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/anon_inodes.h>
  40#include <linux/migrate.h>
  41#include <linux/ramfs.h>
  42#include <linux/percpu-refcount.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;
  56        unsigned        tail;
  57
  58        unsigned        magic;
  59        unsigned        compat_features;
  60        unsigned        incompat_features;
  61        unsigned        header_length;  /* size of aio_ring */
  62
  63
  64        struct io_event         io_events[0];
  65}; /* 128 bytes + ring size */
  66
  67#define AIO_RING_PAGES  8
  68
  69struct kioctx_table {
  70        struct rcu_head rcu;
  71        unsigned        nr;
  72        struct kioctx   *table[];
  73};
  74
  75struct kioctx_cpu {
  76        unsigned                reqs_available;
  77};
  78
  79struct kioctx {
  80        struct percpu_ref       users;
  81        atomic_t                dead;
  82
  83        unsigned long           user_id;
  84
  85        struct __percpu kioctx_cpu *cpu;
  86
  87        /*
  88         * For percpu reqs_available, number of slots we move to/from global
  89         * counter at a time:
  90         */
  91        unsigned                req_batch;
  92        /*
  93         * This is what userspace passed to io_setup(), it's not used for
  94         * anything but counting against the global max_reqs quota.
  95         *
  96         * The real limit is nr_events - 1, which will be larger (see
  97         * aio_setup_ring())
  98         */
  99        unsigned                max_reqs;
 100
 101        /* Size of ringbuffer, in units of struct io_event */
 102        unsigned                nr_events;
 103
 104        unsigned long           mmap_base;
 105        unsigned long           mmap_size;
 106
 107        struct page             **ring_pages;
 108        long                    nr_pages;
 109
 110        struct rcu_head         rcu_head;
 111        struct work_struct      free_work;
 112
 113        struct {
 114                /*
 115                 * This counts the number of available slots in the ringbuffer,
 116                 * so we avoid overflowing it: it's decremented (if positive)
 117                 * when allocating a kiocb and incremented when the resulting
 118                 * io_event is pulled off the ringbuffer.
 119                 *
 120                 * We batch accesses to it with a percpu version.
 121                 */
 122                atomic_t        reqs_available;
 123        } ____cacheline_aligned_in_smp;
 124
 125        struct {
 126                spinlock_t      ctx_lock;
 127                struct list_head active_reqs;   /* used for cancellation */
 128        } ____cacheline_aligned_in_smp;
 129
 130        struct {
 131                struct mutex    ring_lock;
 132                wait_queue_head_t wait;
 133        } ____cacheline_aligned_in_smp;
 134
 135        struct {
 136                unsigned        tail;
 137                spinlock_t      completion_lock;
 138        } ____cacheline_aligned_in_smp;
 139
 140        struct page             *internal_pages[AIO_RING_PAGES];
 141        struct file             *aio_ring_file;
 142
 143        unsigned                id;
 144};
 145
 146/*------ sysctl variables----*/
 147static DEFINE_SPINLOCK(aio_nr_lock);
 148unsigned long aio_nr;           /* current system wide number of aio requests */
 149unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
 150/*----end sysctl variables---*/
 151
 152static struct kmem_cache        *kiocb_cachep;
 153static struct kmem_cache        *kioctx_cachep;
 154
 155/* aio_setup
 156 *      Creates the slab caches used by the aio routines, panic on
 157 *      failure as this is done early during the boot sequence.
 158 */
 159static int __init aio_setup(void)
 160{
 161        kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
 162        kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
 163
 164        pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
 165
 166        return 0;
 167}
 168__initcall(aio_setup);
 169
 170static void put_aio_ring_file(struct kioctx *ctx)
 171{
 172        struct file *aio_ring_file = ctx->aio_ring_file;
 173        if (aio_ring_file) {
 174                truncate_setsize(aio_ring_file->f_inode, 0);
 175
 176                /* Prevent further access to the kioctx from migratepages */
 177                spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
 178                aio_ring_file->f_inode->i_mapping->private_data = NULL;
 179                ctx->aio_ring_file = NULL;
 180                spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
 181
 182                fput(aio_ring_file);
 183        }
 184}
 185
 186static void aio_free_ring(struct kioctx *ctx)
 187{
 188        int i;
 189
 190        for (i = 0; i < ctx->nr_pages; i++) {
 191                pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
 192                                page_count(ctx->ring_pages[i]));
 193                put_page(ctx->ring_pages[i]);
 194        }
 195
 196        put_aio_ring_file(ctx);
 197
 198        if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages)
 199                kfree(ctx->ring_pages);
 200}
 201
 202static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
 203{
 204        vma->vm_ops = &generic_file_vm_ops;
 205        return 0;
 206}
 207
 208static const struct file_operations aio_ring_fops = {
 209        .mmap = aio_ring_mmap,
 210};
 211
 212static int aio_set_page_dirty(struct page *page)
 213{
 214        return 0;
 215}
 216
 217#if IS_ENABLED(CONFIG_MIGRATION)
 218static int aio_migratepage(struct address_space *mapping, struct page *new,
 219                        struct page *old, enum migrate_mode mode)
 220{
 221        struct kioctx *ctx;
 222        unsigned long flags;
 223        int rc;
 224
 225        /* Writeback must be complete */
 226        BUG_ON(PageWriteback(old));
 227        put_page(old);
 228
 229        rc = migrate_page_move_mapping(mapping, new, old, NULL, mode);
 230        if (rc != MIGRATEPAGE_SUCCESS) {
 231                get_page(old);
 232                return rc;
 233        }
 234
 235        get_page(new);
 236
 237        /* We can potentially race against kioctx teardown here.  Use the
 238         * address_space's private data lock to protect the mapping's
 239         * private_data.
