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#include <linux/kernel.h>
  12#include <linux/init.h>
  13#include <linux/errno.h>
  14#include <linux/time.h>
  15#include <linux/aio_abi.h>
  16#include <linux/export.h>
  17#include <linux/syscalls.h>
  18#include <linux/backing-dev.h>
  19#include <linux/uio.h>
  20
  21#define DEBUG 0
  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/slab.h>
  30#include <linux/timer.h>
  31#include <linux/aio.h>
  32#include <linux/highmem.h>
  33#include <linux/workqueue.h>
  34#include <linux/security.h>
  35#include <linux/eventfd.h>
  36#include <linux/blkdev.h>
  37#include <linux/compat.h>
  38
  39#include <asm/kmap_types.h>
  40#include <asm/uaccess.h>
  41
  42#if DEBUG > 1
  43#define dprintk         printk
  44#else
  45#define dprintk(x...)   do { ; } while (0)
  46#endif
  47
  48/*------ sysctl variables----*/
  49static DEFINE_SPINLOCK(aio_nr_lock);
  50unsigned long aio_nr;           /* current system wide number of aio requests */
  51unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
  52/*----end sysctl variables---*/
  53
  54static struct kmem_cache        *kiocb_cachep;
  55static struct kmem_cache        *kioctx_cachep;
  56
  57static struct workqueue_struct *aio_wq;
  58
  59static void aio_kick_handler(struct work_struct *);
  60static void aio_queue_work(struct kioctx *);
  61
  62/* aio_setup
  63 *      Creates the slab caches used by the aio routines, panic on
  64 *      failure as this is done early during the boot sequence.
  65 */
  66static int __init aio_setup(void)
  67{
  68        kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
  69        kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
  70
  71        aio_wq = alloc_workqueue("aio", 0, 1);  /* used to limit concurrency */
  72        BUG_ON(!aio_wq);
  73
  74        pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
  75
  76        return 0;
  77}
  78__initcall(aio_setup);
  79
  80static void aio_free_ring(struct kioctx *ctx)
  81{
  82        struct aio_ring_info *info = &ctx->ring_info;
  83        long i;
  84
  85        for (i=0; i<info->nr_pages; i++)
  86                put_page(info->ring_pages[i]);
  87
  88        if (info->mmap_size) {
  89                BUG_ON(ctx->mm != current->mm);
  90                vm_munmap(info->mmap_base, info->mmap_size);
  91        }
  92
  93        if (info->ring_pages && info->ring_pages != info->internal_pages)
  94                kfree(info->ring_pages);
  95        info->ring_pages = NULL;
  96        info->nr = 0;
  97}
  98
  99static int aio_setup_ring(struct kioctx *ctx)
 100{
 101        struct aio_ring *ring;
 102        struct aio_ring_info *info = &ctx->ring_info;
 103        unsigned nr_events = ctx->max_reqs;
 104        unsigned long size;
 105        int nr_pages;
 106
 107        /* Compensate for the ring buffer's head/tail overlap entry */
 108        nr_events += 2; /* 1 is required, 2 for good luck */
 109
 110        size = sizeof(struct aio_ring);
 111        size += sizeof(struct io_event) * nr_events;
 112        nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
 113
 114        if (nr_pages < 0)
 115                return -EINVAL;
 116
 117        nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
 118
 119        info->nr = 0;
 120        info->ring_pages = info->internal_pages;
 121        if (nr_pages > AIO_RING_PAGES) {
 122                info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
 123                if (!info->ring_pages)
 124                        return -ENOMEM;
 125        }
 126
 127        info->mmap_size = nr_pages * PAGE_SIZE;
 128        dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
 129        down_write(&ctx->mm->mmap_sem);
 130        info->mmap_base = do_mmap_pgoff(NULL, 0, info->mmap_size, 
 131                                        PROT_READ|PROT_WRITE,
 132                                        MAP_ANONYMOUS|MAP_PRIVATE, 0);
 133        if (IS_ERR((void *)info->mmap_base)) {
 134                up_write(&ctx->mm->mmap_sem);
 135                info->mmap_size = 0;
 136                aio_free_ring(ctx);
 137                return -EAGAIN;
 138        }
 139
 140        dprintk("mmap address: 0x%08lx\n", info->mmap_base);
 141        info->nr_pages = get_user_pages(current, ctx->mm,
 142                                        info->mmap_base, nr_pages, 
 143                                        1, 0, info->ring_pages, NULL);
 144        up_write(&ctx->mm->mmap_sem);
 145
 146        if (unlikely(info->nr_pages != nr_pages)) {
 147                aio_free_ring(ctx);
 148                return -EAGAIN;
 149        }
 150
 151        ctx->user_id = info->mmap_base;
 152
 153        info->nr = nr_events;           /* trusted copy */
 154
 155        ring = kmap_atomic(info->ring_pages[0]);
 156        ring->nr = nr_events;   /* user copy */
 157        ring->id = ctx->user_id;
 158        ring->head = ring->tail = 0;
 159        ring->magic = AIO_RING_MAGIC;
 160        ring->compat_features = AIO_RING_COMPAT_FEATURES;
 161        ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
 162        ring->header_length = sizeof(struct aio_ring);
 163        kunmap_atomic(ring);
 164
 165        return 0;
 166}
 167
 168
 169/* aio_ring_event: returns a pointer to the event at the given index from
 170 * kmap_atomic().  Release the pointer with put_aio_ring_event();
 171 */
 172#define AIO_EVENTS_PER_PAGE     (PAGE_SIZE / sizeof(struct io_event))
 173#define AIO_EVENTS_FIRST_PAGE   ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
 174#define AIO_EVENTS_OFFSET       (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
 175
 176#define aio_ring_event(info, nr) ({                                     \
 177        unsigned pos = (nr) + AIO_EVENTS_OFFSET;                        \
 178        struct io_event *__event;                                       \
 179        __event = kmap_atomic(                                          \
 180                        (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE]); \
 181        __event += pos % AIO_EVENTS_PER_PAGE;                           \
 182        __event;                                                        \
 183})
 184
 185#define put_aio_ring_event(event) do {          \
 186        struct io_event *__event = (event);     \
 187        (void)__event;                          \
 188        kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK)); \
 189} while(0)
 190
 191static void ctx_rcu_free(struct rcu_head *head)
 192{
 193        struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
 194        kmem_cache_free(kioctx_cachep, ctx);
 195}
 196
 197/* __put_ioctx
 198 *      Called when the last user of an aio context has gone away,
 199 *      and the struct needs to be freed.
