linux/fs/aio.c
<<
>>
Prefs
   1/*
   2 *      An async IO implementation for Linux
   3 *      Written by Benjamin LaHaise <bcrl@kvack.org>
   4 *
   5 *      Implements an efficient asynchronous io interface.
   6 *
   7 *      Copyright 2000, 2001, 2002 Red Hat, Inc.  All Rights Reserved.
   8 *
   9 *      See ../COPYING for licensing terms.
  10 */
  11#define pr_fmt(fmt) "%s: " fmt, __func__
  12
  13#include <linux/kernel.h>
  14#include <linux/init.h>
  15#include <linux/errno.h>
  16#include <linux/time.h>
  17#include <linux/aio_abi.h>
  18#include <linux/export.h>
  19#include <linux/syscalls.h>
  20#include <linux/backing-dev.h>
  21#include <linux/uio.h>
  22
  23#include <linux/sched.h>
  24#include <linux/fs.h>
  25#include <linux/file.h>
  26#include <linux/mm.h>
  27#include <linux/mman.h>
  28#include <linux/mmu_context.h>
  29#include <linux/percpu.h>
  30#include <linux/slab.h>
  31#include <linux/timer.h>
  32#include <linux/aio.h>
  33#include <linux/highmem.h>
  34#include <linux/workqueue.h>
  35#include <linux/security.h>
  36#include <linux/eventfd.h>
  37#include <linux/blkdev.h>
  38#include <linux/compat.h>
  39#include <linux/migrate.h>
  40#include <linux/ramfs.h>
  41#include <linux/percpu-refcount.h>
  42#include <linux/mount.h>
  43
  44#include <asm/kmap_types.h>
  45#include <asm/uaccess.h>
  46
  47#include "internal.h"
  48
  49#define AIO_RING_MAGIC                  0xa10a10a1
  50#define AIO_RING_COMPAT_FEATURES        1
  51#define AIO_RING_INCOMPAT_FEATURES      0
  52struct aio_ring {
  53        unsigned        id;     /* kernel internal index number */
  54        unsigned        nr;     /* number of io_events */
  55        unsigned        head;   /* Written to by userland or under ring_lock
  56                                 * mutex by aio_read_events_ring(). */
  57        unsigned        tail;
  58
  59        unsigned        magic;
  60        unsigned        compat_features;
  61        unsigned        incompat_features;
  62        unsigned        header_length;  /* size of aio_ring */
  63
  64
  65        struct io_event         io_events[0];
  66}; /* 128 bytes + ring size */
  67
  68#define AIO_RING_PAGES  8
  69
  70struct kioctx_table {
  71        struct rcu_head rcu;
  72        unsigned        nr;
  73        struct kioctx   *table[];
  74};
  75
  76struct kioctx_cpu {
  77        unsigned                reqs_available;
  78};
  79
  80struct ctx_rq_wait {
  81        struct completion comp;
  82        atomic_t count;
  83};
  84
  85struct kioctx {
  86        struct percpu_ref       users;
  87        atomic_t                dead;
  88
  89        struct percpu_ref       reqs;
  90
  91        unsigned long           user_id;
  92
  93        struct __percpu kioctx_cpu *cpu;
  94
  95        /*
  96         * For percpu reqs_available, number of slots we move to/from global
  97         * counter at a time:
  98         */
  99        unsigned                req_batch;
 100        /*
 101         * This is what userspace passed to io_setup(), it's not used for
 102         * anything but counting against the global max_reqs quota.
 103         *
 104         * The real limit is nr_events - 1, which will be larger (see
 105         * aio_setup_ring())
 106         */
 107        unsigned                max_reqs;
 108
 109        /* Size of ringbuffer, in units of struct io_event */
 110        unsigned                nr_events;
 111
 112        unsigned long           mmap_base;
 113        unsigned long           mmap_size;
 114
 115        struct page             **ring_pages;
 116        long                    nr_pages;
 117
 118        struct work_struct      free_work;
 119
 120        /*
 121         * signals when all in-flight requests are done
 122         */
 123        struct ctx_rq_wait      *rq_wait;
 124
 125        struct {
 126                /*
 127                 * This counts the number of available slots in the ringbuffer,
 128                 * so we avoid overflowing it: it's decremented (if positive)
 129                 * when allocating a kiocb and incremented when the resulting
 130                 * io_event is pulled off the ringbuffer.
 131                 *
 132                 * We batch accesses to it with a percpu version.
 133                 */
 134                atomic_t        reqs_available;
 135        } ____cacheline_aligned_in_smp;
 136
 137        struct {
 138                spinlock_t      ctx_lock;
 139                struct list_head active_reqs;   /* used for cancellation */
 140        } ____cacheline_aligned_in_smp;
 141
 142        struct {
 143                struct mutex    ring_lock;
 144                wait_queue_head_t wait;
 145        } ____cacheline_aligned_in_smp;
 146
 147        struct {
 148                unsigned        tail;
 149                unsigned        completed_events;
 150                spinlock_t      completion_lock;
 151        } ____cacheline_aligned_in_smp;
 152
 153        struct page             *internal_pages[AIO_RING_PAGES];
 154        struct file             *aio_ring_file;
 155
 156        unsigned                id;
 157};
 158
 159/*
 160 * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either
 161 * cancelled or completed (this makes a certain amount of sense because
 162 * successful cancellation - io_cancel() - does deliver the completion to
 163 * userspace).
 164 *
 165 * And since most things don't implement kiocb cancellation and we'd really like
 166 * kiocb completion to be lockless when possible, we use ki_cancel to
 167 * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED
 168 * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel().
 169 */
 170#define KIOCB_CANCELLED         ((void *) (~0ULL))
 171
 172struct aio_kiocb {
 173        struct kiocb            common;
 174
 175        struct kioctx           *ki_ctx;
 176        kiocb_cancel_fn         *ki_cancel;
 177
 178        struct iocb __user      *ki_user_iocb;  /* user's aiocb */
 179        __u64                   ki_user_data;   /* user's data for completion */
 180
 181        struct list_head        ki_list;        /* the aio core uses this
 182                                                 * for cancellation */
 183
 184        /*
 185         * If the aio_resfd field of the userspace iocb is not zero,
 186         * this is the underlying eventfd context to deliver events to.
 187         */
 188        struct eventfd_ctx      *ki_eventfd;
 189};
 190
 191/*------ sysctl variables----*/
 192static DEFINE_SPINLOCK(aio_nr_lock);
 193unsigned long aio_nr;           /* current system wide number of aio requests */
 194unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
 195/*----end sysctl variables---*/
 196
 197static struct kmem_cache        *kiocb_cachep;
 198static struct kmem_cache        *kioctx_cachep;
 199
 200static struct vfsmount *aio_mnt;
 201
 202static const struct file_operations aio_ring_fops;
 203static const struct address_space_operations aio_ctx_aops;
 204
 205static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
 206{
 207        struct qstr this = QSTR_INIT("[aio]", 5);
 208        struct file *file;
 209        struct path path;
 210        struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
 211        if (IS_ERR(inode))
 212                return ERR_CAST(inode);
 213
 214        inode->i_mapping->a_ops = &aio_ctx_aops;
 215        inode->i_mapping->private_data = ctx;
 216        inode->i_size = PAGE_SIZE * nr_pages;
 217
 218        path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
 219        if (!path.dentry) {
 220                iput(inode);
 221                return ERR_PTR(-ENOMEM);
 222        }
 223        path.mnt = mntget(aio_mnt);
 224
 225        d_instantiate(path.dentry, inode);
 226        file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
 227        if (IS_ERR(file)) {
 228                path_put(&path);
 229                return file;
 230        }
 231
 232        file->f_flags = O_RDWR;
 233        return file;
 234}
 235
 236static struct dentry *aio_mount(struct file_system_type *fs_type,
 237                                int flags, const char *dev_name, void *data)
 238{
 239        static const struct dentry_operations ops = {
 240                .d_dname        = simple_dname,
 241        };
 242        return mount_pseudo(fs_type, "aio:", NULL, &ops, AIO_RING_MAGIC);
 243}
 244
 245/* aio_setup
 246 *      Creates the slab caches used by the aio routines, panic on
 247 *      failure as this is done early during the boot sequence.
