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/module.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
  59/* Used for rare fput completion. */
  60static void aio_fput_routine(struct work_struct *);
  61static DECLARE_WORK(fput_work, aio_fput_routine);
  62
  63static DEFINE_SPINLOCK(fput_lock);
  64static LIST_HEAD(fput_head);
  65
  66static void aio_kick_handler(struct work_struct *);
  67static void aio_queue_work(struct kioctx *);
  68
  69/* aio_setup
  70 *      Creates the slab caches used by the aio routines, panic on
  71 *      failure as this is done early during the boot sequence.
  72 */
  73static int __init aio_setup(void)
  74{
  75        kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
  76        kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
  77
  78        aio_wq = alloc_workqueue("aio", 0, 1);  /* used to limit concurrency */
  79        BUG_ON(!aio_wq);
  80
  81        pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
  82
  83        return 0;
  84}
  85__initcall(aio_setup);
  86
  87static void aio_free_ring(struct kioctx *ctx)
  88{
  89        struct aio_ring_info *info = &ctx->ring_info;
  90        long i;
  91
  92        for (i=0; i<info->nr_pages; i++)
  93                put_page(info->ring_pages[i]);
  94
  95        if (info->mmap_size) {
  96                down_write(&ctx->mm->mmap_sem);
  97                do_munmap(ctx->mm, info->mmap_base, info->mmap_size);
  98                up_write(&ctx->mm->mmap_sem);
  99        }
 100
 101        if (info->ring_pages && info->ring_pages != info->internal_pages)
 102                kfree(info->ring_pages);
 103        info->ring_pages = NULL;
 104        info->nr = 0;
 105}
 106
 107static int aio_setup_ring(struct kioctx *ctx)
 108{
 109        struct aio_ring *ring;
 110        struct aio_ring_info *info = &ctx->ring_info;
 111        unsigned nr_events = ctx->max_reqs;
 112        unsigned long size;
 113        int nr_pages;
 114
 115        /* Compensate for the ring buffer's head/tail overlap entry */
 116        nr_events += 2; /* 1 is required, 2 for good luck */
 117
 118        size = sizeof(struct aio_ring);
 119        size += sizeof(struct io_event) * nr_events;
 120        nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
 121
 122        if (nr_pages < 0)
 123                return -EINVAL;
 124
 125        nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
 126
 127        info->nr = 0;
 128        info->ring_pages = info->internal_pages;
 129        if (nr_pages > AIO_RING_PAGES) {
 130                info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
 131                if (!info->ring_pages)
 132                        return -ENOMEM;
 133        }
 134
 135        info->mmap_size = nr_pages * PAGE_SIZE;
 136        dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
 137        down_write(&ctx->mm->mmap_sem);
 138        info->mmap_base = do_mmap(NULL, 0, info->mmap_size, 
 139                                  PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE,
 140                                  0);
 141        if (IS_ERR((void *)info->mmap_base)) {
 142                up_write(&ctx->mm->mmap_sem);
 143                info->mmap_size = 0;
 144                aio_free_ring(ctx);
 145                return -EAGAIN;
 146        }
 147
 148        dprintk("mmap address: 0x%08lx\n", info->mmap_base);
 149        info->nr_pages = get_user_pages(current, ctx->mm,
 150                                        info->mmap_base, nr_pages, 
 151                                        1, 0, info->ring_pages, NULL);
 152        up_write(&ctx->mm->mmap_sem);
 153
 154        if (unlikely(info->nr_pages != nr_pages)) {
 155                aio_free_ring(ctx);
 156                return -EAGAIN;
 157        }
 158
 159        ctx->user_id = info->mmap_base;
 160
 161        info->nr = nr_events;           /* trusted copy */
 162
 163        ring = kmap_atomic(info->ring_pages[0], KM_USER0);
 164        ring->nr = nr_events;   /* user copy */
 165        ring->id = ctx->user_id;
 166        ring->head = ring->tail = 0;
 167        ring->magic = AIO_RING_MAGIC;
 168        ring->compat_features = AIO_RING_COMPAT_FEATURES;
 169        ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
 170        ring->header_length = sizeof(struct aio_ring);
 171        kunmap_atomic(ring, KM_USER0);
 172
 173        return 0;
 174}
 175
 176
 177/* aio_ring_event: returns a pointer to the event at the given index from
 178 * kmap_atomic(, km).  Release the pointer with put_aio_ring_event();
 179 */
 180#define AIO_EVENTS_PER_PAGE     (PAGE_SIZE / sizeof(struct io_event))
 181#define AIO_EVENTS_FIRST_PAGE   ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
 182#define AIO_EVENTS_OFFSET       (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
 183
 184#define aio_ring_event(info, nr, km) ({                                 \
 185        unsigned pos = (nr) + AIO_EVENTS_OFFSET;                        \
 186        struct io_event *__event;                                       \
 187        __event = kmap_atomic(                                          \
 188                        (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
 189        __event += pos % AIO_EVENTS_PER_PAGE;                           \
 190        __event;                                                        \
 191})
 192
 193#define put_aio_ring_event(event, km) do {      \
 194        struct io_event *__event = (event);     \
 195        (void)__event;                          \
 196        kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
 197} while(0)
 198
 199static void ctx_rcu_free(struct rcu_head *head)
 200{
 201        struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
 202        unsigned nr_events = ctx->max_reqs;
 203
 204        kmem_cache_free(kioctx_cachep, ctx);
 205
 206        if (nr_events) {
 207                spin_lock(&aio_nr_lock);
 208                BUG_ON(aio_nr - nr_events > aio_nr);
 209                aio_nr -= nr_events;
 210                spin_unlock(&aio_nr_lock);
 211        }
 212}
 213
 214/* __put_ioctx
 215 *      Called when the last user of an aio context has gone away,
 216 *      and the struct needs to be freed.
