy_ubuf->ubuf = -1UL;
1295
1296        if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1297                            PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1298                goto err;
1299
1300        ctx->flags = p->flags;
1301        init_waitqueue_head(&ctx->sqo_sq_wait);
1302        INIT_LIST_HEAD(&ctx->sqd_list);
1303        init_waitqueue_head(&ctx->poll_wait);
1304        INIT_LIST_HEAD(&ctx->cq_overflow_list);
1305        init_completion(&ctx->ref_comp);
1306        xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1307        xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1308        mutex_init(&ctx->uring_lock);
1309        init_waitqueue_head(&ctx->cq_wait);
1310        spin_lock_init(&ctx->completion_lock);
1311        spin_lock_init(&ctx->timeout_lock);
1312        INIT_LIST_HEAD(&ctx->iopoll_list);
1313        INIT_LIST_HEAD(&ctx->defer_list);
1314        INIT_LIST_HEAD(&ctx->timeout_list);
1315        INIT_LIST_HEAD(&ctx->ltimeout_list);
1316        spin_lock_init(&ctx->rsrc_ref_lock);
1317        INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1318        INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1319        init_llist_head(&ctx->rsrc_put_llist);
1320        INIT_LIST_HEAD(&ctx->tctx_list);
1321        INIT_LIST_HEAD(&ctx->submit_state.free_list);
1322        INIT_LIST_HEAD(&ctx->locked_free_list);
1323        INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1324        return ctx;
1325err:
1326        kfree(ctx->dummy_ubuf);
1327        kfree(ctx->cancel_hash);
1328        kfree(ctx);
1329        return NULL;
1330}
1331
1332static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1333{
1334        struct io_rings *r = ctx->rings;
1335
1336        WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1337        ctx->cq_extra--;
1338}
1339
1340static bool req_need_defer(struct io_kiocb *req, u32 seq)
1341{
1342        if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1343                struct io_ring_ctx *ctx = req->ctx;
1344
1345                return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1346        }
1347
1348        return false;
1349}
1350
1351#define FFS_ASYNC_READ          0x1UL
1352#define FFS_ASYNC_WRITE         0x2UL
1353#ifdef CONFIG_64BIT
1354#define FFS_ISREG               0x4UL
1355#else
1356#define FFS_ISREG               0x0UL
1357#endif
1358#define FFS_MASK                ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1359
1360static inline bool io_req_ffs_set(struct io_kiocb *req)
1361{
1362        return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1363}
1364
1365static void io_req_track_inflight(struct io_kiocb *req)
1366{
1367        if (!(req->flags & REQ_F_INFLIGHT)) {
1368                req->flags |= REQ_F_INFLIGHT;
1369                atomic_inc(&current->io_uring->inflight_tracked);
1370        }
1371}
1372
1373static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1374{
1375        if (WARN_ON_ONCE(!req->link))
1376                return NULL;
1377
1378        req->flags &= ~REQ_F_ARM_LTIMEOUT;
1379        req->flags |= REQ_F_LINK_TIMEOUT;
1380
1381        /* linked timeouts should have two refs once prep'ed */
1382        io_req_set_refcount(req);
1383        __io_req_set_refcount(req->link, 2);
1384        return req->link;
1385}
1386
1387static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1388{
1389        if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
1390                return NULL;
1391        return __io_prep_linked_timeout(req);
1392}
1393
1394static void io_prep_async_work(struct io_kiocb *req)
1395{
1396        const struct io_op_def *def = &io_op_defs[req->opcode];
1397        struct io_ring_ctx *ctx = req->ctx;
1398
1399        if (!(req->flags & REQ_F_CREDS)) {
1400                req->flags |= REQ_F_CREDS;
1401                req->creds = get_current_cred();
1402        }
1403
1404        req->work.list.next = NULL;
1405        req->work.flags = 0;
1406        if (req->flags & REQ_F_FORCE_ASYNC)
1407                req->work.flags |= IO_WQ_WORK_CONCURRENT;
1408
1409        if (req->flags & REQ_F_ISREG) {
1410                if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1411                        io_wq_hash_work(&req->work, file_inode(req->file));
1412        } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1413                if (def->unbound_nonreg_file)
1414                        req->work.flags |= IO_WQ_WORK_UNBOUND;
1415        }
1416
1417        switch (req->opcode) {
1418        case IORING_OP_SPLICE:
1419        case IORING_OP_TEE:
1420                if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1421                        req->work.flags |= IO_WQ_WORK_UNBOUND;
1422                break;
1423        }
1424}
1425
1426static void io_prep_async_link(struct io_kiocb *req)
1427{
1428        struct io_kiocb *cur;
1429
1430        if (req->flags & REQ_F_LINK_TIMEOUT) {
1431                struct io_ring_ctx *ctx = req->ctx;
1432
1433                spin_lock(&ctx->completion_lock);
1434                io_for_each_link(cur, req)
1435                        io_prep_async_work(cur);
1436                spin_unlock(&ctx->completion_lock);
1437        } else {
1438                io_for_each_link(cur, req)
1439                        io_prep_async_work(cur);
1440        }
1441}
1442
1443static void io_queue_async_work(struct io_kiocb *req, bool *locked)
1444{
1445        struct io_ring_ctx *ctx = req->ctx;
1446        struct io_kiocb *link = io_prep_linked_timeout(req);
1447        struct io_uring_task *tctx = req->task->io_uring;
1448
1449        /* must not take the lock, NULL it as a precaution */
1450        locked = NULL;
1451
1452        BUG_ON(!tctx);
1453        BUG_ON(!tctx->io_wq);
1454
1455        /* init ->work of the whole link before punting */
1456        io_prep_async_link(req);
1457
1458        /*
1459         * Not expected to happen, but if we do have a bug where this _can_
1460         * happen, catch it here and ensure the request is marked as
1461         * canceled. That will make io-wq go through the usual work cancel
1462         * procedure rather than attempt to run this request (or create a new
1463         * worker for it).
1464         */
1465        if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1466                req->work.flags |= IO_WQ_WORK_CANCEL;
1467
1468        trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1469                                        &req->work, req->flags);
1470        io_wq_enqueue(tctx->io_wq, &req->work);
1471        if (link)
1472                io_queue_linked_timeout(link);
1473}
1474
1475static void io_kill_timeout(struct io_kiocb *req, int status)
1476        __must_hold(&req->ctx->completion_lock)
1477        __must_hold(&req->ctx->timeout_lock)
1478{
1479        struct io_timeout_data *io = req->async_data;
1480
1481        if (hrtimer_try_to_cancel(&io->timer) != -1) {
1482                if (status)
1483                        req_set_fail(req);
1484                atomic_set(&req->ctx->cq_timeouts,
1485                        atomic_read(&req->ctx->cq_timeouts) + 1);
1486                list_del_init(&req->timeout.list);
1487                io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1488                io_put_req_deferred(req);
1489        }
1490}
1491
1492static void io_queue_deferred(struct io_ring_ctx *ctx)
1493{
1494        while (!list_empty(&ctx->defer_list)) {
1495                struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1496                                                struct io_defer_entry, list);
1497
1498                if (req_need_defer(de->req, de->seq))
1499                        break;
1500                list_del_init(&de->list);
1501                io_req_task_queue(de->req);
1502                kfree(de);
1503        }
1504}
1505
1506static void io_flush_timeouts(struct io_ring_ctx *ctx)
1507        __must_hold(&ctx->completion_lock)
1508{
1509        u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1510
1511        spin_lock_irq(&ctx->timeout_lock);
1512        while (!list_empty(&ctx->timeout_list)) {
1513                u32 events_needed, events_got;
1514                struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1515                                                struct io_kiocb, timeout.list);
1516
1517                if (io_is_timeout_noseq(req))
1518                        break;
1519
1520                /*
1521                 * Since seq can easily wrap around over time, subtract
1522                 * the last seq at which timeouts were flushed before comparing.
1523                 * Assuming not more than 2^31-1 events have happened since,
1524                 * these subtractions won't have wrapped, so we can check if
1525                 * target is in [last_seq, current_seq] by comparing the two.
1526                 */
1527                events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1528                events_got = seq - ctx->cq_last_tm_flush;
1529                if (events_got < events_needed)
1530                        break;
1531
1532                list_del_init(&req->timeout.list);
1533                io_kill_timeout(req, 0);
1534        }
1535        ctx->cq_last_tm_flush = seq;
1536        spin_unlock_irq(&ctx->timeout_lock);
1537}
1538
1539static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1540{
1541        if (ctx->off_timeout_used)
1542                io_flush_timeouts(ctx);
1543        if (ctx->drain_active)
1544                io_queue_deferred(ctx);
1545}
1546
1547static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1548{
1549        if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1550                __io_commit_cqring_flush(ctx);
1551        /* order cqe stores with ring update */
1552        smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1553}
1554
1555static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1556{
1557        struct io_rings *r = ctx->rings;
1558
1559        return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1560}
1561
1562static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1563{
1564        return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1565}
1566
1567static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1568{
1569        struct io_rings *rings = ctx->rings;
1570        unsigned tail, mask = ctx->cq_entries - 1;
1571
1572        /*
1573         * writes to the cq entry need to come after reading head; the
1574         * control dependency is enough as we're using WRITE_ONCE to
1575         * fill the cq entry
1576         */
1577        if (__io_cqring_events(ctx) == ctx->cq_entries)
1578                return NULL;
1579
1580        tail = ctx->cached_cq_tail++;
1581        return &rings->cqes[tail & mask];
1582}
1583
1584static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1585{
1586        if (likely(!ctx->cq_ev_fd))
1587                return false;
1588        if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1589                return false;
1590        return !ctx->eventfd_async || io_wq_current_is_worker();
1591}
1592
1593/*
1594 * This should only get called when at least one event has been posted.
1595 * Some applications rely on the eventfd notification count only changing
1596 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1597 * 1:1 relationship between how many times this function is called (and
1598 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1599 */
1600static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1601{
1602        /*
1603         * wake_up_all() may seem excessive, but io_wake_function() and
1604         * io_should_wake() handle the termination of the loop and only
1605         * wake as many waiters as we need to.
1606         */
1607        if (wq_has_sleeper(&ctx->cq_wait))
1608                wake_up_all(&ctx->cq_wait);
1609        if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1610                wake_up(&ctx->sq_data->wait);
1611        if (io_should_trigger_evfd(ctx))
1612                eventfd_signal(ctx->cq_ev_fd, 1);
1613        if (waitqueue_active(&ctx->poll_wait))
1614                wake_up_interruptible(&ctx->poll_wait);
1615}
1616
1617static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1618{
1619        /* see waitqueue_active() comment */
1620        smp_mb();
1621
1622        if (ctx->flags & IORING_SETUP_SQPOLL) {
1623                if (waitqueue_active(&ctx->cq_wait))
1624                        wake_up_all(&ctx->cq_wait);
1625        }
1626        if (io_should_trigger_evfd(ctx))
1627                eventfd_signal(ctx->cq_ev_fd, 1);
1628        if (waitqueue_active(&ctx->poll_wait))
1629                wake_up_interruptible(&ctx->poll_wait);
1630}
1631
1632/* Returns true if there are no backlogged entries after the flush */
1633static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1634{
1635        bool all_flushed, posted;
1636
1637        if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1638                return false;
1639
1640        posted = false;
1641        spin_lock(&ctx->completion_lock);
1642        while (!list_empty(&ctx->cq_overflow_list)) {
1643                struct io_uring_cqe *cqe = io_get_cqe(ctx);
1644                struct io_overflow_cqe *ocqe;
1645
1646                if (!cqe && !force)
1647                        break;
1648                ocqe = list_first_entry(&ctx->cq_overflow_list,
1649                                        struct io_overflow_cqe, list);
1650                if (cqe)
1651                        memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1652                else
1653                        io_account_cq_overflow(ctx);
1654
1655                posted = true;
1656                list_del(&ocqe->list);
1657                kfree(ocqe);
1658        }
1659
1660        all_flushed = list_empty(&ctx->cq_overflow_list);
1661        if (all_flushed) {
1662                clear_bit(0, &ctx->check_cq_overflow);
1663                WRITE_ONCE(ctx->rings->sq_flags,
1664                           ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1665        }
1666
1667        if (posted)
1668                io_commit_cqring(ctx);
1669        spin_unlock(&ctx->completion_lock);
1670        if (posted)
1671                io_cqring_ev_posted(ctx);
1672        return all_flushed;
1673}
1674
1675static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1676{
1677        bool ret = true;
1678
1679        if (test_bit(0, &ctx->check_cq_overflow)) {
1680                /* iopoll syncs against uring_lock, not completion_lock */
1681                if (ctx->flags & IORING_SETUP_IOPOLL)
1682                        mutex_lock(&ctx->uring_lock);
1683                ret = __io_cqring_overflow_flush(ctx, false);
1684                if (ctx->flags & IORING_SETUP_IOPOLL)
1685                        mutex_unlock(&ctx->uring_lock);
1686        }
1687
1688        return ret;
1689}
1690
1691/* must to be called somewhat shortly after putting a request */
1692static inline void io_put_task(struct task_struct *task, int nr)
1693{
1694        struct io_uring_task *tctx = task->io_uring;
1695
1696        if (likely(task == current)) {
1697                tctx->cached_refs += nr;
1698        } else {
1699                percpu_counter_sub(&tctx->inflight, nr);
1700                if (unlikely(atomic_read(&tctx->in_idle)))
1701                        wake_up(&tctx->wait);
1702                put_task_struct_many(task, nr);
1703        }
1704}
1705
1706static void io_task_refs_refill(struct io_uring_task *tctx)
1707{
1708        unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
1709
1710        percpu_counter_add(&tctx->inflight, refill);
1711        refcount_add(refill, &current->usage);
1712        tctx->cached_refs += refill;
1713}
1714
1715static inline void io_get_task_refs(int nr)
1716{
1717        struct io_uring_task *tctx = current->io_uring;
1718
1719        tctx->cached_refs -= nr;
1720        if (unlikely(tctx->cached_refs < 0))
1721                io_task_refs_refill(tctx);
1722}
1723
1724static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1725                                     long res, unsigned int cflags)
1726{
1727        struct io_overflow_cqe *ocqe;
1728
1729        ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1730        if (!ocqe) {
1731                /*
1732                 * If we're in ring overflow flush mode, or in task cancel mode,
1733                 * or cannot allocate an overflow entry, then we need to drop it
1734                 * on the floor.
1735                 */
1736                io_account_cq_overflow(ctx);
1737                return false;
1738        }
1739        if (list_empty(&ctx->cq_overflow_list)) {
1740                set_bit(0, &ctx->check_cq_overflow);
1741                WRITE_ONCE(ctx->rings->sq_flags,
1742                           ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1743
1744        }
1745        ocqe->cqe.user_data = user_data;
1746        ocqe->cqe.res = res;
1747        ocqe->cqe.flags = cflags;
1748        list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1749        return true;
1750}
1751
1752static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1753                                          long res, unsigned int cflags)
1754{
1755        struct io_uring_cqe *cqe;
1756
1757        trace_io_uring_complete(ctx, user_data, res, cflags);
1758
1759        /*
1760         * If we can't get a cq entry, userspace overflowed the
1761         * submission (by quite a lot). Increment the overflow count in
1762         * the ring.
1763         */
1764        cqe = io_get_cqe(ctx);
1765        if (likely(cqe)) {
1766                WRITE_ONCE(cqe->user_data, user_data);
1767                WRITE_ONCE(cqe->res, res);
1768                WRITE_ONCE(cqe->flags, cflags);
1769                return true;
1770        }
1771        return io_cqring_event_overflow(ctx, user_data, res, cflags);
1772}
1773
1774/* not as hot to bloat with inlining */
1775static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1776                                          long res, unsigned int cflags)
1777{
1778        return __io_cqring_fill_event(ctx, user_data, res, cflags);
1779}
1780
1781static void io_req_complete_post(struct io_kiocb *req, long res,
1782                                 unsigned int cflags)
1783{
1784        struct io_ring_ctx *ctx = req->ctx;
1785
1786        spin_lock(&ctx->completion_lock);
1787        __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1788        /*
1789         * If we're the last reference to this request, add to our locked
1790         * free_list cache.
1791         */
1792        if (req_ref_put_and_test(req)) {
1793                if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1794                        if (req->flags & IO_DISARM_MASK)
1795                                io_disarm_next(req);
1796                        if (req->link) {
1797                                io_req_task_queue(req->link);
1798                                req->link = NULL;
1799                        }
1800                }
1801                io_dismantle_req(req);
1802                io_put_task(req->task, 1);
1803                list_add(&req->inflight_entry, &ctx->locked_free_list);
1804                ctx->locked_free_nr++;
1805        } else {
1806                if (!percpu_ref_tryget(&ctx->refs))
1807                        req = NULL;
1808        }
1809        io_commit_cqring(ctx);
1810        spin_unlock(&ctx->completion_lock);
1811
1812        if (req) {
1813                io_cqring_ev_posted(ctx);
1814                percpu_ref_put(&ctx->refs);
1815        }
1816}
1817
1818static inline bool io_req_needs_clean(struct io_kiocb *req)
1819{
1820        return req->flags & IO_REQ_CLEAN_FLAGS;
1821}
1822
1823static void io_req_complete_state(struct io_kiocb *req, long res,
1824                                  unsigned int cflags)
1825{
1826        if (io_req_needs_clean(req))
1827                io_clean_op(req);
1828        req->result = res;
1829        req->compl.cflags = cflags;
1830        req->flags |= REQ_F_COMPLETE_INLINE;
1831}
1832
1833static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1834                                     long res, unsigned cflags)
1835{
1836        if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1837                io_req_complete_state(req, res, cflags);
1838        else
1839                io_req_complete_post(req, res, cflags);
1840}
1841
1842static inline void io_req_complete(struct io_kiocb *req, long res)
1843{
1844        __io_req_complete(req, 0, res, 0);
1845}
1846
1847static void io_req_complete_failed(struct io_kiocb *req, long res)
1848{
1849        req_set_fail(req);
1850        io_req_complete_post(req, res, 0);
1851}
1852
1853static void io_req_complete_fail_submit(struct io_kiocb *req)
1854{
1855        /*
1856         * We don't submit, fail them all, for that replace hardlinks with
1857         * normal links. Extra REQ_F_LINK is tolerated.
1858         */
1859        req->flags &= ~REQ_F_HARDLINK;
1860        req->flags |= REQ_F_LINK;
1861        io_req_complete_failed(req, req->result);
1862}
1863
1864/*
1865 * Don't initialise the fields below on every allocation, but do that in
1866 * advance and keep them valid across allocations.
1867 */
1868static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1869{
1870        req->ctx = ctx;
1871        req->link = NULL;
1872        req->async_data = NULL;
1873        /* not necessary, but safer to zero */
1874        req->result = 0;
1875}
1876
1877static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1878                                        struct io_submit_state *state)
1879{
1880        spin_lock(&ctx->completion_lock);
1881        list_splice_init(&ctx->locked_free_list, &state->free_list);
1882        ctx->locked_free_nr = 0;
1883        spin_unlock(&ctx->completion_lock);
1884}
1885
1886/* Returns true IFF there are requests in the cache */
1887static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1888{
1889        struct io_submit_state *state = &ctx->submit_state;
1890        int nr;
1891
1892        /*
1893         * If we have more than a batch's worth of requests in our IRQ side
1894         * locked cache, grab the lock and move them over to our submission
1895         * side cache.
1896         */
1897        if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1898                io_flush_cached_locked_reqs(ctx, state);
1899
1900        nr = state->free_reqs;
1901        while (!list_empty(&state->free_list)) {
1902                struct io_kiocb *req = list_first_entry(&state->free_list,
1903                                        struct io_kiocb, inflight_entry);
1904
1905                list_del(&req->inflight_entry);
1906                state->reqs[nr++] = req;
1907                if (nr == ARRAY_SIZE(state->reqs))
1908                        break;
1909        }
1910
1911        state->free_reqs = nr;
1912        return nr != 0;
1913}
1914
1915/*
1916 * A request might get retired back into the request caches even before opcode
1917 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1918 * Because of that, io_alloc_req() should be called only under ->uring_lock
1919 * and with extra caution to not get a request that is still worked on.
1920 */
1921static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1922        __must_hold(&ctx->uring_lock)
1923{
1924        struct io_submit_state *state = &ctx->submit_state;
1925        gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1926        int ret, i;
1927
1928        BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1929
1930        if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1931                goto got_req;
1932
1933        ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1934                                    state->reqs);
1935
1936        /*
1937         * Bulk alloc is all-or-nothing. If we fail to get a batch,
1938         * retry single alloc to be on the safe side.
1939         */
1940        if (unlikely(ret <= 0)) {
1941                state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1942                if (!state->reqs[0])
1943                        return NULL;
1944                ret = 1;
1945        }
1946
1947        for (i = 0; i < ret; i++)
1948                io_preinit_req(state->reqs[i], ctx);
1949        state->free_reqs = ret;
1950got_req:
1951        state->free_reqs--;
1952        return state->reqs[state->free_reqs];
1953}
1954
1955static inline void io_put_file(struct file *file)
1956{
1957        if (file)
1958                fput(file);
1959}
1960
1961static void io_dismantle_req(struct io_kiocb *req)
1962{
1963        unsigned int flags = req->flags;
1964
1965        if (io_req_needs_clean(req))
1966                io_clean_op(req);
1967        if (!(flags & REQ_F_FIXED_FILE))
1968                io_put_file(req->file);
1969        if (req->fixed_rsrc_refs)
1970                percpu_ref_put(req->fixed_rsrc_refs);
1971        if (req->async_data) {
1972                kfree(req->async_data);
1973                req->async_data = NULL;
1974        }
1975}
1976
1977static void __io_free_req(struct io_kiocb *req)
1978{
1979        struct io_ring_ctx *ctx = req->ctx;
1980
1981        io_dismantle_req(req);
1982        io_put_task(req->task, 1);
1983
1984        spin_lock(&ctx->completion_lock);
1985        list_add(&req->inflight_entry, &ctx->locked_free_list);
1986        ctx->locked_free_nr++;
1987        spin_unlock(&ctx->completion_lock);
1988
1989        percpu_ref_put(&ctx->refs);
1990}
1991
1992static inline void io_remove_next_linked(struct io_kiocb *req)
1993{
1994        struct io_kiocb *nxt = req->link;
1995
1996        req->link = nxt->link;
1997        nxt->link = NULL;
1998}
1999
2000static bool io_kill_linked_timeout(struct io_kiocb *req)
2001        __must_hold(&req->ctx->completion_lock)
2002        __must_hold(&req->ctx->timeout_lock)
2003{
2004        struct io_kiocb *link = req->link;
2005
2006        if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2007                struct io_timeout_data *io = link->async_data;
2008
2009                io_remove_next_linked(req);
2010                link->timeout.head = NULL;
2011                if (hrtimer_try_to_cancel(&io->timer) != -1) {
2012                        list_del(&link->timeout.list);
2013                        io_cqring_fill_event(link->ctx, link->user_data,
2014                                             -ECANCELED, 0);
2015                        io_put_req_deferred(link);
2016                        return true;
2017                }
2018        }
2019        return false;
2020}
2021
2022static void io_fail_links(struct io_kiocb *req)
2023        __must_hold(&req->ctx->completion_lock)
2024{
2025        struct io_kiocb *nxt, *link = req->link;
2026
2027        req->link = NULL;
2028        while (link) {
2029                long res = -ECANCELED;
2030
2031                if (link->flags & REQ_F_FAIL)
2032                        res = link->result;
2033
2034                nxt = link->link;
2035                link->link = NULL;
2036
2037                trace_io_uring_fail_link(req, link);
2038                io_cqring_fill_event(link->ctx, link->user_data, res, 0);
2039                io_put_req_deferred(link);
2040                link = nxt;
2041        }
2042}
2043
2044static bool io_disarm_next(struct io_kiocb *req)
2045        __must_hold(&req->ctx->completion_lock)
2046{
2047        bool posted = false;
2048
2049        if (req->flags & REQ_F_ARM_LTIMEOUT) {
2050                struct io_kiocb *link = req->link;
2051
2052                req->flags &= ~REQ_F_ARM_LTIMEOUT;
2053                if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2054                        io_remove_next_linked(req);
2055                        io_cqring_fill_event(link->ctx, link->user_data,
2056                                             -ECANCELED, 0);
2057                        io_put_req_deferred(link);
2058                        posted = true;
2059                }
2060        } else if (req->flags & REQ_F_LINK_TIMEOUT) {
2061                struct io_ring_ctx *ctx = req->ctx;
2062
2063                spin_lock_irq(&ctx->timeout_lock);
2064                posted = io_kill_linked_timeout(req);
2065                spin_unlock_irq(&ctx->timeout_lock);
2066        }
2067        if (unlikely((req->flags & REQ_F_FAIL) &&
2068                     !(req->flags & REQ_F_HARDLINK))) {
2069                posted |= (req->link != NULL);
2070                io_fail_links(req);
2071        }
2072        return posted;
2073}
2074
2075static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2076{
2077        struct io_kiocb *nxt;
2078
2079        /*
2080         * If LINK is set, we have dependent requests in this chain. If we
2081         * didn't fail this request, queue the first one up, moving any other
2082         * dependencies to the next request. In case of failure, fail the rest
2083         * of the chain.
2084         */
2085        if (req->flags & IO_DISARM_MASK) {
2086                struct io_ring_ctx *ctx = req->ctx;
2087                bool posted;
2088
2089                spin_lock(&ctx->completion_lock);
2090                posted = io_disarm_next(req);
2091                if (posted)
2092                        io_commit_cqring(req->ctx);
2093                spin_unlock(&ctx->completion_lock);
2094                if (posted)
2095                        io_cqring_ev_posted(ctx);
2096        }
2097        nxt = req->link;
2098        req->link = NULL;
2099        return nxt;
2100}
2101
2102static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2103{
2104        if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2105                return NULL;
2106        return __io_req_find_next(req);
2107}
2108
2109static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
2110{
2111        if (!ctx)
2112                return;
2113        if (*locked) {
2114                if (ctx->submit_state.compl_nr)
2115                        io_submit_flush_completions(ctx);
2116                mutex_unlock(&ctx->uring_lock);
2117                *locked = false;
2118        }
2119        percpu_ref_put(&ctx->refs);
2120}
2121
2122static void tctx_task_work(struct callback_head *cb)
2123{
2124        bool locked = false;
2125        struct io_ring_ctx *ctx = NULL;
2126        struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2127                                                  task_work);
2128
2129        while (1) {
2130                struct io_wq_work_node *node;
2131
2132                if (!tctx->task_list.first && locked && ctx->submit_state.compl_nr)
2133                        io_submit_flush_completions(ctx);
2134
2135                spin_lock_irq(&tctx->task_lock);
2136                node = tctx->task_list.first;
2137                INIT_WQ_LIST(&tctx->task_list);
2138                if (!node)
2139                        tctx->task_running = false;
2140                spin_unlock_irq(&tctx->task_lock);
2141                if (!node)
2142                        break;
2143
2144                do {
2145                        struct io_wq_work_node *next = node->next;
2146                        struct io_kiocb *req = container_of(node, struct io_kiocb,
2147                                                            io_task_work.node);
2148
2149                        if (req->ctx != ctx) {
2150                                ctx_flush_and_put(ctx, &locked);
2151                                ctx = req->ctx;
2152                                /* if not contended, grab and improve batching */
2153                                locked = mutex_trylock(&ctx->uring_lock);
2154                                percpu_ref_get(&ctx->refs);
2155                        }
2156                        req->io_task_work.func(req, &locked);
2157                        node = next;
2158                } while (node);
2159
2160                cond_resched();
2161        }
2162
2163        ctx_flush_and_put(ctx, &locked);
2164}
2165
2166static void io_req_task_work_add(struct io_kiocb *req)
2167{
2168        struct task_struct *tsk = req->task;
2169        struct io_uring_task *tctx = tsk->io_uring;
2170        enum task_work_notify_mode notify;
2171        struct io_wq_work_node *node;
2172        unsigned long flags;
2173        bool running;
2174
2175        WARN_ON_ONCE(!tctx);
2176
2177        spin_lock_irqsave(&tctx->task_lock, flags);
2178        wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2179        running = tctx->task_running;
2180        if (!running)
2181                tctx->task_running = true;
2182        spin_unlock_irqrestore(&tctx->task_lock, flags);
2183
2184        /* task_work already pending, we're done */
2185        if (running)
2186                return;
2187
2188        /*
2189         * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2190         * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2191         * processing task_work. There's no reliable way to tell if TWA_RESUME
2192         * will do the job.
2193         */
2194        notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2195        if (!task_work_add(tsk, &tctx->task_work, notify)) {
2196                wake_up_process(tsk);
2197                return;
2198        }
2199
2200        spin_lock_irqsave(&tctx->task_lock, flags);
2201        tctx->task_running = false;
2202        node = tctx->task_list.first;
2203        INIT_WQ_LIST(&tctx->task_list);
2204        spin_unlock_irqrestore(&tctx->task_lock, flags);
2205
2206        while (node) {
2207                req = container_of(node, struct io_kiocb, io_task_work.node);
2208                node = node->next;
2209                if (llist_add(&req->io_task_work.fallback_node,
2210                              &req->ctx->fallback_llist))
2211                        schedule_delayed_work(&req->ctx->fallback_work, 1);
2212        }
2213}
2214
2215static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
2216{
2217        struct io_ring_ctx *ctx = req->ctx;
2218
2219        /* not needed for normal modes, but SQPOLL depends on it */
2220        io_tw_lock(ctx, locked);
2221        io_req_complete_failed(req, req->result);
2222}
2223
2224static void io_req_task_submit(struct io_kiocb *req, bool *locked)
2225{
2226        struct io_ring_ctx *ctx = req->ctx;
2227
2228        io_tw_lock(ctx, locked);
2229        /* req->task == current here, checking PF_EXITING is safe */
2230        if (likely(!(req->task->flags & PF_EXITING)))
2231                __io_queue_sqe(req);
2232        else
2233                io_req_complete_failed(req, -EFAULT);
2234}
2235
2236static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2237{
2238        req->result = ret;
2239        req->io_task_work.func = io_req_task_cancel;
2240        io_req_task_work_add(req);
2241}
2242
2243static void io_req_task_queue(struct io_kiocb *req)
2244{
2245        req->io_task_work.func = io_req_task_submit;
2246        io_req_task_work_add(req);
2247}
2248
2249static void io_req_task_queue_reissue(struct io_kiocb *req)
2250{
2251        req->io_task_work.func = io_queue_async_work;
2252        io_req_task_work_add(req);
2253}
2254
2255static inline void io_queue_next(struct io_kiocb *req)
2256{
2257        struct io_kiocb *nxt = io_req_find_next(req);
2258
2259        if (nxt)
2260                io_req_task_queue(nxt);
2261}
2262
2263static void io_free_req(struct io_kiocb *req)
2264{
2265        io_queue_next(req);
2266        __io_free_req(req);
2267}
2268
2269static void io_free_req_work(struct io_kiocb *req, bool *locked)
2270{
2271        io_free_req(req);
2272}
2273
2274struct req_batch {
2275        struct task_struct      *task;
2276        int                     task_refs;
2277        int                     ctx_refs;
2278};
2279
2280static inline void io_init_req_batch(struct req_batch *rb)
2281{
2282        rb->task_refs = 0;
2283        rb->ctx_refs = 0;
2284        rb->task = NULL;
2285}
2286
2287static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2288                                     struct req_batch *rb)
2289{
2290        if (rb->ctx_refs)
2291                percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2292        if (rb->task)
2293                io_put_task(rb->task, rb->task_refs);
2294}
2295
2296static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2297                              struct io_submit_state *state)
2298{
2299        io_queue_next(req);
2300        io_dismantle_req(req);
2301
2302        if (req->task != rb->task) {
2303                if (rb->task)
2304                        io_put_task(rb->task, rb->task_refs);
2305                rb->task = req->task;
2306                rb->task_refs = 0;
2307        }
2308        rb->task_refs++;
2309        rb->ctx_refs++;
2310
2311        if (state->free_reqs != ARRAY_SIZE(state->reqs))
2312                state->reqs[state->free_reqs++] = req;
2313        else
2314                list_add(&req->inflight_entry, &state->free_list);
2315}
2316
2317static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2318        __must_hold(&ctx->uring_lock)
2319{
2320        struct io_submit_state *state = &ctx->submit_state;
2321        int i, nr = state->compl_nr;
2322        struct req_batch rb;
2323
2324        spin_lock(&ctx->completion_lock);
2325        for (i = 0; i < nr; i++) {
2326                struct io_kiocb *req = state->compl_reqs[i];
2327
2328                __io_cqring_fill_event(ctx, req->user_data, req->result,
2329                                        req->compl.cflags);
2330        }
2331        io_commit_cqring(ctx);
2332        spin_unlock(&ctx->completion_lock);
2333        io_cqring_ev_posted(ctx);
2334
2335        io_init_req_batch(&rb);
2336        for (i = 0; i < nr; i++) {
2337                struct io_kiocb *req = state->compl_reqs[i];
2338
2339                if (req_ref_put_and_test(req))
2340                        io_req_free_batch(&rb, req, &ctx->submit_state);
2341        }
2342
2343        io_req_free_batch_finish(ctx, &rb);
2344        state->compl_nr = 0;
2345}
2346
2347/*
2348 * Drop reference to request, return next in chain (if there is one) if this
2349 * was the last reference to this request.
2350 */
2351static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2352{
2353        struct io_kiocb *nxt = NULL;
2354
2355        if (req_ref_put_and_test(req)) {
2356                nxt = io_req_find_next(req);
2357                __io_free_req(req);
2358        }
2359        return nxt;
2360}
2361
2362static inline void io_put_req(struct io_kiocb *req)
2363{
2364        if (req_ref_put_and_test(req))
2365                io_free_req(req);
2366}
2367
2368static inline void io_put_req_deferred(struct io_kiocb *req)
2369{
2370        if (req_ref_put_and_test(req)) {
2371                req->io_task_work.func = io_free_req_work;
2372                io_req_task_work_add(req);
2373        }
2374}
2375
2376static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2377{
2378        /* See comment at the top of this file */
2379        smp_rmb();
2380        return __io_cqring_events(ctx);
2381}
2382
2383static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2384{
2385        struct io_rings *rings = ctx->rings;
2386
2387        /* make sure SQ entry isn't read before tail */
2388        return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2389}
2390
2391static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2392{
2393        unsigned int cflags;
2394
2395        cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2396        cflags |= IORING_CQE_F_BUFFER;
2397        req->flags &= ~REQ_F_BUFFER_SELECTED;
2398        kfree(kbuf);
2399        return cflags;
2400}
2401
2402static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2403{
2404        struct io_buffer *kbuf;
2405
2406        if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2407                return 0;
2408        kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2409        return io_put_kbuf(req, kbuf);
2410}
2411
2412static inline bool io_run_task_work(void)
2413{
2414        if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2415                __set_current_state(TASK_RUNNING);
2416                tracehook_notify_signal();
2417                return true;
2418        }
2419
2420        return false;
2421}
2422
2423/*
2424 * Find and free completed poll iocbs
2425 */
2426static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2427                               struct list_head *done)
2428{
2429        struct req_batch rb;
2430        struct io_kiocb *req;
2431
2432        /* order with ->result store in io_complete_rw_iopoll() */
2433        smp_rmb();
2434
2435        io_init_req_batch(&rb);
2436        while (!list_empty(done)) {
2437                req = list_first_entry(done, struct io_kiocb, inflight_entry);
2438                list_del(&req->inflight_entry);
2439
2440                __io_cqring_fill_event(ctx, req->user_data, req->result,
2441                                        io_put_rw_kbuf(req));
2442                (*nr_events)++;
2443
2444                if (req_ref_put_and_test(req))
2445                        io_req_free_batch(&rb, req, &ctx->submit_state);
2446        }
2447
2448        io_commit_cqring(ctx);
2449        io_cqring_ev_posted_iopoll(ctx);
2450        io_req_free_batch_finish(ctx, &rb);
2451}
2452
2453static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2454                        long min)
2455{
2456        struct io_kiocb *req, *tmp;
2457        LIST_HEAD(done);
2458        bool spin;
2459
2460        /*
2461         * Only spin for completions if we don't have multiple devices hanging
2462         * off our complete list, and we're under the requested amount.
2463         */
2464        spin = !ctx->poll_multi_queue && *nr_events < min;
2465
2466        list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2467                struct kiocb *kiocb = &req->rw.kiocb;
2468                int ret;
2469
2470                /*
2471                 * Move completed and retryable entries to our local lists.
2472                 * If we find a request that requires polling, break out
2473                 * and complete those lists first, if we have entries there.
2474                 */
2475                if (READ_ONCE(req->iopoll_completed)) {
2476                        list_move_tail(&req->inflight_entry, &done);
2477                        continue;
2478                }
2479                if (!list_empty(&done))
2480                        break;
2481
2482                ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2483                if (unlikely(ret < 0))
2484                        return ret;
2485                else if (ret)
2486                        spin = false;
2487
2488                /* iopoll may have completed current req */
2489                if (READ_ONCE(req->iopoll_completed))
2490                        list_move_tail(&req->inflight_entry, &done);
2491        }
2492
2493        if (!list_empty(&done))
2494                io_iopoll_complete(ctx, nr_events, &done);
2495
2496        return 0;
2497}
2498
2499/*
2500 * We can't just wait for polled events to come to us, we have to actively
2501 * find and complete them.