 200 */
 201static void __put_ioctx(struct kioctx *ctx)
 202{
 203        unsigned nr_events = ctx->max_reqs;
 204        BUG_ON(ctx->reqs_active);
 205
 206        cancel_delayed_work_sync(&ctx->wq);
 207        aio_free_ring(ctx);
 208        mmdrop(ctx->mm);
 209        ctx->mm = NULL;
 210        if (nr_events) {
 211                spin_lock(&aio_nr_lock);
 212                BUG_ON(aio_nr - nr_events > aio_nr);
 213                aio_nr -= nr_events;
 214                spin_unlock(&aio_nr_lock);
 215        }
 216        pr_debug("__put_ioctx: freeing %p\n", ctx);
 217        call_rcu(&ctx->rcu_head, ctx_rcu_free);
 218}
 219
 220static inline int try_get_ioctx(struct kioctx *kioctx)
 221{
 222        return atomic_inc_not_zero(&kioctx->users);
 223}
 224
 225static inline void put_ioctx(struct kioctx *kioctx)
 226{
 227        BUG_ON(atomic_read(&kioctx->users) <= 0);
 228        if (unlikely(atomic_dec_and_test(&kioctx->users)))
 229                __put_ioctx(kioctx);
 230}
 231
 232/* ioctx_alloc
 233 *      Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
 234 */
 235static struct kioctx *ioctx_alloc(unsigned nr_events)
 236{
 237        struct mm_struct *mm;
 238        struct kioctx *ctx;
 239        int err = -ENOMEM;
 240
 241        /* Prevent overflows */
 242        if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
 243            (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
 244                pr_debug("ENOMEM: nr_events too high\n");
 245                return ERR_PTR(-EINVAL);
 246        }
 247
 248        if (!nr_events || (unsigned long)nr_events > aio_max_nr)
 249                return ERR_PTR(-EAGAIN);
 250
 251        ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
 252        if (!ctx)
 253                return ERR_PTR(-ENOMEM);
 254
 255        ctx->max_reqs = nr_events;
 256        mm = ctx->mm = current->mm;
 257        atomic_inc(&mm->mm_count);
 258
 259        atomic_set(&ctx->users, 2);
 260        spin_lock_init(&ctx->ctx_lock);
 261        spin_lock_init(&ctx->ring_info.ring_lock);
 262        init_waitqueue_head(&ctx->wait);
 263
 264        INIT_LIST_HEAD(&ctx->active_reqs);
 265        INIT_LIST_HEAD(&ctx->run_list);
 266        INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
 267
 268        if (aio_setup_ring(ctx) < 0)
 269                goto out_freectx;
 270
 271        /* limit the number of system wide aios */
 272        spin_lock(&aio_nr_lock);
 273        if (aio_nr + nr_events > aio_max_nr ||
 274            aio_nr + nr_events < aio_nr) {
 275                spin_unlock(&aio_nr_lock);
 276                goto out_cleanup;
 277        }
 278        aio_nr += ctx->max_reqs;
 279        spin_unlock(&aio_nr_lock);
 280
 281        /* now link into global list. */
 282        spin_lock(&mm->ioctx_lock);
 283        hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
 284        spin_unlock(&mm->ioctx_lock);
 285
 286        dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
 287                ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
 288        return ctx;
 289
 290out_cleanup:
 291        err = -EAGAIN;
 292        aio_free_ring(ctx);
 293out_freectx:
 294        mmdrop(mm);
 295        kmem_cache_free(kioctx_cachep, ctx);
 296        dprintk("aio: error allocating ioctx %d\n", err);
 297        return ERR_PTR(err);
 298}
 299
 300/* kill_ctx
 301 *      Cancels all outstanding aio requests on an aio context.  Used 
 302 *      when the processes owning a context have all exited to encourage 
 303 *      the rapid destruction of the kioctx.
 304 */
 305static void kill_ctx(struct kioctx *ctx)
 306{
 307        int (*cancel)(struct kiocb *, struct io_event *);
 308        struct task_struct *tsk = current;
 309        DECLARE_WAITQUEUE(wait, tsk);
 310        struct io_event res;
 311
 312        spin_lock_irq(&ctx->ctx_lock);
 313        ctx->dead = 1;
 314        while (!list_empty(&ctx->active_reqs)) {
 315                struct list_head *pos = ctx->active_reqs.next;
 316                struct kiocb *iocb = list_kiocb(pos);
 317                list_del_init(&iocb->ki_list);
 318                cancel = iocb->ki_cancel;
 319                kiocbSetCancelled(iocb);
 320                if (cancel) {
 321                        iocb->ki_users++;
 322                        spin_unlock_irq(&ctx->ctx_lock);
 323                        cancel(iocb, &res);
 324                        spin_lock_irq(&ctx->ctx_lock);
 325                }
 326        }
 327
 328        if (!ctx->reqs_active)
 329                goto out;
 330
 331        add_wait_queue(&ctx->wait, &wait);
 332        set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 333        while (ctx->reqs_active) {
 334                spin_unlock_irq(&ctx->ctx_lock);
 335                io_schedule();
 336                set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 337                spin_lock_irq(&ctx->ctx_lock);
 338        }
 339        __set_task_state(tsk, TASK_RUNNING);
 340        remove_wait_queue(&ctx->wait, &wait);
 341
 342out:
 343        spin_unlock_irq(&ctx->ctx_lock);
 344}
 345
 346/* wait_on_sync_kiocb:
 347 *      Waits on the given sync kiocb to complete.
 348 */
 349ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
 350{
 351        while (iocb->ki_users) {
 352                set_current_state(TASK_UNINTERRUPTIBLE);
 353                if (!iocb->ki_users)
 354                        break;
 355                io_schedule();
 356        }
 357        __set_current_state(TASK_RUNNING);
 358        return iocb->ki_user_data;
 359}
 360EXPORT_SYMBOL(wait_on_sync_kiocb);
 361
 362/* exit_aio: called when the last user of mm goes away.  At this point, 
 363 * there is no way for any new requests to be submited or any of the 
 364 * io_* syscalls to be called on the context.  However, there may be 
 365 * outstanding requests which hold references to the context; as they 
 366 * go away, they will call put_ioctx and release any pinned memory
 367 * associated with the request (held via struct page * references).
 368 */
 369void exit_aio(struct mm_struct *mm)
 370{
 371        struct kioctx *ctx;
 372
 373        while (!hlist_empty(&mm->ioctx_list)) {
 374                ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
 375                hlist_del_rcu(&ctx->list);
 376
 377                kill_ctx(ctx);
 378
 379                if (1 != atomic_read(&ctx->users))
 380                        printk(KERN_DEBUG
 381                                "exit_aio:ioctx still alive: %d %d %d\n",
 382                                atomic_read(&ctx->users), ctx->dead,
 383                                ctx->reqs_active);
 384                /*
 385                 * We don't need to bother with munmap() here -
 386                 * exit_mmap(mm) is coming and it'll unmap everything.
 387                 * Since aio_free_ring() uses non-zero ->mmap_size
 388                 * as indicator that it needs to unmap the area,
 389                 * just set it to 0; aio_free_ring() is the only
 390                 * place that uses ->mmap_size, so it's safe.
 391                 * That way we get all munmap done to current->mm -
 392                 * all other callers have ctx->mm == current->mm.
 393                 */
 394                ctx->ring_info.mmap_size = 0;
 395                put_ioctx(ctx);
 396        }
 397}
 398
 399/* aio_get_req
 400 *      Allocate a slot for an aio request.  Increments the users count
 401 * of the kioctx so that the kioctx stays around until all requests are
 402 * complete.  Returns NULL if no requests are free.
 403 *
 404 * Returns with kiocb->users set to 2.  The io submit code path holds
 405 * an extra reference while submitting the i/o.
 406 * This prevents races between the aio code path referencing the
 407 * req (after submitting it) and aio_complete() freeing the req.
 408 */
 409static struct kiocb *__aio_get_req(struct kioctx *ctx)
 410{
 411        struct kiocb *req = NULL;
 412
 413        req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
 414        if (unlikely(!req))
 415                return NULL;
 416
 417        req->ki_flags = 0;
 418        req->ki_users = 2;
 419        req->ki_key = 0;
 420        req->ki_ctx = ctx;
 421        req->ki_cancel = NULL;
 422        req->ki_retry = NULL;
 423        req->ki_dtor = NULL;
 424        req->private = NULL;
 425        req->ki_iovec = NULL;
 426        INIT_LIST_HEAD(&req->ki_run_list);
 427        req->ki_eventfd = NULL;
 428
 429        return req;
 430}
 431
 432/*
 433 * struct kiocb's are allocated in batches to reduce the number of
 434 * times the ctx lock is acquired and released.