 248 */
 249static int __init aio_setup(void)
 250{
 251        static struct file_system_type aio_fs = {
 252                .name           = "aio",
 253                .mount          = aio_mount,
 254                .kill_sb        = kill_anon_super,
 255        };
 256        aio_mnt = kern_mount(&aio_fs);
 257        if (IS_ERR(aio_mnt))
 258                panic("Failed to create aio fs mount.");
 259
 260        kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
 261        kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
 262
 263        pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
 264
 265        return 0;
 266}
 267__initcall(aio_setup);
 268
 269static void put_aio_ring_file(struct kioctx *ctx)
 270{
 271        struct file *aio_ring_file = ctx->aio_ring_file;
 272        if (aio_ring_file) {
 273                truncate_setsize(aio_ring_file->f_inode, 0);
 274
 275                /* Prevent further access to the kioctx from migratepages */
 276                spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
 277                aio_ring_file->f_inode->i_mapping->private_data = NULL;
 278                ctx->aio_ring_file = NULL;
 279                spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
 280
 281                fput(aio_ring_file);
 282        }
 283}
 284
 285static void aio_free_ring(struct kioctx *ctx)
 286{
 287        int i;
 288
 289        /* Disconnect the kiotx from the ring file.  This prevents future
 290         * accesses to the kioctx from page migration.
 291         */
 292        put_aio_ring_file(ctx);
 293
 294        for (i = 0; i < ctx->nr_pages; i++) {
 295                struct page *page;
 296                pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
 297                                page_count(ctx->ring_pages[i]));
 298                page = ctx->ring_pages[i];
 299                if (!page)
 300                        continue;
 301                ctx->ring_pages[i] = NULL;
 302                put_page(page);
 303        }
 304
 305        if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
 306                kfree(ctx->ring_pages);
 307                ctx->ring_pages = NULL;
 308        }
 309}
 310
 311static int aio_ring_mremap(struct vm_area_struct *vma)
 312{
 313        struct file *file = vma->vm_file;
 314        struct mm_struct *mm = vma->vm_mm;
 315        struct kioctx_table *table;
 316        int i, res = -EINVAL;
 317
 318        spin_lock(&mm->ioctx_lock);
 319        rcu_read_lock();
 320        table = rcu_dereference(mm->ioctx_table);
 321        for (i = 0; i < table->nr; i++) {
 322                struct kioctx *ctx;
 323
 324                ctx = table->table[i];
 325                if (ctx && ctx->aio_ring_file == file) {
 326                        if (!atomic_read(&ctx->dead)) {
 327                                ctx->user_id = ctx->mmap_base = vma->vm_start;
 328                                res = 0;
 329                        }
 330                        break;
 331                }
 332        }
 333
 334        rcu_read_unlock();
 335        spin_unlock(&mm->ioctx_lock);
 336        return res;
 337}
 338
 339static const struct vm_operations_struct aio_ring_vm_ops = {
 340        .mremap         = aio_ring_mremap,
 341#if IS_ENABLED(CONFIG_MMU)
 342        .fault          = filemap_fault,
 343        .map_pages      = filemap_map_pages,
 344        .page_mkwrite   = filemap_page_mkwrite,
 345#endif
 346};
 347
 348static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
 349{
 350        vma->vm_flags |= VM_DONTEXPAND;
 351        vma->vm_ops = &aio_ring_vm_ops;
 352        return 0;
 353}
 354
 355static const struct file_operations aio_ring_fops = {
 356        .mmap = aio_ring_mmap,
 357};
 358
 359#if IS_ENABLED(CONFIG_MIGRATION)
 360static int aio_migratepage(struct address_space *mapping, struct page *new,
 361                        struct page *old, enum migrate_mode mode)
 362{
 363        struct kioctx *ctx;
 364        unsigned long flags;
 365        pgoff_t idx;
 366        int rc;
 367
 368        rc = 0;
 369
 370        /* mapping->private_lock here protects against the kioctx teardown.  */
 371        spin_lock(&mapping->private_lock);
 372        ctx = mapping->private_data;
 373        if (!ctx) {
 374                rc = -EINVAL;
 375                goto out;
 376        }
 377
 378        /* The ring_lock mutex.  The prevents aio_read_events() from writing
 379         * to the ring's head, and prevents page migration from mucking in
 380         * a partially initialized kiotx.
 381         */
 382        if (!mutex_trylock(&ctx->ring_lock)) {
 383                rc = -EAGAIN;
 384                goto out;
 385        }
 386
 387        idx = old->index;
 388        if (idx < (pgoff_t)ctx->nr_pages) {
 389                /* Make sure the old page hasn't already been changed */
 390                if (ctx->ring_pages[idx] != old)
 391                        rc = -EAGAIN;
 392        } else
 393                rc = -EINVAL;
 394
 395        if (rc != 0)
 396                goto out_unlock;
 397
 398        /* Writeback must be complete */
 399        BUG_ON(PageWriteback(old));
 400        get_page(new);
 401
 402        rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
 403        if (rc != MIGRATEPAGE_SUCCESS) {
 404                put_page(new);
 405                goto out_unlock;
 406        }
 407
 408        /* Take completion_lock to prevent other writes to the ring buffer
 409         * while the old page is copied to the new.  This prevents new
 410         * events from being lost.