2502 */
2503static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2504{
2505        if (!(ctx->flags & IORING_SETUP_IOPOLL))
2506                return;
2507
2508        mutex_lock(&ctx->uring_lock);
2509        while (!list_empty(&ctx->iopoll_list)) {
2510                unsigned int nr_events = 0;
2511
2512                io_do_iopoll(ctx, &nr_events, 0);
2513
2514                /* let it sleep and repeat later if can't complete a request */
2515                if (nr_events == 0)
2516                        break;
2517                /*
2518                 * Ensure we allow local-to-the-cpu processing to take place,
2519                 * in this case we need to ensure that we reap all events.
2520                 * Also let task_work, etc. to progress by releasing the mutex
2521                 */
2522                if (need_resched()) {
2523                        mutex_unlock(&ctx->uring_lock);
2524                        cond_resched();
2525                        mutex_lock(&ctx->uring_lock);
2526                }
2527        }
2528        mutex_unlock(&ctx->uring_lock);
2529}
2530
2531static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2532{
2533        unsigned int nr_events = 0;
2534        int ret = 0;
2535
2536        /*
2537         * We disallow the app entering submit/complete with polling, but we
2538         * still need to lock the ring to prevent racing with polled issue
2539         * that got punted to a workqueue.
2540         */
2541        mutex_lock(&ctx->uring_lock);
2542        /*
2543         * Don't enter poll loop if we already have events pending.
2544         * If we do, we can potentially be spinning for commands that
2545         * already triggered a CQE (eg in error).
2546         */
2547        if (test_bit(0, &ctx->check_cq_overflow))
2548                __io_cqring_overflow_flush(ctx, false);
2549        if (io_cqring_events(ctx))
2550                goto out;
2551        do {
2552                /*
2553                 * If a submit got punted to a workqueue, we can have the
2554                 * application entering polling for a command before it gets
2555                 * issued. That app will hold the uring_lock for the duration
2556                 * of the poll right here, so we need to take a breather every
2557                 * now and then to ensure that the issue has a chance to add
2558                 * the poll to the issued list. Otherwise we can spin here
2559                 * forever, while the workqueue is stuck trying to acquire the
2560                 * very same mutex.
2561                 */
2562                if (list_empty(&ctx->iopoll_list)) {
2563                        u32 tail = ctx->cached_cq_tail;
2564
2565                        mutex_unlock(&ctx->uring_lock);
2566                        io_run_task_work();
2567                        mutex_lock(&ctx->uring_lock);
2568
2569                        /* some requests don't go through iopoll_list */
2570                        if (tail != ctx->cached_cq_tail ||
2571                            list_empty(&ctx->iopoll_list))
2572                                break;
2573                }
2574                ret = io_do_iopoll(ctx, &nr_events, min);
2575        } while (!ret && nr_events < min && !need_resched());
2576out:
2577        mutex_unlock(&ctx->uring_lock);
2578        return ret;
2579}
2580
2581static void kiocb_end_write(struct io_kiocb *req)
2582{
2583        /*
2584         * Tell lockdep we inherited freeze protection from submission
2585         * thread.
2586         */
2587        if (req->flags & REQ_F_ISREG) {
2588                struct super_block *sb = file_inode(req->file)->i_sb;
2589
2590                __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2591                sb_end_write(sb);
2592        }
2593}
2594
2595#ifdef CONFIG_BLOCK
2596static bool io_resubmit_prep(struct io_kiocb *req)
2597{
2598        struct io_async_rw *rw = req->async_data;
2599
2600        if (!rw)
2601                return !io_req_prep_async(req);
2602        iov_iter_restore(&rw->iter, &rw->iter_state);
2603        return true;
2604}
2605
2606static bool io_rw_should_reissue(struct io_kiocb *req)
2607{
2608        umode_t mode = file_inode(req->file)->i_mode;
2609        struct io_ring_ctx *ctx = req->ctx;
2610
2611        if (!S_ISBLK(mode) && !S_ISREG(mode))
2612                return false;
2613        if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2614            !(ctx->flags & IORING_SETUP_IOPOLL)))
2615                return false;
2616        /*
2617         * If ref is dying, we might be running poll reap from the exit work.
2618         * Don't attempt to reissue from that path, just let it fail with
2619         * -EAGAIN.
2620         */
2621        if (percpu_ref_is_dying(&ctx->refs))
2622                return false;
2623        /*
2624         * Play it safe and assume not safe to re-import and reissue if we're
2625         * not in the original thread group (or in task context).
2626         */
2627        if (!same_thread_group(req->task, current) || !in_task())
2628                return false;
2629        return true;
2630}
2631#else
2632static bool io_resubmit_prep(struct io_kiocb *req)
2633{
2634        return false;
2635}
2636static bool io_rw_should_reissue(struct io_kiocb *req)
2637{
2638        return false;
2639}
2640#endif
2641
2642static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2643{
2644        if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2645                kiocb_end_write(req);
2646        if (res != req->result) {
2647                if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2648                    io_rw_should_reissue(req)) {
2649                        req->flags |= REQ_F_REISSUE;
2650                        return true;
2651                }
2652                req_set_fail(req);
2653                req->result = res;
2654        }
2655        return false;
2656}
2657
2658static void io_req_task_complete(struct io_kiocb *req, bool *locked)
2659{
2660        unsigned int cflags = io_put_rw_kbuf(req);
2661        long res = req->result;
2662
2663        if (*locked) {
2664                struct io_ring_ctx *ctx = req->ctx;
2665                struct io_submit_state *state = &ctx->submit_state;
2666
2667                io_req_complete_state(req, res, cflags);
2668                state->compl_reqs[state->compl_nr++] = req;
2669                if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
2670                        io_submit_flush_completions(ctx);
2671        } else {
2672                io_req_complete_post(req, res, cflags);
2673        }
2674}
2675
2676static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2677                             unsigned int issue_flags)
2678{
2679        if (__io_complete_rw_common(req, res))
2680                return;
2681        __io_req_complete(req, issue_flags, req->result, io_put_rw_kbuf(req));
2682}
2683
2684static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2685{
2686        struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2687
2688        if (__io_complete_rw_common(req, res))
2689                return;
2690        req->result = res;
2691        req->io_task_work.func = io_req_task_complete;
2692        io_req_task_work_add(req);
2693}
2694
2695static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2696{
2697        struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2698
2699        if (kiocb->ki_flags & IOCB_WRITE)
2700                kiocb_end_write(req);
2701        if (unlikely(res != req->result)) {
2702                if (res == -EAGAIN && io_rw_should_reissue(req)) {
2703                        req->flags |= REQ_F_REISSUE;
2704                        return;
2705                }
2706        }
2707
2708        WRITE_ONCE(req->result, res);
2709        /* order with io_iopoll_complete() checking ->result */
2710        smp_wmb();
2711        WRITE_ONCE(req->iopoll_completed, 1);
2712}
2713
2714/*
2715 * After the iocb has been issued, it's safe to be found on the poll list.
2716 * Adding the kiocb to the list AFTER submission ensures that we don't
2717 * find it from a io_do_iopoll() thread before the issuer is done
2718 * accessing the kiocb cookie.
2719 */
2720static void io_iopoll_req_issued(struct io_kiocb *req)
2721{
2722        struct io_ring_ctx *ctx = req->ctx;
2723        const bool in_async = io_wq_current_is_worker();
2724
2725        /* workqueue context doesn't hold uring_lock, grab it now */
2726        if (unlikely(in_async))
2727                mutex_lock(&ctx->uring_lock);
2728
2729        /*
2730         * Track whether we have multiple files in our lists. This will impact
2731         * how we do polling eventually, not spinning if we're on potentially
2732         * different devices.
2733         */
2734        if (list_empty(&ctx->iopoll_list)) {
2735                ctx->poll_multi_queue = false;
2736        } else if (!ctx->poll_multi_queue) {
2737                struct io_kiocb *list_req;
2738                unsigned int queue_num0, queue_num1;
2739
2740                list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2741                                                inflight_entry);
2742
2743                if (list_req->file != req->file) {
2744                        ctx->poll_multi_queue = true;
2745                } else {
2746                        queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2747                        queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2748                        if (queue_num0 != queue_num1)
2749                                ctx->poll_multi_queue = true;
2750                }
2751        }
2752
2753        /*
2754         * For fast devices, IO may have already completed. If it has, add
2755         * it to the front so we find it first.
2756         */
2757        if (READ_ONCE(req->iopoll_completed))
2758                list_add(&req->inflight_entry, &ctx->iopoll_list);
2759        else
2760                list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2761
2762        if (unlikely(in_async)) {
2763                /*
2764                 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2765                 * in sq thread task context or in io worker task context. If
2766                 * current task context is sq thread, we don't need to check
2767                 * whether should wake up sq thread.
2768                 */
2769                if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2770                    wq_has_sleeper(&ctx->sq_data->wait))
2771                        wake_up(&ctx->sq_data->wait);
2772
2773                mutex_unlock(&ctx->uring_lock);
2774        }
2775}
2776
2777static bool io_bdev_nowait(struct block_device *bdev)
2778{
2779        return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2780}
2781
2782/*
2783 * If we tracked the file through the SCM inflight mechanism, we could support
2784 * any file. For now, just ensure that anything potentially problematic is done
2785 * inline.
2786 */
2787static bool __io_file_supports_nowait(struct file *file, int rw)
2788{
2789        umode_t mode = file_inode(file)->i_mode;
2790
2791        if (S_ISBLK(mode)) {
2792                if (IS_ENABLED(CONFIG_BLOCK) &&
2793                    io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2794                        return true;
2795                return false;
2796        }
2797        if (S_ISSOCK(mode))
2798                return true;
2799        if (S_ISREG(mode)) {
2800                if (IS_ENABLED(CONFIG_BLOCK) &&
2801                    io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2802                    file->f_op != &io_uring_fops)
2803                        return true;
2804                return false;
2805        }
2806
2807        /* any ->read/write should understand O_NONBLOCK */
2808        if (file->f_flags & O_NONBLOCK)
2809                return true;
2810
2811        if (!(file->f_mode & FMODE_NOWAIT))
2812                return false;
2813
2814        if (rw == READ)
2815                return file->f_op->read_iter != NULL;
2816
2817        return file->f_op->write_iter != NULL;
2818}
2819
2820static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2821{
2822        if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2823                return true;
2824        else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2825                return true;
2826
2827        return __io_file_supports_nowait(req->file, rw);
2828}
2829
2830static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2831                      int rw)
2832{
2833        struct io_ring_ctx *ctx = req->ctx;
2834        struct kiocb *kiocb = &req->rw.kiocb;
2835        struct file *file = req->file;
2836        unsigned ioprio;
2837        int ret;
2838
2839        if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2840                req->flags |= REQ_F_ISREG;
2841
2842        kiocb->ki_pos = READ_ONCE(sqe->off);
2843        if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2844                req->flags |= REQ_F_CUR_POS;
2845                kiocb->ki_pos = file->f_pos;
2846        }
2847        kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2848        kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2849        ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2850        if (unlikely(ret))
2851                return ret;
2852
2853        /*
2854         * If the file is marked O_NONBLOCK, still allow retry for it if it
2855         * supports async. Otherwise it's impossible to use O_NONBLOCK files
2856         * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
2857         */
2858        if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2859            ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req, rw)))
2860                req->flags |= REQ_F_NOWAIT;
2861
2862        ioprio = READ_ONCE(sqe->ioprio);
2863        if (ioprio) {
2864                ret = ioprio_check_cap(ioprio);
2865                if (ret)
2866                        return ret;
2867
2868                kiocb->ki_ioprio = ioprio;
2869        } else
2870                kiocb->ki_ioprio = get_current_ioprio();
2871
2872        if (ctx->flags & IORING_SETUP_IOPOLL) {
2873                if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2874                    !kiocb->ki_filp->f_op->iopoll)
2875                        return -EOPNOTSUPP;
2876
2877                kiocb->ki_flags |= IOCB_HIPRI | IOCB_ALLOC_CACHE;
2878                kiocb->ki_complete = io_complete_rw_iopoll;
2879                req->iopoll_completed = 0;
2880        } else {
2881                if (kiocb->ki_flags & IOCB_HIPRI)
2882                        return -EINVAL;
2883                kiocb->ki_complete = io_complete_rw;
2884        }
2885
2886        if (req->opcode == IORING_OP_READ_FIXED ||
2887            req->opcode == IORING_OP_WRITE_FIXED) {
2888                req->imu = NULL;
2889                io_req_set_rsrc_node(req);
2890        }
2891
2892        req->rw.addr = READ_ONCE(sqe->addr);
2893        req->rw.len = READ_ONCE(sqe->len);
2894        req->buf_index = READ_ONCE(sqe->buf_index);
2895        return 0;
2896}
2897
2898static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2899{
2900        switch (ret) {
2901        case -EIOCBQUEUED:
2902                break;
2903        case -ERESTARTSYS:
2904        case -ERESTARTNOINTR:
2905        case -ERESTARTNOHAND:
2906        case -ERESTART_RESTARTBLOCK:
2907                /*
2908                 * We can't just restart the syscall, since previously
2909                 * submitted sqes may already be in progress. Just fail this
2910                 * IO with EINTR.
2911                 */
2912                ret = -EINTR;
2913                fallthrough;
2914        default:
2915                kiocb->ki_complete(kiocb, ret, 0);
2916        }
2917}
2918
2919static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2920                       unsigned int issue_flags)
2921{
2922        struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2923        struct io_async_rw *io = req->async_data;
2924
2925        /* add previously done IO, if any */
2926        if (io && io->bytes_done > 0) {
2927                if (ret < 0)
2928                        ret = io->bytes_done;
2929                else
2930                        ret += io->bytes_done;
2931        }
2932
2933        if (req->flags & REQ_F_CUR_POS)
2934                req->file->f_pos = kiocb->ki_pos;
2935        if (ret >= 0 && (kiocb->ki_complete == io_complete_rw))
2936                __io_complete_rw(req, ret, 0, issue_flags);
2937        else
2938                io_rw_done(kiocb, ret);
2939
2940        if (req->flags & REQ_F_REISSUE) {
2941                req->flags &= ~REQ_F_REISSUE;
2942                if (io_resubmit_prep(req)) {
2943                        io_req_task_queue_reissue(req);
2944                } else {
2945                        unsigned int cflags = io_put_rw_kbuf(req);
2946                        struct io_ring_ctx *ctx = req->ctx;
2947
2948                        req_set_fail(req);
2949                        if (!(issue_flags & IO_URING_F_NONBLOCK)) {
2950                                mutex_lock(&ctx->uring_lock);
2951                                __io_req_complete(req, issue_flags, ret, cflags);
2952                                mutex_unlock(&ctx->uring_lock);
2953                        } else {
2954                                __io_req_complete(req, issue_flags, ret, cflags);
2955                        }
2956                }
2957        }
2958}
2959
2960static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2961                             struct io_mapped_ubuf *imu)
2962{
2963        size_t len = req->rw.len;
2964        u64 buf_end, buf_addr = req->rw.addr;
2965        size_t offset;
2966
2967        if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2968                return -EFAULT;
2969        /* not inside the mapped region */
2970        if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2971                return -EFAULT;
2972
2973        /*
2974         * May not be a start of buffer, set size appropriately
2975         * and advance us to the beginning.
2976         */
2977        offset = buf_addr - imu->ubuf;
2978        iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2979
2980        if (offset) {
2981                /*
2982                 * Don't use iov_iter_advance() here, as it's really slow for
2983                 * using the latter parts of a big fixed buffer - it iterates
2984                 * over each segment manually. We can cheat a bit here, because
2985                 * we know that:
2986                 *
2987                 * 1) it's a BVEC iter, we set it up
2988                 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2989                 *    first and last bvec
2990                 *
2991                 * So just find our index, and adjust the iterator afterwards.
2992                 * If the offset is within the first bvec (or the whole first
2993                 * bvec, just use iov_iter_advance(). This makes it easier
2994                 * since we can just skip the first segment, which may not
2995                 * be PAGE_SIZE aligned.
2996                 */
2997                const struct bio_vec *bvec = imu->bvec;
2998
2999                if (offset <= bvec->bv_len) {
3000                        iov_iter_advance(iter, offset);
3001                } else {
3002                        unsigned long seg_skip;
3003
3004                        /* skip first vec */
3005                        offset -= bvec->bv_len;
3006                        seg_skip = 1 + (offset >> PAGE_SHIFT);
3007
3008                        iter->bvec = bvec + seg_skip;
3009                        iter->nr_segs -= seg_skip;
3010                        iter->count -= bvec->bv_len + offset;
3011                        iter->iov_offset = offset & ~PAGE_MASK;
3012                }
3013        }
3014
3015        return 0;
3016}
3017
3018static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
3019{
3020        struct io_ring_ctx *ctx = req->ctx;
3021        struct io_mapped_ubuf *imu = req->imu;
3022        u16 index, buf_index = req->buf_index;
3023
3024        if (likely(!imu)) {
3025                if (unlikely(buf_index >= ctx->nr_user_bufs))
3026                        return -EFAULT;
3027                index = array_index_nospec(buf_index, ctx->nr_user_bufs);
3028                imu = READ_ONCE(ctx->user_bufs[index]);
3029                req->imu = imu;
3030        }
3031        return __io_import_fixed(req, rw, iter, imu);
3032}
3033
3034static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3035{
3036        if (needs_lock)
3037                mutex_unlock(&ctx->uring_lock);
3038}
3039
3040static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3041{
3042        /*
3043         * "Normal" inline submissions always hold the uring_lock, since we
3044         * grab it from the system call. Same is true for the SQPOLL offload.
3045         * The only exception is when we've detached the request and issue it
3046         * from an async worker thread, grab the lock for that case.
3047         */
3048        if (needs_lock)
3049                mutex_lock(&ctx->uring_lock);
3050}
3051
3052static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3053                                          int bgid, struct io_buffer *kbuf,
3054                                          bool needs_lock)
3055{
3056        struct io_buffer *head;
3057
3058        if (req->flags & REQ_F_BUFFER_SELECTED)
3059                return kbuf;
3060
3061        io_ring_submit_lock(req->ctx, needs_lock);
3062
3063        lockdep_assert_held(&req->ctx->uring_lock);
3064
3065        head = xa_load(&req->ctx->io_buffers, bgid);
3066        if (head) {
3067                if (!list_empty(&head->list)) {
3068                        kbuf = list_last_entry(&head->list, struct io_buffer,
3069                                                        list);
3070                        list_del(&kbuf->list);
3071                } else {
3072                        kbuf = head;
3073                        xa_erase(&req->ctx->io_buffers, bgid);
3074                }
3075                if (*len > kbuf->len)
3076                        *len = kbuf->len;
3077        } else {
3078                kbuf = ERR_PTR(-ENOBUFS);
3079        }
3080
3081        io_ring_submit_unlock(req->ctx, needs_lock);
3082
3083        return kbuf;
3084}
3085
3086static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3087                                        bool needs_lock)
3088{
3089        struct io_buffer *kbuf;
3090        u16 bgid;
3091
3092        kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3093        bgid = req->buf_index;
3094        kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3095        if (IS_ERR(kbuf))
3096                return kbuf;
3097        req->rw.addr = (u64) (unsigned long) kbuf;
3098        req->flags |= REQ_F_BUFFER_SELECTED;
3099        return u64_to_user_ptr(kbuf->addr);
3100}
3101
3102#ifdef CONFIG_COMPAT
3103static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3104                                bool needs_lock)
3105{
3106        struct compat_iovec __user *uiov;
3107        compat_ssize_t clen;
3108        void __user *buf;
3109        ssize_t len;
3110
3111        uiov = u64_to_user_ptr(req->rw.addr);
3112        if (!access_ok(uiov, sizeof(*uiov)))
3113                return -EFAULT;
3114        if (__get_user(clen, &uiov->iov_len))
3115                return -EFAULT;
3116        if (clen < 0)
3117                return -EINVAL;
3118
3119        len = clen;
3120        buf = io_rw_buffer_select(req, &len, needs_lock);
3121        if (IS_ERR(buf))
3122                return PTR_ERR(buf);
3123        iov[0].iov_base = buf;
3124        iov[0].iov_len = (compat_size_t) len;
3125        return 0;
3126}
3127#endif
3128
3129static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3130                                      bool needs_lock)
3131{
3132        struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3133        void __user *buf;
3134        ssize_t len;
3135
3136        if (copy_from_user(iov, uiov, sizeof(*uiov)))
3137                return -EFAULT;
3138
3139        len = iov[0].iov_len;
3140        if (len < 0)
3141                return -EINVAL;
3142        buf = io_rw_buffer_select(req, &len, needs_lock);
3143        if (IS_ERR(buf))
3144                return PTR_ERR(buf);
3145        iov[0].iov_base = buf;
3146        iov[0].iov_len = len;
3147        return 0;
3148}
3149
3150static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3151                                    bool needs_lock)
3152{
3153        if (req->flags & REQ_F_BUFFER_SELECTED) {
3154                struct io_buffer *kbuf;
3155
3156                kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3157                iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3158                iov[0].iov_len = kbuf->len;
3159                return 0;
3160        }
3161        if (req->rw.len != 1)
3162                return -EINVAL;
3163
3164#ifdef CONFIG_COMPAT
3165        if (req->ctx->compat)
3166                return io_compat_import(req, iov, needs_lock);
3167#endif
3168
3169        return __io_iov_buffer_select(req, iov, needs_lock);
3170}
3171
3172static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3173                           struct iov_iter *iter, bool needs_lock)
3174{
3175        void __user *buf = u64_to_user_ptr(req->rw.addr);
3176        size_t sqe_len = req->rw.len;
3177        u8 opcode = req->opcode;
3178        ssize_t ret;
3179
3180        if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3181                *iovec = NULL;
3182                return io_import_fixed(req, rw, iter);
3183        }
3184
3185        /* buffer index only valid with fixed read/write, or buffer select  */
3186        if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3187                return -EINVAL;
3188
3189        if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3190                if (req->flags & REQ_F_BUFFER_SELECT) {
3191                        buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3192                        if (IS_ERR(buf))
3193                                return PTR_ERR(buf);
3194                        req->rw.len = sqe_len;
3195                }
3196
3197                ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3198                *iovec = NULL;
3199                return ret;
3200        }
3201
3202        if (req->flags & REQ_F_BUFFER_SELECT) {
3203                ret = io_iov_buffer_select(req, *iovec, needs_lock);
3204                if (!ret)
3205                        iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3206                *iovec = NULL;
3207                return ret;
3208        }
3209
3210        return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3211                              req->ctx->compat);
3212}
3213
3214static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3215{
3216        return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3217}
3218
3219/*
3220 * For files that don't have ->read_iter() and ->write_iter(), handle them
3221 * by looping over ->read() or ->write() manually.
3222 */
3223static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3224{
3225        struct kiocb *kiocb = &req->rw.kiocb;
3226        struct file *file = req->file;
3227        ssize_t ret = 0;
3228
3229        /*
3230         * Don't support polled IO through this interface, and we can't
3231         * support non-blocking either. For the latter, this just causes
3232         * the kiocb to be handled from an async context.
3233         */
3234        if (kiocb->ki_flags & IOCB_HIPRI)
3235                return -EOPNOTSUPP;
3236        if (kiocb->ki_flags & IOCB_NOWAIT)
3237                return -EAGAIN;
3238
3239        while (iov_iter_count(iter)) {
3240                struct iovec iovec;
3241                ssize_t nr;
3242
3243                if (!iov_iter_is_bvec(iter)) {
3244                        iovec = iov_iter_iovec(iter);
3245                } else {
3246                        iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3247                        iovec.iov_len = req->rw.len;
3248                }
3249
3250                if (rw == READ) {
3251                        nr = file->f_op->read(file, iovec.iov_base,
3252                                              iovec.iov_len, io_kiocb_ppos(kiocb));
3253                } else {
3254                        nr = file->f_op->write(file, iovec.iov_base,
3255                                               iovec.iov_len, io_kiocb_ppos(kiocb));
3256                }
3257
3258                if (nr < 0) {
3259                        if (!ret)
3260                                ret = nr;
3261                        break;
3262                }
3263                if (!iov_iter_is_bvec(iter)) {
3264                        iov_iter_advance(iter, nr);
3265                } else {
3266                        req->rw.len -= nr;
3267                        req->rw.addr += nr;
3268                }
3269                ret += nr;
3270                if (nr != iovec.iov_len)
3271                        break;
3272        }
3273
3274        return ret;
3275}
3276
3277static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3278                          const struct iovec *fast_iov, struct iov_iter *iter)
3279{
3280        struct io_async_rw *rw = req->async_data;
3281
3282        memcpy(&rw->iter, iter, sizeof(*iter));
3283        rw->free_iovec = iovec;
3284        rw->bytes_done = 0;
3285        /* can only be fixed buffers, no need to do anything */
3286        if (iov_iter_is_bvec(iter))
3287                return;
3288        if (!iovec) {
3289                unsigned iov_off = 0;
3290
3291                rw->iter.iov = rw->fast_iov;
3292                if (iter->iov != fast_iov) {
3293                        iov_off = iter->iov - fast_iov;
3294                        rw->iter.iov += iov_off;
3295                }
3296                if (rw->fast_iov != fast_iov)
3297                        memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3298                               sizeof(struct iovec) * iter->nr_segs);
3299        } else {
3300                req->flags |= REQ_F_NEED_CLEANUP;
3301        }
3302}
3303
3304static inline int io_alloc_async_data(struct io_kiocb *req)
3305{
3306        WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3307        req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3308        return req->async_data == NULL;
3309}
3310
3311static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3312                             const struct iovec *fast_iov,
3313                             struct iov_iter *iter, bool force)
3314{
3315        if (!force && !io_op_defs[req->opcode].needs_async_setup)
3316                return 0;
3317        if (!req->async_data) {
3318                struct io_async_rw *iorw;
3319
3320                if (io_alloc_async_data(req)) {
3321                        kfree(iovec);
3322                        return -ENOMEM;
3323                }
3324
3325                io_req_map_rw(req, iovec, fast_iov, iter);
3326                iorw = req->async_data;
3327                /* we've copied and mapped the iter, ensure state is saved */
3328                iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3329        }
3330        return 0;
3331}
3332
3333static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3334{
3335        struct io_async_rw *iorw = req->async_data;
3336        struct iovec *iov = iorw->fast_iov;
3337        int ret;
3338
3339        ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3340        if (unlikely(ret < 0))
3341                return ret;
3342
3343        iorw->bytes_done = 0;
3344        iorw->free_iovec = iov;
3345        if (iov)
3346                req->flags |= REQ_F_NEED_CLEANUP;
3347        iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3348        return 0;
3349}
3350
3351static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3352{
3353        if (unlikely(!(req->file->f_mode & FMODE_READ)))
3354                return -EBADF;
3355        return io_prep_rw(req, sqe, READ);
3356}
3357
3358/*
3359 * This is our waitqueue callback handler, registered through lock_page_async()
3360 * when we initially tried to do the IO with the iocb armed our waitqueue.
3361 * This gets called when the page is unlocked, and we generally expect that to
3362 * happen when the page IO is completed and the page is now uptodate. This will
3363 * queue a task_work based retry of the operation, attempting to copy the data
3364 * again. If the latter fails because the page was NOT uptodate, then we will
3365 * do a thread based blocking retry of the operation. That's the unexpected
3366 * slow path.
3367 */
3368static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3369                             int sync, void *arg)
3370{
3371        struct wait_page_queue *wpq;
3372        struct io_kiocb *req = wait->private;
3373        struct wait_page_key *key = arg;
3374
3375        wpq = container_of(wait, struct wait_page_queue, wait);
3376
3377        if (!wake_page_match(wpq, key))
3378                return 0;
3379
3380        req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3381        list_del_init(&wait->entry);
3382        io_req_task_queue(req);
3383        return 1;
3384}
3385
3386/*
3387 * This controls whether a given IO request should be armed for async page
3388 * based retry. If we return false here, the request is handed to the async
3389 * worker threads for retry. If we're doing buffered reads on a regular file,
3390 * we prepare a private wait_page_queue entry and retry the operation. This
3391 * will either succeed because the page is now uptodate and unlocked, or it
3392 * will register a callback when the page is unlocked at IO completion. Through
3393 * that callback, io_uring uses task_work to setup a retry of the operation.
3394 * That retry will attempt the buffered read again. The retry will generally
3395 * succeed, or in rare cases where it fails, we then fall back to using the
3396 * async worker threads for a blocking retry.
3397 */
3398static bool io_rw_should_retry(struct io_kiocb *req)
3399{
3400        struct io_async_rw *rw = req->async_data;
3401        struct wait_page_queue *wait = &rw->wpq;
3402        struct kiocb *kiocb = &req->rw.kiocb;
3403
3404        /* never retry for NOWAIT, we just complete with -EAGAIN */
3405        if (req->flags & REQ_F_NOWAIT)
3406                return false;
3407
3408        /* Only for buffered IO */
3409        if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3410                return false;
3411
3412        /*
3413         * just use poll if we can, and don't attempt if the fs doesn't
3414         * support callback based unlocks
3415         */
3416        if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3417                return false;
3418
3419        wait->wait.func = io_async_buf_func;
3420        wait->wait.private = req;
3421        wait->wait.flags = 0;
3422        INIT_LIST_HEAD(&wait->wait.entry);
3423        kiocb->ki_flags |= IOCB_WAITQ;
3424        kiocb->ki_flags &= ~IOCB_NOWAIT;
3425        kiocb->ki_waitq = wait;
3426        return true;
3427}
3428
3429static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3430{
3431        if (req->file->f_op->read_iter)
3432                return call_read_iter(req->file, &req->rw.kiocb, iter);
3433        else if (req->file->f_op->read)
3434                return loop_rw_iter(READ, req, iter);
3435        else
3436                return -EINVAL;
3437}
3438
3439static bool need_read_all(struct io_kiocb *req)
3440{
3441        return req->flags & REQ_F_ISREG ||
3442                S_ISBLK(file_inode(req->file)->i_mode);
3443}
3444
3445static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3446{
3447        struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3448        struct kiocb *kiocb = &req->rw.kiocb;
3449        struct iov_iter __iter, *iter = &__iter;
3450        struct io_async_rw *rw = req->async_data;
3451        bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3452        struct iov_iter_state __state, *state;
3453        ssize_t ret, ret2;
3454
3455        if (rw) {
3456                iter = &rw->iter;
3457                state = &rw->iter_state;
3458                /*
3459                 * We come here from an earlier attempt, restore our state to
3460                 * match in case it doesn't. It's cheap enough that we don't
3461                 * need to make this conditional.
3462                 */
3463                iov_iter_restore(iter, state);
3464                iovec = NULL;
3465        } else {
3466                ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3467                if (ret < 0)
3468                        return ret;
3469                state = &__state;
3470                iov_iter_save_state(iter, state);
3471        }
3472        req->result = iov_iter_count(iter);
3473
3474        /* Ensure we clear previously set non-block flag */
3475        if (!force_nonblock)
3476                kiocb->ki_flags &= ~IOCB_NOWAIT;
3477        else
3478                kiocb->ki_flags |= IOCB_NOWAIT;
3479
3480        /* If the file doesn't support async, just async punt */
3481        if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3482                ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3483                return ret ?: -EAGAIN;
3484        }
3485
3486        ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), req->result);
3487        if (unlikely(ret)) {
3488                kfree(iovec);
3489                return ret;
3490        }
3491
3492        ret = io_iter_do_read(req, iter);
3493
3494        if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3495                req->flags &= ~REQ_F_REISSUE;
3496                /* IOPOLL retry should happen for io-wq threads */
3497                if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3498                        goto done;
3499                /* no retry on NONBLOCK nor RWF_NOWAIT */
3500                if (req->flags & REQ_F_NOWAIT)
3501                        goto done;
3502                ret = 0;
3503        } else if (ret == -EIOCBQUEUED) {
3504                goto out_free;
3505        } else if (ret <= 0 || ret == req->result || !force_nonblock ||
3506                   (req->flags & REQ_F_NOWAIT) || !need_read_all(req)) {
3507                /* read all, failed, already did sync or don't want to retry */
3508                goto done;
3509        }
3510
3511        /*
3512         * Don't depend on the iter state matching what was consumed, or being
3513         * untouched in case of error. Restore it and we'll advance it
3514         * manually if we need to.
3515         */
3516        iov_iter_restore(iter, state);
3517
3518        ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3519        if (ret2)
3520                return ret2;
3521
3522        iovec = NULL;
3523        rw = req->async_data;
3524        /*
3525         * Now use our persistent iterator and state, if we aren't already.
3526         * We've restored and mapped the iter to match.
3527         */
3528        if (iter != &rw->iter) {
3529                iter = &rw->iter;
3530                state = &rw->iter_state;
3531        }
3532
3533        do {
3534                /*
3535                 * We end up here because of a partial read, either from
3536                 * above or inside this loop. Advance the iter by the bytes
3537                 * that were consumed.
3538                 */
3539                iov_iter_advance(iter, ret);
3540                if (!iov_iter_count(iter))
3541                        break;
3542                rw->bytes_done += ret;
3543                iov_iter_save_state(iter, state);
3544
3545                /* if we can retry, do so with the callbacks armed */
3546                if (!io_rw_should_retry(req)) {
3547                        kiocb->ki_flags &= ~IOCB_WAITQ;
3548                        return -EAGAIN;
3549                }
3550
3551                /*
3552                 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3553                 * we get -EIOCBQUEUED, then we'll get a notification when the
3554                 * desired page gets unlocked. We can also get a partial read
3555                 * here, and if we do, then just retry at the new offset.
3556                 */
3557                ret = io_iter_do_read(req, iter);
3558                if (ret == -EIOCBQUEUED)
3559                        return 0;
3560                /* we got some bytes, but not all. retry. */
3561                kiocb->ki_flags &= ~IOCB_WAITQ;
3562                iov_iter_restore(iter, state);
3563        } while (ret > 0);
3564done:
3565        kiocb_done(kiocb, ret, issue_flags);
3566out_free:
3567        /* it's faster to check here then delegate to kfree */
3568        if (iovec)
3569                kfree(iovec);
3570        return 0;
3571}
3572
3573static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3574{
3575        if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3576                return -EBADF;
3577        return io_prep_rw(req, sqe, WRITE);
3578}
3579
3580static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3581{
3582        struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3583        struct kiocb *kiocb = &req->rw.kiocb;
3584        struct iov_iter __iter, *iter = &__iter;
3585        struct io_async_rw *rw = req->async_data;
3586        bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3587        struct iov_iter_state __state, *state;
3588        ssize_t ret, ret2;
3589
3590        if (rw) {
3591                iter = &rw->iter;
3592                state = &rw->iter_state;
3593                iov_iter_restore(iter, state);
3594                iovec = NULL;
3595        } else {
3596                ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3597                if (ret < 0)
3598                        return ret;
3599                state = &__state;
3600                iov_iter_save_state(iter, state);
3601        }
3602        req->result = iov_iter_count(iter);
3603
3604        /* Ensure we clear previously set non-block flag */
3605        if (!force_nonblock)
3606                kiocb->ki_flags &= ~IOCB_NOWAIT;
3607        else
3608                kiocb->ki_flags |= IOCB_NOWAIT;
3609
3610        /* If the file doesn't support async, just async punt */
3611        if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3612                goto copy_iov;
3613
3614        /* file path doesn't support NOWAIT for non-direct_IO */
3615        if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3616            (req->flags & REQ_F_ISREG))
3617                goto copy_iov;
3618
3619        ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), req->result);
3620        if (unlikely(ret))
3621                goto out_free;
3622
3623        /*
3624         * Open-code file_start_write here to grab freeze protection,
3625         * which will be released by another thread in
3626         * io_complete_rw().  Fool lockdep by telling it the lock got
3627         * released so that it doesn't complain about the held lock when
3628         * we return to userspace.
3629         */
3630        if (req->flags & REQ_F_ISREG) {
3631                sb_start_write(file_inode(req->file)->i_sb);
3632                __sb_writers_release(file_inode(req->file)->i_sb,
3633                                        SB_FREEZE_WRITE);
3634        }
3635        kiocb->ki_flags |= IOCB_WRITE;
3636
3637        if (req->file->f_op->write_iter)
3638                ret2 = call_write_iter(req->file, kiocb, iter);
3639        else if (req->file->f_op->write)
3640                ret2 = loop_rw_iter(WRITE, req, iter);
3641        else
3642                ret2 = -EINVAL;
3643
3644        if (req->flags & REQ_F_REISSUE) {
3645                req->flags &= ~REQ_F_REISSUE;
3646                ret2 = -EAGAIN;
3647        }
3648
3649        /*
3650         * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3651         * retry them without IOCB_NOWAIT.