 435 */
 436#define KIOCB_BATCH_SIZE        32L
 437struct kiocb_batch {
 438        struct list_head head;
 439        long count; /* number of requests left to allocate */
 440};
 441
 442static void kiocb_batch_init(struct kiocb_batch *batch, long total)
 443{
 444        INIT_LIST_HEAD(&batch->head);
 445        batch->count = total;
 446}
 447
 448static void kiocb_batch_free(struct kioctx *ctx, struct kiocb_batch *batch)
 449{
 450        struct kiocb *req, *n;
 451
 452        if (list_empty(&batch->head))
 453                return;
 454
 455        spin_lock_irq(&ctx->ctx_lock);
 456        list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
 457                list_del(&req->ki_batch);
 458                list_del(&req->ki_list);
 459                kmem_cache_free(kiocb_cachep, req);
 460                ctx->reqs_active--;
 461        }
 462        if (unlikely(!ctx->reqs_active && ctx->dead))
 463                wake_up_all(&ctx->wait);
 464        spin_unlock_irq(&ctx->ctx_lock);
 465}
 466
 467/*
 468 * Allocate a batch of kiocbs.  This avoids taking and dropping the
 469 * context lock a lot during setup.
 470 */
 471static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch)
 472{
 473        unsigned short allocated, to_alloc;
 474        long avail;
 475        struct kiocb *req, *n;
 476        struct aio_ring *ring;
 477
 478        to_alloc = min(batch->count, KIOCB_BATCH_SIZE);
 479        for (allocated = 0; allocated < to_alloc; allocated++) {
 480                req = __aio_get_req(ctx);
 481                if (!req)
 482                        /* allocation failed, go with what we've got */
 483                        break;
 484                list_add(&req->ki_batch, &batch->head);
 485        }
 486
 487        if (allocated == 0)
 488                goto out;
 489
 490        spin_lock_irq(&ctx->ctx_lock);
 491        ring = kmap_atomic(ctx->ring_info.ring_pages[0]);
 492
 493        avail = aio_ring_avail(&ctx->ring_info, ring) - ctx->reqs_active;
 494        BUG_ON(avail < 0);
 495        if (avail < allocated) {
 496                /* Trim back the number of requests. */
 497                list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
 498                        list_del(&req->ki_batch);
 499                        kmem_cache_free(kiocb_cachep, req);
 500                        if (--allocated <= avail)
 501                                break;
 502                }
 503        }
 504
 505        batch->count -= allocated;
 506        list_for_each_entry(req, &batch->head, ki_batch) {
 507                list_add(&req->ki_list, &ctx->active_reqs);
 508                ctx->reqs_active++;
 509        }
 510
 511        kunmap_atomic(ring);
 512        spin_unlock_irq(&ctx->ctx_lock);
 513
 514out:
 515        return allocated;
 516}
 517
 518static inline struct kiocb *aio_get_req(struct kioctx *ctx,
 519                                        struct kiocb_batch *batch)
 520{
 521        struct kiocb *req;
 522
 523        if (list_empty(&batch->head))
 524                if (kiocb_batch_refill(ctx, batch) == 0)
 525                        return NULL;
 526        req = list_first_entry(&batch->head, struct kiocb, ki_batch);
 527        list_del(&req->ki_batch);
 528        return req;
 529}
 530
 531static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
 532{
 533        assert_spin_locked(&ctx->ctx_lock);
 534
 535        if (req->ki_eventfd != NULL)
 536                eventfd_ctx_put(req->ki_eventfd);
 537        if (req->ki_dtor)
 538                req->ki_dtor(req);
 539        if (req->ki_iovec != &req->ki_inline_vec)
 540                kfree(req->ki_iovec);
 541        kmem_cache_free(kiocb_cachep, req);
 542        ctx->reqs_active--;
 543
 544        if (unlikely(!ctx->reqs_active && ctx->dead))
 545                wake_up_all(&ctx->wait);
 546}
 547
 548/* __aio_put_req
 549 *      Returns true if this put was the last user of the request.
 550 */
 551static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
 552{
 553        dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
 554                req, atomic_long_read(&req->ki_filp->f_count));
 555
 556        assert_spin_locked(&ctx->ctx_lock);
 557
 558        req->ki_users--;
 559        BUG_ON(req->ki_users < 0);
 560        if (likely(req->ki_users))
 561                return 0;
 562        list_del(&req->ki_list);                /* remove from active_reqs */
 563        req->ki_cancel = NULL;
 564        req->ki_retry = NULL;
 565
 566        fput(req->ki_filp);
 567        req->ki_filp = NULL;
 568        really_put_req(ctx, req);
 569        return 1;
 570}
 571
 572/* aio_put_req
 573 *      Returns true if this put was the last user of the kiocb,
 574 *      false if the request is still in use.
 575 */
 576int aio_put_req(struct kiocb *req)
 577{
 578        struct kioctx *ctx = req->ki_ctx;
 579        int ret;
 580        spin_lock_irq(&ctx->ctx_lock);
 581        ret = __aio_put_req(ctx, req);
 582        spin_unlock_irq(&ctx->ctx_lock);
 583        return ret;
 584}
 585EXPORT_SYMBOL(aio_put_req);
 586
 587static struct kioctx *lookup_ioctx(unsigned long ctx_id)
 588{
 589        struct mm_struct *mm = current->mm;
 590        struct kioctx *ctx, *ret = NULL;
 591        struct hlist_node *n;
 592
 593        rcu_read_lock();
 594
 595        hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) {
 596                /*
 597                 * RCU protects us against accessing freed memory but
 598                 * we have to be careful not to get a reference when the
 599                 * reference count already dropped to 0 (ctx->dead test
 600                 * is unreliable because of races).
 601                 */
 602                if (ctx->user_id == ctx_id && !ctx->dead && try_get_ioctx(ctx)){
 603                        ret = ctx;
 604                        break;
 605                }
 606        }
 607
 608        rcu_read_unlock();
 609        return ret;
 610}
 611
 612/*
 613 * Queue up a kiocb to be retried. Assumes that the kiocb
 614 * has already been marked as kicked, and places it on
 615 * the retry run list for the corresponding ioctx, if it
 616 * isn't already queued. Returns 1 if it actually queued
 617 * the kiocb (to tell the caller to activate the work
 618 * queue to process it), or 0, if it found that it was
 619 * already queued.
 620 */
 621static inline int __queue_kicked_iocb(struct kiocb *iocb)
 622{
 623        struct kioctx *ctx = iocb->ki_ctx;
 624
 625        assert_spin_locked(&ctx->ctx_lock);
 626
 627        if (list_empty(&iocb->ki_run_list)) {
 628                list_add_tail(&iocb->ki_run_list,
 629                        &ctx->run_list);
 630                return 1;
 631        }
 632        return 0;
 633}
 634
 635/* aio_run_iocb
 636 *      This is the core aio execution routine. It is
 637 *      invoked both for initial i/o submission and
 638 *      subsequent retries via the aio_kick_handler.
 639 *      Expects to be invoked with iocb->ki_ctx->lock
 640 *      already held. The lock is released and reacquired
 641 *      as needed during processing.
 642 *
 643 * Calls the iocb retry method (already setup for the
 644 * iocb on initial submission) for operation specific
 645 * handling, but takes care of most of common retry
 646 * execution details for a given iocb. The retry method
 647 * needs to be non-blocking as far as possible, to avoid
 648 * holding up other iocbs waiting to be serviced by the
 649 * retry kernel thread.
 650 *
 651 * The trickier parts in this code have to do with
 652 * ensuring that only one retry instance is in progress
 653 * for a given iocb at any time. Providing that guarantee
 654 * simplifies the coding of individual aio operations as
 655 * it avoids various potential races.