 240         */
 241        spin_lock(&mapping->private_lock);
 242        ctx = mapping->private_data;
 243        if (ctx) {
 244                pgoff_t idx;
 245                spin_lock_irqsave(&ctx->completion_lock, flags);
 246                migrate_page_copy(new, old);
 247                idx = old->index;
 248                if (idx < (pgoff_t)ctx->nr_pages)
 249                        ctx->ring_pages[idx] = new;
 250                spin_unlock_irqrestore(&ctx->completion_lock, flags);
 251        } else
 252                rc = -EBUSY;
 253        spin_unlock(&mapping->private_lock);
 254
 255        return rc;
 256}
 257#endif
 258
 259static const struct address_space_operations aio_ctx_aops = {
 260        .set_page_dirty = aio_set_page_dirty,
 261#if IS_ENABLED(CONFIG_MIGRATION)
 262        .migratepage    = aio_migratepage,
 263#endif
 264};
 265
 266static int aio_setup_ring(struct kioctx *ctx)
 267{
 268        struct aio_ring *ring;
 269        unsigned nr_events = ctx->max_reqs;
 270        struct mm_struct *mm = current->mm;
 271        unsigned long size, populate;
 272        int nr_pages;
 273        int i;
 274        struct file *file;
 275
 276        /* Compensate for the ring buffer's head/tail overlap entry */
 277        nr_events += 2; /* 1 is required, 2 for good luck */
 278
 279        size = sizeof(struct aio_ring);
 280        size += sizeof(struct io_event) * nr_events;
 281
 282        nr_pages = PFN_UP(size);
 283        if (nr_pages < 0)
 284                return -EINVAL;
 285
 286        file = anon_inode_getfile_private("[aio]", &aio_ring_fops, ctx, O_RDWR);
 287        if (IS_ERR(file)) {
 288                ctx->aio_ring_file = NULL;
 289                return -EAGAIN;
 290        }
 291
 292        file->f_inode->i_mapping->a_ops = &aio_ctx_aops;
 293        file->f_inode->i_mapping->private_data = ctx;
 294        file->f_inode->i_size = PAGE_SIZE * (loff_t)nr_pages;
 295
 296        for (i = 0; i < nr_pages; i++) {
 297                struct page *page;
 298                page = find_or_create_page(file->f_inode->i_mapping,
 299                                           i, GFP_HIGHUSER | __GFP_ZERO);
 300                if (!page)
 301                        break;
 302                pr_debug("pid(%d) page[%d]->count=%d\n",
 303                         current->pid, i, page_count(page));
 304                SetPageUptodate(page);
 305                SetPageDirty(page);
 306                unlock_page(page);
 307        }
 308        ctx->aio_ring_file = file;
 309        nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
 310                        / sizeof(struct io_event);
 311
 312        ctx->ring_pages = ctx->internal_pages;
 313        if (nr_pages > AIO_RING_PAGES) {
 314                ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
 315                                          GFP_KERNEL);
 316                if (!ctx->ring_pages)
 317                        return -ENOMEM;
 318        }
 319
 320        ctx->mmap_size = nr_pages * PAGE_SIZE;
 321        pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
 322
 323        down_write(&mm->mmap_sem);
 324        ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
 325                                       PROT_READ | PROT_WRITE,
 326                                       MAP_SHARED | MAP_POPULATE, 0, &populate);
 327        if (IS_ERR((void *)ctx->mmap_base)) {
 328                up_write(&mm->mmap_sem);
 329                ctx->mmap_size = 0;
 330                aio_free_ring(ctx);
 331                return -EAGAIN;
 332        }
 333
 334        pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
 335
 336        /* We must do this while still holding mmap_sem for write, as we
 337         * need to be protected against userspace attempting to mremap()
 338         * or munmap() the ring buffer.
 339         */
 340        ctx->nr_pages = get_user_pages(current, mm, ctx->mmap_base, nr_pages,
 341                                       1, 0, ctx->ring_pages, NULL);
 342
 343        /* Dropping the reference here is safe as the page cache will hold
 344         * onto the pages for us.  It is also required so that page migration
 345         * can unmap the pages and get the right reference count.
 346         */
 347        for (i = 0; i < ctx->nr_pages; i++)
 348                put_page(ctx->ring_pages[i]);
 349
 350        up_write(&mm->mmap_sem);
 351
 352        if (unlikely(ctx->nr_pages != nr_pages)) {
 353                aio_free_ring(ctx);
 354                return -EAGAIN;
 355        }
 356
 357        ctx->user_id = ctx->mmap_base;
 358        ctx->nr_events = nr_events; /* trusted copy */
 359
 360        ring = kmap_atomic(ctx->ring_pages[0]);
 361        ring->nr = nr_events;   /* user copy */
 362        ring->id = ~0U;
 363        ring->head = ring->tail = 0;
 364        ring->magic = AIO_RING_MAGIC;
 365        ring->compat_features = AIO_RING_COMPAT_FEATURES;
 366        ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
 367        ring->header_length = sizeof(struct aio_ring);
 368        kunmap_atomic(ring);
 369        flush_dcache_page(ctx->ring_pages[0]);
 370
 371        return 0;
 372}
 373
 374#define AIO_EVENTS_PER_PAGE     (PAGE_SIZE / sizeof(struct io_event))
 375#define AIO_EVENTS_FIRST_PAGE   ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
 376#define AIO_EVENTS_OFFSET       (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
 377
 378void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
 379{
 380        struct kioctx *ctx = req->ki_ctx;
 381        unsigned long flags;
 382
 383        spin_lock_irqsave(&ctx->ctx_lock, flags);
 384
 385        if (!req->ki_list.next)
 386                list_add(&req->ki_list, &ctx->active_reqs);
 387
 388        req->ki_cancel = cancel;
 389
 390        spin_unlock_irqrestore(&ctx->ctx_lock, flags);
 391}
 392EXPORT_SYMBOL(kiocb_set_cancel_fn);
 393
 394static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb)
 395{
 396        kiocb_cancel_fn *old, *cancel;
 397
 398        /*
 399         * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
 400         * actually has a cancel function, hence the cmpxchg()
 401         */
 402
 403        cancel = ACCESS_ONCE(kiocb->ki_cancel);
 404        do {
 405                if (!cancel || cancel == KIOCB_CANCELLED)
 406                        return -EINVAL;
 407
 408                old = cancel;
 409                cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
 410        } while (cancel != old);
 411
 412        return cancel(kiocb);
 413}
 414
 415static void free_ioctx_rcu(struct rcu_head *head)
 416{
 417        struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
 418
 419        free_percpu(ctx->cpu);
 420        kmem_cache_free(kioctx_cachep, ctx);
 421}
 422
 423/*
 424 * When this function runs, the kioctx has been removed from the "hash table"
 425 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
 426 * now it's safe to cancel any that need to be.