3652         */
3653        if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3654                ret2 = -EAGAIN;
3655        /* no retry on NONBLOCK nor RWF_NOWAIT */
3656        if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3657                goto done;
3658        if (!force_nonblock || ret2 != -EAGAIN) {
3659                /* IOPOLL retry should happen for io-wq threads */
3660                if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3661                        goto copy_iov;
3662done:
3663                kiocb_done(kiocb, ret2, issue_flags);
3664        } else {
3665copy_iov:
3666                iov_iter_restore(iter, state);
3667                ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3668                return ret ?: -EAGAIN;
3669        }
3670out_free:
3671        /* it's reportedly faster than delegating the null check to kfree() */
3672        if (iovec)
3673                kfree(iovec);
3674        return ret;
3675}
3676
3677static int io_renameat_prep(struct io_kiocb *req,
3678                            const struct io_uring_sqe *sqe)
3679{
3680        struct io_rename *ren = &req->rename;
3681        const char __user *oldf, *newf;
3682
3683        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3684                return -EINVAL;
3685        if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3686                return -EINVAL;
3687        if (unlikely(req->flags & REQ_F_FIXED_FILE))
3688                return -EBADF;
3689
3690        ren->old_dfd = READ_ONCE(sqe->fd);
3691        oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3692        newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3693        ren->new_dfd = READ_ONCE(sqe->len);
3694        ren->flags = READ_ONCE(sqe->rename_flags);
3695
3696        ren->oldpath = getname(oldf);
3697        if (IS_ERR(ren->oldpath))
3698                return PTR_ERR(ren->oldpath);
3699
3700        ren->newpath = getname(newf);
3701        if (IS_ERR(ren->newpath)) {
3702                putname(ren->oldpath);
3703                return PTR_ERR(ren->newpath);
3704        }
3705
3706        req->flags |= REQ_F_NEED_CLEANUP;
3707        return 0;
3708}
3709
3710static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3711{
3712        struct io_rename *ren = &req->rename;
3713        int ret;
3714
3715        if (issue_flags & IO_URING_F_NONBLOCK)
3716                return -EAGAIN;
3717
3718        ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3719                                ren->newpath, ren->flags);
3720
3721        req->flags &= ~REQ_F_NEED_CLEANUP;
3722        if (ret < 0)
3723                req_set_fail(req);
3724        io_req_complete(req, ret);
3725        return 0;
3726}
3727
3728static int io_unlinkat_prep(struct io_kiocb *req,
3729                            const struct io_uring_sqe *sqe)
3730{
3731        struct io_unlink *un = &req->unlink;
3732        const char __user *fname;
3733
3734        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3735                return -EINVAL;
3736        if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3737            sqe->splice_fd_in)
3738                return -EINVAL;
3739        if (unlikely(req->flags & REQ_F_FIXED_FILE))
3740                return -EBADF;
3741
3742        un->dfd = READ_ONCE(sqe->fd);
3743
3744        un->flags = READ_ONCE(sqe->unlink_flags);
3745        if (un->flags & ~AT_REMOVEDIR)
3746                return -EINVAL;
3747
3748        fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3749        un->filename = getname(fname);
3750        if (IS_ERR(un->filename))
3751                return PTR_ERR(un->filename);
3752
3753        req->flags |= REQ_F_NEED_CLEANUP;
3754        return 0;
3755}
3756
3757static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3758{
3759        struct io_unlink *un = &req->unlink;
3760        int ret;
3761
3762        if (issue_flags & IO_URING_F_NONBLOCK)
3763                return -EAGAIN;
3764
3765        if (un->flags & AT_REMOVEDIR)
3766                ret = do_rmdir(un->dfd, un->filename);
3767        else
3768                ret = do_unlinkat(un->dfd, un->filename);
3769
3770        req->flags &= ~REQ_F_NEED_CLEANUP;
3771        if (ret < 0)
3772                req_set_fail(req);
3773        io_req_complete(req, ret);
3774        return 0;
3775}
3776
3777static int io_mkdirat_prep(struct io_kiocb *req,
3778                            const struct io_uring_sqe *sqe)
3779{
3780        struct io_mkdir *mkd = &req->mkdir;
3781        const char __user *fname;
3782
3783        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3784                return -EINVAL;
3785        if (sqe->ioprio || sqe->off || sqe->rw_flags || sqe->buf_index ||
3786            sqe->splice_fd_in)
3787                return -EINVAL;
3788        if (unlikely(req->flags & REQ_F_FIXED_FILE))
3789                return -EBADF;
3790
3791        mkd->dfd = READ_ONCE(sqe->fd);
3792        mkd->mode = READ_ONCE(sqe->len);
3793
3794        fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3795        mkd->filename = getname(fname);
3796        if (IS_ERR(mkd->filename))
3797                return PTR_ERR(mkd->filename);
3798
3799        req->flags |= REQ_F_NEED_CLEANUP;
3800        return 0;
3801}
3802
3803static int io_mkdirat(struct io_kiocb *req, int issue_flags)
3804{
3805        struct io_mkdir *mkd = &req->mkdir;
3806        int ret;
3807
3808        if (issue_flags & IO_URING_F_NONBLOCK)
3809                return -EAGAIN;
3810
3811        ret = do_mkdirat(mkd->dfd, mkd->filename, mkd->mode);
3812
3813        req->flags &= ~REQ_F_NEED_CLEANUP;
3814        if (ret < 0)
3815                req_set_fail(req);
3816        io_req_complete(req, ret);
3817        return 0;
3818}
3819
3820static int io_symlinkat_prep(struct io_kiocb *req,
3821                            const struct io_uring_sqe *sqe)
3822{
3823        struct io_symlink *sl = &req->symlink;
3824        const char __user *oldpath, *newpath;
3825
3826        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3827                return -EINVAL;
3828        if (sqe->ioprio || sqe->len || sqe->rw_flags || sqe->buf_index ||
3829            sqe->splice_fd_in)
3830                return -EINVAL;
3831        if (unlikely(req->flags & REQ_F_FIXED_FILE))
3832                return -EBADF;
3833
3834        sl->new_dfd = READ_ONCE(sqe->fd);
3835        oldpath = u64_to_user_ptr(READ_ONCE(sqe->addr));
3836        newpath = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3837
3838        sl->oldpath = getname(oldpath);
3839        if (IS_ERR(sl->oldpath))
3840                return PTR_ERR(sl->oldpath);
3841
3842        sl->newpath = getname(newpath);
3843        if (IS_ERR(sl->newpath)) {
3844                putname(sl->oldpath);
3845                return PTR_ERR(sl->newpath);
3846        }
3847
3848        req->flags |= REQ_F_NEED_CLEANUP;
3849        return 0;
3850}
3851
3852static int io_symlinkat(struct io_kiocb *req, int issue_flags)
3853{
3854        struct io_symlink *sl = &req->symlink;
3855        int ret;
3856
3857        if (issue_flags & IO_URING_F_NONBLOCK)
3858                return -EAGAIN;
3859
3860        ret = do_symlinkat(sl->oldpath, sl->new_dfd, sl->newpath);
3861
3862        req->flags &= ~REQ_F_NEED_CLEANUP;
3863        if (ret < 0)
3864                req_set_fail(req);
3865        io_req_complete(req, ret);
3866        return 0;
3867}
3868
3869static int io_linkat_prep(struct io_kiocb *req,
3870                            const struct io_uring_sqe *sqe)
3871{
3872        struct io_hardlink *lnk = &req->hardlink;
3873        const char __user *oldf, *newf;
3874
3875        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3876                return -EINVAL;
3877        if (sqe->ioprio || sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
3878                return -EINVAL;
3879        if (unlikely(req->flags & REQ_F_FIXED_FILE))
3880                return -EBADF;
3881
3882        lnk->old_dfd = READ_ONCE(sqe->fd);
3883        lnk->new_dfd = READ_ONCE(sqe->len);
3884        oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3885        newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3886        lnk->flags = READ_ONCE(sqe->hardlink_flags);
3887
3888        lnk->oldpath = getname(oldf);
3889        if (IS_ERR(lnk->oldpath))
3890                return PTR_ERR(lnk->oldpath);
3891
3892        lnk->newpath = getname(newf);
3893        if (IS_ERR(lnk->newpath)) {
3894                putname(lnk->oldpath);
3895                return PTR_ERR(lnk->newpath);
3896        }
3897
3898        req->flags |= REQ_F_NEED_CLEANUP;
3899        return 0;
3900}
3901
3902static int io_linkat(struct io_kiocb *req, int issue_flags)
3903{
3904        struct io_hardlink *lnk = &req->hardlink;
3905        int ret;
3906
3907        if (issue_flags & IO_URING_F_NONBLOCK)
3908                return -EAGAIN;
3909
3910        ret = do_linkat(lnk->old_dfd, lnk->oldpath, lnk->new_dfd,
3911                                lnk->newpath, lnk->flags);
3912
3913        req->flags &= ~REQ_F_NEED_CLEANUP;
3914        if (ret < 0)
3915                req_set_fail(req);
3916        io_req_complete(req, ret);
3917        return 0;
3918}
3919
3920static int io_shutdown_prep(struct io_kiocb *req,
3921                            const struct io_uring_sqe *sqe)
3922{
3923#if defined(CONFIG_NET)
3924        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3925                return -EINVAL;
3926        if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3927                     sqe->buf_index || sqe->splice_fd_in))
3928                return -EINVAL;
3929
3930        req->shutdown.how = READ_ONCE(sqe->len);
3931        return 0;
3932#else
3933        return -EOPNOTSUPP;
3934#endif
3935}
3936
3937static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3938{
3939#if defined(CONFIG_NET)
3940        struct socket *sock;
3941        int ret;
3942
3943        if (issue_flags & IO_URING_F_NONBLOCK)
3944                return -EAGAIN;
3945
3946        sock = sock_from_file(req->file);
3947        if (unlikely(!sock))
3948                return -ENOTSOCK;
3949
3950        ret = __sys_shutdown_sock(sock, req->shutdown.how);
3951        if (ret < 0)
3952                req_set_fail(req);
3953        io_req_complete(req, ret);
3954        return 0;
3955#else
3956        return -EOPNOTSUPP;
3957#endif
3958}
3959
3960static int __io_splice_prep(struct io_kiocb *req,
3961                            const struct io_uring_sqe *sqe)
3962{
3963        struct io_splice *sp = &req->splice;
3964        unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3965
3966        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3967                return -EINVAL;
3968
3969        sp->file_in = NULL;
3970        sp->len = READ_ONCE(sqe->len);
3971        sp->flags = READ_ONCE(sqe->splice_flags);
3972
3973        if (unlikely(sp->flags & ~valid_flags))
3974                return -EINVAL;
3975
3976        sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3977                                  (sp->flags & SPLICE_F_FD_IN_FIXED));
3978        if (!sp->file_in)
3979                return -EBADF;
3980        req->flags |= REQ_F_NEED_CLEANUP;
3981        return 0;
3982}
3983
3984static int io_tee_prep(struct io_kiocb *req,
3985                       const struct io_uring_sqe *sqe)
3986{
3987        if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3988                return -EINVAL;
3989        return __io_splice_prep(req, sqe);
3990}
3991
3992static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3993{
3994        struct io_splice *sp = &req->splice;
3995        struct file *in = sp->file_in;
3996        struct file *out = sp->file_out;
3997        unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3998        long ret = 0;
3999
4000        if (issue_flags & IO_URING_F_NONBLOCK)
4001                return -EAGAIN;
4002        if (sp->len)
4003                ret = do_tee(in, out, sp->len, flags);
4004
4005        if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4006                io_put_file(in);
4007        req->flags &= ~REQ_F_NEED_CLEANUP;
4008
4009        if (ret != sp->len)
4010                req_set_fail(req);
4011        io_req_complete(req, ret);
4012        return 0;
4013}
4014
4015static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4016{
4017        struct io_splice *sp = &req->splice;
4018
4019        sp->off_in = READ_ONCE(sqe->splice_off_in);
4020        sp->off_out = READ_ONCE(sqe->off);
4021        return __io_splice_prep(req, sqe);
4022}
4023
4024static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
4025{
4026        struct io_splice *sp = &req->splice;
4027        struct file *in = sp->file_in;
4028        struct file *out = sp->file_out;
4029        unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4030        loff_t *poff_in, *poff_out;
4031        long ret = 0;
4032
4033        if (issue_flags & IO_URING_F_NONBLOCK)
4034                return -EAGAIN;
4035
4036        poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
4037        poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
4038
4039        if (sp->len)
4040                ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
4041
4042        if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4043                io_put_file(in);
4044        req->flags &= ~REQ_F_NEED_CLEANUP;
4045
4046        if (ret != sp->len)
4047                req_set_fail(req);
4048        io_req_complete(req, ret);
4049        return 0;
4050}
4051
4052/*
4053 * IORING_OP_NOP just posts a completion event, nothing else.
4054 */
4055static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4056{
4057        struct io_ring_ctx *ctx = req->ctx;
4058
4059        if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4060                return -EINVAL;
4061
4062        __io_req_complete(req, issue_flags, 0, 0);
4063        return 0;
4064}
4065
4066static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4067{
4068        struct io_ring_ctx *ctx = req->ctx;
4069
4070        if (!req->file)
4071                return -EBADF;
4072
4073        if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4074                return -EINVAL;
4075        if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4076                     sqe->splice_fd_in))
4077                return -EINVAL;
4078
4079        req->sync.flags = READ_ONCE(sqe->fsync_flags);
4080        if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4081                return -EINVAL;
4082
4083        req->sync.off = READ_ONCE(sqe->off);
4084        req->sync.len = READ_ONCE(sqe->len);
4085        return 0;
4086}
4087
4088static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4089{
4090        loff_t end = req->sync.off + req->sync.len;
4091        int ret;
4092
4093        /* fsync always requires a blocking context */
4094        if (issue_flags & IO_URING_F_NONBLOCK)
4095                return -EAGAIN;
4096
4097        ret = vfs_fsync_range(req->file, req->sync.off,
4098                                end > 0 ? end : LLONG_MAX,
4099                                req->sync.flags & IORING_FSYNC_DATASYNC);
4100        if (ret < 0)
4101                req_set_fail(req);
4102        io_req_complete(req, ret);
4103        return 0;
4104}
4105
4106static int io_fallocate_prep(struct io_kiocb *req,
4107                             const struct io_uring_sqe *sqe)
4108{
4109        if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
4110            sqe->splice_fd_in)
4111                return -EINVAL;
4112        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4113                return -EINVAL;
4114
4115        req->sync.off = READ_ONCE(sqe->off);
4116        req->sync.len = READ_ONCE(sqe->addr);
4117        req->sync.mode = READ_ONCE(sqe->len);
4118        return 0;
4119}
4120
4121static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4122{
4123        int ret;
4124
4125        /* fallocate always requiring blocking context */
4126        if (issue_flags & IO_URING_F_NONBLOCK)
4127                return -EAGAIN;
4128        ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4129                                req->sync.len);
4130        if (ret < 0)
4131                req_set_fail(req);
4132        io_req_complete(req, ret);
4133        return 0;
4134}
4135
4136static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4137{
4138        const char __user *fname;
4139        int ret;
4140
4141        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4142                return -EINVAL;
4143        if (unlikely(sqe->ioprio || sqe->buf_index))
4144                return -EINVAL;
4145        if (unlikely(req->flags & REQ_F_FIXED_FILE))
4146                return -EBADF;
4147
4148        /* open.how should be already initialised */
4149        if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4150                req->open.how.flags |= O_LARGEFILE;
4151
4152        req->open.dfd = READ_ONCE(sqe->fd);
4153        fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4154        req->open.filename = getname(fname);
4155        if (IS_ERR(req->open.filename)) {
4156                ret = PTR_ERR(req->open.filename);
4157                req->open.filename = NULL;
4158                return ret;
4159        }
4160
4161        req->open.file_slot = READ_ONCE(sqe->file_index);
4162        if (req->open.file_slot && (req->open.how.flags & O_CLOEXEC))
4163                return -EINVAL;
4164
4165        req->open.nofile = rlimit(RLIMIT_NOFILE);
4166        req->flags |= REQ_F_NEED_CLEANUP;
4167        return 0;
4168}
4169
4170static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4171{
4172        u64 mode = READ_ONCE(sqe->len);
4173        u64 flags = READ_ONCE(sqe->open_flags);
4174
4175        req->open.how = build_open_how(flags, mode);
4176        return __io_openat_prep(req, sqe);
4177}
4178
4179static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4180{
4181        struct open_how __user *how;
4182        size_t len;
4183        int ret;
4184
4185        how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4186        len = READ_ONCE(sqe->len);
4187        if (len < OPEN_HOW_SIZE_VER0)
4188                return -EINVAL;
4189
4190        ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4191                                        len);
4192        if (ret)
4193                return ret;
4194
4195        return __io_openat_prep(req, sqe);
4196}
4197
4198static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4199{
4200        struct open_flags op;
4201        struct file *file;
4202        bool resolve_nonblock, nonblock_set;
4203        bool fixed = !!req->open.file_slot;
4204        int ret;
4205
4206        ret = build_open_flags(&req->open.how, &op);
4207        if (ret)
4208                goto err;
4209        nonblock_set = op.open_flag & O_NONBLOCK;
4210        resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4211        if (issue_flags & IO_URING_F_NONBLOCK) {
4212                /*
4213                 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4214                 * it'll always -EAGAIN
4215                 */
4216                if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4217                        return -EAGAIN;
4218                op.lookup_flags |= LOOKUP_CACHED;
4219                op.open_flag |= O_NONBLOCK;
4220        }
4221
4222        if (!fixed) {
4223                ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4224                if (ret < 0)
4225                        goto err;
4226        }
4227
4228        file = do_filp_open(req->open.dfd, req->open.filename, &op);
4229        if (IS_ERR(file)) {
4230                /*
4231                 * We could hang on to this 'fd' on retrying, but seems like
4232                 * marginal gain for something that is now known to be a slower
4233                 * path. So just put it, and we'll get a new one when we retry.
4234                 */
4235                if (!fixed)
4236                        put_unused_fd(ret);
4237
4238                ret = PTR_ERR(file);
4239                /* only retry if RESOLVE_CACHED wasn't already set by application */
4240                if (ret == -EAGAIN &&
4241                    (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
4242                        return -EAGAIN;
4243                goto err;
4244        }
4245
4246        if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4247                file->f_flags &= ~O_NONBLOCK;
4248        fsnotify_open(file);
4249
4250        if (!fixed)
4251                fd_install(ret, file);
4252        else
4253                ret = io_install_fixed_file(req, file, issue_flags,
4254                                            req->open.file_slot - 1);
4255err:
4256        putname(req->open.filename);
4257        req->flags &= ~REQ_F_NEED_CLEANUP;
4258        if (ret < 0)
4259                req_set_fail(req);
4260        __io_req_complete(req, issue_flags, ret, 0);
4261        return 0;
4262}
4263
4264static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4265{
4266        return io_openat2(req, issue_flags);
4267}
4268
4269static int io_remove_buffers_prep(struct io_kiocb *req,
4270                                  const struct io_uring_sqe *sqe)
4271{
4272        struct io_provide_buf *p = &req->pbuf;
4273        u64 tmp;
4274
4275        if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
4276            sqe->splice_fd_in)
4277                return -EINVAL;
4278
4279        tmp = READ_ONCE(sqe->fd);
4280        if (!tmp || tmp > USHRT_MAX)
4281                return -EINVAL;
4282
4283        memset(p, 0, sizeof(*p));
4284        p->nbufs = tmp;
4285        p->bgid = READ_ONCE(sqe->buf_group);
4286        return 0;
4287}
4288
4289static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4290                               int bgid, unsigned nbufs)
4291{
4292        unsigned i = 0;
4293
4294        /* shouldn't happen */
4295        if (!nbufs)
4296                return 0;
4297
4298        /* the head kbuf is the list itself */
4299        while (!list_empty(&buf->list)) {
4300                struct io_buffer *nxt;
4301
4302                nxt = list_first_entry(&buf->list, struct io_buffer, list);
4303                list_del(&nxt->list);
4304                kfree(nxt);
4305                if (++i == nbufs)
4306                        return i;
4307        }
4308        i++;
4309        kfree(buf);
4310        xa_erase(&ctx->io_buffers, bgid);
4311
4312        return i;
4313}
4314
4315static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4316{
4317        struct io_provide_buf *p = &req->pbuf;
4318        struct io_ring_ctx *ctx = req->ctx;
4319        struct io_buffer *head;
4320        int ret = 0;
4321        bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4322
4323        io_ring_submit_lock(ctx, !force_nonblock);
4324
4325        lockdep_assert_held(&ctx->uring_lock);
4326
4327        ret = -ENOENT;
4328        head = xa_load(&ctx->io_buffers, p->bgid);
4329        if (head)
4330                ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4331        if (ret < 0)
4332                req_set_fail(req);
4333
4334        /* complete before unlock, IOPOLL may need the lock */
4335        __io_req_complete(req, issue_flags, ret, 0);
4336        io_ring_submit_unlock(ctx, !force_nonblock);
4337        return 0;
4338}
4339
4340static int io_provide_buffers_prep(struct io_kiocb *req,
4341                                   const struct io_uring_sqe *sqe)
4342{
4343        unsigned long size, tmp_check;
4344        struct io_provide_buf *p = &req->pbuf;
4345        u64 tmp;
4346
4347        if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4348                return -EINVAL;
4349
4350        tmp = READ_ONCE(sqe->fd);
4351        if (!tmp || tmp > USHRT_MAX)
4352                return -E2BIG;
4353        p->nbufs = tmp;
4354        p->addr = READ_ONCE(sqe->addr);
4355        p->len = READ_ONCE(sqe->len);
4356
4357        if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4358                                &size))
4359                return -EOVERFLOW;
4360        if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4361                return -EOVERFLOW;
4362
4363        size = (unsigned long)p->len * p->nbufs;
4364        if (!access_ok(u64_to_user_ptr(p->addr), size))
4365                return -EFAULT;
4366
4367        p->bgid = READ_ONCE(sqe->buf_group);
4368        tmp = READ_ONCE(sqe->off);
4369        if (tmp > USHRT_MAX)
4370                return -E2BIG;
4371        p->bid = tmp;
4372        return 0;
4373}
4374
4375static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4376{
4377        struct io_buffer *buf;
4378        u64 addr = pbuf->addr;
4379        int i, bid = pbuf->bid;
4380
4381        for (i = 0; i < pbuf->nbufs; i++) {
4382                buf = kmalloc(sizeof(*buf), GFP_KERNEL_ACCOUNT);
4383                if (!buf)
4384                        break;
4385
4386                buf->addr = addr;
4387                buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4388                buf->bid = bid;
4389                addr += pbuf->len;
4390                bid++;
4391                if (!*head) {
4392                        INIT_LIST_HEAD(&buf->list);
4393                        *head = buf;
4394                } else {
4395                        list_add_tail(&buf->list, &(*head)->list);
4396                }
4397        }
4398
4399        return i ? i : -ENOMEM;
4400}
4401
4402static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4403{
4404        struct io_provide_buf *p = &req->pbuf;
4405        struct io_ring_ctx *ctx = req->ctx;
4406        struct io_buffer *head, *list;
4407        int ret = 0;
4408        bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4409
4410        io_ring_submit_lock(ctx, !force_nonblock);
4411
4412        lockdep_assert_held(&ctx->uring_lock);
4413
4414        list = head = xa_load(&ctx->io_buffers, p->bgid);
4415
4416        ret = io_add_buffers(p, &head);
4417        if (ret >= 0 && !list) {
4418                ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4419                if (ret < 0)
4420                        __io_remove_buffers(ctx, head, p->bgid, -1U);
4421        }
4422        if (ret < 0)
4423                req_set_fail(req);
4424        /* complete before unlock, IOPOLL may need the lock */
4425        __io_req_complete(req, issue_flags, ret, 0);
4426        io_ring_submit_unlock(ctx, !force_nonblock);
4427        return 0;
4428}
4429
4430static int io_epoll_ctl_prep(struct io_kiocb *req,
4431                             const struct io_uring_sqe *sqe)
4432{
4433#if defined(CONFIG_EPOLL)
4434        if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4435                return -EINVAL;
4436        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4437                return -EINVAL;
4438
4439        req->epoll.epfd = READ_ONCE(sqe->fd);
4440        req->epoll.op = READ_ONCE(sqe->len);
4441        req->epoll.fd = READ_ONCE(sqe->off);
4442
4443        if (ep_op_has_event(req->epoll.op)) {
4444                struct epoll_event __user *ev;
4445
4446                ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4447                if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4448                        return -EFAULT;
4449        }
4450
4451        return 0;
4452#else
4453        return -EOPNOTSUPP;
4454#endif
4455}
4456
4457static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4458{
4459#if defined(CONFIG_EPOLL)
4460        struct io_epoll *ie = &req->epoll;
4461        int ret;
4462        bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4463
4464        ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4465        if (force_nonblock && ret == -EAGAIN)
4466                return -EAGAIN;
4467
4468        if (ret < 0)
4469                req_set_fail(req);
4470        __io_req_complete(req, issue_flags, ret, 0);
4471        return 0;
4472#else
4473        return -EOPNOTSUPP;
4474#endif
4475}
4476
4477static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4478{
4479#if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4480        if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4481                return -EINVAL;
4482        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4483                return -EINVAL;
4484
4485        req->madvise.addr = READ_ONCE(sqe->addr);
4486        req->madvise.len = READ_ONCE(sqe->len);
4487        req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4488        return 0;
4489#else
4490        return -EOPNOTSUPP;
4491#endif
4492}
4493
4494static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4495{
4496#if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4497        struct io_madvise *ma = &req->madvise;
4498        int ret;
4499
4500        if (issue_flags & IO_URING_F_NONBLOCK)
4501                return -EAGAIN;
4502
4503        ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4504        if (ret < 0)
4505                req_set_fail(req);
4506        io_req_complete(req, ret);
4507        return 0;
4508#else
4509        return -EOPNOTSUPP;
4510#endif
4511}
4512
4513static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4514{
4515        if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4516                return -EINVAL;
4517        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4518                return -EINVAL;
4519
4520        req->fadvise.offset = READ_ONCE(sqe->off);
4521        req->fadvise.len = READ_ONCE(sqe->len);
4522        req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4523        return 0;
4524}
4525
4526static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4527{
4528        struct io_fadvise *fa = &req->fadvise;
4529        int ret;
4530
4531        if (issue_flags & IO_URING_F_NONBLOCK) {
4532                switch (fa->advice) {
4533                case POSIX_FADV_NORMAL:
4534                case POSIX_FADV_RANDOM:
4535                case POSIX_FADV_SEQUENTIAL:
4536                        break;
4537                default:
4538                        return -EAGAIN;
4539                }
4540        }
4541
4542        ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4543        if (ret < 0)
4544                req_set_fail(req);
4545        __io_req_complete(req, issue_flags, ret, 0);
4546        return 0;
4547}
4548
4549static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4550{
4551        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4552                return -EINVAL;
4553        if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4554                return -EINVAL;
4555        if (req->flags & REQ_F_FIXED_FILE)
4556                return -EBADF;
4557
4558        req->statx.dfd = READ_ONCE(sqe->fd);
4559        req->statx.mask = READ_ONCE(sqe->len);
4560        req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4561        req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4562        req->statx.flags = READ_ONCE(sqe->statx_flags);
4563
4564        return 0;
4565}
4566
4567static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4568{
4569        struct io_statx *ctx = &req->statx;
4570        int ret;
4571
4572        if (issue_flags & IO_URING_F_NONBLOCK)
4573                return -EAGAIN;
4574
4575        ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4576                       ctx->buffer);
4577
4578        if (ret < 0)
4579                req_set_fail(req);
4580        io_req_complete(req, ret);
4581        return 0;
4582}
4583
4584static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4585{
4586        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4587                return -EINVAL;
4588        if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4589            sqe->rw_flags || sqe->buf_index)
4590                return -EINVAL;
4591        if (req->flags & REQ_F_FIXED_FILE)
4592                return -EBADF;
4593
4594        req->close.fd = READ_ONCE(sqe->fd);
4595        req->close.file_slot = READ_ONCE(sqe->file_index);
4596        if (req->close.file_slot && req->close.fd)
4597                return -EINVAL;
4598
4599        return 0;
4600}
4601
4602static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4603{
4604        struct files_struct *files = current->files;
4605        struct io_close *close = &req->close;
4606        struct fdtable *fdt;
4607        struct file *file = NULL;
4608        int ret = -EBADF;
4609
4610        if (req->close.file_slot) {
4611                ret = io_close_fixed(req, issue_flags);
4612                goto err;
4613        }
4614
4615        spin_lock(&files->file_lock);
4616        fdt = files_fdtable(files);
4617        if (close->fd >= fdt->max_fds) {
4618                spin_unlock(&files->file_lock);
4619                goto err;
4620        }
4621        file = fdt->fd[close->fd];
4622        if (!file || file->f_op == &io_uring_fops) {
4623                spin_unlock(&files->file_lock);
4624                file = NULL;
4625                goto err;
4626        }
4627
4628        /* if the file has a flush method, be safe and punt to async */
4629        if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4630                spin_unlock(&files->file_lock);
4631                return -EAGAIN;
4632        }
4633
4634        ret = __close_fd_get_file(close->fd, &file);
4635        spin_unlock(&files->file_lock);
4636        if (ret < 0) {
4637                if (ret == -ENOENT)
4638                        ret = -EBADF;
4639                goto err;
4640        }
4641
4642        /* No ->flush() or already async, safely close from here */
4643        ret = filp_close(file, current->files);
4644err:
4645        if (ret < 0)
4646                req_set_fail(req);
4647        if (file)
4648                fput(file);
4649        __io_req_complete(req, issue_flags, ret, 0);
4650        return 0;
4651}
4652
4653static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4654{
4655        struct io_ring_ctx *ctx = req->ctx;
4656
4657        if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4658                return -EINVAL;
4659        if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4660                     sqe->splice_fd_in))
4661                return -EINVAL;
4662
4663        req->sync.off = READ_ONCE(sqe->off);
4664        req->sync.len = READ_ONCE(sqe->len);
4665        req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4666        return 0;
4667}
4668
4669static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4670{
4671        int ret;
4672
4673        /* sync_file_range always requires a blocking context */
4674        if (issue_flags & IO_URING_F_NONBLOCK)
4675                return -EAGAIN;
4676
4677        ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4678                                req->sync.flags);
4679        if (ret < 0)
4680                req_set_fail(req);
4681        io_req_complete(req, ret);
4682        return 0;
4683}
4684
4685#if defined(CONFIG_NET)
4686static int io_setup_async_msg(struct io_kiocb *req,
4687                              struct io_async_msghdr *kmsg)
4688{
4689        struct io_async_msghdr *async_msg = req->async_data;
4690
4691        if (async_msg)
4692                return -EAGAIN;
4693        if (io_alloc_async_data(req)) {
4694                kfree(kmsg->free_iov);
4695                return -ENOMEM;
4696        }
4697        async_msg = req->async_data;
4698        req->flags |= REQ_F_NEED_CLEANUP;
4699        memcpy(async_msg, kmsg, sizeof(*kmsg));
4700        async_msg->msg.msg_name = &async_msg->addr;
4701        /* if were using fast_iov, set it to the new one */
4702        if (!async_msg->free_iov)
4703                async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4704
4705        return -EAGAIN;
4706}
4707
4708static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4709                               struct io_async_msghdr *iomsg)
4710{
4711        iomsg->msg.msg_name = &iomsg->addr;
4712        iomsg->free_iov = iomsg->fast_iov;
4713        return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4714                                   req->sr_msg.msg_flags, &iomsg->free_iov);
4715}
4716
4717static int io_sendmsg_prep_async(struct io_kiocb *req)
4718{
4719        int ret;
4720
4721        ret = io_sendmsg_copy_hdr(req, req->async_data);
4722        if (!ret)
4723                req->flags |= REQ_F_NEED_CLEANUP;
4724        return ret;
4725}
4726
4727static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4728{
4729        struct io_sr_msg *sr = &req->sr_msg;
4730
4731        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4732                return -EINVAL;
4733
4734        sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4735        sr->len = READ_ONCE(sqe->len);
4736        sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4737        if (sr->msg_flags & MSG_DONTWAIT)
4738                req->flags |= REQ_F_NOWAIT;
4739
4740#ifdef CONFIG_COMPAT
4741        if (req->ctx->compat)
4742                sr->msg_flags |= MSG_CMSG_COMPAT;
4743#endif
4744        return 0;
4745}
4746
4747static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4748{
4749        struct io_async_msghdr iomsg, *kmsg;
4750        struct socket *sock;
4751        unsigned flags;
4752        int min_ret = 0;
4753        int ret;
4754
4755        sock = sock_from_file(req->file);
4756        if (unlikely(!sock))
4757                return -ENOTSOCK;
4758
4759        kmsg = req->async_data;
4760        if (!kmsg) {
4761                ret = io_sendmsg_copy_hdr(req, &iomsg);
4762                if (ret)
4763                        return ret;
4764                kmsg = &iomsg;
4765        }
4766
4767        flags = req->sr_msg.msg_flags;
4768        if (issue_flags & IO_URING_F_NONBLOCK)
4769                flags |= MSG_DONTWAIT;
4770        if (flags & MSG_WAITALL)
4771                min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4772
4773        ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4774        if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4775                return io_setup_async_msg(req, kmsg);
4776        if (ret == -ERESTARTSYS)
4777                ret = -EINTR;
4778
4779        /* fast path, check for non-NULL to avoid function call */
4780        if (kmsg->free_iov)
4781                kfree(kmsg->free_iov);
4782        req->flags &= ~REQ_F_NEED_CLEANUP;
4783        if (ret < min_ret)
4784                req_set_fail(req);
4785        __io_req_complete(req, issue_flags, ret, 0);
4786        return 0;
4787}
4788
4789static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4790{
4791        struct io_sr_msg *sr = &req->sr_msg;
4792        struct msghdr msg;
4793        struct iovec iov;
4794        struct socket *sock;
4795        unsigned flags;
4796        int min_ret = 0;
4797        int ret;
4798
4799        sock = sock_from_file(req->file);
4800        if (unlikely(!sock))
4801                return -ENOTSOCK;
4802
4803        ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4804        if (unlikely(ret))
4805                return ret;
4806
4807        msg.msg_name = NULL;
4808        msg.msg_control = NULL;
4809        msg.msg_controllen = 0;
4810        msg.msg_namelen = 0;
4811
4812        flags = req->sr_msg.msg_flags;
4813        if (issue_flags & IO_URING_F_NONBLOCK)
4814                flags |= MSG_DONTWAIT;
4815        if (flags & MSG_WAITALL)
4816                min_ret = iov_iter_count(&msg.msg_iter);
4817
4818        msg.msg_flags = flags;
4819        ret = sock_sendmsg(sock, &msg);
4820        if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4821                return -EAGAIN;
4822        if (ret == -ERESTARTSYS)
4823                ret = -EINTR;
4824
4825        if (ret < min_ret)
4826                req_set_fail(req);
4827        __io_req_complete(req, issue_flags, ret, 0);
4828        return 0;
4829}
4830
4831static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4832                                 struct io_async_msghdr *iomsg)
4833{
4834        struct io_sr_msg *sr = &req->sr_msg;
4835        struct iovec __user *uiov;
4836        size_t iov_len;
4837        int ret;
4838
4839        ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4840                                        &iomsg->uaddr, &uiov, &iov_len);
4841        if (ret)
4842                return ret;
4843
4844        if (req->flags & REQ_F_BUFFER_SELECT) {
4845                if (iov_len > 1)
4846                        return -EINVAL;
4847                if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4848                        return -EFAULT;
4849                sr->len = iomsg->fast_iov[0].iov_len;
4850                iomsg->free_iov = NULL;
4851        } else {
4852                iomsg->free_iov = iomsg->fast_iov;
4853                ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4854                                     &iomsg->free_iov, &iomsg->msg.msg_iter,
4855                                     false);
4856                if (ret > 0)
4857                        ret = 0;
4858        }
4859
4860        return ret;
4861}
4862
4863#ifdef CONFIG_COMPAT
4864static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4865                                        struct io_async_msghdr *iomsg)
4866{
4867        struct io_sr_msg *sr = &req->sr_msg;
4868        struct compat_iovec __user *uiov;
4869        compat_uptr_t ptr;
4870        compat_size_t len;
4871        int ret;
4872
4873        ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4874                                  &ptr, &len);
4875        if (ret)
4876                return ret;
4877
4878        uiov = compat_ptr(ptr);
4879        if (req->flags & REQ_F_BUFFER_SELECT) {
4880                compat_ssize_t clen;
4881
4882                if (len > 1)
4883                        return -EINVAL;
4884                if (!access_ok(uiov, sizeof(*uiov)))
4885                        return -EFAULT;
4886                if (__get_user(clen, &uiov->iov_len))
4887                        return -EFAULT;
4888                if (clen < 0)
4889                        return -EINVAL;
4890                sr->len = clen;
4891                iomsg->free_iov = NULL;
4892        } else {
4893                iomsg->free_iov = iomsg->fast_iov;
4894                ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4895                                   UIO_FASTIOV, &iomsg->free_iov,
4896                                   &iomsg->msg.msg_iter, true);
4897                if (ret < 0)
4898                        return ret;
4899        }
4900
4901        return 0;
4902}
4903#endif
4904
4905static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4906                               struct io_async_msghdr *iomsg)
4907{
4908        iomsg->msg.msg_name = &iomsg->addr;
4909
4910#ifdef CONFIG_COMPAT
4911        if (req->ctx->compat)
4912                return __io_compat_recvmsg_copy_hdr(req, iomsg);
4913#endif
4914
4915        return __io_recvmsg_copy_hdr(req, iomsg);
4916}
4917
4918static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4919                                               bool needs_lock)
4920{
4921        struct io_sr_msg *sr = &req->sr_msg;
4922        struct io_buffer *kbuf;
4923
4924        kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4925        if (IS_ERR(kbuf))
4926                return kbuf;
4927
4928        sr->kbuf = kbuf;
4929        req->flags |= REQ_F_BUFFER_SELECTED;
4930        return kbuf;
4931}
4932
4933static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4934{
4935        return io_put_kbuf(req, req->sr_msg.kbuf);
4936}
4937
4938static int io_recvmsg_prep_async(struct io_kiocb *req)
4939{
4940        int ret;
4941
4942        ret = io_recvmsg_copy_hdr(req, req->async_data);
4943        if (!ret)
4944                req->flags |= REQ_F_NEED_CLEANUP;
4945        return ret;
4946}
4947
4948static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4949{
4950        struct io_sr_msg *sr = &req->sr_msg;
4951
4952        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4953                return -EINVAL;
4954
4955        sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4956        sr->len = READ_ONCE(sqe->len);
4957        sr->bgid = READ_ONCE(sqe->buf_group);
4958        sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4959        if (sr->msg_flags & MSG_DONTWAIT)
4960                req->flags |= REQ_F_NOWAIT;
4961
4962#ifdef CONFIG_COMPAT
4963        if (req->ctx->compat)
4964                sr->msg_flags |= MSG_CMSG_COMPAT;
4965#endif
4966        return 0;
4967}
4968
4969static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4970{
4971        struct io_async_msghdr iomsg, *kmsg;
4972        struct socket *sock;
4973        struct io_buffer *kbuf;
4974        unsigned flags;
4975        int min_ret = 0;
4976        int ret, cflags = 0;
4977        bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4978
4979        sock = sock_from_file(req->file);
4980        if (unlikely(!sock))
4981                return -ENOTSOCK;
4982
4983        kmsg = req->async_data;
4984        if (!kmsg) {
4985                ret = io_recvmsg_copy_hdr(req, &iomsg);
4986                if (ret)
4987                        return ret;
4988                kmsg = &iomsg;
4989        }
4990
4991        if (req->flags & REQ_F_BUFFER_SELECT) {
4992                kbuf = io_recv_buffer_select(req, !force_nonblock);
4993                if (IS_ERR(kbuf))
4994                        return PTR_ERR(kbuf);
4995                kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4996                kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4997                iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4998                                1, req->sr_msg.len);
4999        }
5000
5001        flags = req->sr_msg.msg_flags;
5002        if (force_nonblock)
5003                flags |= MSG_DONTWAIT;
5004        if (flags & MSG_WAITALL)
5005                min_ret = iov_iter_count(&kmsg->msg.msg_iter);
5006
5007        ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
5008                                        kmsg->uaddr, flags);
5009        if (force_nonblock && ret == -EAGAIN)
5010                return io_setup_async_msg(req, kmsg);
5011        if (ret == -ERESTARTSYS)
5012                ret = -EINTR;
5013
5014        if (req->flags & REQ_F_BUFFER_SELECTED)
5015                cflags = io_put_recv_kbuf(req);
5016        /* fast path, check for non-NULL to avoid function call */
5017        if (kmsg->free_iov)
5018                kfree(kmsg->free_iov);
5019        req->flags &= ~REQ_F_NEED_CLEANUP;
5020        if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5021                req_set_fail(req);
5022        __io_req_complete(req, issue_flags, ret, cflags);
5023        return 0;
5024}
5025
5026static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
5027{
5028        struct io_buffer *kbuf;
5029        struct io_sr_msg *sr = &req->sr_msg;
5030        struct msghdr msg;
5031        void __user *buf = sr->buf;
5032        struct socket *sock;
5033        struct iovec iov;
5034        unsigned flags;
5035        int min_ret = 0;
5036        int ret, cflags = 0;
5037        bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5038
5039        sock = sock_from_file(req->file);
5040        if (unlikely(!sock))
5041                return -ENOTSOCK;
5042
5043        if (req->flags & REQ_F_BUFFER_SELECT) {
5044                kbuf = io_recv_buffer_select(req, !force_nonblock);
5045                if (IS_ERR(kbuf))
5046                        return PTR_ERR(kbuf);
5047                buf = u64_to_user_ptr(kbuf->addr);
5048        }
5049
5050        ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5051        if (unlikely(ret))
5052                goto out_free;
5053
5054        msg.msg_name = NULL;
5055        msg.msg_control = NULL;
5056        msg.msg_controllen = 0;
5057        msg.msg_namelen = 0;
5058        msg.msg_iocb = NULL;
5059        msg.msg_flags = 0;
5060
5061        flags = req->sr_msg.msg_flags;
5062        if (force_nonblock)
5063                flags |= MSG_DONTWAIT;
5064        if (flags & MSG_WAITALL)
5065                min_ret = iov_iter_count(&msg.msg_iter);
5066
5067        ret = sock_recvmsg(sock, &msg, flags);
5068        if (force_nonblock && ret == -EAGAIN)
5069                return -EAGAIN;
5070        if (ret == -ERESTARTSYS)
5071                ret = -EINTR;
5072out_free:
5073        if (req->flags & REQ_F_BUFFER_SELECTED)
5074                cflags = io_put_recv_kbuf(req);
5075        if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5076                req_set_fail(req);
5077        __io_req_complete(req, issue_flags, ret, cflags);
5078        return 0;
5079}
5080
5081static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5082{
5083        struct io_accept *accept = &req->accept;
5084
5085        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5086                return -EINVAL;
5087        if (sqe->ioprio || sqe->len || sqe->buf_index)
5088                return -EINVAL;
5089
5090        accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5091        accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5092        accept->flags = READ_ONCE(sqe->accept_flags);
5093        accept->nofile = rlimit(RLIMIT_NOFILE);
5094
5095        accept->file_slot = READ_ONCE(sqe->file_index);
5096        if (accept->file_slot && ((req->open.how.flags & O_CLOEXEC) ||
5097                                  (accept->flags & SOCK_CLOEXEC)))
5098                return -EINVAL;
5099        if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
5100                return -EINVAL;
5101        if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
5102                accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
5103        return 0;
5104}
5105
5106static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5107{
5108        struct io_accept *accept = &req->accept;
5109        bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5110        unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5111        bool fixed = !!accept->file_slot;
5112        struct file *file;
5113        int ret, fd;
5114
5115        if (req->file->f_flags & O_NONBLOCK)
5116                req->flags |= REQ_F_NOWAIT;
5117
5118        if (!fixed) {
5119                fd = __get_unused_fd_flags(accept->flags, accept->nofile);
5120                if (unlikely(fd < 0))
5121                        return fd;
5122        }
5123        file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
5124                         accept->flags);
5125        if (IS_ERR(file)) {
5126                if (!fixed)
5127                        put_unused_fd(fd);
5128                ret = PTR_ERR(file);
5129                if (ret == -EAGAIN && force_nonblock)
5130                        return -EAGAIN;
5131                if (ret == -ERESTARTSYS)
5132                        ret = -EINTR;
5133                req_set_fail(req);
5134        } else if (!fixed) {
5135                fd_install(fd, file);
5136                ret = fd;
5137        } else {
5138                ret = io_install_fixed_file(req, file, issue_flags,
5139                                            accept->file_slot - 1);
5140        }
5141        __io_req_complete(req, issue_flags, ret, 0);
5142        return 0;
5143}
5144
5145static int io_connect_prep_async(struct io_kiocb *req)
5146{
5147        struct io_async_connect *io = req->async_data;
5148        struct io_connect *conn = &req->connect;
5149
5150        return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5151}
5152
5153static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5154{
5155        struct io_connect *conn = &req->connect;
5156
5157        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5158                return -EINVAL;
5159        if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
5160            sqe->splice_fd_in)
5161                return -EINVAL;
5162
5163        conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5164        conn->addr_len =  READ_ONCE(sqe->addr2);
5165        return 0;
5166}
5167
5168static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5169{
5170        struct io_async_connect __io, *io;
5171        unsigned file_flags;
5172        int ret;
5173        bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5174
5175        if (req->async_data) {
5176                io = req->async_data;
5177        } else {
5178                ret = move_addr_to_kernel(req->connect.addr,
5179                                                req->connect.addr_len,
5180                                                &__io.address);
5181                if (ret)
5182                        goto out;
5183                io = &__io;
5184        }
5185
5186        file_flags = force_nonblock ? O_NONBLOCK : 0;
5187
5188        ret = __sys_connect_file(req->file, &io->address,
5189                                        req->connect.addr_len, file_flags);
5190        if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5191                if (req->async_data)
5192                        return -EAGAIN;
5193                if (io_alloc_async_data(req)) {
5194                        ret = -ENOMEM;
5195                        goto out;
5196                }
5197                memcpy(req->async_data, &__io, sizeof(__io));
5198                return -EAGAIN;
5199        }
5200        if (ret == -ERESTARTSYS)
5201                ret = -EINTR;
5202out:
5203        if (ret < 0)
5204                req_set_fail(req);
5205        __io_req_complete(req, issue_flags, ret, 0);
5206        return 0;
5207}
5208#else /* !CONFIG_NET */
5209#define IO_NETOP_FN(op)                                                 \
5210static int io_##op(struct io_kiocb *req, unsigned int issue_flags)      \
5211{                                                                       \
5212        return -EOPNOTSUPP;                                             \
5213}
5214
5215#define IO_NETOP_PREP(op)                                               \
5216IO_NETOP_FN(op)                                                         \
5217static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5218{                                                                       \
5219        return -EOPNOTSUPP;                                             \
5220}                                                                       \
5221
5222#define IO_NETOP_PREP_ASYNC(op)                                         \
5223IO_NETOP_PREP(op)                                                       \
5224static int io_##op##_prep_async(struct io_kiocb *req)                   \
5225{                                                                       \
5226        return -EOPNOTSUPP;                                             \
5227}
5228
5229IO_NETOP_PREP_ASYNC(sendmsg);
5230IO_NETOP_PREP_ASYNC(recvmsg);
5231IO_NETOP_PREP_ASYNC(connect);
5232IO_NETOP_PREP(accept);
5233IO_NETOP_FN(send);
5234IO_NETOP_FN(recv);
5235#endif /* CONFIG_NET */
5236
5237struct io_poll_table {
5238        struct poll_table_struct pt;
5239        struct io_kiocb *req;
5240        int nr_entries;
5241        int error;
5242};
5243
5244static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5245                           __poll_t mask, io_req_tw_func_t func)
5246{
5247        /* for instances that support it check for an event match first: */
5248        if (mask && !(mask & poll->events))
5249                return 0;
5250
5251        trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5252
5253        list_del_init(&poll->wait.entry);
5254
5255        req->result = mask;
5256        req->io_task_work.func = func;
5257
5258        /*
5259         * If this fails, then the task is exiting. When a task exits, the
5260         * work gets canceled, so just cancel this request as well instead
5261         * of executing it. We can't safely execute it anyway, as we may not
5262         * have the needed state needed for it anyway.