 411         */
 412        spin_lock_irqsave(&ctx->completion_lock, flags);
 413        migrate_page_copy(new, old);
 414        BUG_ON(ctx->ring_pages[idx] != old);
 415        ctx->ring_pages[idx] = new;
 416        spin_unlock_irqrestore(&ctx->completion_lock, flags);
 417
 418        /* The old page is no longer accessible. */
 419        put_page(old);
 420
 421out_unlock:
 422        mutex_unlock(&ctx->ring_lock);
 423out:
 424        spin_unlock(&mapping->private_lock);
 425        return rc;
 426}
 427#endif
 428
 429static const struct address_space_operations aio_ctx_aops = {
 430        .set_page_dirty = __set_page_dirty_no_writeback,
 431#if IS_ENABLED(CONFIG_MIGRATION)
 432        .migratepage    = aio_migratepage,
 433#endif
 434};
 435
 436static int aio_setup_ring(struct kioctx *ctx)
 437{
 438        struct aio_ring *ring;
 439        unsigned nr_events = ctx->max_reqs;
 440        struct mm_struct *mm = current->mm;
 441        unsigned long size, unused;
 442        int nr_pages;
 443        int i;
 444        struct file *file;
 445
 446        /* Compensate for the ring buffer's head/tail overlap entry */
 447        nr_events += 2; /* 1 is required, 2 for good luck */
 448
 449        size = sizeof(struct aio_ring);
 450        size += sizeof(struct io_event) * nr_events;
 451
 452        nr_pages = PFN_UP(size);
 453        if (nr_pages < 0)
 454                return -EINVAL;
 455
 456        file = aio_private_file(ctx, nr_pages);
 457        if (IS_ERR(file)) {
 458                ctx->aio_ring_file = NULL;
 459                return -ENOMEM;
 460        }
 461
 462        ctx->aio_ring_file = file;
 463        nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
 464                        / sizeof(struct io_event);
 465
 466        ctx->ring_pages = ctx->internal_pages;
 467        if (nr_pages > AIO_RING_PAGES) {
 468                ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
 469                                          GFP_KERNEL);
 470                if (!ctx->ring_pages) {
 471                        put_aio_ring_file(ctx);
 472                        return -ENOMEM;
 473                }
 474        }
 475
 476        for (i = 0; i < nr_pages; i++) {
 477                struct page *page;
 478                page = find_or_create_page(file->f_inode->i_mapping,
 479                                           i, GFP_HIGHUSER | __GFP_ZERO);
 480                if (!page)
 481                        break;
 482                pr_debug("pid(%d) page[%d]->count=%d\n",
 483                         current->pid, i, page_count(page));
 484                SetPageUptodate(page);
 485                unlock_page(page);
 486
 487                ctx->ring_pages[i] = page;
 488        }
 489        ctx->nr_pages = i;
 490
 491        if (unlikely(i != nr_pages)) {
 492                aio_free_ring(ctx);
 493                return -ENOMEM;
 494        }
 495
 496        ctx->mmap_size = nr_pages * PAGE_SIZE;
 497        pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
 498
 499        down_write(&mm->mmap_sem);
 500        ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
 501                                       PROT_READ | PROT_WRITE,
 502                                       MAP_SHARED, 0, &unused);
 503        up_write(&mm->mmap_sem);
 504        if (IS_ERR((void *)ctx->mmap_base)) {
 505                ctx->mmap_size = 0;
 506                aio_free_ring(ctx);
 507                return -ENOMEM;
 508        }
 509
 510        pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
 511
 512        ctx->user_id = ctx->mmap_base;
 513        ctx->nr_events = nr_events; /* trusted copy */
 514
 515        ring = kmap_atomic(ctx->ring_pages[0]);
 516        ring->nr = nr_events;   /* user copy */
 517        ring->id = ~0U;
 518        ring->head = ring->tail = 0;
 519        ring->magic = AIO_RING_MAGIC;
 520        ring->compat_features = AIO_RING_COMPAT_FEATURES;
 521        ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
 522        ring->header_length = sizeof(struct aio_ring);
 523        kunmap_atomic(ring);
 524        flush_dcache_page(ctx->ring_pages[0]);
 525
 526        return 0;
 527}
 528
 529#define AIO_EVENTS_PER_PAGE     (PAGE_SIZE / sizeof(struct io_event))
 530#define AIO_EVENTS_FIRST_PAGE   ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
 531#define AIO_EVENTS_OFFSET       (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
 532
 533void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
 534{
 535        struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common);
 536        struct kioctx *ctx = req->ki_ctx;
 537        unsigned long flags;
 538
 539        spin_lock_irqsave(&ctx->ctx_lock, flags);
 540
 541        if (!req->ki_list.next)
 542                list_add(&req->ki_list, &ctx->active_reqs);
 543
 544        req->ki_cancel = cancel;
 545
 546        spin_unlock_irqrestore(&ctx->ctx_lock, flags);
 547}
 548EXPORT_SYMBOL(kiocb_set_cancel_fn);
 549
 550static int kiocb_cancel(struct aio_kiocb *kiocb)
 551{
 552        kiocb_cancel_fn *old, *cancel;
 553
 554        /*
 555         * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
 556         * actually has a cancel function, hence the cmpxchg()
 557         */
 558
 559        cancel = ACCESS_ONCE(kiocb->ki_cancel);
 560        do {
 561                if (!cancel || cancel == KIOCB_CANCELLED)
 562                        return -EINVAL;
 563
 564                old = cancel;
 565                cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
 566        } while (cancel != old);
 567
 568        return cancel(&kiocb->common);
 569}
 570
 571static void free_ioctx(struct work_struct *work)
 572{
 573        struct kioctx *ctx = container_of(work, struct kioctx, free_work);
 574
 575        pr_debug("freeing %p\n", ctx);
 576
 577        aio_free_ring(ctx);
 578        free_percpu(ctx->cpu);
 579        percpu_ref_exit(&ctx->reqs);
 580        percpu_ref_exit(&ctx->users);
 581        kmem_cache_free(kioctx_cachep, ctx);
 582}
 583
 584static void free_ioctx_reqs(struct percpu_ref *ref)
 585{
 586        struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
 587
 588        /* At this point we know that there are no any in-flight requests */
 589        if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
 590                complete(&ctx->rq_wait->comp);
 591
 592        INIT_WORK(&ctx->free_work, free_ioctx);
 593        schedule_work(&ctx->free_work);
 594}
 595
 596/*
 597 * When this function runs, the kioctx has been removed from the "hash table"
 598 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
 599 * now it's safe to cancel any that need to be.
 600 */
 601static void free_ioctx_users(struct percpu_ref *ref)
 602{
 603        struct kioctx *ctx = container_of(ref, struct kioctx, users);
 604        struct aio_kiocb *req;
 605
 606        spin_lock_irq(&ctx->ctx_lock);
 607
 608        while (!list_empty(&ctx->active_reqs)) {
 609                req = list_first_entry(&ctx->active_reqs,
 610                                       struct aio_kiocb, ki_list);
 611
 612                list_del_init(&req->ki_list);
 613                kiocb_cancel(req);
 614        }
 615
 616        spin_unlock_irq(&ctx->ctx_lock);
 617
 618        percpu_ref_kill(&ctx->reqs);
 619        percpu_ref_put(&ctx->reqs);
 620}
 621
 622static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
 623{
 624        unsigned i, new_nr;
 625        struct kioctx_table *table, *old;
 626        struct aio_ring *ring;
 627
 628        spin_lock(&mm->ioctx_lock);
 629        table = rcu_dereference_raw(mm->ioctx_table);
 630
 631        while (1) {
 632                if (table)
 633                        for (i = 0; i < table->nr; i++)
 634                                if (!table->table[i]) {
 635                                        ctx->id = i;
 636                                        table->table[i] = ctx;
 637                                        spin_unlock(&mm->ioctx_lock);
 638
 639                                        /* While kioctx setup is in progress,
 640                                         * we are protected from page migration
 641                                         * changes ring_pages by ->ring_lock.