 217 */
 218static void __put_ioctx(struct kioctx *ctx)
 219{
 220        BUG_ON(ctx->reqs_active);
 221
 222        cancel_delayed_work(&ctx->wq);
 223        cancel_work_sync(&ctx->wq.work);
 224        aio_free_ring(ctx);
 225        mmdrop(ctx->mm);
 226        ctx->mm = NULL;
 227        pr_debug("__put_ioctx: freeing %p\n", ctx);
 228        call_rcu(&ctx->rcu_head, ctx_rcu_free);
 229}
 230
 231static inline void get_ioctx(struct kioctx *kioctx)
 232{
 233        BUG_ON(atomic_read(&kioctx->users) <= 0);
 234        atomic_inc(&kioctx->users);
 235}
 236
 237static inline int try_get_ioctx(struct kioctx *kioctx)
 238{
 239        return atomic_inc_not_zero(&kioctx->users);
 240}
 241
 242static inline void put_ioctx(struct kioctx *kioctx)
 243{
 244        BUG_ON(atomic_read(&kioctx->users) <= 0);
 245        if (unlikely(atomic_dec_and_test(&kioctx->users)))
 246                __put_ioctx(kioctx);
 247}
 248
 249/* ioctx_alloc
 250 *      Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
 251 */
 252static struct kioctx *ioctx_alloc(unsigned nr_events)
 253{
 254        struct mm_struct *mm;
 255        struct kioctx *ctx;
 256        int did_sync = 0;
 257
 258        /* Prevent overflows */
 259        if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
 260            (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
 261                pr_debug("ENOMEM: nr_events too high\n");
 262                return ERR_PTR(-EINVAL);
 263        }
 264
 265        if ((unsigned long)nr_events > aio_max_nr)
 266                return ERR_PTR(-EAGAIN);
 267
 268        ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
 269        if (!ctx)
 270                return ERR_PTR(-ENOMEM);
 271
 272        ctx->max_reqs = nr_events;
 273        mm = ctx->mm = current->mm;
 274        atomic_inc(&mm->mm_count);
 275
 276        atomic_set(&ctx->users, 1);
 277        spin_lock_init(&ctx->ctx_lock);
 278        spin_lock_init(&ctx->ring_info.ring_lock);
 279        init_waitqueue_head(&ctx->wait);
 280
 281        INIT_LIST_HEAD(&ctx->active_reqs);
 282        INIT_LIST_HEAD(&ctx->run_list);
 283        INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
 284
 285        if (aio_setup_ring(ctx) < 0)
 286                goto out_freectx;
 287
 288        /* limit the number of system wide aios */
 289        do {
 290                spin_lock_bh(&aio_nr_lock);
 291                if (aio_nr + nr_events > aio_max_nr ||
 292                    aio_nr + nr_events < aio_nr)
 293                        ctx->max_reqs = 0;
 294                else
 295                        aio_nr += ctx->max_reqs;
 296                spin_unlock_bh(&aio_nr_lock);
 297                if (ctx->max_reqs || did_sync)
 298                        break;
 299
 300                /* wait for rcu callbacks to have completed before giving up */
 301                synchronize_rcu();
 302                did_sync = 1;
 303                ctx->max_reqs = nr_events;
 304        } while (1);
 305
 306        if (ctx->max_reqs == 0)
 307                goto out_cleanup;
 308
 309        /* now link into global list. */
 310        spin_lock(&mm->ioctx_lock);
 311        hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
 312        spin_unlock(&mm->ioctx_lock);
 313
 314        dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
 315                ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
 316        return ctx;
 317
 318out_cleanup:
 319        __put_ioctx(ctx);
 320        return ERR_PTR(-EAGAIN);
 321
 322out_freectx:
 323        mmdrop(mm);
 324        kmem_cache_free(kioctx_cachep, ctx);
 325        ctx = ERR_PTR(-ENOMEM);
 326
 327        dprintk("aio: error allocating ioctx %p\n", ctx);
 328        return ctx;
 329}
 330
 331/* aio_cancel_all
 332 *      Cancels all outstanding aio requests on an aio context.  Used 
 333 *      when the processes owning a context have all exited to encourage 
 334 *      the rapid destruction of the kioctx.