5263         */
5264        io_req_task_work_add(req);
5265        return 1;
5266}
5267
5268static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5269        __acquires(&req->ctx->completion_lock)
5270{
5271        struct io_ring_ctx *ctx = req->ctx;
5272
5273        /* req->task == current here, checking PF_EXITING is safe */
5274        if (unlikely(req->task->flags & PF_EXITING))
5275                WRITE_ONCE(poll->canceled, true);
5276
5277        if (!req->result && !READ_ONCE(poll->canceled)) {
5278                struct poll_table_struct pt = { ._key = poll->events };
5279
5280                req->result = vfs_poll(req->file, &pt) & poll->events;
5281        }
5282
5283        spin_lock(&ctx->completion_lock);
5284        if (!req->result && !READ_ONCE(poll->canceled)) {
5285                add_wait_queue(poll->head, &poll->wait);
5286                return true;
5287        }
5288
5289        return false;
5290}
5291
5292static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5293{
5294        /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5295        if (req->opcode == IORING_OP_POLL_ADD)
5296                return req->async_data;
5297        return req->apoll->double_poll;
5298}
5299
5300static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5301{
5302        if (req->opcode == IORING_OP_POLL_ADD)
5303                return &req->poll;
5304        return &req->apoll->poll;
5305}
5306
5307static void io_poll_remove_double(struct io_kiocb *req)
5308        __must_hold(&req->ctx->completion_lock)
5309{
5310        struct io_poll_iocb *poll = io_poll_get_double(req);
5311
5312        lockdep_assert_held(&req->ctx->completion_lock);
5313
5314        if (poll && poll->head) {
5315                struct wait_queue_head *head = poll->head;
5316
5317                spin_lock_irq(&head->lock);
5318                list_del_init(&poll->wait.entry);
5319                if (poll->wait.private)
5320                        req_ref_put(req);
5321                poll->head = NULL;
5322                spin_unlock_irq(&head->lock);
5323        }
5324}
5325
5326static bool __io_poll_complete(struct io_kiocb *req, __poll_t mask)
5327        __must_hold(&req->ctx->completion_lock)
5328{
5329        struct io_ring_ctx *ctx = req->ctx;
5330        unsigned flags = IORING_CQE_F_MORE;
5331        int error;
5332
5333        if (READ_ONCE(req->poll.canceled)) {
5334                error = -ECANCELED;
5335                req->poll.events |= EPOLLONESHOT;
5336        } else {
5337                error = mangle_poll(mask);
5338        }
5339        if (req->poll.events & EPOLLONESHOT)
5340                flags = 0;
5341        if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
5342                req->poll.events |= EPOLLONESHOT;
5343                flags = 0;
5344        }
5345        if (flags & IORING_CQE_F_MORE)
5346                ctx->cq_extra++;
5347
5348        return !(flags & IORING_CQE_F_MORE);
5349}
5350
5351static inline bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
5352        __must_hold(&req->ctx->completion_lock)
5353{
5354        bool done;
5355
5356        done = __io_poll_complete(req, mask);
5357        io_commit_cqring(req->ctx);
5358        return done;
5359}
5360
5361static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5362{
5363        struct io_ring_ctx *ctx = req->ctx;
5364        struct io_kiocb *nxt;
5365
5366        if (io_poll_rewait(req, &req->poll)) {
5367                spin_unlock(&ctx->completion_lock);
5368        } else {
5369                bool done;
5370
5371                if (req->poll.done) {
5372                        spin_unlock(&ctx->completion_lock);
5373                        return;
5374                }
5375                done = __io_poll_complete(req, req->result);
5376                if (done) {
5377                        io_poll_remove_double(req);
5378                        hash_del(&req->hash_node);
5379                        req->poll.done = true;
5380                } else {
5381                        req->result = 0;
5382                        add_wait_queue(req->poll.head, &req->poll.wait);
5383                }
5384                io_commit_cqring(ctx);
5385                spin_unlock(&ctx->completion_lock);
5386                io_cqring_ev_posted(ctx);
5387
5388                if (done) {
5389                        nxt = io_put_req_find_next(req);
5390                        if (nxt)
5391                                io_req_task_submit(nxt, locked);
5392                }
5393        }
5394}
5395
5396static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5397                               int sync, void *key)
5398{
5399        struct io_kiocb *req = wait->private;
5400        struct io_poll_iocb *poll = io_poll_get_single(req);
5401        __poll_t mask = key_to_poll(key);
5402        unsigned long flags;
5403
5404        /* for instances that support it check for an event match first: */
5405        if (mask && !(mask & poll->events))
5406                return 0;
5407        if (!(poll->events & EPOLLONESHOT))
5408                return poll->wait.func(&poll->wait, mode, sync, key);
5409
5410        list_del_init(&wait->entry);
5411
5412        if (poll->head) {
5413                bool done;
5414
5415                spin_lock_irqsave(&poll->head->lock, flags);
5416                done = list_empty(&poll->wait.entry);
5417                if (!done)
5418                        list_del_init(&poll->wait.entry);
5419                /* make sure double remove sees this as being gone */
5420                wait->private = NULL;
5421                spin_unlock_irqrestore(&poll->head->lock, flags);
5422                if (!done) {
5423                        /* use wait func handler, so it matches the rq type */
5424                        poll->wait.func(&poll->wait, mode, sync, key);
5425                }
5426        }
5427        req_ref_put(req);
5428        return 1;
5429}
5430
5431static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5432                              wait_queue_func_t wake_func)
5433{
5434        poll->head = NULL;
5435        poll->done = false;
5436        poll->canceled = false;
5437#define IO_POLL_UNMASK  (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5438        /* mask in events that we always want/need */
5439        poll->events = events | IO_POLL_UNMASK;
5440        INIT_LIST_HEAD(&poll->wait.entry);
5441        init_waitqueue_func_entry(&poll->wait, wake_func);
5442}
5443
5444static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5445                            struct wait_queue_head *head,
5446                            struct io_poll_iocb **poll_ptr)
5447{
5448        struct io_kiocb *req = pt->req;
5449
5450        /*
5451         * The file being polled uses multiple waitqueues for poll handling
5452         * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5453         * if this happens.
5454         */
5455        if (unlikely(pt->nr_entries)) {
5456                struct io_poll_iocb *poll_one = poll;
5457
5458                /* double add on the same waitqueue head, ignore */
5459                if (poll_one->head == head)
5460                        return;
5461                /* already have a 2nd entry, fail a third attempt */
5462                if (*poll_ptr) {
5463                        if ((*poll_ptr)->head == head)
5464                                return;
5465                        pt->error = -EINVAL;
5466                        return;
5467                }
5468                /*
5469                 * Can't handle multishot for double wait for now, turn it
5470                 * into one-shot mode.
5471                 */
5472                if (!(poll_one->events & EPOLLONESHOT))
5473                        poll_one->events |= EPOLLONESHOT;
5474                poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5475                if (!poll) {
5476                        pt->error = -ENOMEM;
5477                        return;
5478                }
5479                io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5480                req_ref_get(req);
5481                poll->wait.private = req;
5482                *poll_ptr = poll;
5483        }
5484
5485        pt->nr_entries++;
5486        poll->head = head;
5487
5488        if (poll->events & EPOLLEXCLUSIVE)
5489                add_wait_queue_exclusive(head, &poll->wait);
5490        else
5491                add_wait_queue(head, &poll->wait);
5492}
5493
5494static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5495                               struct poll_table_struct *p)
5496{
5497        struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5498        struct async_poll *apoll = pt->req->apoll;
5499
5500        __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5501}
5502
5503static void io_async_task_func(struct io_kiocb *req, bool *locked)
5504{
5505        struct async_poll *apoll = req->apoll;
5506        struct io_ring_ctx *ctx = req->ctx;
5507
5508        trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5509
5510        if (io_poll_rewait(req, &apoll->poll)) {
5511                spin_unlock(&ctx->completion_lock);
5512                return;
5513        }
5514
5515        hash_del(&req->hash_node);
5516        io_poll_remove_double(req);
5517        apoll->poll.done = true;
5518        spin_unlock(&ctx->completion_lock);
5519
5520        if (!READ_ONCE(apoll->poll.canceled))
5521                io_req_task_submit(req, locked);
5522        else
5523                io_req_complete_failed(req, -ECANCELED);
5524}
5525
5526static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5527                        void *key)
5528{
5529        struct io_kiocb *req = wait->private;
5530        struct io_poll_iocb *poll = &req->apoll->poll;
5531
5532        trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5533                                        key_to_poll(key));
5534
5535        return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5536}
5537
5538static void io_poll_req_insert(struct io_kiocb *req)
5539{
5540        struct io_ring_ctx *ctx = req->ctx;
5541        struct hlist_head *list;
5542
5543        list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5544        hlist_add_head(&req->hash_node, list);
5545}
5546
5547static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5548                                      struct io_poll_iocb *poll,
5549                                      struct io_poll_table *ipt, __poll_t mask,
5550                                      wait_queue_func_t wake_func)
5551        __acquires(&ctx->completion_lock)
5552{
5553        struct io_ring_ctx *ctx = req->ctx;
5554        bool cancel = false;
5555
5556        INIT_HLIST_NODE(&req->hash_node);
5557        io_init_poll_iocb(poll, mask, wake_func);
5558        poll->file = req->file;
5559        poll->wait.private = req;
5560
5561        ipt->pt._key = mask;
5562        ipt->req = req;
5563        ipt->error = 0;
5564        ipt->nr_entries = 0;
5565
5566        mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5567        if (unlikely(!ipt->nr_entries) && !ipt->error)
5568                ipt->error = -EINVAL;
5569
5570        spin_lock(&ctx->completion_lock);
5571        if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5572                io_poll_remove_double(req);
5573        if (likely(poll->head)) {
5574                spin_lock_irq(&poll->head->lock);
5575                if (unlikely(list_empty(&poll->wait.entry))) {
5576                        if (ipt->error)
5577                                cancel = true;
5578                        ipt->error = 0;
5579                        mask = 0;
5580                }
5581                if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5582                        list_del_init(&poll->wait.entry);
5583                else if (cancel)
5584                        WRITE_ONCE(poll->canceled, true);
5585                else if (!poll->done) /* actually waiting for an event */
5586                        io_poll_req_insert(req);
5587                spin_unlock_irq(&poll->head->lock);
5588        }
5589
5590        return mask;
5591}
5592
5593enum {
5594        IO_APOLL_OK,
5595        IO_APOLL_ABORTED,
5596        IO_APOLL_READY
5597};
5598
5599static int io_arm_poll_handler(struct io_kiocb *req)
5600{
5601        const struct io_op_def *def = &io_op_defs[req->opcode];
5602        struct io_ring_ctx *ctx = req->ctx;
5603        struct async_poll *apoll;
5604        struct io_poll_table ipt;
5605        __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5606        int rw;
5607
5608        if (!req->file || !file_can_poll(req->file))
5609                return IO_APOLL_ABORTED;
5610        if (req->flags & REQ_F_POLLED)
5611                return IO_APOLL_ABORTED;
5612        if (!def->pollin && !def->pollout)
5613                return IO_APOLL_ABORTED;
5614
5615        if (def->pollin) {
5616                rw = READ;
5617                mask |= POLLIN | POLLRDNORM;
5618
5619                /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5620                if ((req->opcode == IORING_OP_RECVMSG) &&
5621                    (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5622                        mask &= ~POLLIN;
5623        } else {
5624                rw = WRITE;
5625                mask |= POLLOUT | POLLWRNORM;
5626        }
5627
5628        /* if we can't nonblock try, then no point in arming a poll handler */
5629        if (!io_file_supports_nowait(req, rw))
5630                return IO_APOLL_ABORTED;
5631
5632        apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5633        if (unlikely(!apoll))
5634                return IO_APOLL_ABORTED;
5635        apoll->double_poll = NULL;
5636        req->apoll = apoll;
5637        req->flags |= REQ_F_POLLED;
5638        ipt.pt._qproc = io_async_queue_proc;
5639        io_req_set_refcount(req);
5640
5641        ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5642                                        io_async_wake);
5643        spin_unlock(&ctx->completion_lock);
5644        if (ret || ipt.error)
5645                return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5646
5647        trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5648                                mask, apoll->poll.events);
5649        return IO_APOLL_OK;
5650}
5651
5652static bool __io_poll_remove_one(struct io_kiocb *req,
5653                                 struct io_poll_iocb *poll, bool do_cancel)
5654        __must_hold(&req->ctx->completion_lock)
5655{
5656        bool do_complete = false;
5657
5658        if (!poll->head)
5659                return false;
5660        spin_lock_irq(&poll->head->lock);
5661        if (do_cancel)
5662                WRITE_ONCE(poll->canceled, true);
5663        if (!list_empty(&poll->wait.entry)) {
5664                list_del_init(&poll->wait.entry);
5665                do_complete = true;
5666        }
5667        spin_unlock_irq(&poll->head->lock);
5668        hash_del(&req->hash_node);
5669        return do_complete;
5670}
5671
5672static bool io_poll_remove_one(struct io_kiocb *req)
5673        __must_hold(&req->ctx->completion_lock)
5674{
5675        bool do_complete;
5676
5677        io_poll_remove_double(req);
5678        do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5679
5680        if (do_complete) {
5681                io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5682                io_commit_cqring(req->ctx);
5683                req_set_fail(req);
5684                io_put_req_deferred(req);
5685        }
5686        return do_complete;
5687}
5688
5689/*
5690 * Returns true if we found and killed one or more poll requests
5691 */
5692static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5693                               bool cancel_all)
5694{
5695        struct hlist_node *tmp;
5696        struct io_kiocb *req;
5697        int posted = 0, i;
5698
5699        spin_lock(&ctx->completion_lock);
5700        for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5701                struct hlist_head *list;
5702
5703                list = &ctx->cancel_hash[i];
5704                hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5705                        if (io_match_task(req, tsk, cancel_all))
5706                                posted += io_poll_remove_one(req);
5707                }
5708        }
5709        spin_unlock(&ctx->completion_lock);
5710
5711        if (posted)
5712                io_cqring_ev_posted(ctx);
5713
5714        return posted != 0;
5715}
5716
5717static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5718                                     bool poll_only)
5719        __must_hold(&ctx->completion_lock)
5720{
5721        struct hlist_head *list;
5722        struct io_kiocb *req;
5723
5724        list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5725        hlist_for_each_entry(req, list, hash_node) {
5726                if (sqe_addr != req->user_data)
5727                        continue;
5728                if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5729                        continue;
5730                return req;
5731        }
5732        return NULL;
5733}
5734
5735static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5736                          bool poll_only)
5737        __must_hold(&ctx->completion_lock)
5738{
5739        struct io_kiocb *req;
5740
5741        req = io_poll_find(ctx, sqe_addr, poll_only);
5742        if (!req)
5743                return -ENOENT;
5744        if (io_poll_remove_one(req))
5745                return 0;
5746
5747        return -EALREADY;
5748}
5749
5750static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5751                                     unsigned int flags)
5752{
5753        u32 events;
5754
5755        events = READ_ONCE(sqe->poll32_events);
5756#ifdef __BIG_ENDIAN
5757        events = swahw32(events);
5758#endif
5759        if (!(flags & IORING_POLL_ADD_MULTI))
5760                events |= EPOLLONESHOT;
5761        return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5762}
5763
5764static int io_poll_update_prep(struct io_kiocb *req,
5765                               const struct io_uring_sqe *sqe)
5766{
5767        struct io_poll_update *upd = &req->poll_update;
5768        u32 flags;
5769
5770        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5771                return -EINVAL;
5772        if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5773                return -EINVAL;
5774        flags = READ_ONCE(sqe->len);
5775        if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5776                      IORING_POLL_ADD_MULTI))
5777                return -EINVAL;
5778        /* meaningless without update */
5779        if (flags == IORING_POLL_ADD_MULTI)
5780                return -EINVAL;
5781
5782        upd->old_user_data = READ_ONCE(sqe->addr);
5783        upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5784        upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5785
5786        upd->new_user_data = READ_ONCE(sqe->off);
5787        if (!upd->update_user_data && upd->new_user_data)
5788                return -EINVAL;
5789        if (upd->update_events)
5790                upd->events = io_poll_parse_events(sqe, flags);
5791        else if (sqe->poll32_events)
5792                return -EINVAL;
5793
5794        return 0;
5795}
5796
5797static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5798                        void *key)
5799{
5800        struct io_kiocb *req = wait->private;
5801        struct io_poll_iocb *poll = &req->poll;
5802
5803        return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5804}
5805
5806static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5807                               struct poll_table_struct *p)
5808{
5809        struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5810
5811        __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5812}
5813
5814static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5815{
5816        struct io_poll_iocb *poll = &req->poll;
5817        u32 flags;
5818
5819        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5820                return -EINVAL;
5821        if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5822                return -EINVAL;
5823        flags = READ_ONCE(sqe->len);
5824        if (flags & ~IORING_POLL_ADD_MULTI)
5825                return -EINVAL;
5826
5827        io_req_set_refcount(req);
5828        poll->events = io_poll_parse_events(sqe, flags);
5829        return 0;
5830}
5831
5832static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5833{
5834        struct io_poll_iocb *poll = &req->poll;
5835        struct io_ring_ctx *ctx = req->ctx;
5836        struct io_poll_table ipt;
5837        __poll_t mask;
5838        bool done;
5839
5840        ipt.pt._qproc = io_poll_queue_proc;
5841
5842        mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5843                                        io_poll_wake);
5844
5845        if (mask) { /* no async, we'd stolen it */
5846                ipt.error = 0;
5847                done = io_poll_complete(req, mask);
5848        }
5849        spin_unlock(&ctx->completion_lock);
5850
5851        if (mask) {
5852                io_cqring_ev_posted(ctx);
5853                if (done)
5854                        io_put_req(req);
5855        }
5856        return ipt.error;
5857}
5858
5859static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5860{
5861        struct io_ring_ctx *ctx = req->ctx;
5862        struct io_kiocb *preq;
5863        bool completing;
5864        int ret;
5865
5866        spin_lock(&ctx->completion_lock);
5867        preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5868        if (!preq) {
5869                ret = -ENOENT;
5870                goto err;
5871        }
5872
5873        if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5874                completing = true;
5875                ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5876                goto err;
5877        }
5878
5879        /*
5880         * Don't allow racy completion with singleshot, as we cannot safely
5881         * update those. For multishot, if we're racing with completion, just
5882         * let completion re-add it.
5883         */
5884        completing = !__io_poll_remove_one(preq, &preq->poll, false);
5885        if (completing && (preq->poll.events & EPOLLONESHOT)) {
5886                ret = -EALREADY;
5887                goto err;
5888        }
5889        /* we now have a detached poll request. reissue. */
5890        ret = 0;
5891err:
5892        if (ret < 0) {
5893                spin_unlock(&ctx->completion_lock);
5894                req_set_fail(req);
5895                io_req_complete(req, ret);
5896                return 0;
5897        }
5898        /* only mask one event flags, keep behavior flags */
5899        if (req->poll_update.update_events) {
5900                preq->poll.events &= ~0xffff;
5901                preq->poll.events |= req->poll_update.events & 0xffff;
5902                preq->poll.events |= IO_POLL_UNMASK;
5903        }
5904        if (req->poll_update.update_user_data)
5905                preq->user_data = req->poll_update.new_user_data;
5906        spin_unlock(&ctx->completion_lock);
5907
5908        /* complete update request, we're done with it */
5909        io_req_complete(req, ret);
5910
5911        if (!completing) {
5912                ret = io_poll_add(preq, issue_flags);
5913                if (ret < 0) {
5914                        req_set_fail(preq);
5915                        io_req_complete(preq, ret);
5916                }
5917        }
5918        return 0;
5919}
5920
5921static void io_req_task_timeout(struct io_kiocb *req, bool *locked)
5922{
5923        req_set_fail(req);
5924        io_req_complete_post(req, -ETIME, 0);
5925}
5926
5927static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5928{
5929        struct io_timeout_data *data = container_of(timer,
5930                                                struct io_timeout_data, timer);
5931        struct io_kiocb *req = data->req;
5932        struct io_ring_ctx *ctx = req->ctx;
5933        unsigned long flags;
5934
5935        spin_lock_irqsave(&ctx->timeout_lock, flags);
5936        list_del_init(&req->timeout.list);
5937        atomic_set(&req->ctx->cq_timeouts,
5938                atomic_read(&req->ctx->cq_timeouts) + 1);
5939        spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5940
5941        req->io_task_work.func = io_req_task_timeout;
5942        io_req_task_work_add(req);
5943        return HRTIMER_NORESTART;
5944}
5945
5946static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5947                                           __u64 user_data)
5948        __must_hold(&ctx->timeout_lock)
5949{
5950        struct io_timeout_data *io;
5951        struct io_kiocb *req;
5952        bool found = false;
5953
5954        list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5955                found = user_data == req->user_data;
5956                if (found)
5957                        break;
5958        }
5959        if (!found)
5960                return ERR_PTR(-ENOENT);
5961
5962        io = req->async_data;
5963        if (hrtimer_try_to_cancel(&io->timer) == -1)
5964                return ERR_PTR(-EALREADY);
5965        list_del_init(&req->timeout.list);
5966        return req;
5967}
5968
5969static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5970        __must_hold(&ctx->completion_lock)
5971        __must_hold(&ctx->timeout_lock)
5972{
5973        struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5974
5975        if (IS_ERR(req))
5976                return PTR_ERR(req);
5977
5978        req_set_fail(req);
5979        io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5980        io_put_req_deferred(req);
5981        return 0;
5982}
5983
5984static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
5985{
5986        switch (data->flags & IORING_TIMEOUT_CLOCK_MASK) {
5987        case IORING_TIMEOUT_BOOTTIME:
5988                return CLOCK_BOOTTIME;
5989        case IORING_TIMEOUT_REALTIME:
5990                return CLOCK_REALTIME;
5991        default:
5992                /* can't happen, vetted at prep time */
5993                WARN_ON_ONCE(1);
5994                fallthrough;
5995        case 0:
5996                return CLOCK_MONOTONIC;
5997        }
5998}
5999
6000static int io_linked_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6001                                    struct timespec64 *ts, enum hrtimer_mode mode)
6002        __must_hold(&ctx->timeout_lock)
6003{
6004        struct io_timeout_data *io;
6005        struct io_kiocb *req;
6006        bool found = false;
6007
6008        list_for_each_entry(req, &ctx->ltimeout_list, timeout.list) {
6009                found = user_data == req->user_data;
6010                if (found)
6011                        break;
6012        }
6013        if (!found)
6014                return -ENOENT;
6015
6016        io = req->async_data;
6017        if (hrtimer_try_to_cancel(&io->timer) == -1)
6018                return -EALREADY;
6019        hrtimer_init(&io->timer, io_timeout_get_clock(io), mode);
6020        io->timer.function = io_link_timeout_fn;
6021        hrtimer_start(&io->timer, timespec64_to_ktime(*ts), mode);
6022        return 0;
6023}
6024
6025static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6026                             struct timespec64 *ts, enum hrtimer_mode mode)
6027        __must_hold(&ctx->timeout_lock)
6028{
6029        struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6030        struct io_timeout_data *data;
6031
6032        if (IS_ERR(req))
6033                return PTR_ERR(req);
6034
6035        req->timeout.off = 0; /* noseq */
6036        data = req->async_data;
6037        list_add_tail(&req->timeout.list, &ctx->timeout_list);
6038        hrtimer_init(&data->timer, io_timeout_get_clock(data), mode);
6039        data->timer.function = io_timeout_fn;
6040        hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
6041        return 0;
6042}
6043
6044static int io_timeout_remove_prep(struct io_kiocb *req,
6045                                  const struct io_uring_sqe *sqe)
6046{
6047        struct io_timeout_rem *tr = &req->timeout_rem;
6048
6049        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6050                return -EINVAL;
6051        if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6052                return -EINVAL;
6053        if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
6054                return -EINVAL;
6055
6056        tr->ltimeout = false;
6057        tr->addr = READ_ONCE(sqe->addr);
6058        tr->flags = READ_ONCE(sqe->timeout_flags);
6059        if (tr->flags & IORING_TIMEOUT_UPDATE_MASK) {
6060                if (hweight32(tr->flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6061                        return -EINVAL;
6062                if (tr->flags & IORING_LINK_TIMEOUT_UPDATE)
6063                        tr->ltimeout = true;
6064                if (tr->flags & ~(IORING_TIMEOUT_UPDATE_MASK|IORING_TIMEOUT_ABS))
6065                        return -EINVAL;
6066                if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
6067                        return -EFAULT;
6068        } else if (tr->flags) {
6069                /* timeout removal doesn't support flags */
6070                return -EINVAL;
6071        }
6072
6073        return 0;
6074}
6075
6076static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
6077{
6078        return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
6079                                            : HRTIMER_MODE_REL;
6080}
6081
6082/*
6083 * Remove or update an existing timeout command
6084 */
6085static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
6086{
6087        struct io_timeout_rem *tr = &req->timeout_rem;
6088        struct io_ring_ctx *ctx = req->ctx;
6089        int ret;
6090
6091        if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
6092                spin_lock(&ctx->completion_lock);
6093                spin_lock_irq(&ctx->timeout_lock);
6094                ret = io_timeout_cancel(ctx, tr->addr);
6095                spin_unlock_irq(&ctx->timeout_lock);
6096                spin_unlock(&ctx->completion_lock);
6097        } else {
6098                enum hrtimer_mode mode = io_translate_timeout_mode(tr->flags);
6099
6100                spin_lock_irq(&ctx->timeout_lock);
6101                if (tr->ltimeout)
6102                        ret = io_linked_timeout_update(ctx, tr->addr, &tr->ts, mode);
6103                else
6104                        ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
6105                spin_unlock_irq(&ctx->timeout_lock);
6106        }
6107
6108        if (ret < 0)
6109                req_set_fail(req);
6110        io_req_complete_post(req, ret, 0);
6111        return 0;
6112}
6113
6114static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6115                           bool is_timeout_link)
6116{
6117        struct io_timeout_data *data;
6118        unsigned flags;
6119        u32 off = READ_ONCE(sqe->off);
6120
6121        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6122                return -EINVAL;
6123        if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
6124            sqe->splice_fd_in)
6125                return -EINVAL;
6126        if (off && is_timeout_link)
6127                return -EINVAL;
6128        flags = READ_ONCE(sqe->timeout_flags);
6129        if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK))
6130                return -EINVAL;
6131        /* more than one clock specified is invalid, obviously */
6132        if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6133                return -EINVAL;
6134
6135        INIT_LIST_HEAD(&req->timeout.list);
6136        req->timeout.off = off;
6137        if (unlikely(off && !req->ctx->off_timeout_used))
6138                req->ctx->off_timeout_used = true;
6139
6140        if (!req->async_data && io_alloc_async_data(req))
6141                return -ENOMEM;
6142
6143        data = req->async_data;
6144        data->req = req;
6145        data->flags = flags;
6146
6147        if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
6148                return -EFAULT;
6149
6150        data->mode = io_translate_timeout_mode(flags);
6151        hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
6152
6153        if (is_timeout_link) {
6154                struct io_submit_link *link = &req->ctx->submit_state.link;
6155
6156                if (!link->head)
6157                        return -EINVAL;
6158                if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
6159                        return -EINVAL;
6160                req->timeout.head = link->last;
6161                link->last->flags |= REQ_F_ARM_LTIMEOUT;
6162        }
6163        return 0;
6164}
6165
6166static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
6167{
6168        struct io_ring_ctx *ctx = req->ctx;
6169        struct io_timeout_data *data = req->async_data;
6170        struct list_head *entry;
6171        u32 tail, off = req->timeout.off;
6172
6173        spin_lock_irq(&ctx->timeout_lock);
6174
6175        /*
6176         * sqe->off holds how many events that need to occur for this
6177         * timeout event to be satisfied. If it isn't set, then this is
6178         * a pure timeout request, sequence isn't used.