 642                                         */
 643                                        ring = kmap_atomic(ctx->ring_pages[0]);
 644                                        ring->id = ctx->id;
 645                                        kunmap_atomic(ring);
 646                                        return 0;
 647                                }
 648
 649                new_nr = (table ? table->nr : 1) * 4;
 650                spin_unlock(&mm->ioctx_lock);
 651
 652                table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
 653                                new_nr, GFP_KERNEL);
 654                if (!table)
 655                        return -ENOMEM;
 656
 657                table->nr = new_nr;
 658
 659                spin_lock(&mm->ioctx_lock);
 660                old = rcu_dereference_raw(mm->ioctx_table);
 661
 662                if (!old) {
 663                        rcu_assign_pointer(mm->ioctx_table, table);
 664                } else if (table->nr > old->nr) {
 665                        memcpy(table->table, old->table,
 666                               old->nr * sizeof(struct kioctx *));
 667
 668                        rcu_assign_pointer(mm->ioctx_table, table);
 669                        kfree_rcu(old, rcu);
 670                } else {
 671                        kfree(table);
 672                        table = old;
 673                }
 674        }
 675}
 676
 677static void aio_nr_sub(unsigned nr)
 678{
 679        spin_lock(&aio_nr_lock);
 680        if (WARN_ON(aio_nr - nr > aio_nr))
 681                aio_nr = 0;
 682        else
 683                aio_nr -= nr;
 684        spin_unlock(&aio_nr_lock);
 685}
 686
 687/* ioctx_alloc
 688 *      Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
 689 */
 690static struct kioctx *ioctx_alloc(unsigned nr_events)
 691{
 692        struct mm_struct *mm = current->mm;
 693        struct kioctx *ctx;
 694        int err = -ENOMEM;
 695
 696        /*
 697         * We keep track of the number of available ringbuffer slots, to prevent
 698         * overflow (reqs_available), and we also use percpu counters for this.
 699         *
 700         * So since up to half the slots might be on other cpu's percpu counters
 701         * and unavailable, double nr_events so userspace sees what they
 702         * expected: additionally, we move req_batch slots to/from percpu
 703         * counters at a time, so make sure that isn't 0:
 704         */
 705        nr_events = max(nr_events, num_possible_cpus() * 4);
 706        nr_events *= 2;
 707
 708        /* Prevent overflows */
 709        if (nr_events > (0x10000000U / sizeof(struct io_event))) {
 710                pr_debug("ENOMEM: nr_events too high\n");
 711                return ERR_PTR(-EINVAL);
 712        }
 713
 714        if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
 715                return ERR_PTR(-EAGAIN);
 716
 717        ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
 718        if (!ctx)
 719                return ERR_PTR(-ENOMEM);
 720
 721        ctx->max_reqs = nr_events;
 722
 723        spin_lock_init(&ctx->ctx_lock);
 724        spin_lock_init(&ctx->completion_lock);
 725        mutex_init(&ctx->ring_lock);
 726        /* Protect against page migration throughout kiotx setup by keeping
 727         * the ring_lock mutex held until setup is complete. */
 728        mutex_lock(&ctx->ring_lock);
 729        init_waitqueue_head(&ctx->wait);
 730
 731        INIT_LIST_HEAD(&ctx->active_reqs);
 732
 733        if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
 734                goto err;
 735
 736        if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
 737                goto err;
 738
 739        ctx->cpu = alloc_percpu(struct kioctx_cpu);
 740        if (!ctx->cpu)
 741                goto err;
 742
 743        err = aio_setup_ring(ctx);
 744        if (err < 0)
 745                goto err;
 746
 747        atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
 748        ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
 749        if (ctx->req_batch < 1)
 750                ctx->req_batch = 1;
 751
 752        /* limit the number of system wide aios */
 753        spin_lock(&aio_nr_lock);
 754        if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
 755            aio_nr + nr_events < aio_nr) {
 756                spin_unlock(&aio_nr_lock);
 757                err = -EAGAIN;
 758                goto err_ctx;
 759        }
 760        aio_nr += ctx->max_reqs;
 761        spin_unlock(&aio_nr_lock);
 762
 763        percpu_ref_get(&ctx->users);    /* io_setup() will drop this ref */
 764        percpu_ref_get(&ctx->reqs);     /* free_ioctx_users() will drop this */
 765
 766        err = ioctx_add_table(ctx, mm);
 767        if (err)
 768                goto err_cleanup;
 769
 770        /* Release the ring_lock mutex now that all setup is complete. */
 771        mutex_unlock(&ctx->ring_lock);
 772
 773        pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
 774                 ctx, ctx->user_id, mm, ctx->nr_events);
 775        return ctx;
 776
 777err_cleanup:
 778        aio_nr_sub(ctx->max_reqs);
 779err_ctx:
 780        atomic_set(&ctx->dead, 1);
 781        if (ctx->mmap_size)
 782                vm_munmap(ctx->mmap_base, ctx->mmap_size);
 783        aio_free_ring(ctx);
 784err:
 785        mutex_unlock(&ctx->ring_lock);
 786        free_percpu(ctx->cpu);
 787        percpu_ref_exit(&ctx->reqs);
 788        percpu_ref_exit(&ctx->users);
 789        kmem_cache_free(kioctx_cachep, ctx);
 790        pr_debug("error allocating ioctx %d\n", err);
 791        return ERR_PTR(err);
 792}
 793
 794/* kill_ioctx
 795 *      Cancels all outstanding aio requests on an aio context.  Used
 796 *      when the processes owning a context have all exited to encourage
 797 *      the rapid destruction of the kioctx.
 798 */
 799static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
 800                      struct ctx_rq_wait *wait)
 801{
 802        struct kioctx_table *table;
 803
 804        spin_lock(&mm->ioctx_lock);
 805        if (atomic_xchg(&ctx->dead, 1)) {
 806                spin_unlock(&mm->ioctx_lock);
 807                return -EINVAL;
 808        }
 809
 810        table = rcu_dereference_raw(mm->ioctx_table);
 811        WARN_ON(ctx != table->table[ctx->id]);
 812        table->table[ctx->id] = NULL;
 813        spin_unlock(&mm->ioctx_lock);
 814
 815        /* percpu_ref_kill() will do the necessary call_rcu() */
 816        wake_up_all(&ctx->wait);
 817
 818        /*
 819         * It'd be more correct to do this in free_ioctx(), after all
 820         * the outstanding kiocbs have finished - but by then io_destroy
 821         * has already returned, so io_setup() could potentially return
 822         * -EAGAIN with no ioctxs actually in use (as far as userspace
 823         *  could tell).
 824         */
 825        aio_nr_sub(ctx->max_reqs);
 826
 827        if (ctx->mmap_size)
 828                vm_munmap(ctx->mmap_base, ctx->mmap_size);
 829
 830        ctx->rq_wait = wait;
 831        percpu_ref_kill(&ctx->users);
 832        return 0;
 833}
 834
 835/*
 836 * exit_aio: called when the last user of mm goes away.  At this point, there is
 837 * no way for any new requests to be submited or any of the io_* syscalls to be
 838 * called on the context.
 839 *
 840 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
 841 * them.
 842 */
 843void exit_aio(struct mm_struct *mm)
 844{
 845        struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
 846        struct ctx_rq_wait wait;
 847        int i, skipped;
 848
 849        if (!table)
 850                return;
 851
 852        atomic_set(&wait.count, table->nr);
 853        init_completion(&wait.comp);
 854
 855        skipped = 0;
 856        for (i = 0; i < table->nr; ++i) {
 857                struct kioctx *ctx = table->table[i];
 858
 859                if (!ctx) {
 860                        skipped++;
 861                        continue;
 862                }
 863
 864                /*
 865                 * We don't need to bother with munmap() here - exit_mmap(mm)
 866                 * is coming and it'll unmap everything. And we simply can't,
 867                 * this is not necessarily our ->mm.