 335 */
 336static void aio_cancel_all(struct kioctx *ctx)
 337{
 338        int (*cancel)(struct kiocb *, struct io_event *);
 339        struct io_event res;
 340        spin_lock_irq(&ctx->ctx_lock);
 341        ctx->dead = 1;
 342        while (!list_empty(&ctx->active_reqs)) {
 343                struct list_head *pos = ctx->active_reqs.next;
 344                struct kiocb *iocb = list_kiocb(pos);
 345                list_del_init(&iocb->ki_list);
 346                cancel = iocb->ki_cancel;
 347                kiocbSetCancelled(iocb);
 348                if (cancel) {
 349                        iocb->ki_users++;
 350                        spin_unlock_irq(&ctx->ctx_lock);
 351                        cancel(iocb, &res);
 352                        spin_lock_irq(&ctx->ctx_lock);
 353                }
 354        }
 355        spin_unlock_irq(&ctx->ctx_lock);
 356}
 357
 358static void wait_for_all_aios(struct kioctx *ctx)
 359{
 360        struct task_struct *tsk = current;
 361        DECLARE_WAITQUEUE(wait, tsk);
 362
 363        spin_lock_irq(&ctx->ctx_lock);
 364        if (!ctx->reqs_active)
 365                goto out;
 366
 367        add_wait_queue(&ctx->wait, &wait);
 368        set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 369        while (ctx->reqs_active) {
 370                spin_unlock_irq(&ctx->ctx_lock);
 371                io_schedule();
 372                set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 373                spin_lock_irq(&ctx->ctx_lock);
 374        }
 375        __set_task_state(tsk, TASK_RUNNING);
 376        remove_wait_queue(&ctx->wait, &wait);
 377
 378out:
 379        spin_unlock_irq(&ctx->ctx_lock);
 380}
 381
 382/* wait_on_sync_kiocb:
 383 *      Waits on the given sync kiocb to complete.
 384 */
 385ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
 386{
 387        while (iocb->ki_users) {
 388                set_current_state(TASK_UNINTERRUPTIBLE);
 389                if (!iocb->ki_users)
 390                        break;
 391                io_schedule();
 392        }
 393        __set_current_state(TASK_RUNNING);
 394        return iocb->ki_user_data;
 395}
 396EXPORT_SYMBOL(wait_on_sync_kiocb);
 397
 398/* exit_aio: called when the last user of mm goes away.  At this point, 
 399 * there is no way for any new requests to be submited or any of the 
 400 * io_* syscalls to be called on the context.  However, there may be 
 401 * outstanding requests which hold references to the context; as they 
 402 * go away, they will call put_ioctx and release any pinned memory
 403 * associated with the request (held via struct page * references).
 404 */
 405void exit_aio(struct mm_struct *mm)
 406{
 407        struct kioctx *ctx;
 408
 409        while (!hlist_empty(&mm->ioctx_list)) {
 410                ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
 411                hlist_del_rcu(&ctx->list);
 412
 413                aio_cancel_all(ctx);
 414
 415                wait_for_all_aios(ctx);
 416                /*
 417                 * Ensure we don't leave the ctx on the aio_wq
 418                 */
 419                cancel_work_sync(&ctx->wq.work);
 420
 421                if (1 != atomic_read(&ctx->users))
 422                        printk(KERN_DEBUG
 423                                "exit_aio:ioctx still alive: %d %d %d\n",
 424                                atomic_read(&ctx->users), ctx->dead,
 425                                ctx->reqs_active);
 426                put_ioctx(ctx);
 427        }
 428}
 429
 430/* aio_get_req
 431 *      Allocate a slot for an aio request.  Increments the users count
 432 * of the kioctx so that the kioctx stays around until all requests are
 433 * complete.  Returns NULL if no requests are free.
 434 *
 435 * Returns with kiocb->users set to 2.  The io submit code path holds
 436 * an extra reference while submitting the i/o.
 437 * This prevents races between the aio code path referencing the
 438 * req (after submitting it) and aio_complete() freeing the req.
