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