 868                 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
 869                 * that it needs to unmap the area, just set it to 0.
 870                 */
 871                ctx->mmap_size = 0;
 872                kill_ioctx(mm, ctx, &wait);
 873        }
 874
 875        if (!atomic_sub_and_test(skipped, &wait.count)) {
 876                /* Wait until all IO for the context are done. */
 877                wait_for_completion(&wait.comp);
 878        }
 879
 880        RCU_INIT_POINTER(mm->ioctx_table, NULL);
 881        kfree(table);
 882}
 883
 884static void put_reqs_available(struct kioctx *ctx, unsigned nr)
 885{
 886        struct kioctx_cpu *kcpu;
 887        unsigned long flags;
 888
 889        local_irq_save(flags);
 890        kcpu = this_cpu_ptr(ctx->cpu);
 891        kcpu->reqs_available += nr;
 892
 893        while (kcpu->reqs_available >= ctx->req_batch * 2) {
 894                kcpu->reqs_available -= ctx->req_batch;
 895                atomic_add(ctx->req_batch, &ctx->reqs_available);
 896        }
 897
 898        local_irq_restore(flags);
 899}
 900
 901static bool get_reqs_available(struct kioctx *ctx)
 902{
 903        struct kioctx_cpu *kcpu;
 904        bool ret = false;
 905        unsigned long flags;
 906
 907        local_irq_save(flags);
 908        kcpu = this_cpu_ptr(ctx->cpu);
 909        if (!kcpu->reqs_available) {
 910                int old, avail = atomic_read(&ctx->reqs_available);
 911
 912                do {
 913                        if (avail < ctx->req_batch)
 914                                goto out;
 915
 916                        old = avail;
 917                        avail = atomic_cmpxchg(&ctx->reqs_available,
 918                                               avail, avail - ctx->req_batch);
 919                } while (avail != old);
 920
 921                kcpu->reqs_available += ctx->req_batch;
 922        }
 923
 924        ret = true;
 925        kcpu->reqs_available--;
 926out:
 927        local_irq_restore(flags);
 928        return ret;
 929}
 930
 931/* refill_reqs_available
 932 *      Updates the reqs_available reference counts used for tracking the
 933 *      number of free slots in the completion ring.  This can be called
 934 *      from aio_complete() (to optimistically update reqs_available) or
 935 *      from aio_get_req() (the we're out of events case).  It must be
 936 *      called holding ctx->completion_lock.
 937 */
 938static void refill_reqs_available(struct kioctx *ctx, unsigned head,
 939                                  unsigned tail)
 940{
 941        unsigned events_in_ring, completed;
 942
 943        /* Clamp head since userland can write to it. */
 944        head %= ctx->nr_events;
 945        if (head <= tail)
 946                events_in_ring = tail - head;
 947        else
 948                events_in_ring = ctx->nr_events - (head - tail);
 949
 950        completed = ctx->completed_events;
 951        if (events_in_ring < completed)
 952                completed -= events_in_ring;
 953        else
 954                completed = 0;
 955
 956        if (!completed)
 957                return;
 958
 959        ctx->completed_events -= completed;
 960        put_reqs_available(ctx, completed);
 961}
 962
 963/* user_refill_reqs_available
 964 *      Called to refill reqs_available when aio_get_req() encounters an
 965 *      out of space in the completion ring.
 966 */
 967static void user_refill_reqs_available(struct kioctx *ctx)
 968{
 969        spin_lock_irq(&ctx->completion_lock);
 970        if (ctx->completed_events) {
 971                struct aio_ring *ring;
 972                unsigned head;
 973
 974                /* Access of ring->head may race with aio_read_events_ring()
 975                 * here, but that's okay since whether we read the old version
 976                 * or the new version, and either will be valid.  The important
 977                 * part is that head cannot pass tail since we prevent
 978                 * aio_complete() from updating tail by holding
 979                 * ctx->completion_lock.  Even if head is invalid, the check
 980                 * against ctx->completed_events below will make sure we do the
 981                 * safe/right thing.
 982                 */
 983                ring = kmap_atomic(ctx->ring_pages[0]);
 984                head = ring->head;
 985                kunmap_atomic(ring);
 986
 987                refill_reqs_available(ctx, head, ctx->tail);
 988        }
 989
 990        spin_unlock_irq(&ctx->completion_lock);
 991}
 992
 993/* aio_get_req
 994 *      Allocate a slot for an aio request.
 995 * Returns NULL if no requests are free.
 996 */
 997static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
 998{
 999        struct aio_kiocb *req;
1000
1001        if (!get_reqs_available(ctx)) {
1002                user_refill_reqs_available(ctx);
1003                if (!get_reqs_available(ctx))
1004                        return NULL;
1005        }
1006
1007        req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
1008        if (unlikely(!req))
1009                goto out_put;
1010
1011        percpu_ref_get(&ctx->reqs);
1012
1013        req->ki_ctx = ctx;
1014        return req;
1015out_put:
1016        put_reqs_available(ctx, 1);
1017        return NULL;
1018}
1019
1020static void kiocb_free(struct aio_kiocb *req)
1021{
1022        if (req->common.ki_filp)
1023                fput(req->common.ki_filp);
1024        if (req->ki_eventfd != NULL)
1025                eventfd_ctx_put(req->ki_eventfd);
1026        kmem_cache_free(kiocb_cachep, req);
1027}
1028
1029static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1030{
1031        struct aio_ring __user *ring  = (void __user *)ctx_id;
1032        struct mm_struct *mm = current->mm;
1033        struct kioctx *ctx, *ret = NULL;
1034        struct kioctx_table *table;
1035        unsigned id;
1036
1037        if (get_user(id, &ring->id))
1038                return NULL;
1039
1040        rcu_read_lock();
1041        table = rcu_dereference(mm->ioctx_table);
1042
1043        if (!table || id >= table->nr)
1044                goto out;
1045
1046        ctx = table->table[id];
1047        if (ctx && ctx->user_id == ctx_id) {
1048                percpu_ref_get(&ctx->users);
1049                ret = ctx;
1050        }
1051out:
1052        rcu_read_unlock();
1053        return ret;
1054}
1055
1056/* aio_complete
1057 *      Called when the io request on the given iocb is complete.
1058 */
1059static void aio_complete(struct kiocb *kiocb, long res, long res2)
1060{
1061        struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common);
1062        struct kioctx   *ctx = iocb->ki_ctx;
1063        struct aio_ring *ring;
1064        struct io_event *ev_page, *event;
1065        unsigned tail, pos, head;
1066        unsigned long   flags;
1067
1068        /*
1069         * Special case handling for sync iocbs:
1070         *  - events go directly into the iocb for fast handling
1071         *  - the sync task with the iocb in its stack holds the single iocb
1072         *    ref, no other paths have a way to get another ref
1073         *  - the sync task helpfully left a reference to itself in the iocb
1074         */
1075        BUG_ON(is_sync_kiocb(kiocb));
1076
1077        if (iocb->ki_list.next) {
1078                unsigned long flags;
1079
1080                spin_lock_irqsave(&ctx->ctx_lock, flags);
1081                list_del(&iocb->ki_list);
1082                spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1083        }
1084
1085        /*
1086         * Add a completion event to the ring buffer. Must be done holding
1087         * ctx->completion_lock to prevent other code from messing with the tail
1088         * pointer since we might be called from irq context.