 427 */
 428static void free_ioctx(struct work_struct *work)
 429{
 430        struct kioctx *ctx = container_of(work, struct kioctx, free_work);
 431        struct aio_ring *ring;
 432        struct kiocb *req;
 433        unsigned cpu, avail;
 434        DEFINE_WAIT(wait);
 435
 436        spin_lock_irq(&ctx->ctx_lock);
 437
 438        while (!list_empty(&ctx->active_reqs)) {
 439                req = list_first_entry(&ctx->active_reqs,
 440                                       struct kiocb, ki_list);
 441
 442                list_del_init(&req->ki_list);
 443                kiocb_cancel(ctx, req);
 444        }
 445
 446        spin_unlock_irq(&ctx->ctx_lock);
 447
 448        for_each_possible_cpu(cpu) {
 449                struct kioctx_cpu *kcpu = per_cpu_ptr(ctx->cpu, cpu);
 450
 451                atomic_add(kcpu->reqs_available, &ctx->reqs_available);
 452                kcpu->reqs_available = 0;
 453        }
 454
 455        while (1) {
 456                prepare_to_wait(&ctx->wait, &wait, TASK_UNINTERRUPTIBLE);
 457
 458                ring = kmap_atomic(ctx->ring_pages[0]);
 459                avail = (ring->head <= ring->tail)
 460                         ? ring->tail - ring->head
 461                         : ctx->nr_events - ring->head + ring->tail;
 462
 463                atomic_add(avail, &ctx->reqs_available);
 464                ring->head = ring->tail;
 465                kunmap_atomic(ring);
 466
 467                if (atomic_read(&ctx->reqs_available) >= ctx->nr_events - 1)
 468                        break;
 469
 470                schedule();
 471        }
 472        finish_wait(&ctx->wait, &wait);
 473
 474        WARN_ON(atomic_read(&ctx->reqs_available) > ctx->nr_events - 1);
 475
 476        aio_free_ring(ctx);
 477
 478        pr_debug("freeing %p\n", ctx);
 479
 480        /*
 481         * Here the call_rcu() is between the wait_event() for reqs_active to
 482         * hit 0, and freeing the ioctx.
 483         *
 484         * aio_complete() decrements reqs_active, but it has to touch the ioctx
 485         * after to issue a wakeup so we use rcu.
 486         */
 487        call_rcu(&ctx->rcu_head, free_ioctx_rcu);
 488}
 489
 490static void free_ioctx_ref(struct percpu_ref *ref)
 491{
 492        struct kioctx *ctx = container_of(ref, struct kioctx, users);
 493
 494        INIT_WORK(&ctx->free_work, free_ioctx);
 495        schedule_work(&ctx->free_work);
 496}
 497
 498static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
 499{
 500        unsigned i, new_nr;
 501        struct kioctx_table *table, *old;
 502        struct aio_ring *ring;
 503
 504        spin_lock(&mm->ioctx_lock);
 505        rcu_read_lock();
 506        table = rcu_dereference(mm->ioctx_table);
 507
 508        while (1) {
 509                if (table)
 510                        for (i = 0; i < table->nr; i++)
 511                                if (!table->table[i]) {
 512                                        ctx->id = i;
 513                                        table->table[i] = ctx;
 514                                        rcu_read_unlock();
 515                                        spin_unlock(&mm->ioctx_lock);
 516
 517                                        ring = kmap_atomic(ctx->ring_pages[0]);
 518                                        ring->id = ctx->id;
 519                                        kunmap_atomic(ring);
 520                                        return 0;
 521                                }
 522
 523                new_nr = (table ? table->nr : 1) * 4;
 524
 525                rcu_read_unlock();
 526                spin_unlock(&mm->ioctx_lock);
 527
 528                table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
 529                                new_nr, GFP_KERNEL);
 530                if (!table)
 531                        return -ENOMEM;
 532
 533                table->nr = new_nr;
 534
 535                spin_lock(&mm->ioctx_lock);
 536                rcu_read_lock();
 537                old = rcu_dereference(mm->ioctx_table);
 538
 539                if (!old) {
 540                        rcu_assign_pointer(mm->ioctx_table, table);
 541                } else if (table->nr > old->nr) {
 542                        memcpy(table->table, old->table,
 543                               old->nr * sizeof(struct kioctx *));
 544
 545                        rcu_assign_pointer(mm->ioctx_table, table);
 546                        kfree_rcu(old, rcu);
 547                } else {
 548                        kfree(table);
 549                        table = old;
 550                }
 551        }
 552}
 553
 554/* ioctx_alloc
 555 *      Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
 556 */
 557static struct kioctx *ioctx_alloc(unsigned nr_events)
 558{
 559        struct mm_struct *mm = current->mm;
 560        struct kioctx *ctx;
 561        int err = -ENOMEM;
 562
 563        /*
 564         * We keep track of the number of available ringbuffer slots, to prevent
 565         * overflow (reqs_available), and we also use percpu counters for this.