6179         */
6180        if (io_is_timeout_noseq(req)) {
6181                entry = ctx->timeout_list.prev;
6182                goto add;
6183        }
6184
6185        tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
6186        req->timeout.target_seq = tail + off;
6187
6188        /* Update the last seq here in case io_flush_timeouts() hasn't.
6189         * This is safe because ->completion_lock is held, and submissions
6190         * and completions are never mixed in the same ->completion_lock section.
6191         */
6192        ctx->cq_last_tm_flush = tail;
6193
6194        /*
6195         * Insertion sort, ensuring the first entry in the list is always
6196         * the one we need first.
6197         */
6198        list_for_each_prev(entry, &ctx->timeout_list) {
6199                struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
6200                                                  timeout.list);
6201
6202                if (io_is_timeout_noseq(nxt))
6203                        continue;
6204                /* nxt.seq is behind @tail, otherwise would've been completed */
6205                if (off >= nxt->timeout.target_seq - tail)
6206                        break;
6207        }
6208add:
6209        list_add(&req->timeout.list, entry);
6210        data->timer.function = io_timeout_fn;
6211        hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
6212        spin_unlock_irq(&ctx->timeout_lock);
6213        return 0;
6214}
6215
6216struct io_cancel_data {
6217        struct io_ring_ctx *ctx;
6218        u64 user_data;
6219};
6220
6221static bool io_cancel_cb(struct io_wq_work *work, void *data)
6222{
6223        struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6224        struct io_cancel_data *cd = data;
6225
6226        return req->ctx == cd->ctx && req->user_data == cd->user_data;
6227}
6228
6229static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
6230                               struct io_ring_ctx *ctx)
6231{
6232        struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
6233        enum io_wq_cancel cancel_ret;
6234        int ret = 0;
6235
6236        if (!tctx || !tctx->io_wq)
6237                return -ENOENT;
6238
6239        cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
6240        switch (cancel_ret) {
6241        case IO_WQ_CANCEL_OK:
6242                ret = 0;
6243                break;
6244        case IO_WQ_CANCEL_RUNNING:
6245                ret = -EALREADY;
6246                break;
6247        case IO_WQ_CANCEL_NOTFOUND:
6248                ret = -ENOENT;
6249                break;
6250        }
6251
6252        return ret;
6253}
6254
6255static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
6256{
6257        struct io_ring_ctx *ctx = req->ctx;
6258        int ret;
6259
6260        WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
6261
6262        ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
6263        if (ret != -ENOENT)
6264                return ret;
6265
6266        spin_lock(&ctx->completion_lock);
6267        spin_lock_irq(&ctx->timeout_lock);
6268        ret = io_timeout_cancel(ctx, sqe_addr);
6269        spin_unlock_irq(&ctx->timeout_lock);
6270        if (ret != -ENOENT)
6271                goto out;
6272        ret = io_poll_cancel(ctx, sqe_addr, false);
6273out:
6274        spin_unlock(&ctx->completion_lock);
6275        return ret;
6276}
6277
6278static int io_async_cancel_prep(struct io_kiocb *req,
6279                                const struct io_uring_sqe *sqe)
6280{
6281        if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6282                return -EINVAL;
6283        if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6284                return -EINVAL;
6285        if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
6286            sqe->splice_fd_in)
6287                return -EINVAL;
6288
6289        req->cancel.addr = READ_ONCE(sqe->addr);
6290        return 0;
6291}
6292
6293static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6294{
6295        struct io_ring_ctx *ctx = req->ctx;
6296        u64 sqe_addr = req->cancel.addr;
6297        struct io_tctx_node *node;
6298        int ret;
6299
6300        ret = io_try_cancel_userdata(req, sqe_addr);
6301        if (ret != -ENOENT)
6302                goto done;
6303
6304        /* slow path, try all io-wq's */
6305        io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6306        ret = -ENOENT;
6307        list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
6308                struct io_uring_task *tctx = node->task->io_uring;
6309
6310                ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
6311                if (ret != -ENOENT)
6312                        break;
6313        }
6314        io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6315done:
6316        if (ret < 0)
6317                req_set_fail(req);
6318        io_req_complete_post(req, ret, 0);
6319        return 0;
6320}
6321
6322static int io_rsrc_update_prep(struct io_kiocb *req,
6323                                const struct io_uring_sqe *sqe)
6324{
6325        if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6326                return -EINVAL;
6327        if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
6328                return -EINVAL;
6329
6330        req->rsrc_update.offset = READ_ONCE(sqe->off);
6331        req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6332        if (!req->rsrc_update.nr_args)
6333                return -EINVAL;
6334        req->rsrc_update.arg = READ_ONCE(sqe->addr);
6335        return 0;
6336}
6337
6338static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6339{
6340        struct io_ring_ctx *ctx = req->ctx;
6341        struct io_uring_rsrc_update2 up;
6342        int ret;
6343
6344        up.offset = req->rsrc_update.offset;
6345        up.data = req->rsrc_update.arg;
6346        up.nr = 0;
6347        up.tags = 0;
6348        up.resv = 0;
6349
6350        io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6351        ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
6352                                        &up, req->rsrc_update.nr_args);
6353        io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6354
6355        if (ret < 0)
6356                req_set_fail(req);
6357        __io_req_complete(req, issue_flags, ret, 0);
6358        return 0;
6359}
6360
6361static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6362{
6363        switch (req->opcode) {
6364        case IORING_OP_NOP:
6365                return 0;
6366        case IORING_OP_READV:
6367        case IORING_OP_READ_FIXED:
6368        case IORING_OP_READ:
6369                return io_read_prep(req, sqe);
6370        case IORING_OP_WRITEV:
6371        case IORING_OP_WRITE_FIXED:
6372        case IORING_OP_WRITE:
6373                return io_write_prep(req, sqe);
6374        case IORING_OP_POLL_ADD:
6375                return io_poll_add_prep(req, sqe);
6376        case IORING_OP_POLL_REMOVE:
6377                return io_poll_update_prep(req, sqe);
6378        case IORING_OP_FSYNC:
6379                return io_fsync_prep(req, sqe);
6380        case IORING_OP_SYNC_FILE_RANGE:
6381                return io_sfr_prep(req, sqe);
6382        case IORING_OP_SENDMSG:
6383        case IORING_OP_SEND:
6384                return io_sendmsg_prep(req, sqe);
6385        case IORING_OP_RECVMSG:
6386        case IORING_OP_RECV:
6387                return io_recvmsg_prep(req, sqe);
6388        case IORING_OP_CONNECT:
6389                return io_connect_prep(req, sqe);
6390        case IORING_OP_TIMEOUT:
6391                return io_timeout_prep(req, sqe, false);
6392        case IORING_OP_TIMEOUT_REMOVE:
6393                return io_timeout_remove_prep(req, sqe);
6394        case IORING_OP_ASYNC_CANCEL:
6395                return io_async_cancel_prep(req, sqe);
6396        case IORING_OP_LINK_TIMEOUT:
6397                return io_timeout_prep(req, sqe, true);
6398        case IORING_OP_ACCEPT:
6399                return io_accept_prep(req, sqe);
6400        case IORING_OP_FALLOCATE:
6401                return io_fallocate_prep(req, sqe);
6402        case IORING_OP_OPENAT:
6403                return io_openat_prep(req, sqe);
6404        case IORING_OP_CLOSE:
6405                return io_close_prep(req, sqe);
6406        case IORING_OP_FILES_UPDATE:
6407                return io_rsrc_update_prep(req, sqe);
6408        case IORING_OP_STATX:
6409                return io_statx_prep(req, sqe);
6410        case IORING_OP_FADVISE:
6411                return io_fadvise_prep(req, sqe);
6412        case IORING_OP_MADVISE:
6413                return io_madvise_prep(req, sqe);
6414        case IORING_OP_OPENAT2:
6415                return io_openat2_prep(req, sqe);
6416        case IORING_OP_EPOLL_CTL:
6417                return io_epoll_ctl_prep(req, sqe);
6418        case IORING_OP_SPLICE:
6419                return io_splice_prep(req, sqe);
6420        case IORING_OP_PROVIDE_BUFFERS:
6421                return io_provide_buffers_prep(req, sqe);
6422        case IORING_OP_REMOVE_BUFFERS:
6423                return io_remove_buffers_prep(req, sqe);
6424        case IORING_OP_TEE:
6425                return io_tee_prep(req, sqe);
6426        case IORING_OP_SHUTDOWN:
6427                return io_shutdown_prep(req, sqe);
6428        case IORING_OP_RENAMEAT:
6429                return io_renameat_prep(req, sqe);
6430        case IORING_OP_UNLINKAT:
6431                return io_unlinkat_prep(req, sqe);
6432        case IORING_OP_MKDIRAT:
6433                return io_mkdirat_prep(req, sqe);
6434        case IORING_OP_SYMLINKAT:
6435                return io_symlinkat_prep(req, sqe);
6436        case IORING_OP_LINKAT:
6437                return io_linkat_prep(req, sqe);
6438        }
6439
6440        printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6441                        req->opcode);
6442        return -EINVAL;
6443}
6444
6445static int io_req_prep_async(struct io_kiocb *req)
6446{
6447        if (!io_op_defs[req->opcode].needs_async_setup)
6448                return 0;
6449        if (WARN_ON_ONCE(req->async_data))
6450                return -EFAULT;
6451        if (io_alloc_async_data(req))
6452                return -EAGAIN;
6453
6454        switch (req->opcode) {
6455        case IORING_OP_READV:
6456                return io_rw_prep_async(req, READ);
6457        case IORING_OP_WRITEV:
6458                return io_rw_prep_async(req, WRITE);
6459        case IORING_OP_SENDMSG:
6460                return io_sendmsg_prep_async(req);
6461        case IORING_OP_RECVMSG:
6462                return io_recvmsg_prep_async(req);
6463        case IORING_OP_CONNECT:
6464                return io_connect_prep_async(req);
6465        }
6466        printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6467                    req->opcode);
6468        return -EFAULT;
6469}
6470
6471static u32 io_get_sequence(struct io_kiocb *req)
6472{
6473        u32 seq = req->ctx->cached_sq_head;
6474
6475        /* need original cached_sq_head, but it was increased for each req */
6476        io_for_each_link(req, req)
6477                seq--;
6478        return seq;
6479}
6480
6481static bool io_drain_req(struct io_kiocb *req)
6482{
6483        struct io_kiocb *pos;
6484        struct io_ring_ctx *ctx = req->ctx;
6485        struct io_defer_entry *de;
6486        int ret;
6487        u32 seq;
6488
6489        if (req->flags & REQ_F_FAIL) {
6490                io_req_complete_fail_submit(req);
6491                return true;
6492        }
6493
6494        /*
6495         * If we need to drain a request in the middle of a link, drain the
6496         * head request and the next request/link after the current link.
6497         * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6498         * maintained for every request of our link.
6499         */
6500        if (ctx->drain_next) {
6501                req->flags |= REQ_F_IO_DRAIN;
6502                ctx->drain_next = false;
6503        }
6504        /* not interested in head, start from the first linked */
6505        io_for_each_link(pos, req->link) {
6506                if (pos->flags & REQ_F_IO_DRAIN) {
6507                        ctx->drain_next = true;
6508                        req->flags |= REQ_F_IO_DRAIN;
6509                        break;
6510                }
6511        }
6512
6513        /* Still need defer if there is pending req in defer list. */
6514        if (likely(list_empty_careful(&ctx->defer_list) &&
6515                !(req->flags & REQ_F_IO_DRAIN))) {
6516                ctx->drain_active = false;
6517                return false;
6518        }
6519
6520        seq = io_get_sequence(req);
6521        /* Still a chance to pass the sequence check */
6522        if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6523                return false;
6524
6525        ret = io_req_prep_async(req);
6526        if (ret)
6527                goto fail;
6528        io_prep_async_link(req);
6529        de = kmalloc(sizeof(*de), GFP_KERNEL);
6530        if (!de) {
6531                ret = -ENOMEM;
6532fail:
6533                io_req_complete_failed(req, ret);
6534                return true;
6535        }
6536
6537        spin_lock(&ctx->completion_lock);
6538        if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6539                spin_unlock(&ctx->completion_lock);
6540                kfree(de);
6541                io_queue_async_work(req, NULL);
6542                return true;
6543        }
6544
6545        trace_io_uring_defer(ctx, req, req->user_data);
6546        de->req = req;
6547        de->seq = seq;
6548        list_add_tail(&de->list, &ctx->defer_list);
6549        spin_unlock(&ctx->completion_lock);
6550        return true;
6551}
6552
6553static void io_clean_op(struct io_kiocb *req)
6554{
6555        if (req->flags & REQ_F_BUFFER_SELECTED) {
6556                switch (req->opcode) {
6557                case IORING_OP_READV:
6558                case IORING_OP_READ_FIXED:
6559                case IORING_OP_READ:
6560                        kfree((void *)(unsigned long)req->rw.addr);
6561                        break;
6562                case IORING_OP_RECVMSG:
6563                case IORING_OP_RECV:
6564                        kfree(req->sr_msg.kbuf);
6565                        break;
6566                }
6567        }
6568
6569        if (req->flags & REQ_F_NEED_CLEANUP) {
6570                switch (req->opcode) {
6571                case IORING_OP_READV:
6572                case IORING_OP_READ_FIXED:
6573                case IORING_OP_READ:
6574                case IORING_OP_WRITEV:
6575                case IORING_OP_WRITE_FIXED:
6576                case IORING_OP_WRITE: {
6577                        struct io_async_rw *io = req->async_data;
6578
6579                        kfree(io->free_iovec);
6580                        break;
6581                        }
6582                case IORING_OP_RECVMSG:
6583                case IORING_OP_SENDMSG: {
6584                        struct io_async_msghdr *io = req->async_data;
6585
6586                        kfree(io->free_iov);
6587                        break;
6588                        }
6589                case IORING_OP_SPLICE:
6590                case IORING_OP_TEE:
6591                        if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6592                                io_put_file(req->splice.file_in);
6593                        break;
6594                case IORING_OP_OPENAT:
6595                case IORING_OP_OPENAT2:
6596                        if (req->open.filename)
6597                                putname(req->open.filename);
6598                        break;
6599                case IORING_OP_RENAMEAT:
6600                        putname(req->rename.oldpath);
6601                        putname(req->rename.newpath);
6602                        break;
6603                case IORING_OP_UNLINKAT:
6604                        putname(req->unlink.filename);
6605                        break;
6606                case IORING_OP_MKDIRAT:
6607                        putname(req->mkdir.filename);
6608                        break;
6609                case IORING_OP_SYMLINKAT:
6610                        putname(req->symlink.oldpath);
6611                        putname(req->symlink.newpath);
6612                        break;
6613                case IORING_OP_LINKAT:
6614                        putname(req->hardlink.oldpath);
6615                        putname(req->hardlink.newpath);
6616                        break;
6617                }
6618        }
6619        if ((req->flags & REQ_F_POLLED) && req->apoll) {
6620                kfree(req->apoll->double_poll);
6621                kfree(req->apoll);
6622                req->apoll = NULL;
6623        }
6624        if (req->flags & REQ_F_INFLIGHT) {
6625                struct io_uring_task *tctx = req->task->io_uring;
6626
6627                atomic_dec(&tctx->inflight_tracked);
6628        }
6629        if (req->flags & REQ_F_CREDS)
6630                put_cred(req->creds);
6631
6632        req->flags &= ~IO_REQ_CLEAN_FLAGS;
6633}
6634
6635static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6636{
6637        struct io_ring_ctx *ctx = req->ctx;
6638        const struct cred *creds = NULL;
6639        int ret;
6640
6641        if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6642                creds = override_creds(req->creds);
6643
6644        switch (req->opcode) {
6645        case IORING_OP_NOP:
6646                ret = io_nop(req, issue_flags);
6647                break;
6648        case IORING_OP_READV:
6649        case IORING_OP_READ_FIXED:
6650        case IORING_OP_READ:
6651                ret = io_read(req, issue_flags);
6652                break;
6653        case IORING_OP_WRITEV:
6654        case IORING_OP_WRITE_FIXED:
6655        case IORING_OP_WRITE:
6656                ret = io_write(req, issue_flags);
6657                break;
6658        case IORING_OP_FSYNC:
6659                ret = io_fsync(req, issue_flags);
6660                break;
6661        case IORING_OP_POLL_ADD:
6662                ret = io_poll_add(req, issue_flags);
6663                break;
6664        case IORING_OP_POLL_REMOVE:
6665                ret = io_poll_update(req, issue_flags);
6666                break;
6667        case IORING_OP_SYNC_FILE_RANGE:
6668                ret = io_sync_file_range(req, issue_flags);
6669                break;
6670        case IORING_OP_SENDMSG:
6671                ret = io_sendmsg(req, issue_flags);
6672                break;
6673        case IORING_OP_SEND:
6674                ret = io_send(req, issue_flags);
6675                break;
6676        case IORING_OP_RECVMSG:
6677                ret = io_recvmsg(req, issue_flags);
6678                break;
6679        case IORING_OP_RECV:
6680                ret = io_recv(req, issue_flags);
6681                break;
6682        case IORING_OP_TIMEOUT:
6683                ret = io_timeout(req, issue_flags);
6684                break;
6685        case IORING_OP_TIMEOUT_REMOVE:
6686                ret = io_timeout_remove(req, issue_flags);
6687                break;
6688        case IORING_OP_ACCEPT:
6689                ret = io_accept(req, issue_flags);
6690                break;
6691        case IORING_OP_CONNECT:
6692                ret = io_connect(req, issue_flags);
6693                break;
6694        case IORING_OP_ASYNC_CANCEL:
6695                ret = io_async_cancel(req, issue_flags);
6696                break;
6697        case IORING_OP_FALLOCATE:
6698                ret = io_fallocate(req, issue_flags);
6699                break;
6700        case IORING_OP_OPENAT:
6701                ret = io_openat(req, issue_flags);
6702                break;
6703        case IORING_OP_CLOSE:
6704                ret = io_close(req, issue_flags);
6705                break;
6706        case IORING_OP_FILES_UPDATE:
6707                ret = io_files_update(req, issue_flags);
6708                break;
6709        case IORING_OP_STATX:
6710                ret = io_statx(req, issue_flags);
6711                break;
6712        case IORING_OP_FADVISE:
6713                ret = io_fadvise(req, issue_flags);
6714                break;
6715        case IORING_OP_MADVISE:
6716                ret = io_madvise(req, issue_flags);
6717                break;
6718        case IORING_OP_OPENAT2:
6719                ret = io_openat2(req, issue_flags);
6720                break;
6721        case IORING_OP_EPOLL_CTL:
6722                ret = io_epoll_ctl(req, issue_flags);
6723                break;
6724        case IORING_OP_SPLICE:
6725                ret = io_splice(req, issue_flags);
6726                break;
6727        case IORING_OP_PROVIDE_BUFFERS:
6728                ret = io_provide_buffers(req, issue_flags);
6729                break;
6730        case IORING_OP_REMOVE_BUFFERS:
6731                ret = io_remove_buffers(req, issue_flags);
6732                break;
6733        case IORING_OP_TEE:
6734                ret = io_tee(req, issue_flags);
6735                break;
6736        case IORING_OP_SHUTDOWN:
6737                ret = io_shutdown(req, issue_flags);
6738                break;
6739        case IORING_OP_RENAMEAT:
6740                ret = io_renameat(req, issue_flags);
6741                break;
6742        case IORING_OP_UNLINKAT:
6743                ret = io_unlinkat(req, issue_flags);
6744                break;
6745        case IORING_OP_MKDIRAT:
6746                ret = io_mkdirat(req, issue_flags);
6747                break;
6748        case IORING_OP_SYMLINKAT:
6749                ret = io_symlinkat(req, issue_flags);
6750                break;
6751        case IORING_OP_LINKAT:
6752                ret = io_linkat(req, issue_flags);
6753                break;
6754        default:
6755                ret = -EINVAL;
6756                break;
6757        }
6758
6759        if (creds)
6760                revert_creds(creds);
6761        if (ret)
6762                return ret;
6763        /* If the op doesn't have a file, we're not polling for it */
6764        if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6765                io_iopoll_req_issued(req);
6766
6767        return 0;
6768}
6769
6770static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6771{
6772        struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6773
6774        req = io_put_req_find_next(req);
6775        return req ? &req->work : NULL;
6776}
6777
6778static void io_wq_submit_work(struct io_wq_work *work)
6779{
6780        struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6781        struct io_kiocb *timeout;
6782        int ret = 0;
6783
6784        /* one will be dropped by ->io_free_work() after returning to io-wq */
6785        if (!(req->flags & REQ_F_REFCOUNT))
6786                __io_req_set_refcount(req, 2);
6787        else
6788                req_ref_get(req);
6789
6790        timeout = io_prep_linked_timeout(req);
6791        if (timeout)
6792                io_queue_linked_timeout(timeout);
6793
6794        /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6795        if (work->flags & IO_WQ_WORK_CANCEL)
6796                ret = -ECANCELED;
6797
6798        if (!ret) {
6799                do {
6800                        ret = io_issue_sqe(req, 0);
6801                        /*
6802                         * We can get EAGAIN for polled IO even though we're
6803                         * forcing a sync submission from here, since we can't
6804                         * wait for request slots on the block side.
6805                         */
6806                        if (ret != -EAGAIN)
6807                                break;
6808                        cond_resched();
6809                } while (1);
6810        }
6811
6812        /* avoid locking problems by failing it from a clean context */
6813        if (ret)
6814                io_req_task_queue_fail(req, ret);
6815}
6816
6817static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6818                                                       unsigned i)
6819{
6820        return &table->files[i];
6821}
6822
6823static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6824                                              int index)
6825{
6826        struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6827
6828        return (struct file *) (slot->file_ptr & FFS_MASK);
6829}
6830
6831static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6832{
6833        unsigned long file_ptr = (unsigned long) file;
6834
6835        if (__io_file_supports_nowait(file, READ))
6836                file_ptr |= FFS_ASYNC_READ;
6837        if (__io_file_supports_nowait(file, WRITE))
6838                file_ptr |= FFS_ASYNC_WRITE;
6839        if (S_ISREG(file_inode(file)->i_mode))
6840                file_ptr |= FFS_ISREG;
6841        file_slot->file_ptr = file_ptr;
6842}
6843
6844static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6845                                             struct io_kiocb *req, int fd)
6846{
6847        struct file *file;
6848        unsigned long file_ptr;
6849
6850        if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6851                return NULL;
6852        fd = array_index_nospec(fd, ctx->nr_user_files);
6853        file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6854        file = (struct file *) (file_ptr & FFS_MASK);
6855        file_ptr &= ~FFS_MASK;
6856        /* mask in overlapping REQ_F and FFS bits */
6857        req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6858        io_req_set_rsrc_node(req);
6859        return file;
6860}
6861
6862static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6863                                       struct io_kiocb *req, int fd)
6864{
6865        struct file *file = fget(fd);
6866
6867        trace_io_uring_file_get(ctx, fd);
6868
6869        /* we don't allow fixed io_uring files */
6870        if (file && unlikely(file->f_op == &io_uring_fops))
6871                io_req_track_inflight(req);
6872        return file;
6873}
6874
6875static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6876                                       struct io_kiocb *req, int fd, bool fixed)
6877{
6878        if (fixed)
6879                return io_file_get_fixed(ctx, req, fd);
6880        else
6881                return io_file_get_normal(ctx, req, fd);
6882}
6883
6884static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
6885{
6886        struct io_kiocb *prev = req->timeout.prev;
6887        int ret;
6888
6889        if (prev) {
6890                ret = io_try_cancel_userdata(req, prev->user_data);
6891                io_req_complete_post(req, ret ?: -ETIME, 0);
6892                io_put_req(prev);
6893        } else {
6894                io_req_complete_post(req, -ETIME, 0);
6895        }
6896}
6897
6898static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6899{
6900        struct io_timeout_data *data = container_of(timer,
6901                                                struct io_timeout_data, timer);
6902        struct io_kiocb *prev, *req = data->req;
6903        struct io_ring_ctx *ctx = req->ctx;
6904        unsigned long flags;
6905
6906        spin_lock_irqsave(&ctx->timeout_lock, flags);
6907        prev = req->timeout.head;
6908        req->timeout.head = NULL;
6909
6910        /*
6911         * We don't expect the list to be empty, that will only happen if we
6912         * race with the completion of the linked work.
6913         */
6914        if (prev) {
6915                io_remove_next_linked(prev);
6916                if (!req_ref_inc_not_zero(prev))
6917                        prev = NULL;
6918        }
6919        list_del(&req->timeout.list);
6920        req->timeout.prev = prev;
6921        spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6922
6923        req->io_task_work.func = io_req_task_link_timeout;
6924        io_req_task_work_add(req);
6925        return HRTIMER_NORESTART;
6926}
6927
6928static void io_queue_linked_timeout(struct io_kiocb *req)
6929{
6930        struct io_ring_ctx *ctx = req->ctx;
6931
6932        spin_lock_irq(&ctx->timeout_lock);
6933        /*
6934         * If the back reference is NULL, then our linked request finished
6935         * before we got a chance to setup the timer
6936         */
6937        if (req->timeout.head) {
6938                struct io_timeout_data *data = req->async_data;
6939
6940                data->timer.function = io_link_timeout_fn;
6941                hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6942                                data->mode);
6943                list_add_tail(&req->timeout.list, &ctx->ltimeout_list);
6944        }
6945        spin_unlock_irq(&ctx->timeout_lock);
6946        /* drop submission reference */
6947        io_put_req(req);
6948}
6949
6950static void __io_queue_sqe(struct io_kiocb *req)
6951        __must_hold(&req->ctx->uring_lock)
6952{
6953        struct io_kiocb *linked_timeout;
6954        int ret;
6955
6956issue_sqe:
6957        ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6958
6959        /*
6960         * We async punt it if the file wasn't marked NOWAIT, or if the file
6961         * doesn't support non-blocking read/write attempts
6962         */
6963        if (likely(!ret)) {
6964                if (req->flags & REQ_F_COMPLETE_INLINE) {
6965                        struct io_ring_ctx *ctx = req->ctx;
6966                        struct io_submit_state *state = &ctx->submit_state;
6967
6968                        state->compl_reqs[state->compl_nr++] = req;
6969                        if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6970                                io_submit_flush_completions(ctx);
6971                        return;
6972                }
6973
6974                linked_timeout = io_prep_linked_timeout(req);
6975                if (linked_timeout)
6976                        io_queue_linked_timeout(linked_timeout);
6977        } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6978                linked_timeout = io_prep_linked_timeout(req);
6979
6980                switch (io_arm_poll_handler(req)) {
6981                case IO_APOLL_READY:
6982                        if (linked_timeout)
6983                                io_queue_linked_timeout(linked_timeout);
6984                        goto issue_sqe;
6985                case IO_APOLL_ABORTED:
6986                        /*
6987                         * Queued up for async execution, worker will release
6988                         * submit reference when the iocb is actually submitted.
6989                         */
6990                        io_queue_async_work(req, NULL);
6991                        break;
6992                }
6993
6994                if (linked_timeout)
6995                        io_queue_linked_timeout(linked_timeout);
6996        } else {
6997                io_req_complete_failed(req, ret);
6998        }
6999}
7000
7001static inline void io_queue_sqe(struct io_kiocb *req)
7002        __must_hold(&req->ctx->uring_lock)
7003{
7004        if (unlikely(req->ctx->drain_active) && io_drain_req(req))
7005                return;
7006
7007        if (likely(!(req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)))) {
7008                __io_queue_sqe(req);
7009        } else if (req->flags & REQ_F_FAIL) {
7010                io_req_complete_fail_submit(req);
7011        } else {
7012                int ret = io_req_prep_async(req);
7013
7014                if (unlikely(ret))
7015                        io_req_complete_failed(req, ret);
7016                else
7017                        io_queue_async_work(req, NULL);
7018        }
7019}
7020
7021/*
7022 * Check SQE restrictions (opcode and flags).
7023 *
7024 * Returns 'true' if SQE is allowed, 'false' otherwise.
7025 */
7026static inline bool io_check_restriction(struct io_ring_ctx *ctx,
7027                                        struct io_kiocb *req,
7028                                        unsigned int sqe_flags)
7029{
7030        if (likely(!ctx->restricted))
7031                return true;
7032
7033        if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
7034                return false;
7035
7036        if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
7037            ctx->restrictions.sqe_flags_required)
7038                return false;
7039
7040        if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
7041                          ctx->restrictions.sqe_flags_required))
7042                return false;
7043
7044        return true;
7045}
7046
7047static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
7048                       const struct io_uring_sqe *sqe)
7049        __must_hold(&ctx->uring_lock)
7050{
7051        struct io_submit_state *state;
7052        unsigned int sqe_flags;
7053        int personality, ret = 0;
7054
7055        /* req is partially pre-initialised, see io_preinit_req() */
7056        req->opcode = READ_ONCE(sqe->opcode);
7057        /* same numerical values with corresponding REQ_F_*, safe to copy */
7058        req->flags = sqe_flags = READ_ONCE(sqe->flags);
7059        req->user_data = READ_ONCE(sqe->user_data);
7060        req->file = NULL;
7061        req->fixed_rsrc_refs = NULL;
7062        req->task = current;
7063
7064        /* enforce forwards compatibility on users */
7065        if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
7066                return -EINVAL;
7067        if (unlikely(req->opcode >= IORING_OP_LAST))
7068                return -EINVAL;
7069        if (!io_check_restriction(ctx, req, sqe_flags))
7070                return -EACCES;
7071
7072        if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
7073            !io_op_defs[req->opcode].buffer_select)
7074                return -EOPNOTSUPP;
7075        if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
7076                ctx->drain_active = true;
7077
7078        personality = READ_ONCE(sqe->personality);
7079        if (personality) {
7080                req->creds = xa_load(&ctx->personalities, personality);
7081                if (!req->creds)
7082                        return -EINVAL;
7083                get_cred(req->creds);
7084                req->flags |= REQ_F_CREDS;
7085        }
7086        state = &ctx->submit_state;
7087
7088        /*
7089         * Plug now if we have more than 1 IO left after this, and the target
7090         * is potentially a read/write to block based storage.
7091         */
7092        if (!state->plug_started && state->ios_left > 1 &&
7093            io_op_defs[req->opcode].plug) {
7094                blk_start_plug(&state->plug);
7095                state->plug_started = true;
7096        }
7097
7098        if (io_op_defs[req->opcode].needs_file) {
7099                req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
7100                                        (sqe_flags & IOSQE_FIXED_FILE));
7101                if (unlikely(!req->file))
7102                        ret = -EBADF;
7103        }
7104
7105        state->ios_left--;
7106        return ret;
7107}
7108
7109static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
7110                         const struct io_uring_sqe *sqe)
7111        __must_hold(&ctx->uring_lock)
7112{
7113        struct io_submit_link *link = &ctx->submit_state.link;
7114        int ret;
7115
7116        ret = io_init_req(ctx, req, sqe);
7117        if (unlikely(ret)) {
7118fail_req:
7119                /* fail even hard links since we don't submit */
7120                if (link->head) {
7121                        /*
7122                         * we can judge a link req is failed or cancelled by if
7123                         * REQ_F_FAIL is set, but the head is an exception since
7124                         * it may be set REQ_F_FAIL because of other req's failure
7125                         * so let's leverage req->result to distinguish if a head
7126                         * is set REQ_F_FAIL because of its failure or other req's
7127                         * failure so that we can set the correct ret code for it.
7128                         * init result here to avoid affecting the normal path.
7129                         */
7130                        if (!(link->head->flags & REQ_F_FAIL))
7131                                req_fail_link_node(link->head, -ECANCELED);
7132                } else if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7133                        /*
7134                         * the current req is a normal req, we should return
7135                         * error and thus break the submittion loop.
7136                         */
7137                        io_req_complete_failed(req, ret);
7138                        return ret;
7139                }
7140                req_fail_link_node(req, ret);
7141        } else {
7142                ret = io_req_prep(req, sqe);
7143                if (unlikely(ret))
7144                        goto fail_req;
7145        }
7146
7147        /* don't need @sqe from now on */
7148        trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
7149                                  req->flags, true,
7150                                  ctx->flags & IORING_SETUP_SQPOLL);
7151
7152        /*
7153         * If we already have a head request, queue this one for async
7154         * submittal once the head completes. If we don't have a head but
7155         * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7156         * submitted sync once the chain is complete. If none of those
7157         * conditions are true (normal request), then just queue it.
7158         */
7159        if (link->head) {
7160                struct io_kiocb *head = link->head;
7161
7162                if (!(req->flags & REQ_F_FAIL)) {
7163                        ret = io_req_prep_async(req);
7164                        if (unlikely(ret)) {
7165                                req_fail_link_node(req, ret);
7166                                if (!(head->flags & REQ_F_FAIL))
7167                                        req_fail_link_node(head, -ECANCELED);
7168                        }
7169                }
7170                trace_io_uring_link(ctx, req, head);
7171                link->last->link = req;
7172                link->last = req;
7173
7174                /* last request of a link, enqueue the link */
7175                if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7176                        link->head = NULL;
7177                        io_queue_sqe(head);
7178                }
7179        } else {
7180                if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
7181                        link->head = req;
7182                        link->last = req;
7183                } else {
7184                        io_queue_sqe(req);
7185                }
7186        }
7187
7188        return 0;
7189}
7190
7191/*
7192 * Batched submission is done, ensure local IO is flushed out.
7193 */
7194static void io_submit_state_end(struct io_submit_state *state,
7195                                struct io_ring_ctx *ctx)
7196{
7197        if (state->link.head)
7198                io_queue_sqe(state->link.head);
7199        if (state->compl_nr)
7200                io_submit_flush_completions(ctx);
7201        if (state->plug_started)
7202                blk_finish_plug(&state->plug);
7203}
7204
7205/*
7206 * Start submission side cache.
7207 */
7208static void io_submit_state_start(struct io_submit_state *state,
7209                                  unsigned int max_ios)
7210{
7211        state->plug_started = false;
7212        state->ios_left = max_ios;
7213        /* set only head, no need to init link_last in advance */
7214        state->link.head = NULL;
7215}
7216
7217static void io_commit_sqring(struct io_ring_ctx *ctx)
7218{
7219        struct io_rings *rings = ctx->rings;
7220
7221        /*
7222         * Ensure any loads from the SQEs are done at this point,
7223         * since once we write the new head, the application could
7224         * write new data to them.
7225         */
7226        smp_store_release(&rings->sq.head, ctx->cached_sq_head);
7227}
7228
7229/*
7230 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7231 * that is mapped by userspace. This means that care needs to be taken to
7232 * ensure that reads are stable, as we cannot rely on userspace always
7233 * being a good citizen. If members of the sqe are validated and then later
7234 * used, it's important that those reads are done through READ_ONCE() to
7235 * prevent a re-load down the line.
7236 */
7237static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
7238{
7239        unsigned head, mask = ctx->sq_entries - 1;
7240        unsigned sq_idx = ctx->cached_sq_head++ & mask;
7241
7242        /*
7243         * The cached sq head (or cq tail) serves two purposes:
7244         *
7245         * 1) allows us to batch the cost of updating the user visible
7246         *    head updates.
7247         * 2) allows the kernel side to track the head on its own, even
7248         *    though the application is the one updating it.