 439 */
 440static struct kiocb *__aio_get_req(struct kioctx *ctx)
 441{
 442        struct kiocb *req = NULL;
 443        struct aio_ring *ring;
 444        int okay = 0;
 445
 446        req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
 447        if (unlikely(!req))
 448                return NULL;
 449
 450        req->ki_flags = 0;
 451        req->ki_users = 2;
 452        req->ki_key = 0;
 453        req->ki_ctx = ctx;
 454        req->ki_cancel = NULL;
 455        req->ki_retry = NULL;
 456        req->ki_dtor = NULL;
 457        req->private = NULL;
 458        req->ki_iovec = NULL;
 459        INIT_LIST_HEAD(&req->ki_run_list);
 460        req->ki_eventfd = NULL;
 461
 462        /* Check if the completion queue has enough free space to
 463         * accept an event from this io.
 464         */
 465        spin_lock_irq(&ctx->ctx_lock);
 466        ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
 467        if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
 468                list_add(&req->ki_list, &ctx->active_reqs);
 469                ctx->reqs_active++;
 470                okay = 1;
 471        }
 472        kunmap_atomic(ring, KM_USER0);
 473        spin_unlock_irq(&ctx->ctx_lock);
 474
 475        if (!okay) {
 476                kmem_cache_free(kiocb_cachep, req);
 477                req = NULL;
 478        }
 479
 480        return req;
 481}
 482
 483static inline struct kiocb *aio_get_req(struct kioctx *ctx)
 484{
 485        struct kiocb *req;
 486        /* Handle a potential starvation case -- should be exceedingly rare as 
 487         * requests will be stuck on fput_head only if the aio_fput_routine is 
 488         * delayed and the requests were the last user of the struct file.
 489         */
 490        req = __aio_get_req(ctx);
 491        if (unlikely(NULL == req)) {
 492                aio_fput_routine(NULL);
 493                req = __aio_get_req(ctx);
 494        }
 495        return req;
 496}
 497
 498static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
 499{
 500        assert_spin_locked(&ctx->ctx_lock);
 501
 502        if (req->ki_eventfd != NULL)
 503                eventfd_ctx_put(req->ki_eventfd);
 504        if (req->ki_dtor)
 505                req->ki_dtor(req);
 506        if (req->ki_iovec != &req->ki_inline_vec)
 507                kfree(req->ki_iovec);
 508        kmem_cache_free(kiocb_cachep, req);
 509        ctx->reqs_active--;
 510
 511        if (unlikely(!ctx->reqs_active && ctx->dead))
 512                wake_up_all(&ctx->wait);
 513}
 514
 515static void aio_fput_routine(struct work_struct *data)
 516{
 517        spin_lock_irq(&fput_lock);
 518        while (likely(!list_empty(&fput_head))) {
 519                struct kiocb *req = list_kiocb(fput_head.next);
 520                struct kioctx *ctx = req->ki_ctx;
 521
 522                list_del(&req->ki_list);
 523                spin_unlock_irq(&fput_lock);
 524
 525                /* Complete the fput(s) */
 526                if (req->ki_filp != NULL)
 527                        fput(req->ki_filp);
 528
 529                /* Link the iocb into the context's free list */
 530                spin_lock_irq(&ctx->ctx_lock);
 531                really_put_req(ctx, req);
 532                spin_unlock_irq(&ctx->ctx_lock);
 533
 534                put_ioctx(ctx);
 535                spin_lock_irq(&fput_lock);
 536        }
 537        spin_unlock_irq(&fput_lock);
 538}
 539
 540/* __aio_put_req
 541 *      Returns true if this put was the last user of the request.
 542 */
 543static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
 544{
 545        dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
 546                req, atomic_long_read(&req->ki_filp->f_count));
 547
 548        assert_spin_locked(&ctx->ctx_lock);
 549
 550        req->ki_users--;
 551        BUG_ON(req->ki_users < 0);
 552        if (likely(req->ki_users))
 553                return 0;
 554        list_del(&req->ki_list);                /* remove from active_reqs */
 555        req->ki_cancel = NULL;
 556        req->ki_retry = NULL;
 557
 558        /*
 559         * Try to optimize the aio and eventfd file* puts, by avoiding to
 560         * schedule work in case it is not final fput() time. In normal cases,
 561         * we would not be holding the last reference to the file*, so
 562         * this function will be executed w/out any aio kthread wakeup.
 563         */
 564        if (unlikely(!fput_atomic(req->ki_filp))) {
 565                get_ioctx(ctx);
 566                spin_lock(&fput_lock);
 567                list_add(&req->ki_list, &fput_head);
 568                spin_unlock(&fput_lock);
 569                schedule_work(&fput_work);
 570        } else {
 571                req->ki_filp = NULL;
 572                really_put_req(ctx, req);
 573        }
 574        return 1;
 575}
 576
 577/* aio_put_req
 578 *      Returns true if this put was the last user of the kiocb,
 579 *      false if the request is still in use.