 566         *
 567         * So since up to half the slots might be on other cpu's percpu counters
 568         * and unavailable, double nr_events so userspace sees what they
 569         * expected: additionally, we move req_batch slots to/from percpu
 570         * counters at a time, so make sure that isn't 0:
 571         */
 572        nr_events = max(nr_events, num_possible_cpus() * 4);
 573        nr_events *= 2;
 574
 575        /* Prevent overflows */
 576        if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
 577            (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
 578                pr_debug("ENOMEM: nr_events too high\n");
 579                return ERR_PTR(-EINVAL);
 580        }
 581
 582        if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
 583                return ERR_PTR(-EAGAIN);
 584
 585        ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
 586        if (!ctx)
 587                return ERR_PTR(-ENOMEM);
 588
 589        ctx->max_reqs = nr_events;
 590
 591        if (percpu_ref_init(&ctx->users, free_ioctx_ref))
 592                goto out_freectx;
 593
 594        spin_lock_init(&ctx->ctx_lock);
 595        spin_lock_init(&ctx->completion_lock);
 596        mutex_init(&ctx->ring_lock);
 597        init_waitqueue_head(&ctx->wait);
 598
 599        INIT_LIST_HEAD(&ctx->active_reqs);
 600
 601        ctx->cpu = alloc_percpu(struct kioctx_cpu);
 602        if (!ctx->cpu)
 603                goto out_freeref;
 604
 605        if (aio_setup_ring(ctx) < 0)
 606                goto out_freepcpu;
 607
 608        atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
 609        ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
 610        if (ctx->req_batch < 1)
 611                ctx->req_batch = 1;
 612
 613        /* limit the number of system wide aios */
 614        spin_lock(&aio_nr_lock);
 615        if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
 616            aio_nr + nr_events < aio_nr) {
 617                spin_unlock(&aio_nr_lock);
 618                goto out_cleanup;
 619        }
 620        aio_nr += ctx->max_reqs;
 621        spin_unlock(&aio_nr_lock);
 622
 623        percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
 624
 625        err = ioctx_add_table(ctx, mm);
 626        if (err)
 627                goto out_cleanup_put;
 628
 629        pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
 630                 ctx, ctx->user_id, mm, ctx->nr_events);
 631        return ctx;
 632
 633out_cleanup_put:
 634        percpu_ref_put(&ctx->users);
 635out_cleanup:
 636        err = -EAGAIN;
 637        aio_free_ring(ctx);
 638out_freepcpu:
 639        free_percpu(ctx->cpu);
 640out_freeref:
 641        free_percpu(ctx->users.pcpu_count);
 642out_freectx:
 643        put_aio_ring_file(ctx);
 644        kmem_cache_free(kioctx_cachep, ctx);
 645        pr_debug("error allocating ioctx %d\n", err);
 646        return ERR_PTR(err);
 647}
 648
 649/* kill_ioctx
 650 *      Cancels all outstanding aio requests on an aio context.  Used
 651 *      when the processes owning a context have all exited to encourage
 652 *      the rapid destruction of the kioctx.
 653 */
 654static void kill_ioctx(struct mm_struct *mm, struct kioctx *ctx)
 655{
 656        if (!atomic_xchg(&ctx->dead, 1)) {
 657                struct kioctx_table *table;
 658
 659                spin_lock(&mm->ioctx_lock);
 660                rcu_read_lock();
 661                table = rcu_dereference(mm->ioctx_table);
 662
 663                WARN_ON(ctx != table->table[ctx->id]);
 664                table->table[ctx->id] = NULL;
 665                rcu_read_unlock();
 666                spin_unlock(&mm->ioctx_lock);
 667
 668                /* percpu_ref_kill() will do the necessary call_rcu() */
 669                wake_up_all(&ctx->wait);
 670
 671                /*
 672                 * It'd be more correct to do this in free_ioctx(), after all
 673                 * the outstanding kiocbs have finished - but by then io_destroy
 674                 * has already returned, so io_setup() could potentially return
 675                 * -EAGAIN with no ioctxs actually in use (as far as userspace
 676                 *  could tell).
 677                 */
 678                spin_lock(&aio_nr_lock);
 679                BUG_ON(aio_nr - ctx->max_reqs > aio_nr);
 680                aio_nr -= ctx->max_reqs;
 681                spin_unlock(&aio_nr_lock);
 682
 683                if (ctx->mmap_size)
 684                        vm_munmap(ctx->mmap_base, ctx->mmap_size);
 685
 686                percpu_ref_kill(&ctx->users);
 687        }
 688}
 689
 690/* wait_on_sync_kiocb:
 691 *      Waits on the given sync kiocb to complete.
 692 */
 693ssize_t wait_on_sync_kiocb(struct kiocb *req)
 694{
 695        while (!req->ki_ctx) {
 696                set_current_state(TASK_UNINTERRUPTIBLE);
 697                if (req->ki_ctx)
 698                        break;
 699                io_schedule();
 700        }
 701        __set_current_state(TASK_RUNNING);
 702        return req->ki_user_data;
 703}
 704EXPORT_SYMBOL(wait_on_sync_kiocb);
 705
 706/*
 707 * exit_aio: called when the last user of mm goes away.  At this point, there is
 708 * no way for any new requests to be submited or any of the io_* syscalls to be
 709 * called on the context.
 710 *
 711 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
 712 * them.
 713 */
 714void exit_aio(struct mm_struct *mm)
 715{
 716        struct kioctx_table *table;
 717        struct kioctx *ctx;
 718        unsigned i = 0;
 719
 720        while (1) {
 721                rcu_read_lock();
 722                table = rcu_dereference(mm->ioctx_table);
 723
 724                do {
 725                        if (!table || i >= table->nr) {
 726                                rcu_read_unlock();
 727                                rcu_assign_pointer(mm->ioctx_table, NULL);
 728                                if (table)
 729                                        kfree(table);
 730                                return;
 731                        }
 732
 733                        ctx = table->table[i++];
 734                } while (!ctx);
 735
 736                rcu_read_unlock();
 737
 738                /*
 739                 * We don't need to bother with munmap() here -
 740                 * exit_mmap(mm) is coming and it'll unmap everything.
 741                 * Since aio_free_ring() uses non-zero ->mmap_size
 742                 * as indicator that it needs to unmap the area,
 743                 * just set it to 0; aio_free_ring() is the only
 744                 * place that uses ->mmap_size, so it's safe.
 745                 */
 746                ctx->mmap_size = 0;
 747
 748                kill_ioctx(mm, ctx);
 749        }
 750}
 751
 752static void put_reqs_available(struct kioctx *ctx, unsigned nr)
 753{
 754        struct kioctx_cpu *kcpu;
 755
 756        preempt_disable();
 757        kcpu = this_cpu_ptr(ctx->cpu);
 758
 759        kcpu->reqs_available += nr;
 760        while (kcpu->reqs_available >= ctx->req_batch * 2) {
 761                kcpu->reqs_available -= ctx->req_batch;
 762                atomic_add(ctx->req_batch, &ctx->reqs_available);
 763        }
 764
 765        preempt_enable();
 766}
 767
 768static bool get_reqs_available(struct kioctx *ctx)
 769{
 770        struct kioctx_cpu *kcpu;
 771        bool ret = false;
 772
 773        preempt_disable();
 774        kcpu = this_cpu_ptr(ctx->cpu);
 775
 776        if (!kcpu->reqs_available) {
 777                int old, avail = atomic_read(&ctx->reqs_available);
 778
 779                do {
 780                        if (avail < ctx->req_batch)
 781                                goto out;
 782
 783                        old = avail;
 784                        avail = atomic_cmpxchg(&ctx->reqs_available,
 785                                               avail, avail - ctx->req_batch);
 786                } while (avail != old);
 787
 788                kcpu->reqs_available += ctx->req_batch;
 789        }
 790
 791        ret = true;
 792        kcpu->reqs_available--;
 793out:
 794        preempt_enable();
 795        return ret;
 796}
 797
 798/* aio_get_req
 799 *      Allocate a slot for an aio request.