1089         */
1090        spin_lock_irqsave(&ctx->completion_lock, flags);
1091
1092        tail = ctx->tail;
1093        pos = tail + AIO_EVENTS_OFFSET;
1094
1095        if (++tail >= ctx->nr_events)
1096                tail = 0;
1097
1098        ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1099        event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1100
1101        event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
1102        event->data = iocb->ki_user_data;
1103        event->res = res;
1104        event->res2 = res2;
1105
1106        kunmap_atomic(ev_page);
1107        flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1108
1109        pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1110                 ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
1111                 res, res2);
1112
1113        /* after flagging the request as done, we
1114         * must never even look at it again
1115         */
1116        smp_wmb();      /* make event visible before updating tail */
1117
1118        ctx->tail = tail;
1119
1120        ring = kmap_atomic(ctx->ring_pages[0]);
1121        head = ring->head;
1122        ring->tail = tail;
1123        kunmap_atomic(ring);
1124        flush_dcache_page(ctx->ring_pages[0]);
1125
1126        ctx->completed_events++;
1127        if (ctx->completed_events > 1)
1128                refill_reqs_available(ctx, head, tail);
1129        spin_unlock_irqrestore(&ctx->completion_lock, flags);
1130
1131        pr_debug("added to ring %p at [%u]\n", iocb, tail);
1132
1133        /*
1134         * Check if the user asked us to deliver the result through an
1135         * eventfd. The eventfd_signal() function is safe to be called
1136         * from IRQ context.
1137         */
1138        if (iocb->ki_eventfd != NULL)
1139                eventfd_signal(iocb->ki_eventfd, 1);
1140
1141        /* everything turned out well, dispose of the aiocb. */
1142        kiocb_free(iocb);
1143
1144        /*
1145         * We have to order our ring_info tail store above and test
1146         * of the wait list below outside the wait lock.  This is
1147         * like in wake_up_bit() where clearing a bit has to be
1148         * ordered with the unlocked test.
1149         */
1150        smp_mb();
1151
1152        if (waitqueue_active(&ctx->wait))
1153                wake_up(&ctx->wait);
1154
1155        percpu_ref_put(&ctx->reqs);
1156}
1157
1158/* aio_read_events_ring
1159 *      Pull an event off of the ioctx's event ring.  Returns the number of
1160 *      events fetched
1161 */
1162static long aio_read_events_ring(struct kioctx *ctx,
1163                                 struct io_event __user *event, long nr)
1164{
1165        struct aio_ring *ring;
1166        unsigned head, tail, pos;
1167        long ret = 0;
1168        int copy_ret;
1169
1170        /*
1171         * The mutex can block and wake us up and that will cause
1172         * wait_event_interruptible_hrtimeout() to schedule without sleeping
1173         * and repeat. This should be rare enough that it doesn't cause
1174         * peformance issues. See the comment in read_events() for more detail.
1175         */
1176        sched_annotate_sleep();
1177        mutex_lock(&ctx->ring_lock);
1178
1179        /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1180        ring = kmap_atomic(ctx->ring_pages[0]);
1181        head = ring->head;
1182        tail = ring->tail;
1183        kunmap_atomic(ring);
1184
1185        /*
1186         * Ensure that once we've read the current tail pointer, that
1187         * we also see the events that were stored up to the tail.
1188         */
1189        smp_rmb();
1190
1191        pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1192
1193        if (head == tail)
1194                goto out;
1195
1196        head %= ctx->nr_events;
1197        tail %= ctx->nr_events;
1198
1199        while (ret < nr) {
1200                long avail;
1201                struct io_event *ev;
1202                struct page *page;
1203
1204                avail = (head <= tail ?  tail : ctx->nr_events) - head;
1205                if (head == tail)
1206                        break;
1207
1208                avail = min(avail, nr - ret);
1209                avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1210                            ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1211
1212                pos = head + AIO_EVENTS_OFFSET;
1213                page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1214                pos %= AIO_EVENTS_PER_PAGE;
1215
1216                ev = kmap(page);
1217                copy_ret = copy_to_user(event + ret, ev + pos,
1218                                        sizeof(*ev) * avail);
1219                kunmap(page);
1220
1221                if (unlikely(copy_ret)) {
1222                        ret = -EFAULT;
1223                        goto out;
1224                }
1225
1226                ret += avail;
1227                head += avail;
1228                head %= ctx->nr_events;
1229        }
1230
1231        ring = kmap_atomic(ctx->ring_pages[0]);
1232        ring->head = head;
1233        kunmap_atomic(ring);
1234        flush_dcache_page(ctx->ring_pages[0]);
1235
1236        pr_debug("%li  h%u t%u\n", ret, head, tail);
1237out:
1238        mutex_unlock(&ctx->ring_lock);
1239
1240        return ret;
1241}
1242
1243static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1244                            struct io_event __user *event, long *i)
1245{
1246        long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1247
1248        if (ret > 0)
1249                *i += ret;
1250
1251        if (unlikely(atomic_read(&ctx->dead)))
1252                ret = -EINVAL;
1253
1254        if (!*i)
1255                *i = ret;
1256
1257        return ret < 0 || *i >= min_nr;
1258}
1259
1260static long read_events(struct kioctx *ctx, long min_nr, long nr,
1261                        struct io_event __user *event,
1262                        struct timespec __user *timeout)
1263{
1264        ktime_t until = { .tv64 = KTIME_MAX };
1265        long ret = 0;
1266
1267        if (timeout) {
1268                struct timespec ts;
1269
1270                if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1271                        return -EFAULT;
1272
1273                until = timespec_to_ktime(ts);
1274        }
1275
1276        /*
1277         * Note that aio_read_events() is being called as the conditional - i.e.
1278         * we're calling it after prepare_to_wait() has set task state to
1279         * TASK_INTERRUPTIBLE.
1280         *
1281         * But aio_read_events() can block, and if it blocks it's going to flip
1282         * the task state back to TASK_RUNNING.
1283         *
1284         * This should be ok, provided it doesn't flip the state back to
1285         * TASK_RUNNING and return 0 too much - that causes us to spin. That
1286         * will only happen if the mutex_lock() call blocks, and we then find
1287         * the ringbuffer empty. So in practice we should be ok, but it's
1288         * something to be aware of when touching this code.
1289         */
1290        if (until.tv64 == 0)
1291                aio_read_events(ctx, min_nr, nr, event, &ret);
1292        else
1293                wait_event_interruptible_hrtimeout(ctx->wait,
1294                                aio_read_events(ctx, min_nr, nr, event, &ret),
1295                                until);
1296
1297        if (!ret && signal_pending(current))
1298                ret = -EINTR;
1299
1300        return ret;
1301}
1302
1303/* sys_io_setup:
1304 *      Create an aio_context capable of receiving at least nr_events.