 656 */
 657static ssize_t aio_run_iocb(struct kiocb *iocb)
 658{
 659        struct kioctx   *ctx = iocb->ki_ctx;
 660        ssize_t (*retry)(struct kiocb *);
 661        ssize_t ret;
 662
 663        if (!(retry = iocb->ki_retry)) {
 664                printk("aio_run_iocb: iocb->ki_retry = NULL\n");
 665                return 0;
 666        }
 667
 668        /*
 669         * We don't want the next retry iteration for this
 670         * operation to start until this one has returned and
 671         * updated the iocb state. However, wait_queue functions
 672         * can trigger a kick_iocb from interrupt context in the
 673         * meantime, indicating that data is available for the next
 674         * iteration. We want to remember that and enable the
 675         * next retry iteration _after_ we are through with
 676         * this one.
 677         *
 678         * So, in order to be able to register a "kick", but
 679         * prevent it from being queued now, we clear the kick
 680         * flag, but make the kick code *think* that the iocb is
 681         * still on the run list until we are actually done.
 682         * When we are done with this iteration, we check if
 683         * the iocb was kicked in the meantime and if so, queue
 684         * it up afresh.
 685         */
 686
 687        kiocbClearKicked(iocb);
 688
 689        /*
 690         * This is so that aio_complete knows it doesn't need to
 691         * pull the iocb off the run list (We can't just call
 692         * INIT_LIST_HEAD because we don't want a kick_iocb to
 693         * queue this on the run list yet)
 694         */
 695        iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
 696        spin_unlock_irq(&ctx->ctx_lock);
 697
 698        /* Quit retrying if the i/o has been cancelled */
 699        if (kiocbIsCancelled(iocb)) {
 700                ret = -EINTR;
 701                aio_complete(iocb, ret, 0);
 702                /* must not access the iocb after this */
 703                goto out;
 704        }
 705
 706        /*
 707         * Now we are all set to call the retry method in async
 708         * context.
 709         */
 710        ret = retry(iocb);
 711
 712        if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
 713                /*
 714                 * There's no easy way to restart the syscall since other AIO's
 715                 * may be already running. Just fail this IO with EINTR.
 716                 */
 717                if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
 718                             ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
 719                        ret = -EINTR;
 720                aio_complete(iocb, ret, 0);
 721        }
 722out:
 723        spin_lock_irq(&ctx->ctx_lock);
 724
 725        if (-EIOCBRETRY == ret) {
 726                /*
 727                 * OK, now that we are done with this iteration
 728                 * and know that there is more left to go,
 729                 * this is where we let go so that a subsequent
 730                 * "kick" can start the next iteration
 731                 */
 732
 733                /* will make __queue_kicked_iocb succeed from here on */
 734                INIT_LIST_HEAD(&iocb->ki_run_list);
 735                /* we must queue the next iteration ourselves, if it
 736                 * has already been kicked */
 737                if (kiocbIsKicked(iocb)) {
 738                        __queue_kicked_iocb(iocb);
 739
 740                        /*
 741                         * __queue_kicked_iocb will always return 1 here, because
 742                         * iocb->ki_run_list is empty at this point so it should
 743                         * be safe to unconditionally queue the context into the
 744                         * work queue.
 745                         */
 746                        aio_queue_work(ctx);
 747                }
 748        }
 749        return ret;
 750}
 751
 752/*
 753 * __aio_run_iocbs:
 754 *      Process all pending retries queued on the ioctx
 755 *      run list.
 756 * Assumes it is operating within the aio issuer's mm
 757 * context.
 758 */
 759static int __aio_run_iocbs(struct kioctx *ctx)
 760{
 761        struct kiocb *iocb;
 762        struct list_head run_list;
 763
 764        assert_spin_locked(&ctx->ctx_lock);
 765
 766        list_replace_init(&ctx->run_list, &run_list);
 767        while (!list_empty(&run_list)) {
 768                iocb = list_entry(run_list.next, struct kiocb,
 769                        ki_run_list);
 770                list_del(&iocb->ki_run_list);
 771                /*
 772                 * Hold an extra reference while retrying i/o.
 773                 */
 774                iocb->ki_users++;       /* grab extra reference */
 775                aio_run_iocb(iocb);
 776                __aio_put_req(ctx, iocb);
 777        }
 778        if (!list_empty(&ctx->run_list))
 779                return 1;
 780        return 0;
 781}
 782
 783static void aio_queue_work(struct kioctx * ctx)
 784{
 785        unsigned long timeout;
 786        /*
 787         * if someone is waiting, get the work started right
 788         * away, otherwise, use a longer delay
 789         */
 790        smp_mb();
 791        if (waitqueue_active(&ctx->wait))
 792                timeout = 1;
 793        else
 794                timeout = HZ/10;
 795        queue_delayed_work(aio_wq, &ctx->wq, timeout);
 796}
 797
 798/*
 799 * aio_run_all_iocbs:
 800 *      Process all pending retries queued on the ioctx
 801 *      run list, and keep running them until the list
 802 *      stays empty.
 803 * Assumes it is operating within the aio issuer's mm context.
 804 */
 805static inline void aio_run_all_iocbs(struct kioctx *ctx)
 806{
 807        spin_lock_irq(&ctx->ctx_lock);
 808        while (__aio_run_iocbs(ctx))
 809                ;
 810        spin_unlock_irq(&ctx->ctx_lock);
 811}
 812
 813/*
 814 * aio_kick_handler:
 815 *      Work queue handler triggered to process pending
 816 *      retries on an ioctx. Takes on the aio issuer's
 817 *      mm context before running the iocbs, so that
 818 *      copy_xxx_user operates on the issuer's address
 819 *      space.
 820 * Run on aiod's context.
 821 */
 822static void aio_kick_handler(struct work_struct *work)
 823{
 824        struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
 825        mm_segment_t oldfs = get_fs();
 826        struct mm_struct *mm;
 827        int requeue;
 828
 829        set_fs(USER_DS);
 830        use_mm(ctx->mm);
 831        spin_lock_irq(&ctx->ctx_lock);
 832        requeue =__aio_run_iocbs(ctx);
 833        mm = ctx->mm;
 834        spin_unlock_irq(&ctx->ctx_lock);
 835        unuse_mm(mm);
 836        set_fs(oldfs);
 837        /*
 838         * we're in a worker thread already; no point using non-zero delay
 839         */
 840        if (requeue)
 841                queue_delayed_work(aio_wq, &ctx->wq, 0);
 842}
 843
 844
 845/*
 846 * Called by kick_iocb to queue the kiocb for retry
 847 * and if required activate the aio work queue to process
 848 * it
 849 */
 850static void try_queue_kicked_iocb(struct kiocb *iocb)
 851{
 852        struct kioctx   *ctx = iocb->ki_ctx;
 853        unsigned long flags;
 854        int run = 0;
 855
 856        spin_lock_irqsave(&ctx->ctx_lock, flags);
 857        /* set this inside the lock so that we can't race with aio_run_iocb()
 858         * testing it and putting the iocb on the run list under the lock */
 859        if (!kiocbTryKick(iocb))
 860                run = __queue_kicked_iocb(iocb);
 861        spin_unlock_irqrestore(&ctx->ctx_lock, flags);
 862        if (run)
 863                aio_queue_work(ctx);
 864}
 865
 866/*
 867 * kick_iocb:
 868 *      Called typically from a wait queue callback context
 869 *      to trigger a retry of the iocb.
 870 *      The retry is usually executed by aio workqueue
 871 *      threads (See aio_kick_handler).
 872 */
 873void kick_iocb(struct kiocb *iocb)
 874{
 875        /* sync iocbs are easy: they can only ever be executing from a 
 876         * single context. */
 877        if (is_sync_kiocb(iocb)) {
 878                kiocbSetKicked(iocb);
 879                wake_up_process(iocb->ki_obj.tsk);
 880                return;
 881        }
 882
 883        try_queue_kicked_iocb(iocb);
 884}
 885EXPORT_SYMBOL(kick_iocb);
 886
 887/* aio_complete
 888 *      Called when the io request on the given iocb is complete.