 800 * Returns NULL if no requests are free.
 801 */
 802static inline struct kiocb *aio_get_req(struct kioctx *ctx)
 803{
 804        struct kiocb *req;
 805
 806        if (!get_reqs_available(ctx))
 807                return NULL;
 808
 809        req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
 810        if (unlikely(!req))
 811                goto out_put;
 812
 813        req->ki_ctx = ctx;
 814        return req;
 815out_put:
 816        put_reqs_available(ctx, 1);
 817        return NULL;
 818}
 819
 820static void kiocb_free(struct kiocb *req)
 821{
 822        if (req->ki_filp)
 823                fput(req->ki_filp);
 824        if (req->ki_eventfd != NULL)
 825                eventfd_ctx_put(req->ki_eventfd);
 826        kmem_cache_free(kiocb_cachep, req);
 827}
 828
 829static struct kioctx *lookup_ioctx(unsigned long ctx_id)
 830{
 831        struct aio_ring __user *ring  = (void __user *)ctx_id;
 832        struct mm_struct *mm = current->mm;
 833        struct kioctx *ctx, *ret = NULL;
 834        struct kioctx_table *table;
 835        unsigned id;
 836
 837        if (get_user(id, &ring->id))
 838                return NULL;
 839
 840        rcu_read_lock();
 841        table = rcu_dereference(mm->ioctx_table);
 842
 843        if (!table || id >= table->nr)
 844                goto out;
 845
 846        ctx = table->table[id];
 847        if (ctx && ctx->user_id == ctx_id) {
 848                percpu_ref_get(&ctx->users);
 849                ret = ctx;
 850        }
 851out:
 852        rcu_read_unlock();
 853        return ret;
 854}
 855
 856/* aio_complete
 857 *      Called when the io request on the given iocb is complete.
 858 */
 859void aio_complete(struct kiocb *iocb, long res, long res2)
 860{
 861        struct kioctx   *ctx = iocb->ki_ctx;
 862        struct aio_ring *ring;
 863        struct io_event *ev_page, *event;
 864        unsigned long   flags;
 865        unsigned tail, pos;
 866
 867        /*
 868         * Special case handling for sync iocbs:
 869         *  - events go directly into the iocb for fast handling
 870         *  - the sync task with the iocb in its stack holds the single iocb
 871         *    ref, no other paths have a way to get another ref
 872         *  - the sync task helpfully left a reference to itself in the iocb
 873         */
 874        if (is_sync_kiocb(iocb)) {
 875                iocb->ki_user_data = res;
 876                smp_wmb();
 877                iocb->ki_ctx = ERR_PTR(-EXDEV);
 878                wake_up_process(iocb->ki_obj.tsk);
 879                return;
 880        }
 881
 882        /*
 883         * Take rcu_read_lock() in case the kioctx is being destroyed, as we
 884         * need to issue a wakeup after incrementing reqs_available.
 885         */
 886        rcu_read_lock();
 887
 888        if (iocb->ki_list.next) {
 889                unsigned long flags;
 890
 891                spin_lock_irqsave(&ctx->ctx_lock, flags);
 892                list_del(&iocb->ki_list);
 893                spin_unlock_irqrestore(&ctx->ctx_lock, flags);
 894        }
 895
 896        /*
 897         * Add a completion event to the ring buffer. Must be done holding
 898         * ctx->completion_lock to prevent other code from messing with the tail
 899         * pointer since we might be called from irq context.
 900         */
 901        spin_lock_irqsave(&ctx->completion_lock, flags);
 902
 903        tail = ctx->tail;
 904        pos = tail + AIO_EVENTS_OFFSET;
 905
 906        if (++tail >= ctx->nr_events)
 907                tail = 0;
 908
 909        ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
 910        event = ev_page + pos % AIO_EVENTS_PER_PAGE;
 911
 912        event->obj = (u64)(unsigned long)iocb->ki_obj.user;
 913        event->data = iocb->ki_user_data;
 914        event->res = res;
 915        event->res2 = res2;
 916
 917        kunmap_atomic(ev_page);
 918        flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
 919
 920        pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
 921                 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
 922                 res, res2);
 923
 924        /* after flagging the request as done, we
 925         * must never even look at it again
 926         */
 927        smp_wmb();      /* make event visible before updating tail */
 928
 929        ctx->tail = tail;
 930
 931        ring = kmap_atomic(ctx->ring_pages[0]);
 932        ring->tail = tail;
 933        kunmap_atomic(ring);
 934        flush_dcache_page(ctx->ring_pages[0]);
 935
 936        spin_unlock_irqrestore(&ctx->completion_lock, flags);
 937
 938        pr_debug("added to ring %p at [%u]\n", iocb, tail);
 939
 940        /*
 941         * Check if the user asked us to deliver the result through an
 942         * eventfd. The eventfd_signal() function is safe to be called
 943         * from IRQ context.
 944         */
 945        if (iocb->ki_eventfd != NULL)
 946                eventfd_signal(iocb->ki_eventfd, 1);
 947
 948        /* everything turned out well, dispose of the aiocb. */
 949        kiocb_free(iocb);
 950
 951        /*
 952         * We have to order our ring_info tail store above and test
 953         * of the wait list below outside the wait lock.  This is
 954         * like in wake_up_bit() where clearing a bit has to be
 955         * ordered with the unlocked test.