7249         */
7250        head = READ_ONCE(ctx->sq_array[sq_idx]);
7251        if (likely(head < ctx->sq_entries))
7252                return &ctx->sq_sqes[head];
7253
7254        /* drop invalid entries */
7255        ctx->cq_extra--;
7256        WRITE_ONCE(ctx->rings->sq_dropped,
7257                   READ_ONCE(ctx->rings->sq_dropped) + 1);
7258        return NULL;
7259}
7260
7261static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
7262        __must_hold(&ctx->uring_lock)
7263{
7264        int submitted = 0;
7265
7266        /* make sure SQ entry isn't read before tail */
7267        nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
7268        if (!percpu_ref_tryget_many(&ctx->refs, nr))
7269                return -EAGAIN;
7270        io_get_task_refs(nr);
7271
7272        io_submit_state_start(&ctx->submit_state, nr);
7273        while (submitted < nr) {
7274                const struct io_uring_sqe *sqe;
7275                struct io_kiocb *req;
7276
7277                req = io_alloc_req(ctx);
7278                if (unlikely(!req)) {
7279                        if (!submitted)
7280                                submitted = -EAGAIN;
7281                        break;
7282                }
7283                sqe = io_get_sqe(ctx);
7284                if (unlikely(!sqe)) {
7285                        list_add(&req->inflight_entry, &ctx->submit_state.free_list);
7286                        break;
7287                }
7288                /* will complete beyond this point, count as submitted */
7289                submitted++;
7290                if (io_submit_sqe(ctx, req, sqe))
7291                        break;
7292        }
7293
7294        if (unlikely(submitted != nr)) {
7295                int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
7296                int unused = nr - ref_used;
7297
7298                current->io_uring->cached_refs += unused;
7299                percpu_ref_put_many(&ctx->refs, unused);
7300        }
7301
7302        io_submit_state_end(&ctx->submit_state, ctx);
7303         /* Commit SQ ring head once we've consumed and submitted all SQEs */
7304        io_commit_sqring(ctx);
7305
7306        return submitted;
7307}
7308
7309static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
7310{
7311        return READ_ONCE(sqd->state);
7312}
7313
7314static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
7315{
7316        /* Tell userspace we may need a wakeup call */
7317        spin_lock(&ctx->completion_lock);
7318        WRITE_ONCE(ctx->rings->sq_flags,
7319                   ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
7320        spin_unlock(&ctx->completion_lock);
7321}
7322
7323static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
7324{
7325        spin_lock(&ctx->completion_lock);
7326        WRITE_ONCE(ctx->rings->sq_flags,
7327                   ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
7328        spin_unlock(&ctx->completion_lock);
7329}
7330
7331static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7332{
7333        unsigned int to_submit;
7334        int ret = 0;
7335
7336        to_submit = io_sqring_entries(ctx);
7337        /* if we're handling multiple rings, cap submit size for fairness */
7338        if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
7339                to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
7340
7341        if (!list_empty(&ctx->iopoll_list) || to_submit) {
7342                unsigned nr_events = 0;
7343                const struct cred *creds = NULL;
7344
7345                if (ctx->sq_creds != current_cred())
7346                        creds = override_creds(ctx->sq_creds);
7347
7348                mutex_lock(&ctx->uring_lock);
7349                if (!list_empty(&ctx->iopoll_list))
7350                        io_do_iopoll(ctx, &nr_events, 0);
7351
7352                /*
7353                 * Don't submit if refs are dying, good for io_uring_register(),
7354                 * but also it is relied upon by io_ring_exit_work()
7355                 */
7356                if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
7357                    !(ctx->flags & IORING_SETUP_R_DISABLED))
7358                        ret = io_submit_sqes(ctx, to_submit);
7359                mutex_unlock(&ctx->uring_lock);
7360
7361                if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
7362                        wake_up(&ctx->sqo_sq_wait);
7363                if (creds)
7364                        revert_creds(creds);
7365        }
7366
7367        return ret;
7368}
7369
7370static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7371{
7372        struct io_ring_ctx *ctx;
7373        unsigned sq_thread_idle = 0;
7374
7375        list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7376                sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
7377        sqd->sq_thread_idle = sq_thread_idle;
7378}
7379
7380static bool io_sqd_handle_event(struct io_sq_data *sqd)
7381{
7382        bool did_sig = false;
7383        struct ksignal ksig;
7384
7385        if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
7386            signal_pending(current)) {
7387                mutex_unlock(&sqd->lock);
7388                if (signal_pending(current))
7389                        did_sig = get_signal(&ksig);
7390                cond_resched();
7391                mutex_lock(&sqd->lock);
7392        }
7393        return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7394}
7395
7396static int io_sq_thread(void *data)
7397{
7398        struct io_sq_data *sqd = data;
7399        struct io_ring_ctx *ctx;
7400        unsigned long timeout = 0;
7401        char buf[TASK_COMM_LEN];
7402        DEFINE_WAIT(wait);
7403
7404        snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
7405        set_task_comm(current, buf);
7406
7407        if (sqd->sq_cpu != -1)
7408                set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
7409        else
7410                set_cpus_allowed_ptr(current, cpu_online_mask);
7411        current->flags |= PF_NO_SETAFFINITY;
7412
7413        mutex_lock(&sqd->lock);
7414        while (1) {
7415                bool cap_entries, sqt_spin = false;
7416
7417                if (io_sqd_events_pending(sqd) || signal_pending(current)) {
7418                        if (io_sqd_handle_event(sqd))
7419                                break;
7420                        timeout = jiffies + sqd->sq_thread_idle;
7421                }
7422
7423                cap_entries = !list_is_singular(&sqd->ctx_list);
7424                list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7425                        int ret = __io_sq_thread(ctx, cap_entries);
7426
7427                        if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7428                                sqt_spin = true;
7429                }
7430                if (io_run_task_work())
7431                        sqt_spin = true;
7432
7433                if (sqt_spin || !time_after(jiffies, timeout)) {
7434                        cond_resched();
7435                        if (sqt_spin)
7436                                timeout = jiffies + sqd->sq_thread_idle;
7437                        continue;
7438                }
7439
7440                prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7441                if (!io_sqd_events_pending(sqd) && !current->task_works) {
7442                        bool needs_sched = true;
7443
7444                        list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7445                                io_ring_set_wakeup_flag(ctx);
7446
7447                                if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7448                                    !list_empty_careful(&ctx->iopoll_list)) {
7449                                        needs_sched = false;
7450                                        break;
7451                                }
7452                                if (io_sqring_entries(ctx)) {
7453                                        needs_sched = false;
7454                                        break;
7455                                }
7456                        }
7457
7458                        if (needs_sched) {
7459                                mutex_unlock(&sqd->lock);
7460                                schedule();
7461                                mutex_lock(&sqd->lock);
7462                        }
7463                        list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7464                                io_ring_clear_wakeup_flag(ctx);
7465                }
7466
7467                finish_wait(&sqd->wait, &wait);
7468                timeout = jiffies + sqd->sq_thread_idle;
7469        }
7470
7471        io_uring_cancel_generic(true, sqd);
7472        sqd->thread = NULL;
7473        list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7474                io_ring_set_wakeup_flag(ctx);
7475        io_run_task_work();
7476        mutex_unlock(&sqd->lock);
7477
7478        complete(&sqd->exited);
7479        do_exit(0);
7480}
7481
7482struct io_wait_queue {
7483        struct wait_queue_entry wq;
7484        struct io_ring_ctx *ctx;
7485        unsigned cq_tail;
7486        unsigned nr_timeouts;
7487};
7488
7489static inline bool io_should_wake(struct io_wait_queue *iowq)
7490{
7491        struct io_ring_ctx *ctx = iowq->ctx;
7492        int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7493
7494        /*
7495         * Wake up if we have enough events, or if a timeout occurred since we
7496         * started waiting. For timeouts, we always want to return to userspace,
7497         * regardless of event count.
7498         */
7499        return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7500}
7501
7502static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7503                            int wake_flags, void *key)
7504{
7505        struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7506                                                        wq);
7507
7508        /*
7509         * Cannot safely flush overflowed CQEs from here, ensure we wake up
7510         * the task, and the next invocation will do it.
7511         */
7512        if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7513                return autoremove_wake_function(curr, mode, wake_flags, key);
7514        return -1;
7515}
7516
7517static int io_run_task_work_sig(void)
7518{
7519        if (io_run_task_work())
7520                return 1;
7521        if (!signal_pending(current))
7522                return 0;
7523        if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7524                return -ERESTARTSYS;
7525        return -EINTR;
7526}
7527
7528/* when returns >0, the caller should retry */
7529static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7530                                          struct io_wait_queue *iowq,
7531                                          signed long *timeout)
7532{
7533        int ret;
7534
7535        /* make sure we run task_work before checking for signals */
7536        ret = io_run_task_work_sig();
7537        if (ret || io_should_wake(iowq))
7538                return ret;
7539        /* let the caller flush overflows, retry */
7540        if (test_bit(0, &ctx->check_cq_overflow))
7541                return 1;
7542
7543        *timeout = schedule_timeout(*timeout);
7544        return !*timeout ? -ETIME : 1;
7545}
7546
7547/*
7548 * Wait until events become available, if we don't already have some. The
7549 * application must reap them itself, as they reside on the shared cq ring.
7550 */
7551static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7552                          const sigset_t __user *sig, size_t sigsz,
7553                          struct __kernel_timespec __user *uts)
7554{
7555        struct io_wait_queue iowq;
7556        struct io_rings *rings = ctx->rings;
7557        signed long timeout = MAX_SCHEDULE_TIMEOUT;
7558        int ret;
7559
7560        do {
7561                io_cqring_overflow_flush(ctx);
7562                if (io_cqring_events(ctx) >= min_events)
7563                        return 0;
7564                if (!io_run_task_work())
7565                        break;
7566        } while (1);
7567
7568        if (uts) {
7569                struct timespec64 ts;
7570
7571                if (get_timespec64(&ts, uts))
7572                        return -EFAULT;
7573                timeout = timespec64_to_jiffies(&ts);
7574        }
7575
7576        if (sig) {
7577#ifdef CONFIG_COMPAT
7578                if (in_compat_syscall())
7579                        ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7580                                                      sigsz);
7581                else
7582#endif
7583                        ret = set_user_sigmask(sig, sigsz);
7584
7585                if (ret)
7586                        return ret;
7587        }
7588
7589        init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7590        iowq.wq.private = current;
7591        INIT_LIST_HEAD(&iowq.wq.entry);
7592        iowq.ctx = ctx;
7593        iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7594        iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7595
7596        trace_io_uring_cqring_wait(ctx, min_events);
7597        do {
7598                /* if we can't even flush overflow, don't wait for more */
7599                if (!io_cqring_overflow_flush(ctx)) {
7600                        ret = -EBUSY;
7601                        break;
7602                }
7603                prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7604                                                TASK_INTERRUPTIBLE);
7605                ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7606                finish_wait(&ctx->cq_wait, &iowq.wq);
7607                cond_resched();
7608        } while (ret > 0);
7609
7610        restore_saved_sigmask_unless(ret == -EINTR);
7611
7612        return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7613}
7614
7615static void io_free_page_table(void **table, size_t size)
7616{
7617        unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7618
7619        for (i = 0; i < nr_tables; i++)
7620                kfree(table[i]);
7621        kfree(table);
7622}
7623
7624static void **io_alloc_page_table(size_t size)
7625{
7626        unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7627        size_t init_size = size;
7628        void **table;
7629
7630        table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7631        if (!table)
7632                return NULL;
7633
7634        for (i = 0; i < nr_tables; i++) {
7635                unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7636
7637                table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7638                if (!table[i]) {
7639                        io_free_page_table(table, init_size);
7640                        return NULL;
7641                }
7642                size -= this_size;
7643        }
7644        return table;
7645}
7646
7647static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7648{
7649        percpu_ref_exit(&ref_node->refs);
7650        kfree(ref_node);
7651}
7652
7653static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7654{
7655        struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7656        struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7657        unsigned long flags;
7658        bool first_add = false;
7659
7660        spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7661        node->done = true;
7662
7663        while (!list_empty(&ctx->rsrc_ref_list)) {
7664                node = list_first_entry(&ctx->rsrc_ref_list,
7665                                            struct io_rsrc_node, node);
7666                /* recycle ref nodes in order */
7667                if (!node->done)
7668                        break;
7669                list_del(&node->node);
7670                first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7671        }
7672        spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7673
7674        if (first_add)
7675                mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7676}
7677
7678static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7679{
7680        struct io_rsrc_node *ref_node;
7681
7682        ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7683        if (!ref_node)
7684                return NULL;
7685
7686        if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7687                            0, GFP_KERNEL)) {
7688                kfree(ref_node);
7689                return NULL;
7690        }
7691        INIT_LIST_HEAD(&ref_node->node);
7692        INIT_LIST_HEAD(&ref_node->rsrc_list);
7693        ref_node->done = false;
7694        return ref_node;
7695}
7696
7697static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7698                                struct io_rsrc_data *data_to_kill)
7699{
7700        WARN_ON_ONCE(!ctx->rsrc_backup_node);
7701        WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7702
7703        if (data_to_kill) {
7704                struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7705
7706                rsrc_node->rsrc_data = data_to_kill;
7707                spin_lock_irq(&ctx->rsrc_ref_lock);
7708                list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7709                spin_unlock_irq(&ctx->rsrc_ref_lock);
7710
7711                atomic_inc(&data_to_kill->refs);
7712                percpu_ref_kill(&rsrc_node->refs);
7713                ctx->rsrc_node = NULL;
7714        }
7715
7716        if (!ctx->rsrc_node) {
7717                ctx->rsrc_node = ctx->rsrc_backup_node;
7718                ctx->rsrc_backup_node = NULL;
7719        }
7720}
7721
7722static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7723{
7724        if (ctx->rsrc_backup_node)
7725                return 0;
7726        ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7727        return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7728}
7729
7730static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7731{
7732        int ret;
7733
7734        /* As we may drop ->uring_lock, other task may have started quiesce */
7735        if (data->quiesce)
7736                return -ENXIO;
7737
7738        data->quiesce = true;
7739        do {
7740                ret = io_rsrc_node_switch_start(ctx);
7741                if (ret)
7742                        break;
7743                io_rsrc_node_switch(ctx, data);
7744
7745                /* kill initial ref, already quiesced if zero */
7746                if (atomic_dec_and_test(&data->refs))
7747                        break;
7748                mutex_unlock(&ctx->uring_lock);
7749                flush_delayed_work(&ctx->rsrc_put_work);
7750                ret = wait_for_completion_interruptible(&data->done);
7751                if (!ret) {
7752                        mutex_lock(&ctx->uring_lock);
7753                        break;
7754                }
7755
7756                atomic_inc(&data->refs);
7757                /* wait for all works potentially completing data->done */
7758                flush_delayed_work(&ctx->rsrc_put_work);
7759                reinit_completion(&data->done);
7760
7761                ret = io_run_task_work_sig();
7762                mutex_lock(&ctx->uring_lock);
7763        } while (ret >= 0);
7764        data->quiesce = false;
7765
7766        return ret;
7767}
7768
7769static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7770{
7771        unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7772        unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7773
7774        return &data->tags[table_idx][off];
7775}
7776
7777static void io_rsrc_data_free(struct io_rsrc_data *data)
7778{
7779        size_t size = data->nr * sizeof(data->tags[0][0]);
7780
7781        if (data->tags)
7782                io_free_page_table((void **)data->tags, size);
7783        kfree(data);
7784}
7785
7786static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7787                              u64 __user *utags, unsigned nr,
7788                              struct io_rsrc_data **pdata)
7789{
7790        struct io_rsrc_data *data;
7791        int ret = -ENOMEM;
7792        unsigned i;
7793
7794        data = kzalloc(sizeof(*data), GFP_KERNEL);
7795        if (!data)
7796                return -ENOMEM;
7797        data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7798        if (!data->tags) {
7799                kfree(data);
7800                return -ENOMEM;
7801        }
7802
7803        data->nr = nr;
7804        data->ctx = ctx;
7805        data->do_put = do_put;
7806        if (utags) {
7807                ret = -EFAULT;
7808                for (i = 0; i < nr; i++) {
7809                        u64 *tag_slot = io_get_tag_slot(data, i);
7810
7811                        if (copy_from_user(tag_slot, &utags[i],
7812                                           sizeof(*tag_slot)))
7813                                goto fail;
7814                }
7815        }
7816
7817        atomic_set(&data->refs, 1);
7818        init_completion(&data->done);
7819        *pdata = data;
7820        return 0;
7821fail:
7822        io_rsrc_data_free(data);
7823        return ret;
7824}
7825
7826static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7827{
7828        table->files = kvcalloc(nr_files, sizeof(table->files[0]),
7829                                GFP_KERNEL_ACCOUNT);
7830        return !!table->files;
7831}
7832
7833static void io_free_file_tables(struct io_file_table *table)
7834{
7835        kvfree(table->files);
7836        table->files = NULL;
7837}
7838
7839static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7840{
7841#if defined(CONFIG_UNIX)
7842        if (ctx->ring_sock) {
7843                struct sock *sock = ctx->ring_sock->sk;
7844                struct sk_buff *skb;
7845
7846                while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7847                        kfree_skb(skb);
7848        }
7849#else
7850        int i;
7851
7852        for (i = 0; i < ctx->nr_user_files; i++) {
7853                struct file *file;
7854
7855                file = io_file_from_index(ctx, i);
7856                if (file)
7857                        fput(file);
7858        }
7859#endif
7860        io_free_file_tables(&ctx->file_table);
7861        io_rsrc_data_free(ctx->file_data);
7862        ctx->file_data = NULL;
7863        ctx->nr_user_files = 0;
7864}
7865
7866static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7867{
7868        int ret;
7869
7870        if (!ctx->file_data)
7871                return -ENXIO;
7872        ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7873        if (!ret)
7874                __io_sqe_files_unregister(ctx);
7875        return ret;
7876}
7877
7878static void io_sq_thread_unpark(struct io_sq_data *sqd)
7879        __releases(&sqd->lock)
7880{
7881        WARN_ON_ONCE(sqd->thread == current);
7882
7883        /*
7884         * Do the dance but not conditional clear_bit() because it'd race with
7885         * other threads incrementing park_pending and setting the bit.
7886         */
7887        clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7888        if (atomic_dec_return(&sqd->park_pending))
7889                set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7890        mutex_unlock(&sqd->lock);
7891}
7892
7893static void io_sq_thread_park(struct io_sq_data *sqd)
7894        __acquires(&sqd->lock)
7895{
7896        WARN_ON_ONCE(sqd->thread == current);
7897
7898        atomic_inc(&sqd->park_pending);
7899        set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7900        mutex_lock(&sqd->lock);
7901        if (sqd->thread)
7902                wake_up_process(sqd->thread);
7903}
7904
7905static void io_sq_thread_stop(struct io_sq_data *sqd)
7906{
7907        WARN_ON_ONCE(sqd->thread == current);
7908        WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7909
7910        set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7911        mutex_lock(&sqd->lock);
7912        if (sqd->thread)
7913                wake_up_process(sqd->thread);
7914        mutex_unlock(&sqd->lock);
7915        wait_for_completion(&sqd->exited);
7916}
7917
7918static void io_put_sq_data(struct io_sq_data *sqd)
7919{
7920        if (refcount_dec_and_test(&sqd->refs)) {
7921                WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7922
7923                io_sq_thread_stop(sqd);
7924                kfree(sqd);
7925        }
7926}
7927
7928static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7929{
7930        struct io_sq_data *sqd = ctx->sq_data;
7931
7932        if (sqd) {
7933                io_sq_thread_park(sqd);
7934                list_del_init(&ctx->sqd_list);
7935                io_sqd_update_thread_idle(sqd);
7936                io_sq_thread_unpark(sqd);
7937
7938                io_put_sq_data(sqd);
7939                ctx->sq_data = NULL;
7940        }
7941}
7942
7943static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7944{
7945        struct io_ring_ctx *ctx_attach;
7946        struct io_sq_data *sqd;
7947        struct fd f;
7948
7949        f = fdget(p->wq_fd);
7950        if (!f.file)
7951                return ERR_PTR(-ENXIO);
7952        if (f.file->f_op != &io_uring_fops) {
7953                fdput(f);
7954                return ERR_PTR(-EINVAL);
7955        }
7956
7957        ctx_attach = f.file->private_data;
7958        sqd = ctx_attach->sq_data;
7959        if (!sqd) {
7960                fdput(f);
7961                return ERR_PTR(-EINVAL);
7962        }
7963        if (sqd->task_tgid != current->tgid) {
7964                fdput(f);
7965                return ERR_PTR(-EPERM);
7966        }
7967
7968        refcount_inc(&sqd->refs);
7969        fdput(f);
7970        return sqd;
7971}
7972
7973static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7974                                         bool *attached)
7975{
7976        struct io_sq_data *sqd;
7977
7978        *attached = false;
7979        if (p->flags & IORING_SETUP_ATTACH_WQ) {
7980                sqd = io_attach_sq_data(p);
7981                if (!IS_ERR(sqd)) {
7982                        *attached = true;
7983                        return sqd;
7984                }
7985                /* fall through for EPERM case, setup new sqd/task */
7986                if (PTR_ERR(sqd) != -EPERM)
7987                        return sqd;
7988        }
7989
7990        sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7991        if (!sqd)
7992                return ERR_PTR(-ENOMEM);
7993
7994        atomic_set(&sqd->park_pending, 0);
7995        refcount_set(&sqd->refs, 1);
7996        INIT_LIST_HEAD(&sqd->ctx_list);
7997        mutex_init(&sqd->lock);
7998        init_waitqueue_head(&sqd->wait);
7999        init_completion(&sqd->exited);
8000        return sqd;
8001}
8002
8003#if defined(CONFIG_UNIX)
8004/*
8005 * Ensure the UNIX gc is aware of our file set, so we are certain that
8006 * the io_uring can be safely unregistered on process exit, even if we have
8007 * loops in the file referencing.
8008 */
8009static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
8010{
8011        struct sock *sk = ctx->ring_sock->sk;
8012        struct scm_fp_list *fpl;
8013        struct sk_buff *skb;
8014        int i, nr_files;
8015
8016        fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
8017        if (!fpl)
8018                return -ENOMEM;
8019
8020        skb = alloc_skb(0, GFP_KERNEL);
8021        if (!skb) {
8022                kfree(fpl);
8023                return -ENOMEM;
8024        }
8025
8026        skb->sk = sk;
8027
8028        nr_files = 0;
8029        fpl->user = get_uid(current_user());
8030        for (i = 0; i < nr; i++) {
8031                struct file *file = io_file_from_index(ctx, i + offset);
8032
8033                if (!file)
8034                        continue;
8035                fpl->fp[nr_files] = get_file(file);
8036                unix_inflight(fpl->user, fpl->fp[nr_files]);
8037                nr_files++;
8038        }
8039
8040        if (nr_files) {
8041                fpl->max = SCM_MAX_FD;
8042                fpl->count = nr_files;
8043                UNIXCB(skb).fp = fpl;
8044                skb->destructor = unix_destruct_scm;
8045                refcount_add(skb->truesize, &sk->sk_wmem_alloc);
8046                skb_queue_head(&sk->sk_receive_queue, skb);
8047
8048                for (i = 0; i < nr_files; i++)
8049                        fput(fpl->fp[i]);
8050        } else {
8051                kfree_skb(skb);
8052                kfree(fpl);
8053        }
8054
8055        return 0;
8056}
8057
8058/*
8059 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8060 * causes regular reference counting to break down. We rely on the UNIX
8061 * garbage collection to take care of this problem for us.
8062 */
8063static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8064{
8065        unsigned left, total;
8066        int ret = 0;
8067
8068        total = 0;
8069        left = ctx->nr_user_files;
8070        while (left) {
8071                unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
8072
8073                ret = __io_sqe_files_scm(ctx, this_files, total);
8074                if (ret)
8075                        break;
8076                left -= this_files;
8077                total += this_files;
8078        }
8079
8080        if (!ret)
8081                return 0;
8082
8083        while (total < ctx->nr_user_files) {
8084                struct file *file = io_file_from_index(ctx, total);
8085
8086                if (file)
8087                        fput(file);
8088                total++;
8089        }
8090
8091        return ret;
8092}
8093#else
8094static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8095{
8096        return 0;
8097}
8098#endif
8099
8100static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8101{
8102        struct file *file = prsrc->file;
8103#if defined(CONFIG_UNIX)
8104        struct sock *sock = ctx->ring_sock->sk;
8105        struct sk_buff_head list, *head = &sock->sk_receive_queue;
8106        struct sk_buff *skb;
8107        int i;
8108
8109        __skb_queue_head_init(&list);
8110
8111        /*
8112         * Find the skb that holds this file in its SCM_RIGHTS. When found,
8113         * remove this entry and rearrange the file array.
8114         */
8115        skb = skb_dequeue(head);
8116        while (skb) {
8117                struct scm_fp_list *fp;
8118
8119                fp = UNIXCB(skb).fp;
8120                for (i = 0; i < fp->count; i++) {
8121                        int left;
8122
8123                        if (fp->fp[i] != file)
8124                                continue;
8125
8126                        unix_notinflight(fp->user, fp->fp[i]);
8127                        left = fp->count - 1 - i;
8128                        if (left) {
8129                                memmove(&fp->fp[i], &fp->fp[i + 1],
8130                                                left * sizeof(struct file *));
8131                        }
8132                        fp->count--;
8133                        if (!fp->count) {
8134                                kfree_skb(skb);
8135                                skb = NULL;
8136                        } else {
8137                                __skb_queue_tail(&list, skb);
8138                        }
8139                        fput(file);
8140                        file = NULL;
8141                        break;
8142                }
8143
8144                if (!file)
8145                        break;
8146
8147                __skb_queue_tail(&list, skb);
8148
8149                skb = skb_dequeue(head);
8150        }
8151
8152        if (skb_peek(&list)) {
8153                spin_lock_irq(&head->lock);
8154                while ((skb = __skb_dequeue(&list)) != NULL)
8155                        __skb_queue_tail(head, skb);
8156                spin_unlock_irq(&head->lock);
8157        }
8158#else
8159        fput(file);
8160#endif
8161}
8162
8163static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
8164{
8165        struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
8166        struct io_ring_ctx *ctx = rsrc_data->ctx;
8167        struct io_rsrc_put *prsrc, *tmp;
8168
8169        list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
8170                list_del(&prsrc->list);
8171
8172                if (prsrc->tag) {
8173                        bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
8174
8175                        io_ring_submit_lock(ctx, lock_ring);
8176                        spin_lock(&ctx->completion_lock);
8177                        io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
8178                        ctx->cq_extra++;
8179                        io_commit_cqring(ctx);
8180                        spin_unlock(&ctx->completion_lock);
8181                        io_cqring_ev_posted(ctx);
8182                        io_ring_submit_unlock(ctx, lock_ring);
8183                }
8184
8185                rsrc_data->do_put(ctx, prsrc);
8186                kfree(prsrc);
8187        }
8188
8189        io_rsrc_node_destroy(ref_node);
8190        if (atomic_dec_and_test(&rsrc_data->refs))
8191                complete(&rsrc_data->done);
8192}
8193
8194static void io_rsrc_put_work(struct work_struct *work)
8195{
8196        struct io_ring_ctx *ctx;
8197        struct llist_node *node;
8198
8199        ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
8200        node = llist_del_all(&ctx->rsrc_put_llist);
8201
8202        while (node) {
8203                struct io_rsrc_node *ref_node;
8204                struct llist_node *next = node->next;
8205
8206                ref_node = llist_entry(node, struct io_rsrc_node, llist);
8207                __io_rsrc_put_work(ref_node);
8208                node = next;
8209        }
8210}
8211
8212static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
8213                                 unsigned nr_args, u64 __user *tags)
8214{
8215        __s32 __user *fds = (__s32 __user *) arg;
8216        struct file *file;
8217        int fd, ret;
8218        unsigned i;
8219
8220        if (ctx->file_data)
8221                return -EBUSY;
8222        if (!nr_args)
8223                return -EINVAL;
8224        if (nr_args > IORING_MAX_FIXED_FILES)
8225                return -EMFILE;
8226        if (nr_args > rlimit(RLIMIT_NOFILE))
8227                return -EMFILE;
8228        ret = io_rsrc_node_switch_start(ctx);
8229        if (ret)
8230                return ret;
8231        ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
8232                                 &ctx->file_data);
8233        if (ret)
8234                return ret;
8235
8236        ret = -ENOMEM;
8237        if (!io_alloc_file_tables(&ctx->file_table, nr_args))
8238                goto out_free;
8239
8240        for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
8241                if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
8242                        ret = -EFAULT;
8243                        goto out_fput;
8244                }
8245                /* allow sparse sets */
8246                if (fd == -1) {
8247                        ret = -EINVAL;
8248                        if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
8249                                goto out_fput;
8250                        continue;
8251                }
8252
8253                file = fget(fd);
8254                ret = -EBADF;
8255                if (unlikely(!file))
8256                        goto out_fput;
8257
8258                /*
8259                 * Don't allow io_uring instances to be registered. If UNIX
8260                 * isn't enabled, then this causes a reference cycle and this
8261                 * instance can never get freed. If UNIX is enabled we'll
8262                 * handle it just fine, but there's still no point in allowing
8263                 * a ring fd as it doesn't support regular read/write anyway.
8264                 */
8265                if (file->f_op == &io_uring_fops) {
8266                        fput(file);
8267                        goto out_fput;
8268                }
8269                io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
8270        }
8271
8272        ret = io_sqe_files_scm(ctx);
8273        if (ret) {
8274                __io_sqe_files_unregister(ctx);
8275                return ret;
8276        }
8277
8278        io_rsrc_node_switch(ctx, NULL);
8279        return ret;
8280out_fput:
8281        for (i = 0; i < ctx->nr_user_files; i++) {
8282                file = io_file_from_index(ctx, i);
8283                if (file)
8284                        fput(file);
8285        }
8286        io_free_file_tables(&ctx->file_table);
8287        ctx->nr_user_files = 0;
8288out_free:
8289        io_rsrc_data_free(ctx->file_data);
8290        ctx->file_data = NULL;
8291        return ret;
8292}
8293
8294static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
8295                                int index)
8296{
8297#if defined(CONFIG_UNIX)
8298        struct sock *sock = ctx->ring_sock->sk;
8299        struct sk_buff_head *head = &sock->sk_receive_queue;
8300        struct sk_buff *skb;
8301
8302        /*
8303         * See if we can merge this file into an existing skb SCM_RIGHTS
8304         * file set. If there's no room, fall back to allocating a new skb
8305         * and filling it in.
8306         */
8307        spin_lock_irq(&head->lock);
8308        skb = skb_peek(head);
8309        if (skb) {
8310                struct scm_fp_list *fpl = UNIXCB(skb).fp;
8311
8312                if (fpl->count < SCM_MAX_FD) {
8313                        __skb_unlink(skb, head);
8314                        spin_unlock_irq(&head->lock);
8315                        fpl->fp[fpl->count] = get_file(file);
8316                        unix_inflight(fpl->user, fpl->fp[fpl->count]);
8317                        fpl->count++;
8318                        spin_lock_irq(&head->lock);
8319                        __skb_queue_head(head, skb);
8320                } else {
8321                        skb = NULL;
8322                }
8323        }
8324        spin_unlock_irq(&head->lock);
8325
8326        if (skb) {
8327                fput(file);
8328                return 0;
8329        }
8330
8331        return __io_sqe_files_scm(ctx, 1, index);
8332#else
8333        return 0;
8334#endif
8335}
8336
8337static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8338                                 struct io_rsrc_node *node, void *rsrc)
8339{
8340        struct io_rsrc_put *prsrc;
8341
8342        prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8343        if (!prsrc)
8344                return -ENOMEM;
8345
8346        prsrc->tag = *io_get_tag_slot(data, idx);
8347        prsrc->rsrc = rsrc;
8348        list_add(&prsrc->list, &node->rsrc_list);
8349        return 0;
8350}
8351
8352static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
8353                                 unsigned int issue_flags, u32 slot_index)
8354{
8355        struct io_ring_ctx *ctx = req->ctx;
8356        bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
8357        bool needs_switch = false;
8358        struct io_fixed_file *file_slot;
8359        int ret = -EBADF;
8360
8361        io_ring_submit_lock(ctx, !force_nonblock);
8362        if (file->f_op == &io_uring_fops)
8363                goto err;
8364        ret = -ENXIO;
8365        if (!ctx->file_data)
8366                goto err;
8367        ret = -EINVAL;
8368        if (slot_index >= ctx->nr_user_files)
8369                goto err;
8370
8371        slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8372        file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8373
8374        if (file_slot->file_ptr) {
8375                struct file *old_file;
8376
8377                ret = io_rsrc_node_switch_start(ctx);
8378                if (ret)
8379                        goto err;
8380
8381                old_file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8382                ret = io_queue_rsrc_removal(ctx->file_data, slot_index,
8383                                            ctx->rsrc_node, old_file);
8384                if (ret)
8385                        goto err;
8386                file_slot->file_ptr = 0;
8387                needs_switch = true;
8388        }
8389
8390        *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8391        io_fixed_file_set(file_slot, file);
8392        ret = io_sqe_file_register(ctx, file, slot_index);
8393        if (ret) {
8394                file_slot->file_ptr = 0;
8395                goto err;
8396        }
8397
8398        ret = 0;
8399err:
8400        if (needs_switch)
8401                io_rsrc_node_switch(ctx, ctx->file_data);
8402        io_ring_submit_unlock(ctx, !force_nonblock);
8403        if (ret)
8404                fput(file);
8405        return ret;
8406}
8407
8408static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags)
8409{
8410        unsigned int offset = req->close.file_slot - 1;
8411        struct io_ring_ctx *ctx = req->ctx;
8412        struct io_fixed_file *file_slot;
8413        struct file *file;
8414        int ret, i;
8415
8416        io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8417        ret = -ENXIO;
8418        if (unlikely(!ctx->file_data))
8419                goto out;
8420        ret = -EINVAL;
8421        if (offset >= ctx->nr_user_files)
8422                goto out;
8423        ret = io_rsrc_node_switch_start(ctx);
8424        if (ret)
8425                goto out;
8426
8427        i = array_index_nospec(offset, ctx->nr_user_files);
8428        file_slot = io_fixed_file_slot(&ctx->file_table, i);
8429        ret = -EBADF;
8430        if (!file_slot->file_ptr)
8431                goto out;
8432
8433        file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8434        ret = io_queue_rsrc_removal(ctx->file_data, offset, ctx->rsrc_node, file);
8435        if (ret)
8436                goto out;
8437
8438        file_slot->file_ptr = 0;
8439        io_rsrc_node_switch(ctx, ctx->file_data);
8440        ret = 0;
8441out:
8442        io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8443        return ret;
8444}
8445
8446static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8447                                 struct io_uring_rsrc_update2 *up,
8448                                 unsigned nr_args)
8449{
8450        u64 __user *tags = u64_to_user_ptr(up->tags);
8451        __s32 __user *fds = u64_to_user_ptr(up->data);
8452        struct io_rsrc_data *data = ctx->file_data;
8453        struct io_fixed_file *file_slot;
8454        struct file *file;
8455        int fd, i, err = 0;
8456        unsigned int done;
8457        bool needs_switch = false;
8458
8459        if (!ctx->file_data)
8460                return -ENXIO;
8461        if (up->offset + nr_args > ctx->nr_user_files)
8462                return -EINVAL;
8463
8464        for (done = 0; done < nr_args; done++) {
8465                u64 tag = 0;
8466
8467                if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
8468                    copy_from_user(&fd, &fds[done], sizeof(fd))) {
8469                        err = -EFAULT;
8470                        break;
8471                }
8472                if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
8473                        err = -EINVAL;
8474                        break;
8475                }
8476                if (fd == IORING_REGISTER_FILES_SKIP)
8477                        continue;
8478
8479                i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8480                file_slot = io_fixed_file_slot(&ctx->file_table, i);
8481
8482                if (file_slot->file_ptr) {
8483                        file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8484                        err = io_queue_rsrc_removal(data, up->offset + done,
8485                                                    ctx->rsrc_node, file);
8486                        if (err)
8487                                break;
8488                        file_slot->file_ptr = 0;
8489                        needs_switch = true;
8490                }
8491                if (fd != -1) {
8492                        file = fget(fd);
8493                        if (!file) {
8494                                err = -EBADF;
8495                                break;
8496                        }
8497                        /*
8498                         * Don't allow io_uring instances to be registered. If
8499                         * UNIX isn't enabled, then this causes a reference
8500                         * cycle and this instance can never get freed. If UNIX
8501                         * is enabled we'll handle it just fine, but there's
8502                         * still no point in allowing a ring fd as it doesn't
8503                         * support regular read/write anyway.