 889 *      Returns true if this is the last user of the request.  The 
 890 *      only other user of the request can be the cancellation code.
 891 */
 892int aio_complete(struct kiocb *iocb, long res, long res2)
 893{
 894        struct kioctx   *ctx = iocb->ki_ctx;
 895        struct aio_ring_info    *info;
 896        struct aio_ring *ring;
 897        struct io_event *event;
 898        unsigned long   flags;
 899        unsigned long   tail;
 900        int             ret;
 901
 902        /*
 903         * Special case handling for sync iocbs:
 904         *  - events go directly into the iocb for fast handling
 905         *  - the sync task with the iocb in its stack holds the single iocb
 906         *    ref, no other paths have a way to get another ref
 907         *  - the sync task helpfully left a reference to itself in the iocb
 908         */
 909        if (is_sync_kiocb(iocb)) {
 910                BUG_ON(iocb->ki_users != 1);
 911                iocb->ki_user_data = res;
 912                iocb->ki_users = 0;
 913                wake_up_process(iocb->ki_obj.tsk);
 914                return 1;
 915        }
 916
 917        info = &ctx->ring_info;
 918
 919        /* add a completion event to the ring buffer.
 920         * must be done holding ctx->ctx_lock to prevent
 921         * other code from messing with the tail
 922         * pointer since we might be called from irq
 923         * context.
 924         */
 925        spin_lock_irqsave(&ctx->ctx_lock, flags);
 926
 927        if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
 928                list_del_init(&iocb->ki_run_list);
 929
 930        /*
 931         * cancelled requests don't get events, userland was given one
 932         * when the event got cancelled.
 933         */
 934        if (kiocbIsCancelled(iocb))
 935                goto put_rq;
 936
 937        ring = kmap_atomic(info->ring_pages[0]);
 938
 939        tail = info->tail;
 940        event = aio_ring_event(info, tail);
 941        if (++tail >= info->nr)
 942                tail = 0;
 943
 944        event->obj = (u64)(unsigned long)iocb->ki_obj.user;
 945        event->data = iocb->ki_user_data;
 946        event->res = res;
 947        event->res2 = res2;
 948
 949        dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
 950                ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
 951                res, res2);
 952
 953        /* after flagging the request as done, we
 954         * must never even look at it again
 955         */
 956        smp_wmb();      /* make event visible before updating tail */
 957
 958        info->tail = tail;
 959        ring->tail = tail;
 960
 961        put_aio_ring_event(event);
 962        kunmap_atomic(ring);
 963
 964        pr_debug("added to ring %p at [%lu]\n", iocb, tail);
 965
 966        /*
 967         * Check if the user asked us to deliver the result through an
 968         * eventfd. The eventfd_signal() function is safe to be called
 969         * from IRQ context.
 970         */
 971        if (iocb->ki_eventfd != NULL)
 972                eventfd_signal(iocb->ki_eventfd, 1);
 973
 974put_rq:
 975        /* everything turned out well, dispose of the aiocb. */
 976        ret = __aio_put_req(ctx, iocb);
 977
 978        /*
 979         * We have to order our ring_info tail store above and test
 980         * of the wait list below outside the wait lock.  This is
 981         * like in wake_up_bit() where clearing a bit has to be
 982         * ordered with the unlocked test.
 983         */
 984        smp_mb();
 985
 986        if (waitqueue_active(&ctx->wait))
 987                wake_up(&ctx->wait);
 988
 989        spin_unlock_irqrestore(&ctx->ctx_lock, flags);
 990        return ret;
 991}
 992EXPORT_SYMBOL(aio_complete);
 993
 994/* aio_read_evt
 995 *      Pull an event off of the ioctx's event ring.  Returns the number of 
 996 *      events fetched (0 or 1 ;-)
 997 *      FIXME: make this use cmpxchg.
 998 *      TODO: make the ringbuffer user mmap()able (requires FIXME).
 999 */
1000static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1001{
1002        struct aio_ring_info *info = &ioctx->ring_info;
1003        struct aio_ring *ring;
1004        unsigned long head;
1005        int ret = 0;
1006
1007        ring = kmap_atomic(info->ring_pages[0]);
1008        dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1009                 (unsigned long)ring->head, (unsigned long)ring->tail,
1010                 (unsigned long)ring->nr);
1011
1012        if (ring->head == ring->tail)
1013                goto out;
1014
1015        spin_lock(&info->ring_lock);
1016
1017        head = ring->head % info->nr;
1018        if (head != ring->tail) {
1019                struct io_event *evp = aio_ring_event(info, head);
1020                *ent = *evp;
1021                head = (head + 1) % info->nr;
1022                smp_mb(); /* finish reading the event before updatng the head */
1023                ring->head = head;
1024                ret = 1;
1025                put_aio_ring_event(evp);
1026        }
1027        spin_unlock(&info->ring_lock);
1028
1029out:
1030        kunmap_atomic(ring);
1031        dprintk("leaving aio_read_evt: %d  h%lu t%lu\n", ret,
1032                 (unsigned long)ring->head, (unsigned long)ring->tail);
1033        return ret;
1034}
1035
1036struct aio_timeout {
1037        struct timer_list       timer;
1038        int                     timed_out;
1039        struct task_struct      *p;
1040};
1041
1042static void timeout_func(unsigned long data)
1043{
1044        struct aio_timeout *to = (struct aio_timeout *)data;
1045
1046        to->timed_out = 1;
1047        wake_up_process(to->p);
1048}
1049
1050static inline void init_timeout(struct aio_timeout *to)
1051{
1052        setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
1053        to->timed_out = 0;
1054        to->p = current;
1055}
1056
1057static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1058                               const struct timespec *ts)
1059{
1060        to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1061        if (time_after(to->timer.expires, jiffies))
1062                add_timer(&to->timer);
1063        else
1064                to->timed_out = 1;
1065}
1066
1067static inline void clear_timeout(struct aio_timeout *to)
1068{
1069        del_singleshot_timer_sync(&to->timer);
1070}
1071
1072static int read_events(struct kioctx *ctx,
1073                        long min_nr, long nr,
1074                        struct io_event __user *event,
1075                        struct timespec __user *timeout)
1076{
1077        long                    start_jiffies = jiffies;
1078        struct task_struct      *tsk = current;
1079        DECLARE_WAITQUEUE(wait, tsk);
1080        int                     ret;
1081        int                     i = 0;
1082        struct io_event         ent;
1083        struct aio_timeout      to;
1084        int                     retry = 0;
1085
1086        /* needed to zero any padding within an entry (there shouldn't be 
1087         * any, but C is fun!