 580 */
 581int aio_put_req(struct kiocb *req)
 582{
 583        struct kioctx *ctx = req->ki_ctx;
 584        int ret;
 585        spin_lock_irq(&ctx->ctx_lock);
 586        ret = __aio_put_req(ctx, req);
 587        spin_unlock_irq(&ctx->ctx_lock);
 588        return ret;
 589}
 590EXPORT_SYMBOL(aio_put_req);
 591
 592static struct kioctx *lookup_ioctx(unsigned long ctx_id)
 593{
 594        struct mm_struct *mm = current->mm;
 595        struct kioctx *ctx, *ret = NULL;
 596        struct hlist_node *n;
 597
 598        rcu_read_lock();
 599
 600        hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) {
 601                /*
 602                 * RCU protects us against accessing freed memory but
 603                 * we have to be careful not to get a reference when the
 604                 * reference count already dropped to 0 (ctx->dead test
 605                 * is unreliable because of races).
 606                 */
 607                if (ctx->user_id == ctx_id && !ctx->dead && try_get_ioctx(ctx)){
 608                        ret = ctx;
 609                        break;
 610                }
 611        }
 612
 613        rcu_read_unlock();
 614        return ret;
 615}
 616
 617/*
 618 * Queue up a kiocb to be retried. Assumes that the kiocb
 619 * has already been marked as kicked, and places it on
 620 * the retry run list for the corresponding ioctx, if it
 621 * isn't already queued. Returns 1 if it actually queued
 622 * the kiocb (to tell the caller to activate the work
 623 * queue to process it), or 0, if it found that it was
 624 * already queued.
 625 */
 626static inline int __queue_kicked_iocb(struct kiocb *iocb)
 627{
 628        struct kioctx *ctx = iocb->ki_ctx;
 629
 630        assert_spin_locked(&ctx->ctx_lock);
 631
 632        if (list_empty(&iocb->ki_run_list)) {
 633                list_add_tail(&iocb->ki_run_list,
 634                        &ctx->run_list);
 635                return 1;
 636        }
 637        return 0;
 638}
 639
 640/* aio_run_iocb
 641 *      This is the core aio execution routine. It is
 642 *      invoked both for initial i/o submission and
 643 *      subsequent retries via the aio_kick_handler.
 644 *      Expects to be invoked with iocb->ki_ctx->lock
 645 *      already held. The lock is released and reacquired
 646 *      as needed during processing.
 647 *
 648 * Calls the iocb retry method (already setup for the
 649 * iocb on initial submission) for operation specific
 650 * handling, but takes care of most of common retry
 651 * execution details for a given iocb. The retry method
 652 * needs to be non-blocking as far as possible, to avoid
 653 * holding up other iocbs waiting to be serviced by the
 654 * retry kernel thread.
 655 *
 656 * The trickier parts in this code have to do with
 657 * ensuring that only one retry instance is in progress
 658 * for a given iocb at any time. Providing that guarantee
 659 * simplifies the coding of individual aio operations as
 660 * it avoids various potential races.
 661 */
 662static ssize_t aio_run_iocb(struct kiocb *iocb)
 663{
 664        struct kioctx   *ctx = iocb->ki_ctx;
 665        ssize_t (*retry)(struct kiocb *);
 666        ssize_t ret;
 667
 668        if (!(retry = iocb->ki_retry)) {
 669                printk("aio_run_iocb: iocb->ki_retry = NULL\n");
 670                return 0;
 671        }
 672
 673        /*
 674         * We don't want the next retry iteration for this
 675         * operation to start until this one has returned and
 676         * updated the iocb state. However, wait_queue functions
 677         * can trigger a kick_iocb from interrupt context in the
 678         * meantime, indicating that data is available for the next
 679         * iteration. We want to remember that and enable the
 680         * next retry iteration _after_ we are through with
 681         * this one.
 682         *
 683         * So, in order to be able to register a "kick", but
 684         * prevent it from being queued now, we clear the kick
 685         * flag, but make the kick code *think* that the iocb is
 686         * still on the run list until we are actually done.
 687         * When we are done with this iteration, we check if
 688         * the iocb was kicked in the meantime and if so, queue
 689         * it up afresh.
 690         */
 691
 692        kiocbClearKicked(iocb);
 693
 694        /*
 695         * This is so that aio_complete knows it doesn't need to
 696         * pull the iocb off the run list (We can't just call
 697         * INIT_LIST_HEAD because we don't want a kick_iocb to
 698         * queue this on the run list yet)
 699         */
 700        iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
 701        spin_unlock_irq(&ctx->ctx_lock);
 702
 703        /* Quit retrying if the i/o has been cancelled */
 704        if (kiocbIsCancelled(iocb)) {
 705                ret = -EINTR;
 706                aio_complete(iocb, ret, 0);
 707                /* must not access the iocb after this */
 708                goto out;
 709        }
 710
 711        /*
 712         * Now we are all set to call the retry method in async
 713         * context.