8504                         */
8505                        if (file->f_op == &io_uring_fops) {
8506                                fput(file);
8507                                err = -EBADF;
8508                                break;
8509                        }
8510                        *io_get_tag_slot(data, up->offset + done) = tag;
8511                        io_fixed_file_set(file_slot, file);
8512                        err = io_sqe_file_register(ctx, file, i);
8513                        if (err) {
8514                                file_slot->file_ptr = 0;
8515                                fput(file);
8516                                break;
8517                        }
8518                }
8519        }
8520
8521        if (needs_switch)
8522                io_rsrc_node_switch(ctx, data);
8523        return done ? done : err;
8524}
8525
8526static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
8527                                        struct task_struct *task)
8528{
8529        struct io_wq_hash *hash;
8530        struct io_wq_data data;
8531        unsigned int concurrency;
8532
8533        mutex_lock(&ctx->uring_lock);
8534        hash = ctx->hash_map;
8535        if (!hash) {
8536                hash = kzalloc(sizeof(*hash), GFP_KERNEL);
8537                if (!hash) {
8538                        mutex_unlock(&ctx->uring_lock);
8539                        return ERR_PTR(-ENOMEM);
8540                }
8541                refcount_set(&hash->refs, 1);
8542                init_waitqueue_head(&hash->wait);
8543                ctx->hash_map = hash;
8544        }
8545        mutex_unlock(&ctx->uring_lock);
8546
8547        data.hash = hash;
8548        data.task = task;
8549        data.free_work = io_wq_free_work;
8550        data.do_work = io_wq_submit_work;
8551
8552        /* Do QD, or 4 * CPUS, whatever is smallest */
8553        concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8554
8555        return io_wq_create(concurrency, &data);
8556}
8557
8558static int io_uring_alloc_task_context(struct task_struct *task,
8559                                       struct io_ring_ctx *ctx)
8560{
8561        struct io_uring_task *tctx;
8562        int ret;
8563
8564        tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8565        if (unlikely(!tctx))
8566                return -ENOMEM;
8567
8568        ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8569        if (unlikely(ret)) {
8570                kfree(tctx);
8571                return ret;
8572        }
8573
8574        tctx->io_wq = io_init_wq_offload(ctx, task);
8575        if (IS_ERR(tctx->io_wq)) {
8576                ret = PTR_ERR(tctx->io_wq);
8577                percpu_counter_destroy(&tctx->inflight);
8578                kfree(tctx);
8579                return ret;
8580        }
8581
8582        xa_init(&tctx->xa);
8583        init_waitqueue_head(&tctx->wait);
8584        atomic_set(&tctx->in_idle, 0);
8585        atomic_set(&tctx->inflight_tracked, 0);
8586        task->io_uring = tctx;
8587        spin_lock_init(&tctx->task_lock);
8588        INIT_WQ_LIST(&tctx->task_list);
8589        init_task_work(&tctx->task_work, tctx_task_work);
8590        return 0;
8591}
8592
8593void __io_uring_free(struct task_struct *tsk)
8594{
8595        struct io_uring_task *tctx = tsk->io_uring;
8596
8597        WARN_ON_ONCE(!xa_empty(&tctx->xa));
8598        WARN_ON_ONCE(tctx->io_wq);
8599        WARN_ON_ONCE(tctx->cached_refs);
8600
8601        percpu_counter_destroy(&tctx->inflight);
8602        kfree(tctx);
8603        tsk->io_uring = NULL;
8604}
8605
8606static int io_sq_offload_create(struct io_ring_ctx *ctx,
8607                                struct io_uring_params *p)
8608{
8609        int ret;
8610
8611        /* Retain compatibility with failing for an invalid attach attempt */
8612        if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8613                                IORING_SETUP_ATTACH_WQ) {
8614                struct fd f;
8615
8616                f = fdget(p->wq_fd);
8617                if (!f.file)
8618                        return -ENXIO;
8619                if (f.file->f_op != &io_uring_fops) {
8620                        fdput(f);
8621                        return -EINVAL;
8622                }
8623                fdput(f);
8624        }
8625        if (ctx->flags & IORING_SETUP_SQPOLL) {
8626                struct task_struct *tsk;
8627                struct io_sq_data *sqd;
8628                bool attached;
8629
8630                sqd = io_get_sq_data(p, &attached);
8631                if (IS_ERR(sqd)) {
8632                        ret = PTR_ERR(sqd);
8633                        goto err;
8634                }
8635
8636                ctx->sq_creds = get_current_cred();
8637                ctx->sq_data = sqd;
8638                ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8639                if (!ctx->sq_thread_idle)
8640                        ctx->sq_thread_idle = HZ;
8641
8642                io_sq_thread_park(sqd);
8643                list_add(&ctx->sqd_list, &sqd->ctx_list);
8644                io_sqd_update_thread_idle(sqd);
8645                /* don't attach to a dying SQPOLL thread, would be racy */
8646                ret = (attached && !sqd->thread) ? -ENXIO : 0;
8647                io_sq_thread_unpark(sqd);
8648
8649                if (ret < 0)
8650                        goto err;
8651                if (attached)
8652                        return 0;
8653
8654                if (p->flags & IORING_SETUP_SQ_AFF) {
8655                        int cpu = p->sq_thread_cpu;
8656
8657                        ret = -EINVAL;
8658                        if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8659                                goto err_sqpoll;
8660                        sqd->sq_cpu = cpu;
8661                } else {
8662                        sqd->sq_cpu = -1;
8663                }
8664
8665                sqd->task_pid = current->pid;
8666                sqd->task_tgid = current->tgid;
8667                tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8668                if (IS_ERR(tsk)) {
8669                        ret = PTR_ERR(tsk);
8670                        goto err_sqpoll;
8671                }
8672
8673                sqd->thread = tsk;
8674                ret = io_uring_alloc_task_context(tsk, ctx);
8675                wake_up_new_task(tsk);
8676                if (ret)
8677                        goto err;
8678        } else if (p->flags & IORING_SETUP_SQ_AFF) {
8679                /* Can't have SQ_AFF without SQPOLL */
8680                ret = -EINVAL;
8681                goto err;
8682        }
8683
8684        return 0;
8685err_sqpoll:
8686        complete(&ctx->sq_data->exited);
8687err:
8688        io_sq_thread_finish(ctx);
8689        return ret;
8690}
8691
8692static inline void __io_unaccount_mem(struct user_struct *user,
8693                                      unsigned long nr_pages)
8694{
8695        atomic_long_sub(nr_pages, &user->locked_vm);
8696}
8697
8698static inline int __io_account_mem(struct user_struct *user,
8699                                   unsigned long nr_pages)
8700{
8701        unsigned long page_limit, cur_pages, new_pages;
8702
8703        /* Don't allow more pages than we can safely lock */
8704        page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8705
8706        do {
8707                cur_pages = atomic_long_read(&user->locked_vm);
8708                new_pages = cur_pages + nr_pages;
8709                if (new_pages > page_limit)
8710                        return -ENOMEM;
8711        } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8712                                        new_pages) != cur_pages);
8713
8714        return 0;
8715}
8716
8717static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8718{
8719        if (ctx->user)
8720                __io_unaccount_mem(ctx->user, nr_pages);
8721
8722        if (ctx->mm_account)
8723                atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8724}
8725
8726static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8727{
8728        int ret;
8729
8730        if (ctx->user) {
8731                ret = __io_account_mem(ctx->user, nr_pages);
8732                if (ret)
8733                        return ret;
8734        }
8735
8736        if (ctx->mm_account)
8737                atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8738
8739        return 0;
8740}
8741
8742static void io_mem_free(void *ptr)
8743{
8744        struct page *page;
8745
8746        if (!ptr)
8747                return;
8748
8749        page = virt_to_head_page(ptr);
8750        if (put_page_testzero(page))
8751                free_compound_page(page);
8752}
8753
8754static void *io_mem_alloc(size_t size)
8755{
8756        gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8757                                __GFP_NORETRY | __GFP_ACCOUNT;
8758
8759        return (void *) __get_free_pages(gfp_flags, get_order(size));
8760}
8761
8762static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8763                                size_t *sq_offset)
8764{
8765        struct io_rings *rings;
8766        size_t off, sq_array_size;
8767
8768        off = struct_size(rings, cqes, cq_entries);
8769        if (off == SIZE_MAX)
8770                return SIZE_MAX;
8771
8772#ifdef CONFIG_SMP
8773        off = ALIGN(off, SMP_CACHE_BYTES);
8774        if (off == 0)
8775                return SIZE_MAX;
8776#endif
8777
8778        if (sq_offset)
8779                *sq_offset = off;
8780
8781        sq_array_size = array_size(sizeof(u32), sq_entries);
8782        if (sq_array_size == SIZE_MAX)
8783                return SIZE_MAX;
8784
8785        if (check_add_overflow(off, sq_array_size, &off))
8786                return SIZE_MAX;
8787
8788        return off;
8789}
8790
8791static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8792{
8793        struct io_mapped_ubuf *imu = *slot;
8794        unsigned int i;
8795
8796        if (imu != ctx->dummy_ubuf) {
8797                for (i = 0; i < imu->nr_bvecs; i++)
8798                        unpin_user_page(imu->bvec[i].bv_page);
8799                if (imu->acct_pages)
8800                        io_unaccount_mem(ctx, imu->acct_pages);
8801                kvfree(imu);
8802        }
8803        *slot = NULL;
8804}
8805
8806static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8807{
8808        io_buffer_unmap(ctx, &prsrc->buf);
8809        prsrc->buf = NULL;
8810}
8811
8812static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8813{
8814        unsigned int i;
8815
8816        for (i = 0; i < ctx->nr_user_bufs; i++)
8817                io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8818        kfree(ctx->user_bufs);
8819        io_rsrc_data_free(ctx->buf_data);
8820        ctx->user_bufs = NULL;
8821        ctx->buf_data = NULL;
8822        ctx->nr_user_bufs = 0;
8823}
8824
8825static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8826{
8827        int ret;
8828
8829        if (!ctx->buf_data)
8830                return -ENXIO;
8831
8832        ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8833        if (!ret)
8834                __io_sqe_buffers_unregister(ctx);
8835        return ret;
8836}
8837
8838static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8839                       void __user *arg, unsigned index)
8840{
8841        struct iovec __user *src;
8842
8843#ifdef CONFIG_COMPAT
8844        if (ctx->compat) {
8845                struct compat_iovec __user *ciovs;
8846                struct compat_iovec ciov;
8847
8848                ciovs = (struct compat_iovec __user *) arg;
8849                if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8850                        return -EFAULT;
8851
8852                dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8853                dst->iov_len = ciov.iov_len;
8854                return 0;
8855        }
8856#endif
8857        src = (struct iovec __user *) arg;
8858        if (copy_from_user(dst, &src[index], sizeof(*dst)))
8859                return -EFAULT;
8860        return 0;
8861}
8862
8863/*
8864 * Not super efficient, but this is just a registration time. And we do cache
8865 * the last compound head, so generally we'll only do a full search if we don't
8866 * match that one.
8867 *
8868 * We check if the given compound head page has already been accounted, to
8869 * avoid double accounting it. This allows us to account the full size of the
8870 * page, not just the constituent pages of a huge page.
8871 */
8872static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8873                                  int nr_pages, struct page *hpage)
8874{
8875        int i, j;
8876
8877        /* check current page array */
8878        for (i = 0; i < nr_pages; i++) {
8879                if (!PageCompound(pages[i]))
8880                        continue;
8881                if (compound_head(pages[i]) == hpage)
8882                        return true;
8883        }
8884
8885        /* check previously registered pages */
8886        for (i = 0; i < ctx->nr_user_bufs; i++) {
8887                struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8888
8889                for (j = 0; j < imu->nr_bvecs; j++) {
8890                        if (!PageCompound(imu->bvec[j].bv_page))
8891                                continue;
8892                        if (compound_head(imu->bvec[j].bv_page) == hpage)
8893                                return true;
8894                }
8895        }
8896
8897        return false;
8898}
8899
8900static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8901                                 int nr_pages, struct io_mapped_ubuf *imu,
8902                                 struct page **last_hpage)
8903{
8904        int i, ret;
8905
8906        imu->acct_pages = 0;
8907        for (i = 0; i < nr_pages; i++) {
8908                if (!PageCompound(pages[i])) {
8909                        imu->acct_pages++;
8910                } else {
8911                        struct page *hpage;
8912
8913                        hpage = compound_head(pages[i]);
8914                        if (hpage == *last_hpage)
8915                                continue;
8916                        *last_hpage = hpage;
8917                        if (headpage_already_acct(ctx, pages, i, hpage))
8918                                continue;
8919                        imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8920                }
8921        }
8922
8923        if (!imu->acct_pages)
8924                return 0;
8925
8926        ret = io_account_mem(ctx, imu->acct_pages);
8927        if (ret)
8928                imu->acct_pages = 0;
8929        return ret;
8930}
8931
8932static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8933                                  struct io_mapped_ubuf **pimu,
8934                                  struct page **last_hpage)
8935{
8936        struct io_mapped_ubuf *imu = NULL;
8937        struct vm_area_struct **vmas = NULL;
8938        struct page **pages = NULL;
8939        unsigned long off, start, end, ubuf;
8940        size_t size;
8941        int ret, pret, nr_pages, i;
8942
8943        if (!iov->iov_base) {
8944                *pimu = ctx->dummy_ubuf;
8945                return 0;
8946        }
8947
8948        ubuf = (unsigned long) iov->iov_base;
8949        end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8950        start = ubuf >> PAGE_SHIFT;
8951        nr_pages = end - start;
8952
8953        *pimu = NULL;
8954        ret = -ENOMEM;
8955
8956        pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8957        if (!pages)
8958                goto done;
8959
8960        vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8961                              GFP_KERNEL);
8962        if (!vmas)
8963                goto done;
8964
8965        imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8966        if (!imu)
8967                goto done;
8968
8969        ret = 0;
8970        mmap_read_lock(current->mm);
8971        pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8972                              pages, vmas);
8973        if (pret == nr_pages) {
8974                /* don't support file backed memory */
8975                for (i = 0; i < nr_pages; i++) {
8976                        struct vm_area_struct *vma = vmas[i];
8977
8978                        if (vma_is_shmem(vma))
8979                                continue;
8980                        if (vma->vm_file &&
8981                            !is_file_hugepages(vma->vm_file)) {
8982                                ret = -EOPNOTSUPP;
8983                                break;
8984                        }
8985                }
8986        } else {
8987                ret = pret < 0 ? pret : -EFAULT;
8988        }
8989        mmap_read_unlock(current->mm);
8990        if (ret) {
8991                /*
8992                 * if we did partial map, or found file backed vmas,
8993                 * release any pages we did get
8994                 */
8995                if (pret > 0)
8996                        unpin_user_pages(pages, pret);
8997                goto done;
8998        }
8999
9000        ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
9001        if (ret) {
9002                unpin_user_pages(pages, pret);
9003                goto done;
9004        }
9005
9006        off = ubuf & ~PAGE_MASK;
9007        size = iov->iov_len;
9008        for (i = 0; i < nr_pages; i++) {
9009                size_t vec_len;
9010
9011                vec_len = min_t(size_t, size, PAGE_SIZE - off);
9012                imu->bvec[i].bv_page = pages[i];
9013                imu->bvec[i].bv_len = vec_len;
9014                imu->bvec[i].bv_offset = off;
9015                off = 0;
9016                size -= vec_len;
9017        }
9018        /* store original address for later verification */
9019        imu->ubuf = ubuf;
9020        imu->ubuf_end = ubuf + iov->iov_len;
9021        imu->nr_bvecs = nr_pages;
9022        *pimu = imu;
9023        ret = 0;
9024done:
9025        if (ret)
9026                kvfree(imu);
9027        kvfree(pages);
9028        kvfree(vmas);
9029        return ret;
9030}
9031
9032static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
9033{
9034        ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
9035        return ctx->user_bufs ? 0 : -ENOMEM;
9036}
9037
9038static int io_buffer_validate(struct iovec *iov)
9039{
9040        unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
9041
9042        /*
9043         * Don't impose further limits on the size and buffer
9044         * constraints here, we'll -EINVAL later when IO is
9045         * submitted if they are wrong.
9046         */
9047        if (!iov->iov_base)
9048                return iov->iov_len ? -EFAULT : 0;
9049        if (!iov->iov_len)
9050                return -EFAULT;
9051
9052        /* arbitrary limit, but we need something */
9053        if (iov->iov_len > SZ_1G)
9054                return -EFAULT;
9055
9056        if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
9057                return -EOVERFLOW;
9058
9059        return 0;
9060}
9061
9062static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
9063                                   unsigned int nr_args, u64 __user *tags)
9064{
9065        struct page *last_hpage = NULL;
9066        struct io_rsrc_data *data;
9067        int i, ret;
9068        struct iovec iov;
9069
9070        if (ctx->user_bufs)
9071                return -EBUSY;
9072        if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
9073                return -EINVAL;
9074        ret = io_rsrc_node_switch_start(ctx);
9075        if (ret)
9076                return ret;
9077        ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
9078        if (ret)
9079                return ret;
9080        ret = io_buffers_map_alloc(ctx, nr_args);
9081        if (ret) {
9082                io_rsrc_data_free(data);
9083                return ret;
9084        }
9085
9086        for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
9087                ret = io_copy_iov(ctx, &iov, arg, i);
9088                if (ret)
9089                        break;
9090                ret = io_buffer_validate(&iov);
9091                if (ret)
9092                        break;
9093                if (!iov.iov_base && *io_get_tag_slot(data, i)) {
9094                        ret = -EINVAL;
9095                        break;
9096                }
9097
9098                ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
9099                                             &last_hpage);
9100                if (ret)
9101                        break;
9102        }
9103
9104        WARN_ON_ONCE(ctx->buf_data);
9105
9106        ctx->buf_data = data;
9107        if (ret)
9108                __io_sqe_buffers_unregister(ctx);
9109        else
9110                io_rsrc_node_switch(ctx, NULL);
9111        return ret;
9112}
9113
9114static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
9115                                   struct io_uring_rsrc_update2 *up,
9116                                   unsigned int nr_args)
9117{
9118        u64 __user *tags = u64_to_user_ptr(up->tags);
9119        struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
9120        struct page *last_hpage = NULL;
9121        bool needs_switch = false;
9122        __u32 done;
9123        int i, err;
9124
9125        if (!ctx->buf_data)
9126                return -ENXIO;
9127        if (up->offset + nr_args > ctx->nr_user_bufs)
9128                return -EINVAL;
9129
9130        for (done = 0; done < nr_args; done++) {
9131                struct io_mapped_ubuf *imu;
9132                int offset = up->offset + done;
9133                u64 tag = 0;
9134
9135                err = io_copy_iov(ctx, &iov, iovs, done);
9136                if (err)
9137                        break;
9138                if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
9139                        err = -EFAULT;
9140                        break;
9141                }
9142                err = io_buffer_validate(&iov);
9143                if (err)
9144                        break;
9145                if (!iov.iov_base && tag) {
9146                        err = -EINVAL;
9147                        break;
9148                }
9149                err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
9150                if (err)
9151                        break;
9152
9153                i = array_index_nospec(offset, ctx->nr_user_bufs);
9154                if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
9155                        err = io_queue_rsrc_removal(ctx->buf_data, offset,
9156                                                    ctx->rsrc_node, ctx->user_bufs[i]);
9157                        if (unlikely(err)) {
9158                                io_buffer_unmap(ctx, &imu);
9159                                break;
9160                        }
9161                        ctx->user_bufs[i] = NULL;
9162                        needs_switch = true;
9163                }
9164
9165                ctx->user_bufs[i] = imu;
9166                *io_get_tag_slot(ctx->buf_data, offset) = tag;
9167        }
9168
9169        if (needs_switch)
9170                io_rsrc_node_switch(ctx, ctx->buf_data);
9171        return done ? done : err;
9172}
9173
9174static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
9175{
9176        __s32 __user *fds = arg;
9177        int fd;
9178
9179        if (ctx->cq_ev_fd)
9180                return -EBUSY;
9181
9182        if (copy_from_user(&fd, fds, sizeof(*fds)))
9183                return -EFAULT;
9184
9185        ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
9186        if (IS_ERR(ctx->cq_ev_fd)) {
9187                int ret = PTR_ERR(ctx->cq_ev_fd);
9188
9189                ctx->cq_ev_fd = NULL;
9190                return ret;
9191        }
9192
9193        return 0;
9194}
9195
9196static int io_eventfd_unregister(struct io_ring_ctx *ctx)
9197{
9198        if (ctx->cq_ev_fd) {
9199                eventfd_ctx_put(ctx->cq_ev_fd);
9200                ctx->cq_ev_fd = NULL;
9201                return 0;
9202        }
9203
9204        return -ENXIO;
9205}
9206
9207static void io_destroy_buffers(struct io_ring_ctx *ctx)
9208{
9209        struct io_buffer *buf;
9210        unsigned long index;
9211
9212        xa_for_each(&ctx->io_buffers, index, buf) {
9213                __io_remove_buffers(ctx, buf, index, -1U);
9214                cond_resched();
9215        }
9216}
9217
9218static void io_req_cache_free(struct list_head *list)
9219{
9220        struct io_kiocb *req, *nxt;
9221
9222        list_for_each_entry_safe(req, nxt, list, inflight_entry) {
9223                list_del(&req->inflight_entry);
9224                kmem_cache_free(req_cachep, req);
9225        }
9226}
9227
9228static void io_req_caches_free(struct io_ring_ctx *ctx)
9229{
9230        struct io_submit_state *state = &ctx->submit_state;
9231
9232        mutex_lock(&ctx->uring_lock);
9233
9234        if (state->free_reqs) {
9235                kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
9236                state->free_reqs = 0;
9237        }
9238
9239        io_flush_cached_locked_reqs(ctx, state);
9240        io_req_cache_free(&state->free_list);
9241        mutex_unlock(&ctx->uring_lock);
9242}
9243
9244static void io_wait_rsrc_data(struct io_rsrc_data *data)
9245{
9246        if (data && !atomic_dec_and_test(&data->refs))
9247                wait_for_completion(&data->done);
9248}
9249
9250static void io_ring_ctx_free(struct io_ring_ctx *ctx)
9251{
9252        io_sq_thread_finish(ctx);
9253
9254        if (ctx->mm_account) {
9255                mmdrop(ctx->mm_account);
9256                ctx->mm_account = NULL;
9257        }
9258
9259        /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9260        io_wait_rsrc_data(ctx->buf_data);
9261        io_wait_rsrc_data(ctx->file_data);
9262
9263        mutex_lock(&ctx->uring_lock);
9264        if (ctx->buf_data)
9265                __io_sqe_buffers_unregister(ctx);
9266        if (ctx->file_data)
9267                __io_sqe_files_unregister(ctx);
9268        if (ctx->rings)
9269                __io_cqring_overflow_flush(ctx, true);
9270        mutex_unlock(&ctx->uring_lock);
9271        io_eventfd_unregister(ctx);
9272        io_destroy_buffers(ctx);
9273        if (ctx->sq_creds)
9274                put_cred(ctx->sq_creds);
9275
9276        /* there are no registered resources left, nobody uses it */
9277        if (ctx->rsrc_node)
9278                io_rsrc_node_destroy(ctx->rsrc_node);
9279        if (ctx->rsrc_backup_node)
9280                io_rsrc_node_destroy(ctx->rsrc_backup_node);
9281        flush_delayed_work(&ctx->rsrc_put_work);
9282
9283        WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
9284        WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
9285
9286#if defined(CONFIG_UNIX)
9287        if (ctx->ring_sock) {
9288                ctx->ring_sock->file = NULL; /* so that iput() is called */
9289                sock_release(ctx->ring_sock);
9290        }
9291#endif
9292        WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
9293
9294        io_mem_free(ctx->rings);
9295        io_mem_free(ctx->sq_sqes);
9296
9297        percpu_ref_exit(&ctx->refs);
9298        free_uid(ctx->user);
9299        io_req_caches_free(ctx);
9300        if (ctx->hash_map)
9301                io_wq_put_hash(ctx->hash_map);
9302        kfree(ctx->cancel_hash);
9303        kfree(ctx->dummy_ubuf);
9304        kfree(ctx);
9305}
9306
9307static __poll_t io_uring_poll(struct file *file, poll_table *wait)
9308{
9309        struct io_ring_ctx *ctx = file->private_data;
9310        __poll_t mask = 0;
9311
9312        poll_wait(file, &ctx->poll_wait, wait);
9313        /*
9314         * synchronizes with barrier from wq_has_sleeper call in
9315         * io_commit_cqring
9316         */
9317        smp_rmb();
9318        if (!io_sqring_full(ctx))
9319                mask |= EPOLLOUT | EPOLLWRNORM;
9320
9321        /*
9322         * Don't flush cqring overflow list here, just do a simple check.
9323         * Otherwise there could possible be ABBA deadlock:
9324         *      CPU0                    CPU1
9325         *      ----                    ----
9326         * lock(&ctx->uring_lock);
9327         *                              lock(&ep->mtx);
9328         *                              lock(&ctx->uring_lock);
9329         * lock(&ep->mtx);
9330         *
9331         * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9332         * pushs them to do the flush.
9333         */
9334        if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
9335                mask |= EPOLLIN | EPOLLRDNORM;
9336
9337        return mask;
9338}
9339
9340static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9341{
9342        const struct cred *creds;
9343
9344        creds = xa_erase(&ctx->personalities, id);
9345        if (creds) {
9346                put_cred(creds);
9347                return 0;
9348        }
9349
9350        return -EINVAL;
9351}
9352
9353struct io_tctx_exit {
9354        struct callback_head            task_work;
9355        struct completion               completion;
9356        struct io_ring_ctx              *ctx;
9357};
9358
9359static void io_tctx_exit_cb(struct callback_head *cb)
9360{
9361        struct io_uring_task *tctx = current->io_uring;
9362        struct io_tctx_exit *work;
9363
9364        work = container_of(cb, struct io_tctx_exit, task_work);
9365        /*
9366         * When @in_idle, we're in cancellation and it's racy to remove the
9367         * node. It'll be removed by the end of cancellation, just ignore it.
9368         */
9369        if (!atomic_read(&tctx->in_idle))
9370                io_uring_del_tctx_node((unsigned long)work->ctx);
9371        complete(&work->completion);
9372}
9373
9374static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
9375{
9376        struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9377
9378        return req->ctx == data;
9379}
9380
9381static void io_ring_exit_work(struct work_struct *work)
9382{
9383        struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
9384        unsigned long timeout = jiffies + HZ * 60 * 5;
9385        unsigned long interval = HZ / 20;
9386        struct io_tctx_exit exit;
9387        struct io_tctx_node *node;
9388        int ret;
9389
9390        /*
9391         * If we're doing polled IO and end up having requests being
9392         * submitted async (out-of-line), then completions can come in while
9393         * we're waiting for refs to drop. We need to reap these manually,
9394         * as nobody else will be looking for them.
9395         */
9396        do {
9397                io_uring_try_cancel_requests(ctx, NULL, true);
9398                if (ctx->sq_data) {
9399                        struct io_sq_data *sqd = ctx->sq_data;
9400                        struct task_struct *tsk;
9401
9402                        io_sq_thread_park(sqd);
9403                        tsk = sqd->thread;
9404                        if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
9405                                io_wq_cancel_cb(tsk->io_uring->io_wq,
9406                                                io_cancel_ctx_cb, ctx, true);
9407                        io_sq_thread_unpark(sqd);
9408                }
9409
9410                if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
9411                        /* there is little hope left, don't run it too often */
9412                        interval = HZ * 60;
9413                }
9414        } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
9415
9416        init_completion(&exit.completion);
9417        init_task_work(&exit.task_work, io_tctx_exit_cb);
9418        exit.ctx = ctx;
9419        /*
9420         * Some may use context even when all refs and requests have been put,
9421         * and they are free to do so while still holding uring_lock or
9422         * completion_lock, see io_req_task_submit(). Apart from other work,
9423         * this lock/unlock section also waits them to finish.
9424         */
9425        mutex_lock(&ctx->uring_lock);
9426        while (!list_empty(&ctx->tctx_list)) {
9427                WARN_ON_ONCE(time_after(jiffies, timeout));
9428
9429                node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
9430                                        ctx_node);
9431                /* don't spin on a single task if cancellation failed */
9432                list_rotate_left(&ctx->tctx_list);
9433                ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
9434                if (WARN_ON_ONCE(ret))
9435                        continue;
9436                wake_up_process(node->task);
9437
9438                mutex_unlock(&ctx->uring_lock);
9439                wait_for_completion(&exit.completion);
9440                mutex_lock(&ctx->uring_lock);
9441        }
9442        mutex_unlock(&ctx->uring_lock);
9443        spin_lock(&ctx->completion_lock);
9444        spin_unlock(&ctx->completion_lock);
9445
9446        io_ring_ctx_free(ctx);
9447}
9448
9449/* Returns true if we found and killed one or more timeouts */
9450static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
9451                             bool cancel_all)
9452{
9453        struct io_kiocb *req, *tmp;
9454        int canceled = 0;
9455
9456        spin_lock(&ctx->completion_lock);
9457        spin_lock_irq(&ctx->timeout_lock);
9458        list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
9459                if (io_match_task(req, tsk, cancel_all)) {
9460                        io_kill_timeout(req, -ECANCELED);
9461                        canceled++;
9462                }
9463        }
9464        spin_unlock_irq(&ctx->timeout_lock);
9465        if (canceled != 0)
9466                io_commit_cqring(ctx);
9467        spin_unlock(&ctx->completion_lock);
9468        if (canceled != 0)
9469                io_cqring_ev_posted(ctx);
9470        return canceled != 0;
9471}
9472
9473static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
9474{
9475        unsigned long index;
9476        struct creds *creds;
9477
9478        mutex_lock(&ctx->uring_lock);
9479        percpu_ref_kill(&ctx->refs);
9480        if (ctx->rings)
9481                __io_cqring_overflow_flush(ctx, true);
9482        xa_for_each(&ctx->personalities, index, creds)
9483                io_unregister_personality(ctx, index);
9484        mutex_unlock(&ctx->uring_lock);
9485
9486        io_kill_timeouts(ctx, NULL, true);
9487        io_poll_remove_all(ctx, NULL, true);
9488
9489        /* if we failed setting up the ctx, we might not have any rings */
9490        io_iopoll_try_reap_events(ctx);
9491
9492        INIT_WORK(&ctx->exit_work, io_ring_exit_work);
9493        /*
9494         * Use system_unbound_wq to avoid spawning tons of event kworkers
9495         * if we're exiting a ton of rings at the same time. It just adds
9496         * noise and overhead, there's no discernable change in runtime
9497         * over using system_wq.
9498         */
9499        queue_work(system_unbound_wq, &ctx->exit_work);
9500}
9501
9502static int io_uring_release(struct inode *inode, struct file *file)
9503{
9504        struct io_ring_ctx *ctx = file->private_data;
9505
9506        file->private_data = NULL;
9507        io_ring_ctx_wait_and_kill(ctx);
9508        return 0;
9509}
9510
9511struct io_task_cancel {
9512        struct task_struct *task;
9513        bool all;
9514};
9515
9516static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
9517{
9518        struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9519        struct io_task_cancel *cancel = data;
9520        bool ret;
9521
9522        if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
9523                struct io_ring_ctx *ctx = req->ctx;
9524
9525                /* protect against races with linked timeouts */
9526                spin_lock(&ctx->completion_lock);
9527                ret = io_match_task(req, cancel->task, cancel->all);
9528                spin_unlock(&ctx->completion_lock);
9529        } else {
9530                ret = io_match_task(req, cancel->task, cancel->all);
9531        }
9532        return ret;
9533}
9534
9535static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
9536                                  struct task_struct *task, bool cancel_all)
9537{
9538        struct io_defer_entry *de;
9539        LIST_HEAD(list);
9540
9541        spin_lock(&ctx->completion_lock);
9542        list_for_each_entry_reverse(de, &ctx->defer_list, list) {
9543                if (io_match_task(de->req, task, cancel_all)) {
9544                        list_cut_position(&list, &ctx->defer_list, &de->list);
9545                        break;
9546                }
9547        }
9548        spin_unlock(&ctx->completion_lock);
9549        if (list_empty(&list))
9550                return false;
9551
9552        while (!list_empty(&list)) {
9553                de = list_first_entry(&list, struct io_defer_entry, list);
9554                list_del_init(&de->list);
9555                io_req_complete_failed(de->req, -ECANCELED);
9556                kfree(de);
9557        }
9558        return true;
9559}
9560
9561static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9562{
9563        struct io_tctx_node *node;
9564        enum io_wq_cancel cret;
9565        bool ret = false;
9566
9567        mutex_lock(&ctx->uring_lock);
9568        list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9569                struct io_uring_task *tctx = node->task->io_uring;
9570
9571                /*
9572                 * io_wq will stay alive while we hold uring_lock, because it's
9573                 * killed after ctx nodes, which requires to take the lock.
9574                 */
9575                if (!tctx || !tctx->io_wq)
9576                        continue;
9577                cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9578                ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9579        }
9580        mutex_unlock(&ctx->uring_lock);
9581
9582        return ret;
9583}
9584
9585static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9586                                         struct task_struct *task,
9587                                         bool cancel_all)
9588{
9589        struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9590        struct io_uring_task *tctx = task ? task->io_uring : NULL;
9591
9592        while (1) {
9593                enum io_wq_cancel cret;
9594                bool ret = false;
9595
9596                if (!task) {
9597                        ret |= io_uring_try_cancel_iowq(ctx);
9598                } else if (tctx && tctx->io_wq) {
9599                        /*
9600                         * Cancels requests of all rings, not only @ctx, but
9601                         * it's fine as the task is in exit/exec.
9602                         */
9603                        cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9604                                               &cancel, true);
9605                        ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9606                }
9607
9608                /* SQPOLL thread does its own polling */
9609                if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9610                    (ctx->sq_data && ctx->sq_data->thread == current)) {
9611                        while (!list_empty_careful(&ctx->iopoll_list)) {
9612                                io_iopoll_try_reap_events(ctx);
9613                                ret = true;
9614                        }
9615                }
9616
9617                ret |= io_cancel_defer_files(ctx, task, cancel_all);
9618                ret |= io_poll_remove_all(ctx, task, cancel_all);
9619                ret |= io_kill_timeouts(ctx, task, cancel_all);
9620                if (task)
9621                        ret |= io_run_task_work();
9622                if (!ret)
9623                        break;
9624                cond_resched();
9625        }
9626}
9627
9628static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9629{
9630        struct io_uring_task *tctx = current->io_uring;
9631        struct io_tctx_node *node;
9632        int ret;
9633
9634        if (unlikely(!tctx)) {
9635                ret = io_uring_alloc_task_context(current, ctx);
9636                if (unlikely(ret))
9637                        return ret;
9638
9639                tctx = current->io_uring;
9640                if (ctx->iowq_limits_set) {
9641                        unsigned int limits[2] = { ctx->iowq_limits[0],
9642                                                   ctx->iowq_limits[1], };
9643
9644                        ret = io_wq_max_workers(tctx->io_wq, limits);
9645                        if (ret)
9646                                return ret;
9647                }
9648        }
9649        if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9650                node = kmalloc(sizeof(*node), GFP_KERNEL);
9651                if (!node)
9652                        return -ENOMEM;
9653                node->ctx = ctx;
9654                node->task = current;
9655
9656                ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9657                                        node, GFP_KERNEL));
9658                if (ret) {
9659                        kfree(node);
9660                        return ret;
9661                }
9662
9663                mutex_lock(&ctx->uring_lock);
9664                list_add(&node->ctx_node, &ctx->tctx_list);
9665                mutex_unlock(&ctx->uring_lock);
9666        }
9667        tctx->last = ctx;
9668        return 0;
9669}
9670
9671/*
9672 * Note that this task has used io_uring. We use it for cancelation purposes.
9673 */
9674static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9675{
9676        struct io_uring_task *tctx = current->io_uring;
9677
9678        if (likely(tctx && tctx->last == ctx))
9679                return 0;
9680        return __io_uring_add_tctx_node(ctx);
9681}
9682
9683/*
9684 * Remove this io_uring_file -> task mapping.
9685 */
9686static void io_uring_del_tctx_node(unsigned long index)
9687{
9688        struct io_uring_task *tctx = current->io_uring;
9689        struct io_tctx_node *node;
9690
9691        if (!tctx)
9692                return;
9693        node = xa_erase(&tctx->xa, index);
9694        if (!node)
9695                return;
9696
9697        WARN_ON_ONCE(current != node->task);
9698        WARN_ON_ONCE(list_empty(&node->ctx_node));
9699
9700        mutex_lock(&node->ctx->uring_lock);
9701        list_del(&node->ctx_node);
9702        mutex_unlock(&node->ctx->uring_lock);
9703
9704        if (tctx->last == node->ctx)
9705                tctx->last = NULL;
9706        kfree(node);
9707}
9708
9709static void io_uring_clean_tctx(struct io_uring_task *tctx)
9710{
9711        struct io_wq *wq = tctx->io_wq;
9712        struct io_tctx_node *node;
9713        unsigned long index;
9714
9715        xa_for_each(&tctx->xa, index, node) {
9716                io_uring_del_tctx_node(index);
9717                cond_resched();
9718        }
9719        if (wq) {
9720                /*
9721                 * Must be after io_uring_del_task_file() (removes nodes under
9722                 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9723                 */
9724                io_wq_put_and_exit(wq);
9725                tctx->io_wq = NULL;
9726        }
9727}
9728
9729static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9730{
9731        if (tracked)
9732                return atomic_read(&tctx->inflight_tracked);
9733        return percpu_counter_sum(&tctx->inflight);
9734}
9735
9736static void io_uring_drop_tctx_refs(struct task_struct *task)
9737{
9738        struct io_uring_task *tctx = task->io_uring;
9739        unsigned int refs = tctx->cached_refs;
9740
9741        if (refs) {
9742                tctx->cached_refs = 0;
9743                percpu_counter_sub(&tctx->inflight, refs);
9744                put_task_struct_many(task, refs);
9745        }
9746}
9747
9748/*
9749 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9750 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9751 */
9752static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9753{
9754        struct io_uring_task *tctx = current->io_uring;
9755        struct io_ring_ctx *ctx;
9756        s64 inflight;
9757        DEFINE_WAIT(wait);
9758
9759        WARN_ON_ONCE(sqd && sqd->thread != current);
9760
9761        if (!current->io_uring)
9762                return;
9763        if (tctx->io_wq)
9764                io_wq_exit_start(tctx->io_wq);
9765
9766        atomic_inc(&tctx->in_idle);
9767        do {
9768                io_uring_drop_tctx_refs(current);
9769                /* read completions before cancelations */
9770                inflight = tctx_inflight(tctx, !cancel_all);
9771                if (!inflight)
9772                        break;
9773
9774                if (!sqd) {
9775                        struct io_tctx_node *node;
9776                        unsigned long index;
9777
9778                        xa_for_each(&tctx->xa, index, node) {
9779                                /* sqpoll task will cancel all its requests */
9780                                if (node->ctx->sq_data)
9781                                        continue;
9782                                io_uring_try_cancel_requests(node->ctx, current,
9783                                                             cancel_all);
9784                        }
9785                } else {
9786                        list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9787                                io_uring_try_cancel_requests(ctx, current,
9788                                                             cancel_all);
9789                }
9790
9791                prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9792                io_uring_drop_tctx_refs(current);
9793                /*
9794                 * If we've seen completions, retry without waiting. This
9795                 * avoids a race where a completion comes in before we did
9796                 * prepare_to_wait().