1305 *      ctxp must not point to an aio_context that already exists, and
1306 *      must be initialized to 0 prior to the call.  On successful
1307 *      creation of the aio_context, *ctxp is filled in with the resulting 
1308 *      handle.  May fail with -EINVAL if *ctxp is not initialized,
1309 *      if the specified nr_events exceeds internal limits.  May fail 
1310 *      with -EAGAIN if the specified nr_events exceeds the user's limit 
1311 *      of available events.  May fail with -ENOMEM if insufficient kernel
1312 *      resources are available.  May fail with -EFAULT if an invalid
1313 *      pointer is passed for ctxp.  Will fail with -ENOSYS if not
1314 *      implemented.
1315 */
1316SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1317{
1318        struct kioctx *ioctx = NULL;
1319        unsigned long ctx;
1320        long ret;
1321
1322        ret = get_user(ctx, ctxp);
1323        if (unlikely(ret))
1324                goto out;
1325
1326        ret = -EINVAL;
1327        if (unlikely(ctx || nr_events == 0)) {
1328                pr_debug("EINVAL: ctx %lu nr_events %u\n",
1329                         ctx, nr_events);
1330                goto out;
1331        }
1332
1333        ioctx = ioctx_alloc(nr_events);
1334        ret = PTR_ERR(ioctx);
1335        if (!IS_ERR(ioctx)) {
1336                ret = put_user(ioctx->user_id, ctxp);
1337                if (ret)
1338                        kill_ioctx(current->mm, ioctx, NULL);
1339                percpu_ref_put(&ioctx->users);
1340        }
1341
1342out:
1343        return ret;
1344}
1345
1346/* sys_io_destroy:
1347 *      Destroy the aio_context specified.  May cancel any outstanding 
1348 *      AIOs and block on completion.  Will fail with -ENOSYS if not
1349 *      implemented.  May fail with -EINVAL if the context pointed to
1350 *      is invalid.
1351 */
1352SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1353{
1354        struct kioctx *ioctx = lookup_ioctx(ctx);
1355        if (likely(NULL != ioctx)) {
1356                struct ctx_rq_wait wait;
1357                int ret;
1358
1359                init_completion(&wait.comp);
1360                atomic_set(&wait.count, 1);
1361
1362                /* Pass requests_done to kill_ioctx() where it can be set
1363                 * in a thread-safe way. If we try to set it here then we have
1364                 * a race condition if two io_destroy() called simultaneously.
1365                 */
1366                ret = kill_ioctx(current->mm, ioctx, &wait);
1367                percpu_ref_put(&ioctx->users);
1368
1369                /* Wait until all IO for the context are done. Otherwise kernel
1370                 * keep using user-space buffers even if user thinks the context
1371                 * is destroyed.
1372                 */
1373                if (!ret)
1374                        wait_for_completion(&wait.comp);
1375
1376                return ret;
1377        }
1378        pr_debug("EINVAL: invalid context id\n");
1379        return -EINVAL;
1380}
1381
1382typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
1383
1384static int aio_setup_vectored_rw(int rw, char __user *buf, size_t len,
1385                                 struct iovec **iovec,
1386                                 bool compat,
1387                                 struct iov_iter *iter)
1388{
1389#ifdef CONFIG_COMPAT
1390        if (compat)
1391                return compat_import_iovec(rw,
1392                                (struct compat_iovec __user *)buf,
1393                                len, UIO_FASTIOV, iovec, iter);
1394#endif
1395        return import_iovec(rw, (struct iovec __user *)buf,
1396                                len, UIO_FASTIOV, iovec, iter);
1397}
1398
1399/*
1400 * aio_run_iocb:
1401 *      Performs the initial checks and io submission.
1402 */
1403static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1404                            char __user *buf, size_t len, bool compat)
1405{
1406        struct file *file = req->ki_filp;
1407        ssize_t ret;
1408        int rw;
1409        fmode_t mode;
1410        rw_iter_op *iter_op;
1411        struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1412        struct iov_iter iter;
1413
1414        switch (opcode) {
1415        case IOCB_CMD_PREAD:
1416        case IOCB_CMD_PREADV:
1417                mode    = FMODE_READ;
1418                rw      = READ;
1419                iter_op = file->f_op->read_iter;
1420                goto rw_common;
1421
1422        case IOCB_CMD_PWRITE:
1423        case IOCB_CMD_PWRITEV:
1424                mode    = FMODE_WRITE;
1425                rw      = WRITE;
1426                iter_op = file->f_op->write_iter;
1427                goto rw_common;
1428rw_common:
1429                if (unlikely(!(file->f_mode & mode)))
1430                        return -EBADF;
1431
1432                if (!iter_op)
1433                        return -EINVAL;
1434
1435                if (opcode == IOCB_CMD_PREADV || opcode == IOCB_CMD_PWRITEV)
1436                        ret = aio_setup_vectored_rw(rw, buf, len,
1437                                                &iovec, compat, &iter);
1438                else {
1439                        ret = import_single_range(rw, buf, len, iovec, &iter);
1440                        iovec = NULL;
1441                }
1442                if (!ret)
1443                        ret = rw_verify_area(rw, file, &req->ki_pos,
1444                                             iov_iter_count(&iter));
1445                if (ret < 0) {
1446                        kfree(iovec);
1447                        return ret;
1448                }
1449
1450                len = ret;
1451
1452                if (rw == WRITE)
1453                        file_start_write(file);
1454
1455                ret = iter_op(req, &iter);
1456
1457                if (rw == WRITE)
1458                        file_end_write(file);
1459                kfree(iovec);
1460                break;
1461
1462        case IOCB_CMD_FDSYNC:
1463                if (!file->f_op->aio_fsync)
1464                        return -EINVAL;
1465
1466                ret = file->f_op->aio_fsync(req, 1);
1467                break;
1468
1469        case IOCB_CMD_FSYNC:
1470                if (!file->f_op->aio_fsync)
1471                        return -EINVAL;
1472
1473                ret = file->f_op->aio_fsync(req, 0);
1474                break;
1475
1476        default:
1477                pr_debug("EINVAL: no operation provided\n");
1478                return -EINVAL;
1479        }
1480
1481        if (ret != -EIOCBQUEUED) {
1482                /*
1483                 * There's no easy way to restart the syscall since other AIO's
1484                 * may be already running. Just fail this IO with EINTR.
1485                 */
1486                if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1487                             ret == -ERESTARTNOHAND ||
1488                             ret == -ERESTART_RESTARTBLOCK))
1489                        ret = -EINTR;
1490                aio_complete(req, ret, 0);
1491        }
1492
1493        return 0;
1494}
1495
1496static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1497                         struct iocb *iocb, bool compat)
1498{
1499        struct aio_kiocb *req;
1500        ssize_t ret;
1501
1502        /* enforce forwards compatibility on users */
1503        if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1504                pr_debug("EINVAL: reserve field set\n");
1505                return -EINVAL;
1506        }
1507
1508        /* prevent overflows */
1509        if (unlikely(
1510            (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1511            (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1512            ((ssize_t)iocb->aio_nbytes < 0)
1513           )) {
1514                pr_debug("EINVAL: overflow check\n");
1515                return -EINVAL;
1516        }
1517
1518        req = aio_get_req(ctx);
1519        if (unlikely(!req))
1520                return -EAGAIN;
1521
1522        req->common.ki_filp = fget(iocb->aio_fildes);
1523        if (unlikely(!req->common.ki_filp)) {
1524                ret = -EBADF;
1525                goto out_put_req;
1526        }
1527        req->common.ki_pos = iocb->aio_offset;
1528        req->common.ki_complete = aio_complete;
1529        req->common.ki_flags = iocb_flags(req->common.ki_filp);
1530
1531        if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1532                /*
1533                 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1534                 * instance of the file* now. The file descriptor must be
1535                 * an eventfd() fd, and will be signaled for each completed
1536                 * event using the eventfd_signal() function.