 714         */
 715        ret = retry(iocb);
 716
 717        if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
 718                /*
 719                 * There's no easy way to restart the syscall since other AIO's
 720                 * may be already running. Just fail this IO with EINTR.
 721                 */
 722                if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
 723                             ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
 724                        ret = -EINTR;
 725                aio_complete(iocb, ret, 0);
 726        }
 727out:
 728        spin_lock_irq(&ctx->ctx_lock);
 729
 730        if (-EIOCBRETRY == ret) {
 731                /*
 732                 * OK, now that we are done with this iteration
 733                 * and know that there is more left to go,
 734                 * this is where we let go so that a subsequent
 735                 * "kick" can start the next iteration
 736                 */
 737
 738                /* will make __queue_kicked_iocb succeed from here on */
 739                INIT_LIST_HEAD(&iocb->ki_run_list);
 740                /* we must queue the next iteration ourselves, if it
 741                 * has already been kicked */
 742                if (kiocbIsKicked(iocb)) {
 743                        __queue_kicked_iocb(iocb);
 744
 745                        /*
 746                         * __queue_kicked_iocb will always return 1 here, because
 747                         * iocb->ki_run_list is empty at this point so it should
 748                         * be safe to unconditionally queue the context into the
 749                         * work queue.
 750                         */
 751                        aio_queue_work(ctx);
 752                }
 753        }
 754        return ret;
 755}
 756
 757/*
 758 * __aio_run_iocbs:
 759 *      Process all pending retries queued on the ioctx
 760 *      run list.
 761 * Assumes it is operating within the aio issuer's mm
 762 * context.
 763 */
 764static int __aio_run_iocbs(struct kioctx *ctx)
 765{
 766        struct kiocb *iocb;
 767        struct list_head run_list;
 768
 769        assert_spin_locked(&ctx->ctx_lock);
 770
 771        list_replace_init(&ctx->run_list, &run_list);
 772        while (!list_empty(&run_list)) {
 773                iocb = list_entry(run_list.next, struct kiocb,
 774                        ki_run_list);
 775                list_del(&iocb->ki_run_list);
 776                /*
 777                 * Hold an extra reference while retrying i/o.
 778                 */
 779                iocb->ki_users++;       /* grab extra reference */
 780                aio_run_iocb(iocb);
 781                __aio_put_req(ctx, iocb);
 782        }
 783        if (!list_empty(&ctx->run_list))
 784                return 1;
 785        return 0;
 786}
 787
 788static void aio_queue_work(struct kioctx * ctx)
 789{
 790        unsigned long timeout;
 791        /*
 792         * if someone is waiting, get the work started right
 793         * away, otherwise, use a longer delay
 794         */
 795        smp_mb();
 796        if (waitqueue_active(&ctx->wait))
 797                timeout = 1;
 798        else
 799                timeout = HZ/10;
 800        queue_delayed_work(aio_wq, &ctx->wq, timeout);
 801}
 802
 803/*
 804 * aio_run_all_iocbs:
 805 *      Process all pending retries queued on the ioctx
 806 *      run list, and keep running them until the list
 807 *      stays empty.
 808 * Assumes it is operating within the aio issuer's mm context.
 809 */
 810static inline void aio_run_all_iocbs(struct kioctx *ctx)
 811{
 812        spin_lock_irq(&ctx->ctx_lock);
 813        while (__aio_run_iocbs(ctx))
 814                ;
 815        spin_unlock_irq(&ctx->ctx_lock);
 816}
 817
 818/*
 819 * aio_kick_handler:
 820 *      Work queue handler triggered to process pending
 821 *      retries on an ioctx. Takes on the aio issuer's
 822 *      mm context before running the iocbs, so that
 823 *      copy_xxx_user operates on the issuer's address
 824 *      space.
 825 * Run on aiod's context.
 826 */
 827static void aio_kick_handler(struct work_struct *work)
 828{
 829        struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
 830        mm_segment_t oldfs = get_fs();
 831        struct mm_struct *mm;
 832        int requeue;
 833
 834        set_fs(USER_DS);
 835        use_mm(ctx->mm);
 836        spin_lock_irq(&ctx->ctx_lock);
 837        requeue =__aio_run_iocbs(ctx);
 838        mm = ctx->mm;
 839        spin_unlock_irq(&ctx->ctx_lock);
 840        unuse_mm(mm);
 841        set_fs(oldfs);
 842        /*
 843         * we're in a worker thread already, don't use queue_delayed_work,
 844         */
 845        if (requeue)
 846                queue_delayed_work(aio_wq, &ctx->wq, 0);
 847}
 848
 849
 850/*
 851 * Called by kick_iocb to queue the kiocb for retry
 852 * and if required activate the aio work queue to process
 853 * it
 854 */
 855static void try_queue_kicked_iocb(struct kiocb *iocb)
 856{
 857        struct kioctx   *ctx = iocb->ki_ctx;
 858        unsigned long flags;
 859        int run = 0;
 860
 861        spin_lock_irqsave(&ctx->ctx_lock, flags);
 862        /* set this inside the lock so that we can't race with aio_run_iocb()
 863         * testing it and putting the iocb on the run list under the lock */
 864        if (!kiocbTryKick(iocb))
 865                run = __queue_kicked_iocb(iocb);
 866        spin_unlock_irqrestore(&ctx->ctx_lock, flags);
 867        if (run)
 868                aio_queue_work(ctx);
 869}
 870
 871/*
 872 * kick_iocb:
 873 *      Called typically from a wait queue callback context
 874 *      to trigger a retry of the iocb.