 956         */
 957        smp_mb();
 958
 959        if (waitqueue_active(&ctx->wait))
 960                wake_up(&ctx->wait);
 961
 962        rcu_read_unlock();
 963}
 964EXPORT_SYMBOL(aio_complete);
 965
 966/* aio_read_events
 967 *      Pull an event off of the ioctx's event ring.  Returns the number of
 968 *      events fetched
 969 */
 970static long aio_read_events_ring(struct kioctx *ctx,
 971                                 struct io_event __user *event, long nr)
 972{
 973        struct aio_ring *ring;
 974        unsigned head, tail, pos;
 975        long ret = 0;
 976        int copy_ret;
 977
 978        mutex_lock(&ctx->ring_lock);
 979
 980        ring = kmap_atomic(ctx->ring_pages[0]);
 981        head = ring->head;
 982        tail = ring->tail;
 983        kunmap_atomic(ring);
 984
 985        pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
 986
 987        if (head == tail)
 988                goto out;
 989
 990        while (ret < nr) {
 991                long avail;
 992                struct io_event *ev;
 993                struct page *page;
 994
 995                avail = (head <= tail ?  tail : ctx->nr_events) - head;
 996                if (head == tail)
 997                        break;
 998
 999                avail = min(avail, nr - ret);
1000                avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1001                            ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1002
1003                pos = head + AIO_EVENTS_OFFSET;
1004                page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1005                pos %= AIO_EVENTS_PER_PAGE;
1006
1007                ev = kmap(page);
1008                copy_ret = copy_to_user(event + ret, ev + pos,
1009                                        sizeof(*ev) * avail);
1010                kunmap(page);
1011
1012                if (unlikely(copy_ret)) {
1013                        ret = -EFAULT;
1014                        goto out;
1015                }
1016
1017                ret += avail;
1018                head += avail;
1019                head %= ctx->nr_events;
1020        }
1021
1022        ring = kmap_atomic(ctx->ring_pages[0]);
1023        ring->head = head;
1024        kunmap_atomic(ring);
1025        flush_dcache_page(ctx->ring_pages[0]);
1026
1027        pr_debug("%li  h%u t%u\n", ret, head, tail);
1028
1029        put_reqs_available(ctx, ret);
1030out:
1031        mutex_unlock(&ctx->ring_lock);
1032
1033        return ret;
1034}
1035
1036static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1037                            struct io_event __user *event, long *i)
1038{
1039        long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1040
1041        if (ret > 0)
1042                *i += ret;
1043
1044        if (unlikely(atomic_read(&ctx->dead)))
1045                ret = -EINVAL;
1046
1047        if (!*i)
1048                *i = ret;
1049
1050        return ret < 0 || *i >= min_nr;
1051}
1052
1053static long read_events(struct kioctx *ctx, long min_nr, long nr,
1054                        struct io_event __user *event,
1055                        struct timespec __user *timeout)
1056{
1057        ktime_t until = { .tv64 = KTIME_MAX };
1058        long ret = 0;
1059
1060        if (timeout) {
1061                struct timespec ts;
1062
1063                if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1064                        return -EFAULT;
1065
1066                until = timespec_to_ktime(ts);
1067        }
1068
1069        /*
1070         * Note that aio_read_events() is being called as the conditional - i.e.
1071         * we're calling it after prepare_to_wait() has set task state to
1072         * TASK_INTERRUPTIBLE.
1073         *
1074         * But aio_read_events() can block, and if it blocks it's going to flip
1075         * the task state back to TASK_RUNNING.
1076         *
1077         * This should be ok, provided it doesn't flip the state back to
1078         * TASK_RUNNING and return 0 too much - that causes us to spin. That
1079         * will only happen if the mutex_lock() call blocks, and we then find
1080         * the ringbuffer empty. So in practice we should be ok, but it's
1081         * something to be aware of when touching this code.
1082         */
1083        wait_event_interruptible_hrtimeout(ctx->wait,
1084                        aio_read_events(ctx, min_nr, nr, event, &ret), until);
1085
1086        if (!ret && signal_pending(current))
1087                ret = -EINTR;
1088
1089        return ret;
1090}
1091
1092/* sys_io_setup:
1093 *      Create an aio_context capable of receiving at least nr_events.
1094 *      ctxp must not point to an aio_context that already exists, and
1095 *      must be initialized to 0 prior to the call.  On successful
1096 *      creation of the aio_context, *ctxp is filled in with the resulting 
1097 *      handle.  May fail with -EINVAL if *ctxp is not initialized,
1098 *      if the specified nr_events exceeds internal limits.  May fail 
1099 *      with -EAGAIN if the specified nr_events exceeds the user's limit 
1100 *      of available events.  May fail with -ENOMEM if insufficient kernel
1101 *      resources are available.  May fail with -EFAULT if an invalid
1102 *      pointer is passed for ctxp.  Will fail with -ENOSYS if not
1103 *      implemented.
1104 */
1105SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1106{
1107        struct kioctx *ioctx = NULL;
1108        unsigned long ctx;
1109        long ret;
1110
1111        ret = get_user(ctx, ctxp);
1112        if (unlikely(ret))
1113                goto out;
1114
1115        ret = -EINVAL;
1116        if (unlikely(ctx || nr_events == 0)) {
1117                pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1118                         ctx, nr_events);
1119                goto out;
1120        }
1121
1122        ioctx = ioctx_alloc(nr_events);
1123        ret = PTR_ERR(ioctx);
1124        if (!IS_ERR(ioctx)) {
1125                ret = put_user(ioctx->user_id, ctxp);
1126                if (ret)
1127                        kill_ioctx(current->mm, ioctx);
1128                percpu_ref_put(&ioctx->users);
1129        }
1130
1131out:
1132        return ret;
1133}
1134
1135/* sys_io_destroy:
1136 *      Destroy the aio_context specified.  May cancel any outstanding 
1137 *      AIOs and block on completion.  Will fail with -ENOSYS if not
1138 *      implemented.  May fail with -EINVAL if the context pointed to
1139 *      is invalid.