1537                 */
1538                req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1539                if (IS_ERR(req->ki_eventfd)) {
1540                        ret = PTR_ERR(req->ki_eventfd);
1541                        req->ki_eventfd = NULL;
1542                        goto out_put_req;
1543                }
1544
1545                req->common.ki_flags |= IOCB_EVENTFD;
1546        }
1547
1548        ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1549        if (unlikely(ret)) {
1550                pr_debug("EFAULT: aio_key\n");
1551                goto out_put_req;
1552        }
1553
1554        req->ki_user_iocb = user_iocb;
1555        req->ki_user_data = iocb->aio_data;
1556
1557        ret = aio_run_iocb(&req->common, iocb->aio_lio_opcode,
1558                           (char __user *)(unsigned long)iocb->aio_buf,
1559                           iocb->aio_nbytes,
1560                           compat);
1561        if (ret)
1562                goto out_put_req;
1563
1564        return 0;
1565out_put_req:
1566        put_reqs_available(ctx, 1);
1567        percpu_ref_put(&ctx->reqs);
1568        kiocb_free(req);
1569        return ret;
1570}
1571
1572long do_io_submit(aio_context_t ctx_id, long nr,
1573                  struct iocb __user *__user *iocbpp, bool compat)
1574{
1575        struct kioctx *ctx;
1576        long ret = 0;
1577        int i = 0;
1578        struct blk_plug plug;
1579
1580        if (unlikely(nr < 0))
1581                return -EINVAL;
1582
1583        if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1584                nr = LONG_MAX/sizeof(*iocbpp);
1585
1586        if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1587                return -EFAULT;
1588
1589        ctx = lookup_ioctx(ctx_id);
1590        if (unlikely(!ctx)) {
1591                pr_debug("EINVAL: invalid context id\n");
1592                return -EINVAL;
1593        }
1594
1595        blk_start_plug(&plug);
1596
1597        /*
1598         * AKPM: should this return a partial result if some of the IOs were
1599         * successfully submitted?
1600         */
1601        for (i=0; i<nr; i++) {
1602                struct iocb __user *user_iocb;
1603                struct iocb tmp;
1604
1605                if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1606                        ret = -EFAULT;
1607                        break;
1608                }
1609
1610                if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1611                        ret = -EFAULT;
1612                        break;
1613                }
1614
1615                ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1616                if (ret)
1617                        break;
1618        }
1619        blk_finish_plug(&plug);
1620
1621        percpu_ref_put(&ctx->users);
1622        return i ? i : ret;
1623}
1624
1625/* sys_io_submit:
1626 *      Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1627 *      the number of iocbs queued.  May return -EINVAL if the aio_context
1628 *      specified by ctx_id is invalid, if nr is < 0, if the iocb at
1629 *      *iocbpp[0] is not properly initialized, if the operation specified
1630 *      is invalid for the file descriptor in the iocb.  May fail with
1631 *      -EFAULT if any of the data structures point to invalid data.  May
1632 *      fail with -EBADF if the file descriptor specified in the first
1633 *      iocb is invalid.  May fail with -EAGAIN if insufficient resources
1634 *      are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1635 *      fail with -ENOSYS if not implemented.
1636 */
1637SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1638                struct iocb __user * __user *, iocbpp)
1639{
1640        return do_io_submit(ctx_id, nr, iocbpp, 0);
1641}
1642
1643/* lookup_kiocb
1644 *      Finds a given iocb for cancellation.
1645 */
1646static struct aio_kiocb *
1647lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key)
1648{
1649        struct aio_kiocb *kiocb;
1650
1651        assert_spin_locked(&ctx->ctx_lock);
1652
1653        if (key != KIOCB_KEY)
1654                return NULL;
1655
1656        /* TODO: use a hash or array, this sucks. */
1657        list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
1658                if (kiocb->ki_user_iocb == iocb)
1659                        return kiocb;
1660        }
1661        return NULL;
1662}
1663
1664/* sys_io_cancel:
1665 *      Attempts to cancel an iocb previously passed to io_submit.  If
1666 *      the operation is successfully cancelled, the resulting event is
1667 *      copied into the memory pointed to by result without being placed
1668 *      into the completion queue and 0 is returned.  May fail with
1669 *      -EFAULT if any of the data structures pointed to are invalid.
1670 *      May fail with -EINVAL if aio_context specified by ctx_id is
1671 *      invalid.  May fail with -EAGAIN if the iocb specified was not
1672 *      cancelled.  Will fail with -ENOSYS if not implemented.
1673 */
1674SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1675                struct io_event __user *, result)
1676{
1677        struct kioctx *ctx;
1678        struct aio_kiocb *kiocb;
1679        u32 key;
1680        int ret;
1681
1682        ret = get_user(key, &iocb->aio_key);
1683        if (unlikely(ret))
1684                return -EFAULT;
1685
1686        ctx = lookup_ioctx(ctx_id);
1687        if (unlikely(!ctx))
1688                return -EINVAL;
1689
1690        spin_lock_irq(&ctx->ctx_lock);
1691
1692        kiocb = lookup_kiocb(ctx, iocb, key);
1693        if (kiocb)
1694                ret = kiocb_cancel(kiocb);
1695        else
1696                ret = -EINVAL;
1697
1698        spin_unlock_irq(&ctx->ctx_lock);
1699
1700        if (!ret) {
1701                /*
1702                 * The result argument is no longer used - the io_event is
1703                 * always delivered via the ring buffer. -EINPROGRESS indicates
1704                 * cancellation is progress:
1705                 */
1706                ret = -EINPROGRESS;
1707        }
1708
1709        percpu_ref_put(&ctx->users);
1710
1711        return ret;
1712}
1713
1714/* io_getevents:
1715 *      Attempts to read at least min_nr events and up to nr events from
1716 *      the completion queue for the aio_context specified by ctx_id. If
1717 *      it succeeds, the number of read events is returned. May fail with
1718 *      -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1719 *      out of range, if timeout is out of range.  May fail with -EFAULT
1720 *      if any of the memory specified is invalid.  May return 0 or
1721 *      < min_nr if the timeout specified by timeout has elapsed
1722 *      before sufficient events are available, where timeout == NULL
1723 *      specifies an infinite timeout. Note that the timeout pointed to by
1724 *      timeout is relative.  Will fail with -ENOSYS if not implemented.
1725 */
1726SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1727                long, min_nr,
1728                long, nr,
1729                struct io_event __user *, events,
1730                struct timespec __user *, timeout)
1731{
1732        struct kioctx *ioctx = lookup_ioctx(ctx_id);
1733        long ret = -EINVAL;
1734
1735        if (likely(ioctx)) {
1736                if (likely(min_nr <= nr && min_nr >= 0))
1737                        ret = read_events(ioctx, min_nr, nr, events, timeout);
1738                percpu_ref_put(&ioctx->users);
1739        }
1740        return ret;
1741}
1742