 875 *      The retry is usually executed by aio workqueue
 876 *      threads (See aio_kick_handler).
 877 */
 878void kick_iocb(struct kiocb *iocb)
 879{
 880        /* sync iocbs are easy: they can only ever be executing from a 
 881         * single context. */
 882        if (is_sync_kiocb(iocb)) {
 883                kiocbSetKicked(iocb);
 884                wake_up_process(iocb->ki_obj.tsk);
 885                return;
 886        }
 887
 888        try_queue_kicked_iocb(iocb);
 889}
 890EXPORT_SYMBOL(kick_iocb);
 891
 892/* aio_complete
 893 *      Called when the io request on the given iocb is complete.
 894 *      Returns true if this is the last user of the request.  The 
 895 *      only other user of the request can be the cancellation code.
 896 */
 897int aio_complete(struct kiocb *iocb, long res, long res2)
 898{
 899        struct kioctx   *ctx = iocb->ki_ctx;
 900        struct aio_ring_info    *info;
 901        struct aio_ring *ring;
 902        struct io_event *event;
 903        unsigned long   flags;
 904        unsigned long   tail;
 905        int             ret;
 906
 907        /*
 908         * Special case handling for sync iocbs:
 909         *  - events go directly into the iocb for fast handling
 910         *  - the sync task with the iocb in its stack holds the single iocb
 911         *    ref, no other paths have a way to get another ref
 912         *  - the sync task helpfully left a reference to itself in the iocb
 913         */
 914        if (is_sync_kiocb(iocb)) {
 915                BUG_ON(iocb->ki_users != 1);
 916                iocb->ki_user_data = res;
 917                iocb->ki_users = 0;
 918                wake_up_process(iocb->ki_obj.tsk);
 919                return 1;
 920        }
 921
 922        info = &ctx->ring_info;
 923
 924        /* add a completion event to the ring buffer.
 925         * must be done holding ctx->ctx_lock to prevent
 926         * other code from messing with the tail
 927         * pointer since we might be called from irq
 928         * context.
 929         */
 930        spin_lock_irqsave(&ctx->ctx_lock, flags);
 931
 932        if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
 933                list_del_init(&iocb->ki_run_list);
 934
 935        /*
 936         * cancelled requests don't get events, userland was given one
 937         * when the event got cancelled.
 938         */
 939        if (kiocbIsCancelled(iocb))
 940                goto put_rq;
 941
 942        ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
 943
 944        tail = info->tail;
 945        event = aio_ring_event(info, tail, KM_IRQ0);
 946        if (++tail >= info->nr)
 947                tail = 0;
 948
 949        event->obj = (u64)(unsigned long)iocb->ki_obj.user;
 950        event->data = iocb->ki_user_data;
 951        event->res = res;
 952        event->res2 = res2;
 953
 954        dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
 955                ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
 956                res, res2);
 957
 958        /* after flagging the request as done, we
 959         * must never even look at it again
 960         */
 961        smp_wmb();      /* make event visible before updating tail */
 962
 963        info->tail = tail;
 964        ring->tail = tail;
 965
 966        put_aio_ring_event(event, KM_IRQ0);
 967        kunmap_atomic(ring, KM_IRQ1);
 968
 969        pr_debug("added to ring %p at [%lu]\n", iocb, tail);
 970
 971        /*
 972         * Check if the user asked us to deliver the result through an
 973         * eventfd. The eventfd_signal() function is safe to be called
 974         * from IRQ context.
 975         */
 976        if (iocb->ki_eventfd != NULL)
 977                eventfd_signal(iocb->ki_eventfd, 1);
 978
 979put_rq:
 980        /* everything turned out well, dispose of the aiocb. */
 981        ret = __aio_put_req(ctx, iocb);
 982
 983        /*
 984         * We have to order our ring_info tail store above and test
 985         * of the wait list below outside the wait lock.  This is
 986         * like in wake_up_bit() where clearing a bit has to be
 987         * ordered with the unlocked test.