1088         */
1089        memset(&ent, 0, sizeof(ent));
1090retry:
1091        ret = 0;
1092        while (likely(i < nr)) {
1093                ret = aio_read_evt(ctx, &ent);
1094                if (unlikely(ret <= 0))
1095                        break;
1096
1097                dprintk("read event: %Lx %Lx %Lx %Lx\n",
1098                        ent.data, ent.obj, ent.res, ent.res2);
1099
1100                /* Could we split the check in two? */
1101                ret = -EFAULT;
1102                if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1103                        dprintk("aio: lost an event due to EFAULT.\n");
1104                        break;
1105                }
1106                ret = 0;
1107
1108                /* Good, event copied to userland, update counts. */
1109                event ++;
1110                i ++;
1111        }
1112
1113        if (min_nr <= i)
1114                return i;
1115        if (ret)
1116                return ret;
1117
1118        /* End fast path */
1119
1120        /* racey check, but it gets redone */
1121        if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1122                retry = 1;
1123                aio_run_all_iocbs(ctx);
1124                goto retry;
1125        }
1126
1127        init_timeout(&to);
1128        if (timeout) {
1129                struct timespec ts;
1130                ret = -EFAULT;
1131                if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1132                        goto out;
1133
1134                set_timeout(start_jiffies, &to, &ts);
1135        }
1136
1137        while (likely(i < nr)) {
1138                add_wait_queue_exclusive(&ctx->wait, &wait);
1139                do {
1140                        set_task_state(tsk, TASK_INTERRUPTIBLE);
1141                        ret = aio_read_evt(ctx, &ent);
1142                        if (ret)
1143                                break;
1144                        if (min_nr <= i)
1145                                break;
1146                        if (unlikely(ctx->dead)) {
1147                                ret = -EINVAL;
1148                                break;
1149                        }
1150                        if (to.timed_out)       /* Only check after read evt */
1151                                break;
1152                        /* Try to only show up in io wait if there are ops
1153                         *  in flight */
1154                        if (ctx->reqs_active)
1155                                io_schedule();
1156                        else
1157                                schedule();
1158                        if (signal_pending(tsk)) {
1159                                ret = -EINTR;
1160                                break;
1161                        }
1162                        /*ret = aio_read_evt(ctx, &ent);*/
1163                } while (1) ;
1164
1165                set_task_state(tsk, TASK_RUNNING);
1166                remove_wait_queue(&ctx->wait, &wait);
1167
1168                if (unlikely(ret <= 0))
1169                        break;
1170
1171                ret = -EFAULT;
1172                if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1173                        dprintk("aio: lost an event due to EFAULT.\n");
1174                        break;
1175                }
1176
1177                /* Good, event copied to userland, update counts. */
1178                event ++;
1179                i ++;
1180        }
1181
1182        if (timeout)
1183                clear_timeout(&to);
1184out:
1185        destroy_timer_on_stack(&to.timer);
1186        return i ? i : ret;
1187}
1188
1189/* Take an ioctx and remove it from the list of ioctx's.  Protects 
1190 * against races with itself via ->dead.
1191 */
1192static void io_destroy(struct kioctx *ioctx)
1193{
1194        struct mm_struct *mm = current->mm;
1195        int was_dead;
1196
1197        /* delete the entry from the list is someone else hasn't already */
1198        spin_lock(&mm->ioctx_lock);
1199        was_dead = ioctx->dead;
1200        ioctx->dead = 1;
1201        hlist_del_rcu(&ioctx->list);
1202        spin_unlock(&mm->ioctx_lock);
1203
1204        dprintk("aio_release(%p)\n", ioctx);
1205        if (likely(!was_dead))
1206                put_ioctx(ioctx);       /* twice for the list */
1207
1208        kill_ctx(ioctx);
1209
1210        /*
1211         * Wake up any waiters.  The setting of ctx->dead must be seen
1212         * by other CPUs at this point.  Right now, we rely on the
1213         * locking done by the above calls to ensure this consistency.
1214         */
1215        wake_up_all(&ioctx->wait);
1216}
1217
1218/* sys_io_setup:
1219 *      Create an aio_context capable of receiving at least nr_events.
1220 *      ctxp must not point to an aio_context that already exists, and
1221 *      must be initialized to 0 prior to the call.  On successful
1222 *      creation of the aio_context, *ctxp is filled in with the resulting 
1223 *      handle.  May fail with -EINVAL if *ctxp is not initialized,
1224 *      if the specified nr_events exceeds internal limits.  May fail 
1225 *      with -EAGAIN if the specified nr_events exceeds the user's limit 
1226 *      of available events.  May fail with -ENOMEM if insufficient kernel
1227 *      resources are available.  May fail with -EFAULT if an invalid
1228 *      pointer is passed for ctxp.  Will fail with -ENOSYS if not
1229 *      implemented.
1230 */
1231SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1232{
1233        struct kioctx *ioctx = NULL;
1234        unsigned long ctx;
1235        long ret;
1236
1237        ret = get_user(ctx, ctxp);
1238        if (unlikely(ret))
1239                goto out;
1240
1241        ret = -EINVAL;
1242        if (unlikely(ctx || nr_events == 0)) {
1243                pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1244                         ctx, nr_events);
1245                goto out;
1246        }
1247
1248        ioctx = ioctx_alloc(nr_events);
1249        ret = PTR_ERR(ioctx);
1250        if (!IS_ERR(ioctx)) {
1251                ret = put_user(ioctx->user_id, ctxp);
1252                if (ret)
1253                        io_destroy(ioctx);
1254                put_ioctx(ioctx);
1255        }
1256
1257out:
1258        return ret;
1259}
1260
1261/* sys_io_destroy:
1262 *      Destroy the aio_context specified.  May cancel any outstanding 
1263 *      AIOs and block on completion.  Will fail with -ENOSYS if not
1264 *      implemented.  May fail with -EINVAL if the context pointed to
1265 *      is invalid.
1266 */
1267SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1268{
1269        struct kioctx *ioctx = lookup_ioctx(ctx);
1270        if (likely(NULL != ioctx)) {
1271                io_destroy(ioctx);
1272                put_ioctx(ioctx);
1273                return 0;
1274        }
1275        pr_debug("EINVAL: io_destroy: invalid context id\n");
1276        return -EINVAL;
1277}
1278
1279static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1280{
1281        struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1282
1283        BUG_ON(ret <= 0);
1284
1285        while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1286                ssize_t this = min((ssize_t)iov->iov_len, ret);
1287                iov->iov_base += this;
1288                iov->iov_len -= this;
1289                iocb->ki_left -= this;
1290                ret -= this;
1291                if (iov->iov_len == 0) {
1292                        iocb->ki_cur_seg++;
1293                        iov++;
1294                }
1295        }
1296
1297        /* the caller should not have done more io than what fit in
1298         * the remaining iovecs */
1299        BUG_ON(ret > 0 && iocb->ki_left == 0);
1300}
1301
1302static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1303{
1304        struct file *file = iocb->ki_filp;
1305        struct address_space *mapping = file->f_mapping;
1306        struct inode *inode = mapping->host;
1307        ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1308                         unsigned long, loff_t);
1309        ssize_t ret = 0;
1310        unsigned short opcode;
1311
1312        if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1313                (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1314                rw_op = file->f_op->aio_read;
1315                opcode = IOCB_CMD_PREADV;
1316        } else {
1317                rw_op = file->f_op->aio_write;
1318                opcode = IOCB_CMD_PWRITEV;
1319        }
1320
1321        /* This matches the pread()/pwrite() logic */
1322        if (iocb->ki_pos < 0)
1323                return -EINVAL;
1324
1325        do {
1326                ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1327                            iocb->ki_nr_segs - iocb->ki_cur_seg,
1328                            iocb->ki_pos);
1329                if (ret > 0)
1330                        aio_advance_iovec(iocb, ret);
1331
1332        /* retry all partial writes.  retry partial reads as long as its a
1333         * regular file. */
1334        } while (ret > 0 && iocb->ki_left > 0 &&
1335                 (opcode == IOCB_CMD_PWRITEV ||
1336                  (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1337
1338        /* This means we must have transferred all that we could */
1339        /* No need to retry anymore */
1340        if ((ret == 0) || (iocb->ki_left == 0))