1140 */
1141SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1142{
1143        struct kioctx *ioctx = lookup_ioctx(ctx);
1144        if (likely(NULL != ioctx)) {
1145                kill_ioctx(current->mm, ioctx);
1146                percpu_ref_put(&ioctx->users);
1147                return 0;
1148        }
1149        pr_debug("EINVAL: io_destroy: invalid context id\n");
1150        return -EINVAL;
1151}
1152
1153typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1154                            unsigned long, loff_t);
1155
1156static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1157                                     int rw, char __user *buf,
1158                                     unsigned long *nr_segs,
1159                                     struct iovec **iovec,
1160                                     bool compat)
1161{
1162        ssize_t ret;
1163
1164        *nr_segs = kiocb->ki_nbytes;
1165
1166#ifdef CONFIG_COMPAT
1167        if (compat)
1168                ret = compat_rw_copy_check_uvector(rw,
1169                                (struct compat_iovec __user *)buf,
1170                                *nr_segs, 1, *iovec, iovec);
1171        else
1172#endif
1173                ret = rw_copy_check_uvector(rw,
1174                                (struct iovec __user *)buf,
1175                                *nr_segs, 1, *iovec, iovec);
1176        if (ret < 0)
1177                return ret;
1178
1179        /* ki_nbytes now reflect bytes instead of segs */
1180        kiocb->ki_nbytes = ret;
1181        return 0;
1182}
1183
1184static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1185                                       int rw, char __user *buf,
1186                                       unsigned long *nr_segs,
1187                                       struct iovec *iovec)
1188{
1189        if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
1190                return -EFAULT;
1191
1192        iovec->iov_base = buf;
1193        iovec->iov_len = kiocb->ki_nbytes;
1194        *nr_segs = 1;
1195        return 0;
1196}
1197
1198/*
1199 * aio_setup_iocb:
1200 *      Performs the initial checks and aio retry method
1201 *      setup for the kiocb at the time of io submission.
1202 */
1203static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1204                            char __user *buf, bool compat)
1205{
1206        struct file *file = req->ki_filp;
1207        ssize_t ret;
1208        unsigned long nr_segs;
1209        int rw;
1210        fmode_t mode;
1211        aio_rw_op *rw_op;
1212        struct iovec inline_vec, *iovec = &inline_vec;
1213
1214        switch (opcode) {
1215        case IOCB_CMD_PREAD:
1216        case IOCB_CMD_PREADV:
1217                mode    = FMODE_READ;
1218                rw      = READ;
1219                rw_op   = file->f_op->aio_read;
1220                goto rw_common;
1221
1222        case IOCB_CMD_PWRITE:
1223        case IOCB_CMD_PWRITEV:
1224                mode    = FMODE_WRITE;
1225                rw      = WRITE;
1226                rw_op   = file->f_op->aio_write;
1227                goto rw_common;
1228rw_common:
1229                if (unlikely(!(file->f_mode & mode)))
1230                        return -EBADF;
1231
1232                if (!rw_op)
1233                        return -EINVAL;
1234
1235                ret = (opcode == IOCB_CMD_PREADV ||
1236                       opcode == IOCB_CMD_PWRITEV)
1237                        ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1238                                                &iovec, compat)
1239                        : aio_setup_single_vector(req, rw, buf, &nr_segs,
1240                                                  iovec);
1241                if (ret)
1242                        return ret;
1243
1244                ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1245                if (ret < 0) {
1246                        if (iovec != &inline_vec)
1247                                kfree(iovec);
1248                        return ret;
1249                }
1250
1251                req->ki_nbytes = ret;
1252
1253                /* XXX: move/kill - rw_verify_area()? */
1254                /* This matches the pread()/pwrite() logic */
1255                if (req->ki_pos < 0) {
1256                        ret = -EINVAL;
1257                        break;
1258                }
1259
1260                if (rw == WRITE)
1261                        file_start_write(file);
1262
1263                ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1264
1265                if (rw == WRITE)
1266                        file_end_write(file);
1267                break;
1268
1269        case IOCB_CMD_FDSYNC:
1270                if (!file->f_op->aio_fsync)
1271                        return -EINVAL;
1272
1273                ret = file->f_op->aio_fsync(req, 1);
1274                break;
1275
1276        case IOCB_CMD_FSYNC:
1277                if (!file->f_op->aio_fsync)
1278                        return -EINVAL;
1279
1280                ret = file->f_op->aio_fsync(req, 0);
1281                break;
1282
1283        default:
1284                pr_debug("EINVAL: no operation provided\n");
1285                return -EINVAL;
1286        }
1287
1288        if (iovec != &inline_vec)
1289                kfree(iovec);
1290
1291        if (ret != -EIOCBQUEUED) {
1292                /*
1293                 * There's no easy way to restart the syscall since other AIO's
1294                 * may be already running. Just fail this IO with EINTR.
1295                 */
1296                if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1297                             ret == -ERESTARTNOHAND ||
1298                             ret == -ERESTART_RESTARTBLOCK))
1299                        ret = -EINTR;
1300                aio_complete(req, ret, 0);
1301        }
1302
1303        return 0;
1304}
1305
1306static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1307                         struct iocb *iocb, bool compat)
1308{
1309        struct kiocb *req;
1310        ssize_t ret;
1311
1312        /* enforce forwards compatibility on users */
1313        if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1314                pr_debug("EINVAL: reserve field set\n");
1315                return -EINVAL;
1316        }
1317
1318        /* prevent overflows */
1319        if (unlikely(
1320            (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1321            (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1322            ((ssize_t)iocb->aio_nbytes < 0)
1323           )) {
1324                pr_debug("EINVAL: io_submit: overflow check\n");
1325                return -EINVAL;
1326        }
1327
1328        req = aio_get_req(ctx);
1329        if (unlikely(!req))
1330                return -EAGAIN;
1331
1332        req->ki_filp = fget(iocb->aio_fildes);
1333        if (unlikely(!req->ki_filp)) {
1334                ret = -EBADF;
1335                goto out_put_req;
1336        }
1337
1338        if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1339                /*
1340                 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1341                 * instance of the file* now. The file descriptor must be
1342                 * an eventfd() fd, and will be signaled for each completed
1343                 * event using the eventfd_signal() function.