9797                 */
9798                if (inflight == tctx_inflight(tctx, !cancel_all))
9799                        schedule();
9800                finish_wait(&tctx->wait, &wait);
9801        } while (1);
9802        atomic_dec(&tctx->in_idle);
9803
9804        io_uring_clean_tctx(tctx);
9805        if (cancel_all) {
9806                /* for exec all current's requests should be gone, kill tctx */
9807                __io_uring_free(current);
9808        }
9809}
9810
9811void __io_uring_cancel(bool cancel_all)
9812{
9813        io_uring_cancel_generic(cancel_all, NULL);
9814}
9815
9816static void *io_uring_validate_mmap_request(struct file *file,
9817                                            loff_t pgoff, size_t sz)
9818{
9819        struct io_ring_ctx *ctx = file->private_data;
9820        loff_t offset = pgoff << PAGE_SHIFT;
9821        struct page *page;
9822        void *ptr;
9823
9824        switch (offset) {
9825        case IORING_OFF_SQ_RING:
9826        case IORING_OFF_CQ_RING:
9827                ptr = ctx->rings;
9828                break;
9829        case IORING_OFF_SQES:
9830                ptr = ctx->sq_sqes;
9831                break;
9832        default:
9833                return ERR_PTR(-EINVAL);
9834        }
9835
9836        page = virt_to_head_page(ptr);
9837        if (sz > page_size(page))
9838                return ERR_PTR(-EINVAL);
9839
9840        return ptr;
9841}
9842
9843#ifdef CONFIG_MMU
9844
9845static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9846{
9847        size_t sz = vma->vm_end - vma->vm_start;
9848        unsigned long pfn;
9849        void *ptr;
9850
9851        ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9852        if (IS_ERR(ptr))
9853                return PTR_ERR(ptr);
9854
9855        pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9856        return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9857}
9858
9859#else /* !CONFIG_MMU */
9860
9861static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9862{
9863        return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9864}
9865
9866static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9867{
9868        return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9869}
9870
9871static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9872        unsigned long addr, unsigned long len,
9873        unsigned long pgoff, unsigned long flags)
9874{
9875        void *ptr;
9876
9877        ptr = io_uring_validate_mmap_request(file, pgoff, len);
9878        if (IS_ERR(ptr))
9879                return PTR_ERR(ptr);
9880
9881        return (unsigned long) ptr;
9882}
9883
9884#endif /* !CONFIG_MMU */
9885
9886static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9887{
9888        DEFINE_WAIT(wait);
9889
9890        do {
9891                if (!io_sqring_full(ctx))
9892                        break;
9893                prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9894
9895                if (!io_sqring_full(ctx))
9896                        break;
9897                schedule();
9898        } while (!signal_pending(current));
9899
9900        finish_wait(&ctx->sqo_sq_wait, &wait);
9901        return 0;
9902}
9903
9904static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9905                          struct __kernel_timespec __user **ts,
9906                          const sigset_t __user **sig)
9907{
9908        struct io_uring_getevents_arg arg;
9909
9910        /*
9911         * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9912         * is just a pointer to the sigset_t.
9913         */
9914        if (!(flags & IORING_ENTER_EXT_ARG)) {
9915                *sig = (const sigset_t __user *) argp;
9916                *ts = NULL;
9917                return 0;
9918        }
9919
9920        /*
9921         * EXT_ARG is set - ensure we agree on the size of it and copy in our
9922         * timespec and sigset_t pointers if good.
9923         */
9924        if (*argsz != sizeof(arg))
9925                return -EINVAL;
9926        if (copy_from_user(&arg, argp, sizeof(arg)))
9927                return -EFAULT;
9928        *sig = u64_to_user_ptr(arg.sigmask);
9929        *argsz = arg.sigmask_sz;
9930        *ts = u64_to_user_ptr(arg.ts);
9931        return 0;
9932}
9933
9934SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9935                u32, min_complete, u32, flags, const void __user *, argp,
9936                size_t, argsz)
9937{
9938        struct io_ring_ctx *ctx;
9939        int submitted = 0;
9940        struct fd f;
9941        long ret;
9942
9943        io_run_task_work();
9944
9945        if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9946                               IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9947                return -EINVAL;
9948
9949        f = fdget(fd);
9950        if (unlikely(!f.file))
9951                return -EBADF;
9952
9953        ret = -EOPNOTSUPP;
9954        if (unlikely(f.file->f_op != &io_uring_fops))
9955                goto out_fput;
9956
9957        ret = -ENXIO;
9958        ctx = f.file->private_data;
9959        if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9960                goto out_fput;
9961
9962        ret = -EBADFD;
9963        if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9964                goto out;
9965
9966        /*
9967         * For SQ polling, the thread will do all submissions and completions.
9968         * Just return the requested submit count, and wake the thread if
9969         * we were asked to.
9970         */
9971        ret = 0;
9972        if (ctx->flags & IORING_SETUP_SQPOLL) {
9973                io_cqring_overflow_flush(ctx);
9974
9975                if (unlikely(ctx->sq_data->thread == NULL)) {
9976                        ret = -EOWNERDEAD;
9977                        goto out;
9978                }
9979                if (flags & IORING_ENTER_SQ_WAKEUP)
9980                        wake_up(&ctx->sq_data->wait);
9981                if (flags & IORING_ENTER_SQ_WAIT) {
9982                        ret = io_sqpoll_wait_sq(ctx);
9983                        if (ret)
9984                                goto out;
9985                }
9986                submitted = to_submit;
9987        } else if (to_submit) {
9988                ret = io_uring_add_tctx_node(ctx);
9989                if (unlikely(ret))
9990                        goto out;
9991                mutex_lock(&ctx->uring_lock);
9992                submitted = io_submit_sqes(ctx, to_submit);
9993                mutex_unlock(&ctx->uring_lock);
9994
9995                if (submitted != to_submit)
9996                        goto out;
9997        }
9998        if (flags & IORING_ENTER_GETEVENTS) {
9999                const sigset_t __user *sig;
10000                struct __kernel_timespec __user *ts;
10001
10002                ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
10003                if (unlikely(ret))
10004                        goto out;
10005
10006                min_complete = min(min_complete, ctx->cq_entries);
10007
10008                /*
10009                 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
10010                 * space applications don't need to do io completion events
10011                 * polling again, they can rely on io_sq_thread to do polling
10012                 * work, which can reduce cpu usage and uring_lock contention.
10013                 */
10014                if (ctx->flags & IORING_SETUP_IOPOLL &&
10015                    !(ctx->flags & IORING_SETUP_SQPOLL)) {
10016                        ret = io_iopoll_check(ctx, min_complete);
10017                } else {
10018                        ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
10019                }
10020        }
10021
10022out:
10023        percpu_ref_put(&ctx->refs);
10024out_fput:
10025        fdput(f);
10026        return submitted ? submitted : ret;
10027}
10028
10029#ifdef CONFIG_PROC_FS
10030static int io_uring_show_cred(struct seq_file *m, unsigned int id,
10031                const struct cred *cred)
10032{
10033        struct user_namespace *uns = seq_user_ns(m);
10034        struct group_info *gi;
10035        kernel_cap_t cap;
10036        unsigned __capi;
10037        int g;
10038
10039        seq_printf(m, "%5d\n", id);
10040        seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
10041        seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
10042        seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
10043        seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
10044        seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
10045        seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
10046        seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
10047        seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
10048        seq_puts(m, "\n\tGroups:\t");
10049        gi = cred->group_info;
10050        for (g = 0; g < gi->ngroups; g++) {
10051                seq_put_decimal_ull(m, g ? " " : "",
10052                                        from_kgid_munged(uns, gi->gid[g]));
10053        }
10054        seq_puts(m, "\n\tCapEff:\t");
10055        cap = cred->cap_effective;
10056        CAP_FOR_EACH_U32(__capi)
10057                seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
10058        seq_putc(m, '\n');
10059        return 0;
10060}
10061
10062static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
10063{
10064        struct io_sq_data *sq = NULL;
10065        bool has_lock;
10066        int i;
10067
10068        /*
10069         * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
10070         * since fdinfo case grabs it in the opposite direction of normal use
10071         * cases. If we fail to get the lock, we just don't iterate any
10072         * structures that could be going away outside the io_uring mutex.
10073         */
10074        has_lock = mutex_trylock(&ctx->uring_lock);
10075
10076        if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
10077                sq = ctx->sq_data;
10078                if (!sq->thread)
10079                        sq = NULL;
10080        }
10081
10082        seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
10083        seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
10084        seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
10085        for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
10086                struct file *f = io_file_from_index(ctx, i);
10087
10088                if (f)
10089                        seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
10090                else
10091                        seq_printf(m, "%5u: <none>\n", i);
10092        }
10093        seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
10094        for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
10095                struct io_mapped_ubuf *buf = ctx->user_bufs[i];
10096                unsigned int len = buf->ubuf_end - buf->ubuf;
10097
10098                seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
10099        }
10100        if (has_lock && !xa_empty(&ctx->personalities)) {
10101                unsigned long index;
10102                const struct cred *cred;
10103
10104                seq_printf(m, "Personalities:\n");
10105                xa_for_each(&ctx->personalities, index, cred)
10106                        io_uring_show_cred(m, index, cred);
10107        }
10108        seq_printf(m, "PollList:\n");
10109        spin_lock(&ctx->completion_lock);
10110        for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
10111                struct hlist_head *list = &ctx->cancel_hash[i];
10112                struct io_kiocb *req;
10113
10114                hlist_for_each_entry(req, list, hash_node)
10115                        seq_printf(m, "  op=%d, task_works=%d\n", req->opcode,
10116                                        req->task->task_works != NULL);
10117        }
10118        spin_unlock(&ctx->completion_lock);
10119        if (has_lock)
10120                mutex_unlock(&ctx->uring_lock);
10121}
10122
10123static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
10124{
10125        struct io_ring_ctx *ctx = f->private_data;
10126
10127        if (percpu_ref_tryget(&ctx->refs)) {
10128                __io_uring_show_fdinfo(ctx, m);
10129                percpu_ref_put(&ctx->refs);
10130        }
10131}
10132#endif
10133
10134static const struct file_operations io_uring_fops = {
10135        .release        = io_uring_release,
10136        .mmap           = io_uring_mmap,
10137#ifndef CONFIG_MMU
10138        .get_unmapped_area = io_uring_nommu_get_unmapped_area,
10139        .mmap_capabilities = io_uring_nommu_mmap_capabilities,
10140#endif
10141        .poll           = io_uring_poll,
10142#ifdef CONFIG_PROC_FS
10143        .show_fdinfo    = io_uring_show_fdinfo,
10144#endif
10145};
10146
10147static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
10148                                  struct io_uring_params *p)
10149{
10150        struct io_rings *rings;
10151        size_t size, sq_array_offset;
10152
10153        /* make sure these are sane, as we already accounted them */
10154        ctx->sq_entries = p->sq_entries;
10155        ctx->cq_entries = p->cq_entries;
10156
10157        size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
10158        if (size == SIZE_MAX)
10159                return -EOVERFLOW;
10160
10161        rings = io_mem_alloc(size);
10162        if (!rings)
10163                return -ENOMEM;
10164
10165        ctx->rings = rings;
10166        ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
10167        rings->sq_ring_mask = p->sq_entries - 1;
10168        rings->cq_ring_mask = p->cq_entries - 1;
10169        rings->sq_ring_entries = p->sq_entries;
10170        rings->cq_ring_entries = p->cq_entries;
10171
10172        size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
10173        if (size == SIZE_MAX) {
10174                io_mem_free(ctx->rings);
10175                ctx->rings = NULL;
10176                return -EOVERFLOW;
10177        }
10178
10179        ctx->sq_sqes = io_mem_alloc(size);
10180        if (!ctx->sq_sqes) {
10181                io_mem_free(ctx->rings);
10182                ctx->rings = NULL;
10183                return -ENOMEM;
10184        }
10185
10186        return 0;
10187}
10188
10189static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
10190{
10191        int ret, fd;
10192
10193        fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
10194        if (fd < 0)
10195                return fd;
10196
10197        ret = io_uring_add_tctx_node(ctx);
10198        if (ret) {
10199                put_unused_fd(fd);
10200                return ret;
10201        }
10202        fd_install(fd, file);
10203        return fd;
10204}
10205
10206/*
10207 * Allocate an anonymous fd, this is what constitutes the application
10208 * visible backing of an io_uring instance. The application mmaps this
10209 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10210 * we have to tie this fd to a socket for file garbage collection purposes.
10211 */
10212static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
10213{
10214        struct file *file;
10215#if defined(CONFIG_UNIX)
10216        int ret;
10217
10218        ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
10219                                &ctx->ring_sock);
10220        if (ret)
10221                return ERR_PTR(ret);
10222#endif
10223
10224        file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
10225                                        O_RDWR | O_CLOEXEC);
10226#if defined(CONFIG_UNIX)
10227        if (IS_ERR(file)) {
10228                sock_release(ctx->ring_sock);
10229                ctx->ring_sock = NULL;
10230        } else {
10231                ctx->ring_sock->file = file;
10232        }
10233#endif
10234        return file;
10235}
10236
10237static int io_uring_create(unsigned entries, struct io_uring_params *p,
10238                           struct io_uring_params __user *params)
10239{
10240        struct io_ring_ctx *ctx;
10241        struct file *file;
10242        int ret;
10243
10244        if (!entries)
10245                return -EINVAL;
10246        if (entries > IORING_MAX_ENTRIES) {
10247                if (!(p->flags & IORING_SETUP_CLAMP))
10248                        return -EINVAL;
10249                entries = IORING_MAX_ENTRIES;
10250        }
10251
10252        /*
10253         * Use twice as many entries for the CQ ring. It's possible for the
10254         * application to drive a higher depth than the size of the SQ ring,
10255         * since the sqes are only used at submission time. This allows for
10256         * some flexibility in overcommitting a bit. If the application has
10257         * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10258         * of CQ ring entries manually.
10259         */
10260        p->sq_entries = roundup_pow_of_two(entries);
10261        if (p->flags & IORING_SETUP_CQSIZE) {
10262                /*
10263                 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10264                 * to a power-of-two, if it isn't already. We do NOT impose
10265                 * any cq vs sq ring sizing.
10266                 */
10267                if (!p->cq_entries)
10268                        return -EINVAL;
10269                if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
10270                        if (!(p->flags & IORING_SETUP_CLAMP))
10271                                return -EINVAL;
10272                        p->cq_entries = IORING_MAX_CQ_ENTRIES;
10273                }
10274                p->cq_entries = roundup_pow_of_two(p->cq_entries);
10275                if (p->cq_entries < p->sq_entries)
10276                        return -EINVAL;
10277        } else {
10278                p->cq_entries = 2 * p->sq_entries;
10279        }
10280
10281        ctx = io_ring_ctx_alloc(p);
10282        if (!ctx)
10283                return -ENOMEM;
10284        ctx->compat = in_compat_syscall();
10285        if (!capable(CAP_IPC_LOCK))
10286                ctx->user = get_uid(current_user());
10287
10288        /*
10289         * This is just grabbed for accounting purposes. When a process exits,
10290         * the mm is exited and dropped before the files, hence we need to hang
10291         * on to this mm purely for the purposes of being able to unaccount
10292         * memory (locked/pinned vm). It's not used for anything else.
10293         */
10294        mmgrab(current->mm);
10295        ctx->mm_account = current->mm;
10296
10297        ret = io_allocate_scq_urings(ctx, p);
10298        if (ret)
10299                goto err;
10300
10301        ret = io_sq_offload_create(ctx, p);
10302        if (ret)
10303                goto err;
10304        /* always set a rsrc node */
10305        ret = io_rsrc_node_switch_start(ctx);
10306        if (ret)
10307                goto err;
10308        io_rsrc_node_switch(ctx, NULL);
10309
10310        memset(&p->sq_off, 0, sizeof(p->sq_off));
10311        p->sq_off.head = offsetof(struct io_rings, sq.head);
10312        p->sq_off.tail = offsetof(struct io_rings, sq.tail);
10313        p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
10314        p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
10315        p->sq_off.flags = offsetof(struct io_rings, sq_flags);
10316        p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
10317        p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
10318
10319        memset(&p->cq_off, 0, sizeof(p->cq_off));
10320        p->cq_off.head = offsetof(struct io_rings, cq.head);
10321        p->cq_off.tail = offsetof(struct io_rings, cq.tail);
10322        p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
10323        p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
10324        p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
10325        p->cq_off.cqes = offsetof(struct io_rings, cqes);
10326        p->cq_off.flags = offsetof(struct io_rings, cq_flags);
10327
10328        p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
10329                        IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
10330                        IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
10331                        IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
10332                        IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
10333                        IORING_FEAT_RSRC_TAGS;
10334
10335        if (copy_to_user(params, p, sizeof(*p))) {
10336                ret = -EFAULT;
10337                goto err;
10338        }
10339
10340        file = io_uring_get_file(ctx);
10341        if (IS_ERR(file)) {
10342                ret = PTR_ERR(file);
10343                goto err;
10344        }
10345
10346        /*
10347         * Install ring fd as the very last thing, so we don't risk someone
10348         * having closed it before we finish setup
10349         */
10350        ret = io_uring_install_fd(ctx, file);
10351        if (ret < 0) {
10352                /* fput will clean it up */
10353                fput(file);
10354                return ret;
10355        }
10356
10357        trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
10358        return ret;
10359err:
10360        io_ring_ctx_wait_and_kill(ctx);
10361        return ret;
10362}
10363
10364/*
10365 * Sets up an aio uring context, and returns the fd. Applications asks for a
10366 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10367 * params structure passed in.
10368 */
10369static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
10370{
10371        struct io_uring_params p;
10372        int i;
10373
10374        if (copy_from_user(&p, params, sizeof(p)))
10375                return -EFAULT;
10376        for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
10377                if (p.resv[i])
10378                        return -EINVAL;
10379        }
10380
10381        if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
10382                        IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
10383                        IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
10384                        IORING_SETUP_R_DISABLED))
10385                return -EINVAL;
10386
10387        return  io_uring_create(entries, &p, params);
10388}
10389
10390SYSCALL_DEFINE2(io_uring_setup, u32, entries,
10391                struct io_uring_params __user *, params)
10392{
10393        return io_uring_setup(entries, params);
10394}
10395
10396static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
10397{
10398        struct io_uring_probe *p;
10399        size_t size;
10400        int i, ret;
10401
10402        size = struct_size(p, ops, nr_args);
10403        if (size == SIZE_MAX)
10404                return -EOVERFLOW;
10405        p = kzalloc(size, GFP_KERNEL);
10406        if (!p)
10407                return -ENOMEM;
10408
10409        ret = -EFAULT;
10410        if (copy_from_user(p, arg, size))
10411                goto out;
10412        ret = -EINVAL;
10413        if (memchr_inv(p, 0, size))
10414                goto out;
10415
10416        p->last_op = IORING_OP_LAST - 1;
10417        if (nr_args > IORING_OP_LAST)
10418                nr_args = IORING_OP_LAST;
10419
10420        for (i = 0; i < nr_args; i++) {
10421                p->ops[i].op = i;
10422                if (!io_op_defs[i].not_supported)
10423                        p->ops[i].flags = IO_URING_OP_SUPPORTED;
10424        }
10425        p->ops_len = i;
10426
10427        ret = 0;
10428        if (copy_to_user(arg, p, size))
10429                ret = -EFAULT;
10430out:
10431        kfree(p);
10432        return ret;
10433}
10434
10435static int io_register_personality(struct io_ring_ctx *ctx)
10436{
10437        const struct cred *creds;
10438        u32 id;
10439        int ret;
10440
10441        creds = get_current_cred();
10442
10443        ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
10444                        XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
10445        if (ret < 0) {
10446                put_cred(creds);
10447                return ret;
10448        }
10449        return id;
10450}
10451
10452static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
10453                                    unsigned int nr_args)
10454{
10455        struct io_uring_restriction *res;
10456        size_t size;
10457        int i, ret;
10458
10459        /* Restrictions allowed only if rings started disabled */
10460        if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10461                return -EBADFD;
10462
10463        /* We allow only a single restrictions registration */
10464        if (ctx->restrictions.registered)
10465                return -EBUSY;
10466
10467        if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
10468                return -EINVAL;
10469
10470        size = array_size(nr_args, sizeof(*res));
10471        if (size == SIZE_MAX)
10472                return -EOVERFLOW;
10473
10474        res = memdup_user(arg, size);
10475        if (IS_ERR(res))
10476                return PTR_ERR(res);
10477
10478        ret = 0;
10479
10480        for (i = 0; i < nr_args; i++) {
10481                switch (res[i].opcode) {
10482                case IORING_RESTRICTION_REGISTER_OP:
10483                        if (res[i].register_op >= IORING_REGISTER_LAST) {
10484                                ret = -EINVAL;
10485                                goto out;
10486                        }
10487
10488                        __set_bit(res[i].register_op,
10489                                  ctx->restrictions.register_op);
10490                        break;
10491                case IORING_RESTRICTION_SQE_OP:
10492                        if (res[i].sqe_op >= IORING_OP_LAST) {
10493                                ret = -EINVAL;
10494                                goto out;
10495                        }
10496
10497                        __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10498                        break;
10499                case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10500                        ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10501                        break;
10502                case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10503                        ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10504                        break;
10505                default:
10506                        ret = -EINVAL;
10507                        goto out;
10508                }
10509        }
10510
10511out:
10512        /* Reset all restrictions if an error happened */
10513        if (ret != 0)
10514                memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10515        else
10516                ctx->restrictions.registered = true;
10517
10518        kfree(res);
10519        return ret;
10520}
10521
10522static int io_register_enable_rings(struct io_ring_ctx *ctx)
10523{
10524        if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10525                return -EBADFD;
10526
10527        if (ctx->restrictions.registered)
10528                ctx->restricted = 1;
10529
10530        ctx->flags &= ~IORING_SETUP_R_DISABLED;
10531        if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
10532                wake_up(&ctx->sq_data->wait);
10533        return 0;
10534}
10535
10536static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
10537                                     struct io_uring_rsrc_update2 *up,
10538                                     unsigned nr_args)
10539{
10540        __u32 tmp;
10541        int err;
10542
10543        if (up->resv)
10544                return -EINVAL;
10545        if (check_add_overflow(up->offset, nr_args, &tmp))
10546                return -EOVERFLOW;
10547        err = io_rsrc_node_switch_start(ctx);
10548        if (err)
10549                return err;
10550
10551        switch (type) {
10552        case IORING_RSRC_FILE:
10553                return __io_sqe_files_update(ctx, up, nr_args);
10554        case IORING_RSRC_BUFFER:
10555                return __io_sqe_buffers_update(ctx, up, nr_args);
10556        }
10557        return -EINVAL;
10558}
10559
10560static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
10561                                    unsigned nr_args)
10562{
10563        struct io_uring_rsrc_update2 up;
10564
10565        if (!nr_args)
10566                return -EINVAL;
10567        memset(&up, 0, sizeof(up));
10568        if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10569                return -EFAULT;
10570        return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10571}
10572
10573static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10574                                   unsigned size, unsigned type)
10575{
10576        struct io_uring_rsrc_update2 up;
10577
10578        if (size != sizeof(up))
10579                return -EINVAL;
10580        if (copy_from_user(&up, arg, sizeof(up)))
10581                return -EFAULT;
10582        if (!up.nr || up.resv)
10583                return -EINVAL;
10584        return __io_register_rsrc_update(ctx, type, &up, up.nr);
10585}
10586
10587static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10588                            unsigned int size, unsigned int type)
10589{
10590        struct io_uring_rsrc_register rr;
10591
10592        /* keep it extendible */
10593        if (size != sizeof(rr))
10594                return -EINVAL;
10595
10596        memset(&rr, 0, sizeof(rr));
10597        if (copy_from_user(&rr, arg, size))
10598                return -EFAULT;
10599        if (!rr.nr || rr.resv || rr.resv2)
10600                return -EINVAL;
10601
10602        switch (type) {
10603        case IORING_RSRC_FILE:
10604                return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10605                                             rr.nr, u64_to_user_ptr(rr.tags));
10606        case IORING_RSRC_BUFFER:
10607                return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10608                                               rr.nr, u64_to_user_ptr(rr.tags));
10609        }
10610        return -EINVAL;
10611}
10612
10613static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10614                                unsigned len)
10615{
10616        struct io_uring_task *tctx = current->io_uring;
10617        cpumask_var_t new_mask;
10618        int ret;
10619
10620        if (!tctx || !tctx->io_wq)
10621                return -EINVAL;
10622
10623        if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10624                return -ENOMEM;
10625
10626        cpumask_clear(new_mask);
10627        if (len > cpumask_size())
10628                len = cpumask_size();
10629
10630        if (copy_from_user(new_mask, arg, len)) {
10631                free_cpumask_var(new_mask);
10632                return -EFAULT;
10633        }
10634
10635        ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10636        free_cpumask_var(new_mask);
10637        return ret;
10638}
10639
10640static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10641{
10642        struct io_uring_task *tctx = current->io_uring;
10643
10644        if (!tctx || !tctx->io_wq)
10645                return -EINVAL;
10646
10647        return io_wq_cpu_affinity(tctx->io_wq, NULL);
10648}
10649
10650static int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
10651                                        void __user *arg)
10652        __must_hold(&ctx->uring_lock)
10653{
10654        struct io_tctx_node *node;
10655        struct io_uring_task *tctx = NULL;
10656        struct io_sq_data *sqd = NULL;
10657        __u32 new_count[2];
10658        int i, ret;
10659
10660        if (copy_from_user(new_count, arg, sizeof(new_count)))
10661                return -EFAULT;
10662        for (i = 0; i < ARRAY_SIZE(new_count); i++)
10663                if (new_count[i] > INT_MAX)
10664                        return -EINVAL;
10665
10666        if (ctx->flags & IORING_SETUP_SQPOLL) {
10667                sqd = ctx->sq_data;
10668                if (sqd) {
10669                        /*
10670                         * Observe the correct sqd->lock -> ctx->uring_lock
10671                         * ordering. Fine to drop uring_lock here, we hold
10672                         * a ref to the ctx.
10673                         */
10674                        refcount_inc(&sqd->refs);
10675                        mutex_unlock(&ctx->uring_lock);
10676                        mutex_lock(&sqd->lock);
10677                        mutex_lock(&ctx->uring_lock);
10678                        if (sqd->thread)
10679                                tctx = sqd->thread->io_uring;
10680                }
10681        } else {
10682                tctx = current->io_uring;
10683        }
10684
10685        BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
10686
10687        memcpy(ctx->iowq_limits, new_count, sizeof(new_count));
10688        ctx->iowq_limits_set = true;
10689
10690        ret = -EINVAL;
10691        if (tctx && tctx->io_wq) {
10692                ret = io_wq_max_workers(tctx->io_wq, new_count);
10693                if (ret)
10694                        goto err;
10695        } else {
10696                memset(new_count, 0, sizeof(new_count));
10697        }
10698
10699        if (sqd) {
10700                mutex_unlock(&sqd->lock);
10701                io_put_sq_data(sqd);
10702        }
10703
10704        if (copy_to_user(arg, new_count, sizeof(new_count)))
10705                return -EFAULT;
10706
10707        /* that's it for SQPOLL, only the SQPOLL task creates requests */
10708        if (sqd)
10709                return 0;
10710
10711        /* now propagate the restriction to all registered users */
10712        list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
10713                struct io_uring_task *tctx = node->task->io_uring;
10714
10715                if (WARN_ON_ONCE(!tctx->io_wq))
10716                        continue;
10717
10718                for (i = 0; i < ARRAY_SIZE(new_count); i++)
10719                        new_count[i] = ctx->iowq_limits[i];
10720                /* ignore errors, it always returns zero anyway */
10721                (void)io_wq_max_workers(tctx->io_wq, new_count);
10722        }
10723        return 0;
10724err:
10725        if (sqd) {
10726                mutex_unlock(&sqd->lock);
10727                io_put_sq_data(sqd);
10728        }
10729        return ret;
10730}
10731
10732static bool io_register_op_must_quiesce(int op)
10733{
10734        switch (op) {
10735        case IORING_REGISTER_BUFFERS:
10736        case IORING_UNREGISTER_BUFFERS:
10737        case IORING_REGISTER_FILES:
10738        case IORING_UNREGISTER_FILES:
10739        case IORING_REGISTER_FILES_UPDATE:
10740        case IORING_REGISTER_PROBE:
10741        case IORING_REGISTER_PERSONALITY:
10742        case IORING_UNREGISTER_PERSONALITY:
10743        case IORING_REGISTER_FILES2:
10744        case IORING_REGISTER_FILES_UPDATE2:
10745        case IORING_REGISTER_BUFFERS2:
10746        case IORING_REGISTER_BUFFERS_UPDATE:
10747        case IORING_REGISTER_IOWQ_AFF:
10748        case IORING_UNREGISTER_IOWQ_AFF:
10749        case IORING_REGISTER_IOWQ_MAX_WORKERS:
10750                return false;
10751        default:
10752                return true;
10753        }
10754}
10755
10756static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10757{
10758        long ret;
10759
10760        percpu_ref_kill(&ctx->refs);
10761
10762        /*
10763         * Drop uring mutex before waiting for references to exit. If another
10764         * thread is currently inside io_uring_enter() it might need to grab the
10765         * uring_lock to make progress. If we hold it here across the drain
10766         * wait, then we can deadlock. It's safe to drop the mutex here, since
10767         * no new references will come in after we've killed the percpu ref.
10768         */
10769        mutex_unlock(&ctx->uring_lock);
10770        do {
10771                ret = wait_for_completion_interruptible(&ctx->ref_comp);
10772                if (!ret)
10773                        break;
10774                ret = io_run_task_work_sig();
10775        } while (ret >= 0);
10776        mutex_lock(&ctx->uring_lock);
10777
10778        if (ret)
10779                io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10780        return ret;
10781}
10782
10783static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10784                               void __user *arg, unsigned nr_args)
10785        __releases(ctx->uring_lock)
10786        __acquires(ctx->uring_lock)
10787{
10788        int ret;
10789
10790        /*
10791         * We're inside the ring mutex, if the ref is already dying, then
10792         * someone else killed the ctx or is already going through
10793         * io_uring_register().
10794         */
10795        if (percpu_ref_is_dying(&ctx->refs))
10796                return -ENXIO;
10797
10798        if (ctx->restricted) {
10799                if (opcode >= IORING_REGISTER_LAST)
10800                        return -EINVAL;
10801                opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10802                if (!test_bit(opcode, ctx->restrictions.register_op))
10803                        return -EACCES;
10804        }
10805
10806        if (io_register_op_must_quiesce(opcode)) {
10807                ret = io_ctx_quiesce(ctx);
10808                if (ret)
10809                        return ret;
10810        }
10811
10812        switch (opcode) {
10813        case IORING_REGISTER_BUFFERS:
10814                ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10815                break;
10816        case IORING_UNREGISTER_BUFFERS:
10817                ret = -EINVAL;
10818                if (arg || nr_args)
10819                        break;
10820                ret = io_sqe_buffers_unregister(ctx);
10821                break;
10822        case IORING_REGISTER_FILES:
10823                ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10824                break;
10825        case IORING_UNREGISTER_FILES:
10826                ret = -EINVAL;
10827                if (arg || nr_args)
10828                        break;
10829                ret = io_sqe_files_unregister(ctx);
10830                break;
10831        case IORING_REGISTER_FILES_UPDATE:
10832                ret = io_register_files_update(ctx, arg, nr_args);
10833                break;
10834        case IORING_REGISTER_EVENTFD:
10835        case IORING_REGISTER_EVENTFD_ASYNC:
10836                ret = -EINVAL;
10837                if (nr_args != 1)
10838                        break;
10839                ret = io_eventfd_register(ctx, arg);
10840                if (ret)
10841                        break;
10842                if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10843                        ctx->eventfd_async = 1;
10844                else
10845                        ctx->eventfd_async = 0;
10846                break;
10847        case IORING_UNREGISTER_EVENTFD:
10848                ret = -EINVAL;
10849                if (arg || nr_args)
10850                        break;
10851                ret = io_eventfd_unregister(ctx);
10852                break;
10853        case IORING_REGISTER_PROBE:
10854                ret = -EINVAL;
10855                if (!arg || nr_args > 256)
10856                        break;
10857                ret = io_probe(ctx, arg, nr_args);
10858                break;
10859        case IORING_REGISTER_PERSONALITY:
10860                ret = -EINVAL;
10861                if (arg || nr_args)
10862                        break;
10863                ret = io_register_personality(ctx);
10864                break;
10865        case IORING_UNREGISTER_PERSONALITY:
10866                ret = -EINVAL;
10867                if (arg)
10868                        break;
10869                ret = io_unregister_personality(ctx, nr_args);
10870                break;
10871        case IORING_REGISTER_ENABLE_RINGS:
10872                ret = -EINVAL;
10873                if (arg || nr_args)
10874                        break;
10875                ret = io_register_enable_rings(ctx);
10876                break;
10877        case IORING_REGISTER_RESTRICTIONS:
10878                ret = io_register_restrictions(ctx, arg, nr_args);
10879                break;
10880        case IORING_REGISTER_FILES2:
10881                ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10882                break;
10883        case IORING_REGISTER_FILES_UPDATE2:
10884                ret = io_register_rsrc_update(ctx, arg, nr_args,
10885                                              IORING_RSRC_FILE);
10886                break;
10887        case IORING_REGISTER_BUFFERS2:
10888                ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10889                break;
10890        case IORING_REGISTER_BUFFERS_UPDATE:
10891                ret = io_register_rsrc_update(ctx, arg, nr_args,
10892                                              IORING_RSRC_BUFFER);
10893                break;
10894        case IORING_REGISTER_IOWQ_AFF:
10895                ret = -EINVAL;
10896                if (!arg || !nr_args)
10897                        break;
10898                ret = io_register_iowq_aff(ctx, arg, nr_args);
10899                break;
10900        case IORING_UNREGISTER_IOWQ_AFF:
10901                ret = -EINVAL;
10902                if (arg || nr_args)
10903                        break;
10904                ret = io_unregister_iowq_aff(ctx);
10905                break;
10906        case IORING_REGISTER_IOWQ_MAX_WORKERS:
10907                ret = -EINVAL;
10908                if (!arg || nr_args != 2)
10909                        break;
10910                ret = io_register_iowq_max_workers(ctx, arg);
10911                break;
10912        default:
10913                ret = -EINVAL;
10914                break;
10915        }
10916
10917        if (io_register_op_must_quiesce(opcode)) {
10918                /* bring the ctx back to life */
10919                percpu_ref_reinit(&ctx->refs);
10920                reinit_completion(&ctx->ref_comp);
10921        }
10922        return ret;
10923}
10924
10925SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10926                void __user *, arg, unsigned int, nr_args)
10927{
10928        struct io_ring_ctx *ctx;
10929        long ret = -EBADF;
10930        struct fd f;
10931
10932        f = fdget(fd);
10933        if (!f.file)
10934                return -EBADF;
10935
10936        ret = -EOPNOTSUPP;
10937        if (f.file->f_op != &io_uring_fops)
10938                goto out_fput;
10939
10940        ctx = f.file->private_data;
10941
10942        io_run_task_work();
10943
10944        mutex_lock(&ctx->uring_lock);
10945        ret = __io_uring_register(ctx, opcode, arg, nr_args);
10946        mutex_unlock(&ctx->uring_lock);
10947        trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10948                                                        ctx->cq_ev_fd != NULL, ret);
10949out_fput:
10950        fdput(f);
10951        return ret;
10952}
10953
10954static int __init io_uring_init(void)
10955{
10956#define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10957        BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10958        BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10959} while (0)
10960
10961#define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10962        __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10963        BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10964        BUILD_BUG_SQE_ELEM(0,  __u8,   opcode);
10965        BUILD_BUG_SQE_ELEM(1,  __u8,   flags);
10966        BUILD_BUG_SQE_ELEM(2,  __u16,  ioprio);
10967        BUILD_BUG_SQE_ELEM(4,  __s32,  fd);
10968        BUILD_BUG_SQE_ELEM(8,  __u64,  off);
10969        BUILD_BUG_SQE_ELEM(8,  __u64,  addr2);
10970        BUILD_BUG_SQE_ELEM(16, __u64,  addr);
10971        BUILD_BUG_SQE_ELEM(16, __u64,  splice_off_in);
10972        BUILD_BUG_SQE_ELEM(24, __u32,  len);
10973        BUILD_BUG_SQE_ELEM(28,     __kernel_rwf_t, rw_flags);
10974        BUILD_BUG_SQE_ELEM(28, /* compat */   int, rw_flags);
10975        BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10976        BUILD_BUG_SQE_ELEM(28, __u32,  fsync_flags);
10977        BUILD_BUG_SQE_ELEM(28, /* compat */ __u16,  poll_events);
10978        BUILD_BUG_SQE_ELEM(28, __u32,  poll32_events);
10979        BUILD_BUG_SQE_ELEM(28, __u32,  sync_range_flags);
10980        BUILD_BUG_SQE_ELEM(28, __u32,  msg_flags);
10981        BUILD_BUG_SQE_ELEM(28, __u32,  timeout_flags);
10982        BUILD_BUG_SQE_ELEM(28, __u32,  accept_flags);
10983        BUILD_BUG_SQE_ELEM(28, __u32,  cancel_flags);
10984        BUILD_BUG_SQE_ELEM(28, __u32,  open_flags);
10985        BUILD_BUG_SQE_ELEM(28, __u32,  statx_flags);
10986        BUILD_BUG_SQE_ELEM(28, __u32,  fadvise_advice);
10987        BUILD_BUG_SQE_ELEM(28, __u32,  splice_flags);
10988        BUILD_BUG_SQE_ELEM(32, __u64,  user_data);
10989        BUILD_BUG_SQE_ELEM(40, __u16,  buf_index);
10990        BUILD_BUG_SQE_ELEM(40, __u16,  buf_group);
10991        BUILD_BUG_SQE_ELEM(42, __u16,  personality);
10992        BUILD_BUG_SQE_ELEM(44, __s32,  splice_fd_in);
10993        BUILD_BUG_SQE_ELEM(44, __u32,  file_index);
10994
10995        BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10996                     sizeof(struct io_uring_rsrc_update));
10997        BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10998                     sizeof(struct io_uring_rsrc_update2));
10999
11000        /* ->buf_index is u16 */
11001        BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
11002
11003        /* should fit into one byte */
11004        BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
11005
11006        BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
11007        BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
11008
11009        req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
11010                                SLAB_ACCOUNT);
11011        return 0;
11012};
11013__initcall(io_uring_init);
11014