1341                ret = iocb->ki_nbytes - iocb->ki_left;
1342
1343        /* If we managed to write some out we return that, rather than
1344         * the eventual error. */
1345        if (opcode == IOCB_CMD_PWRITEV
1346            && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
1347            && iocb->ki_nbytes - iocb->ki_left)
1348                ret = iocb->ki_nbytes - iocb->ki_left;
1349
1350        return ret;
1351}
1352
1353static ssize_t aio_fdsync(struct kiocb *iocb)
1354{
1355        struct file *file = iocb->ki_filp;
1356        ssize_t ret = -EINVAL;
1357
1358        if (file->f_op->aio_fsync)
1359                ret = file->f_op->aio_fsync(iocb, 1);
1360        return ret;
1361}
1362
1363static ssize_t aio_fsync(struct kiocb *iocb)
1364{
1365        struct file *file = iocb->ki_filp;
1366        ssize_t ret = -EINVAL;
1367
1368        if (file->f_op->aio_fsync)
1369                ret = file->f_op->aio_fsync(iocb, 0);
1370        return ret;
1371}
1372
1373static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
1374{
1375        ssize_t ret;
1376
1377#ifdef CONFIG_COMPAT
1378        if (compat)
1379                ret = compat_rw_copy_check_uvector(type,
1380                                (struct compat_iovec __user *)kiocb->ki_buf,
1381                                kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1382                                &kiocb->ki_iovec);
1383        else
1384#endif
1385                ret = rw_copy_check_uvector(type,
1386                                (struct iovec __user *)kiocb->ki_buf,
1387                                kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1388                                &kiocb->ki_iovec);
1389        if (ret < 0)
1390                goto out;
1391
1392        ret = rw_verify_area(type, kiocb->ki_filp, &kiocb->ki_pos, ret);
1393        if (ret < 0)
1394                goto out;
1395
1396        kiocb->ki_nr_segs = kiocb->ki_nbytes;
1397        kiocb->ki_cur_seg = 0;
1398        /* ki_nbytes/left now reflect bytes instead of segs */
1399        kiocb->ki_nbytes = ret;
1400        kiocb->ki_left = ret;
1401
1402        ret = 0;
1403out:
1404        return ret;
1405}
1406
1407static ssize_t aio_setup_single_vector(int type, struct file * file, struct kiocb *kiocb)
1408{
1409        int bytes;
1410
1411        bytes = rw_verify_area(type, file, &kiocb->ki_pos, kiocb->ki_left);
1412        if (bytes < 0)
1413                return bytes;
1414
1415        kiocb->ki_iovec = &kiocb->ki_inline_vec;
1416        kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1417        kiocb->ki_iovec->iov_len = bytes;
1418        kiocb->ki_nr_segs = 1;
1419        kiocb->ki_cur_seg = 0;
1420        return 0;
1421}
1422
1423/*
1424 * aio_setup_iocb:
1425 *      Performs the initial checks and aio retry method
1426 *      setup for the kiocb at the time of io submission.
1427 */
1428static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
1429{
1430        struct file *file = kiocb->ki_filp;
1431        ssize_t ret = 0;
1432
1433        switch (kiocb->ki_opcode) {
1434        case IOCB_CMD_PREAD:
1435                ret = -EBADF;
1436                if (unlikely(!(file->f_mode & FMODE_READ)))
1437                        break;
1438                ret = -EFAULT;
1439                if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1440                        kiocb->ki_left)))
1441                        break;
1442                ret = aio_setup_single_vector(READ, file, kiocb);
1443                if (ret)
1444                        break;
1445                ret = -EINVAL;
1446                if (file->f_op->aio_read)
1447                        kiocb->ki_retry = aio_rw_vect_retry;
1448                break;
1449        case IOCB_CMD_PWRITE:
1450                ret = -EBADF;
1451                if (unlikely(!(file->f_mode & FMODE_WRITE)))
1452                        break;
1453                ret = -EFAULT;
1454                if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1455                        kiocb->ki_left)))
1456                        break;
1457                ret = aio_setup_single_vector(WRITE, file, kiocb);
1458                if (ret)
1459                        break;
1460                ret = -EINVAL;
1461                if (file->f_op->aio_write)
1462                        kiocb->ki_retry = aio_rw_vect_retry;
1463                break;
1464        case IOCB_CMD_PREADV:
1465                ret = -EBADF;
1466                if (unlikely(!(file->f_mode & FMODE_READ)))
1467                        break;
1468                ret = aio_setup_vectored_rw(READ, kiocb, compat);
1469                if (ret)
1470                        break;
1471                ret = -EINVAL;
1472                if (file->f_op->aio_read)
1473                        kiocb->ki_retry = aio_rw_vect_retry;
1474                break;
1475        case IOCB_CMD_PWRITEV:
1476                ret = -EBADF;
1477                if (unlikely(!(file->f_mode & FMODE_WRITE)))
1478                        break;
1479                ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1480                if (ret)
1481                        break;
1482                ret = -EINVAL;
1483                if (file->f_op->aio_write)
1484                        kiocb->ki_retry = aio_rw_vect_retry;
1485                break;
1486        case IOCB_CMD_FDSYNC:
1487                ret = -EINVAL;
1488                if (file->f_op->aio_fsync)
1489                        kiocb->ki_retry = aio_fdsync;
1490                break;
1491        case IOCB_CMD_FSYNC:
1492                ret = -EINVAL;
1493                if (file->f_op->aio_fsync)
1494                        kiocb->ki_retry = aio_fsync;
1495                break;
1496        default:
1497                dprintk("EINVAL: io_submit: no operation provided\n");
1498                ret = -EINVAL;
1499        }
1500
1501        if (!kiocb->ki_retry)
1502                return ret;
1503
1504        return 0;
1505}
1506
1507static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1508                         struct iocb *iocb, struct kiocb_batch *batch,
1509                         bool compat)
1510{
1511        struct kiocb *req;
1512        struct file *file;
1513        ssize_t ret;
1514
1515        /* enforce forwards compatibility on users */
1516        if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1517                pr_debug("EINVAL: io_submit: reserve field set\n");
1518                return -EINVAL;
1519        }
1520
1521        /* prevent overflows */
1522        if (unlikely(
1523            (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1524            (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1525            ((ssize_t)iocb->aio_nbytes < 0)
1526           )) {
1527                pr_debug("EINVAL: io_submit: overflow check\n");
1528                return -EINVAL;
1529        }
1530
1531        file = fget(iocb->aio_fildes);
1532        if (unlikely(!file))
1533                return -EBADF;
1534
1535        req = aio_get_req(ctx, batch);  /* returns with 2 references to req */
1536        if (unlikely(!req)) {
1537                fput(file);
1538                return -EAGAIN;
1539        }
1540        req->ki_filp = file;
1541        if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1542                /*
1543                 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1544                 * instance of the file* now. The file descriptor must be
1545                 * an eventfd() fd, and will be signaled for each completed
1546                 * event using the eventfd_signal() function.
1547                 */
1548                req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1549                if (IS_ERR(req->ki_eventfd)) {
1550                        ret = PTR_ERR(req->ki_eventfd);
1551                        req->ki_eventfd = NULL;
1552                        goto out_put_req;
1553                }
1554        }
1555
1556        ret = put_user(req->ki_key, &user_iocb->aio_key);
1557        if (unlikely(ret)) {
1558                dprintk("EFAULT: aio_key\n");
1559                goto out_put_req;
1560        }
1561
1562        req->ki_obj.user = user_iocb;
1563        req->ki_user_data = iocb->aio_data;
1564        req->ki_pos = iocb->aio_offset;
1565
1566        req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1567        req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1568        req->ki_opcode = iocb->aio_lio_opcode;
1569
1570        ret = aio_setup_iocb(req, compat);
1571
1572        if (ret)
1573                goto out_put_req;
1574
1575        spin_lock_irq(&ctx->ctx_lock);
1576        /*
1577         * We could have raced with io_destroy() and are currently holding a
1578         * reference to ctx which should be destroyed. We cannot submit IO
1579         * since ctx gets freed as soon as io_submit() puts its reference.  The
1580         * check here is reliable: io_destroy() sets ctx->dead before waiting
1581         * for outstanding IO and the barrier between these two is realized by
1582         * unlock of mm->ioctx_lock and lock of ctx->ctx_lock.  Analogously we
1583         * increment ctx->reqs_active before checking for ctx->dead and the
1584         * barrier is realized by unlock and lock of ctx->ctx_lock. Thus if we
1585         * don't see ctx->dead set here, io_destroy() waits for our IO to
1586         * finish.