1344                 */
1345                req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1346                if (IS_ERR(req->ki_eventfd)) {
1347                        ret = PTR_ERR(req->ki_eventfd);
1348                        req->ki_eventfd = NULL;
1349                        goto out_put_req;
1350                }
1351        }
1352
1353        ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1354        if (unlikely(ret)) {
1355                pr_debug("EFAULT: aio_key\n");
1356                goto out_put_req;
1357        }
1358
1359        req->ki_obj.user = user_iocb;
1360        req->ki_user_data = iocb->aio_data;
1361        req->ki_pos = iocb->aio_offset;
1362        req->ki_nbytes = iocb->aio_nbytes;
1363
1364        ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1365                           (char __user *)(unsigned long)iocb->aio_buf,
1366                           compat);
1367        if (ret)
1368                goto out_put_req;
1369
1370        return 0;
1371out_put_req:
1372        put_reqs_available(ctx, 1);
1373        kiocb_free(req);
1374        return ret;
1375}
1376
1377long do_io_submit(aio_context_t ctx_id, long nr,
1378                  struct iocb __user *__user *iocbpp, bool compat)
1379{
1380        struct kioctx *ctx;
1381        long ret = 0;
1382        int i = 0;
1383        struct blk_plug plug;
1384
1385        if (unlikely(nr < 0))
1386                return -EINVAL;
1387
1388        if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1389                nr = LONG_MAX/sizeof(*iocbpp);
1390
1391        if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1392                return -EFAULT;
1393
1394        ctx = lookup_ioctx(ctx_id);
1395        if (unlikely(!ctx)) {
1396                pr_debug("EINVAL: invalid context id\n");
1397                return -EINVAL;
1398        }
1399
1400        blk_start_plug(&plug);
1401
1402        /*
1403         * AKPM: should this return a partial result if some of the IOs were
1404         * successfully submitted?
1405         */
1406        for (i=0; i<nr; i++) {
1407                struct iocb __user *user_iocb;
1408                struct iocb tmp;
1409
1410                if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1411                        ret = -EFAULT;
1412                        break;
1413                }
1414
1415                if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1416                        ret = -EFAULT;
1417                        break;
1418                }
1419
1420                ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1421                if (ret)
1422                        break;
1423        }
1424        blk_finish_plug(&plug);
1425
1426        percpu_ref_put(&ctx->users);
1427        return i ? i : ret;
1428}
1429
1430/* sys_io_submit:
1431 *      Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1432 *      the number of iocbs queued.  May return -EINVAL if the aio_context
1433 *      specified by ctx_id is invalid, if nr is < 0, if the iocb at
1434 *      *iocbpp[0] is not properly initialized, if the operation specified
1435 *      is invalid for the file descriptor in the iocb.  May fail with
1436 *      -EFAULT if any of the data structures point to invalid data.  May
1437 *      fail with -EBADF if the file descriptor specified in the first
1438 *      iocb is invalid.  May fail with -EAGAIN if insufficient resources
1439 *      are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1440 *      fail with -ENOSYS if not implemented.
1441 */
1442SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1443                struct iocb __user * __user *, iocbpp)
1444{
1445        return do_io_submit(ctx_id, nr, iocbpp, 0);
1446}
1447
1448/* lookup_kiocb
1449 *      Finds a given iocb for cancellation.
1450 */
1451static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1452                                  u32 key)
1453{
1454        struct list_head *pos;
1455
1456        assert_spin_locked(&ctx->ctx_lock);
1457
1458        if (key != KIOCB_KEY)
1459                return NULL;
1460
1461        /* TODO: use a hash or array, this sucks. */
1462        list_for_each(pos, &ctx->active_reqs) {
1463                struct kiocb *kiocb = list_kiocb(pos);
1464                if (kiocb->ki_obj.user == iocb)
1465                        return kiocb;
1466        }
1467        return NULL;
1468}
1469
1470/* sys_io_cancel:
1471 *      Attempts to cancel an iocb previously passed to io_submit.  If
1472 *      the operation is successfully cancelled, the resulting event is
1473 *      copied into the memory pointed to by result without being placed
1474 *      into the completion queue and 0 is returned.  May fail with
1475 *      -EFAULT if any of the data structures pointed to are invalid.
1476 *      May fail with -EINVAL if aio_context specified by ctx_id is
1477 *      invalid.  May fail with -EAGAIN if the iocb specified was not
1478 *      cancelled.  Will fail with -ENOSYS if not implemented.
1479 */
1480SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1481                struct io_event __user *, result)
1482{
1483        struct kioctx *ctx;
1484        struct kiocb *kiocb;
1485        u32 key;
1486        int ret;
1487
1488        ret = get_user(key, &iocb->aio_key);
1489        if (unlikely(ret))
1490                return -EFAULT;
1491
1492        ctx = lookup_ioctx(ctx_id);
1493        if (unlikely(!ctx))
1494                return -EINVAL;
1495
1496        spin_lock_irq(&ctx->ctx_lock);
1497
1498        kiocb = lookup_kiocb(ctx, iocb, key);
1499        if (kiocb)
1500                ret = kiocb_cancel(ctx, kiocb);
1501        else
1502                ret = -EINVAL;
1503
1504        spin_unlock_irq(&ctx->ctx_lock);
1505
1506        if (!ret) {
1507                /*
1508                 * The result argument is no longer used - the io_event is
1509                 * always delivered via the ring buffer. -EINPROGRESS indicates
1510                 * cancellation is progress:
1511                 */
1512                ret = -EINPROGRESS;
1513        }
1514
1515        percpu_ref_put(&ctx->users);
1516
1517        return ret;
1518}
1519
1520/* io_getevents:
1521 *      Attempts to read at least min_nr events and up to nr events from
1522 *      the completion queue for the aio_context specified by ctx_id. If
1523 *      it succeeds, the number of read events is returned. May fail with
1524 *      -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1525 *      out of range, if timeout is out of range.  May fail with -EFAULT
1526 *      if any of the memory specified is invalid.  May return 0 or
1527 *      < min_nr if the timeout specified by timeout has elapsed
1528 *      before sufficient events are available, where timeout == NULL
1529 *      specifies an infinite timeout. Note that the timeout pointed to by
1530 *      timeout is relative.  Will fail with -ENOSYS if not implemented.
1531 */
1532SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1533                long, min_nr,
1534                long, nr,
1535                struct io_event __user *, events,
1536                struct timespec __user *, timeout)
1537{
1538        struct kioctx *ioctx = lookup_ioctx(ctx_id);
1539        long ret = -EINVAL;
1540
1541        if (likely(ioctx)) {
1542                if (likely(min_nr <= nr && min_nr >= 0))
1543                        ret = read_events(ioctx, min_nr, nr, events, timeout);
1544                percpu_ref_put(&ioctx->users);
1545        }
1546        return ret;
1547}
1548
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.