1587         */
1588        if (ctx->dead) {
1589                spin_unlock_irq(&ctx->ctx_lock);
1590                ret = -EINVAL;
1591                goto out_put_req;
1592        }
1593        aio_run_iocb(req);
1594        if (!list_empty(&ctx->run_list)) {
1595                /* drain the run list */
1596                while (__aio_run_iocbs(ctx))
1597                        ;
1598        }
1599        spin_unlock_irq(&ctx->ctx_lock);
1600
1601        aio_put_req(req);       /* drop extra ref to req */
1602        return 0;
1603
1604out_put_req:
1605        aio_put_req(req);       /* drop extra ref to req */
1606        aio_put_req(req);       /* drop i/o ref to req */
1607        return ret;
1608}
1609
1610long do_io_submit(aio_context_t ctx_id, long nr,
1611                  struct iocb __user *__user *iocbpp, bool compat)
1612{
1613        struct kioctx *ctx;
1614        long ret = 0;
1615        int i = 0;
1616        struct blk_plug plug;
1617        struct kiocb_batch batch;
1618
1619        if (unlikely(nr < 0))
1620                return -EINVAL;
1621
1622        if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1623                nr = LONG_MAX/sizeof(*iocbpp);
1624
1625        if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1626                return -EFAULT;
1627
1628        ctx = lookup_ioctx(ctx_id);
1629        if (unlikely(!ctx)) {
1630                pr_debug("EINVAL: io_submit: invalid context id\n");
1631                return -EINVAL;
1632        }
1633
1634        kiocb_batch_init(&batch, nr);
1635
1636        blk_start_plug(&plug);
1637
1638        /*
1639         * AKPM: should this return a partial result if some of the IOs were
1640         * successfully submitted?
1641         */
1642        for (i=0; i<nr; i++) {
1643                struct iocb __user *user_iocb;
1644                struct iocb tmp;
1645
1646                if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1647                        ret = -EFAULT;
1648                        break;
1649                }
1650
1651                if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1652                        ret = -EFAULT;
1653                        break;
1654                }
1655
1656                ret = io_submit_one(ctx, user_iocb, &tmp, &batch, compat);
1657                if (ret)
1658                        break;
1659        }
1660        blk_finish_plug(&plug);
1661
1662        kiocb_batch_free(ctx, &batch);
1663        put_ioctx(ctx);
1664        return i ? i : ret;
1665}
1666
1667/* sys_io_submit:
1668 *      Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1669 *      the number of iocbs queued.  May return -EINVAL if the aio_context
1670 *      specified by ctx_id is invalid, if nr is < 0, if the iocb at
1671 *      *iocbpp[0] is not properly initialized, if the operation specified
1672 *      is invalid for the file descriptor in the iocb.  May fail with
1673 *      -EFAULT if any of the data structures point to invalid data.  May
1674 *      fail with -EBADF if the file descriptor specified in the first
1675 *      iocb is invalid.  May fail with -EAGAIN if insufficient resources
1676 *      are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1677 *      fail with -ENOSYS if not implemented.
1678 */
1679SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1680                struct iocb __user * __user *, iocbpp)
1681{
1682        return do_io_submit(ctx_id, nr, iocbpp, 0);
1683}
1684
1685/* lookup_kiocb
1686 *      Finds a given iocb for cancellation.
1687 */
1688static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1689                                  u32 key)
1690{
1691        struct list_head *pos;
1692
1693        assert_spin_locked(&ctx->ctx_lock);
1694
1695        /* TODO: use a hash or array, this sucks. */
1696        list_for_each(pos, &ctx->active_reqs) {
1697                struct kiocb *kiocb = list_kiocb(pos);
1698                if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1699                        return kiocb;
1700        }
1701        return NULL;
1702}
1703
1704/* sys_io_cancel:
1705 *      Attempts to cancel an iocb previously passed to io_submit.  If
1706 *      the operation is successfully cancelled, the resulting event is
1707 *      copied into the memory pointed to by result without being placed
1708 *      into the completion queue and 0 is returned.  May fail with
1709 *      -EFAULT if any of the data structures pointed to are invalid.
1710 *      May fail with -EINVAL if aio_context specified by ctx_id is
1711 *      invalid.  May fail with -EAGAIN if the iocb specified was not
1712 *      cancelled.  Will fail with -ENOSYS if not implemented.
1713 */
1714SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1715                struct io_event __user *, result)
1716{
1717        int (*cancel)(struct kiocb *iocb, struct io_event *res);
1718        struct kioctx *ctx;
1719        struct kiocb *kiocb;
1720        u32 key;
1721        int ret;
1722
1723        ret = get_user(key, &iocb->aio_key);
1724        if (unlikely(ret))
1725                return -EFAULT;
1726
1727        ctx = lookup_ioctx(ctx_id);
1728        if (unlikely(!ctx))
1729                return -EINVAL;
1730
1731        spin_lock_irq(&ctx->ctx_lock);
1732        ret = -EAGAIN;
1733        kiocb = lookup_kiocb(ctx, iocb, key);
1734        if (kiocb && kiocb->ki_cancel) {
1735                cancel = kiocb->ki_cancel;
1736                kiocb->ki_users ++;
1737                kiocbSetCancelled(kiocb);
1738        } else
1739                cancel = NULL;
1740        spin_unlock_irq(&ctx->ctx_lock);
1741
1742        if (NULL != cancel) {
1743                struct io_event tmp;
1744                pr_debug("calling cancel\n");
1745                memset(&tmp, 0, sizeof(tmp));
1746                tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1747                tmp.data = kiocb->ki_user_data;
1748                ret = cancel(kiocb, &tmp);
1749                if (!ret) {
1750                        /* Cancellation succeeded -- copy the result
1751                         * into the user's buffer.
1752                         */
1753                        if (copy_to_user(result, &tmp, sizeof(tmp)))
1754                                ret = -EFAULT;
1755                }
1756        } else
1757                ret = -EINVAL;
1758
1759        put_ioctx(ctx);
1760
1761        return ret;
1762}
1763
1764/* io_getevents:
1765 *      Attempts to read at least min_nr events and up to nr events from
1766 *      the completion queue for the aio_context specified by ctx_id. If
1767 *      it succeeds, the number of read events is returned. May fail with
1768 *      -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1769 *      out of range, if timeout is out of range.  May fail with -EFAULT
1770 *      if any of the memory specified is invalid.  May return 0 or
1771 *      < min_nr if the timeout specified by timeout has elapsed
1772 *      before sufficient events are available, where timeout == NULL
1773 *      specifies an infinite timeout. Note that the timeout pointed to by
1774 *      timeout is relative and will be updated if not NULL and the
1775 *      operation blocks. Will fail with -ENOSYS if not implemented.
1776 */
1777SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1778                long, min_nr,
1779                long, nr,
1780                struct io_event __user *, events,
1781                struct timespec __user *, timeout)
1782{
1783        struct kioctx *ioctx = lookup_ioctx(ctx_id);
1784        long ret = -EINVAL;
1785
1786        if (likely(ioctx)) {
1787                if (likely(min_nr <= nr && min_nr >= 0))
1788                        ret = read_events(ioctx, min_nr, nr, events, timeout);
1789                put_ioctx(ioctx);
1790        }
1791
1792        asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);
1793        return ret;
1794}
1795
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.