 988         */
 989        smp_mb();
 990
 991        if (waitqueue_active(&ctx->wait))
 992                wake_up(&ctx->wait);
 993
 994        spin_unlock_irqrestore(&ctx->ctx_lock, flags);
 995        return ret;
 996}
 997EXPORT_SYMBOL(aio_complete);
 998
 999/* aio_read_evt
1000 *      Pull an event off of the ioctx's event ring.  Returns the number of 
1001 *      events fetched (0 or 1 ;-)
1002 *      FIXME: make this use cmpxchg.
1003 *      TODO: make the ringbuffer user mmap()able (requires FIXME).
1004 */
1005static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1006{
1007        struct aio_ring_info *info = &ioctx->ring_info;
1008        struct aio_ring *ring;
1009        unsigned long head;
1010        int ret = 0;
1011
1012        ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1013        dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1014                 (unsigned long)ring->head, (unsigned long)ring->tail,
1015                 (unsigned long)ring->nr);
1016
1017        if (ring->head == ring->tail)
1018                goto out;
1019
1020        spin_lock(&info->ring_lock);
1021
1022        head = ring->head % info->nr;
1023        if (head != ring->tail) {
1024                struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1025                *ent = *evp;
1026                head = (head + 1) % info->nr;
1027                smp_mb(); /* finish reading the event before updatng the head */
1028                ring->head = head;
1029                ret = 1;
1030                put_aio_ring_event(evp, KM_USER1);
1031        }
1032        spin_unlock(&info->ring_lock);
1033
1034out:
1035        kunmap_atomic(ring, KM_USER0);
1036        dprintk("leaving aio_read_evt: %d  h%lu t%lu\n", ret,
1037                 (unsigned long)ring->head, (unsigned long)ring->tail);
1038        return ret;
1039}
1040
1041struct aio_timeout {
1042        struct timer_list       timer;
1043        int                     timed_out;
1044        struct task_struct      *p;
1045};
1046
1047static void timeout_func(unsigned long data)
1048{
1049        struct aio_timeout *to = (struct aio_timeout *)data;
1050
1051        to->timed_out = 1;
1052        wake_up_process(to->p);
1053}
1054
1055static inline void init_timeout(struct aio_timeout *to)
1056{
1057        setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
1058        to->timed_out = 0;
1059        to->p = current;
1060}
1061
1062static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1063                               const struct timespec *ts)
1064{
1065        to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1066        if (time_after(to->timer.expires, jiffies))
1067                add_timer(&to->timer);
1068        else
1069                to->timed_out = 1;
1070}
1071
1072static inline void clear_timeout(struct aio_timeout *to)
1073{
1074        del_singleshot_timer_sync(&to->timer);
1075}
1076
1077static int read_events(struct kioctx *ctx,
1078                        long min_nr, long nr,
1079                        struct io_event __user *event,
1080                        struct timespec __user *timeout)
1081{
1082        long                    start_jiffies = jiffies;
1083        struct task_struct      *tsk = current;
1084        DECLARE_WAITQUEUE(wait, tsk);
1085        int                     ret;
1086        int                     i = 0;
1087        struct io_event         ent;
1088        struct aio_timeout      to;
1089        int                     retry = 0;
1090
1091        /* needed to zero any padding within an entry (there shouldn't be 
1092         * any, but C is fun!
1093         */
1094        memset(&ent, 0, sizeof(ent));
1095retry:
1096        ret = 0;
1097        while (likely(i < nr)) {
1098                ret = aio_read_evt(ctx, &ent);
1099                if (unlikely(ret <= 0))
1100                        break;
1101
1102                dprintk("read event: %Lx %Lx %Lx %Lx\n",
1103                        ent.data, ent.obj, ent.res, ent.res2);
1104
1105                /* Could we split the check in two? */
1106                ret = -EFAULT;
1107                if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1108                        dprintk("aio: lost an event due to EFAULT.\n");
1109                        break;
1110                }
1111                ret = 0;
1112
1113                /* Good, event copied to userland, update counts. */
1114                event ++;
1115                i ++;
1116        }
1117
1118        if (min_nr <= i)
1119                return i;
1120        if (ret)
1121                return ret;
1122
1123        /* End fast path */
1124
1125        /* racey check, but it gets redone */
1126        if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1127                retry = 1;
1128                aio_run_all_iocbs(ctx);
1129                goto retry;
1130        }
1131
1132        init_timeout(&to);
1133        if (timeout) {
1134                struct timespec ts;
1135                ret = -EFAULT;
1136                if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1137                        goto out;
1138
1139                set_timeout(start_jiffies, &to, &ts);
1140        }
1141
1142        while (likely(i < nr)) {
1143                add_wait_queue_exclusive(&ctx->wait, &wait);
1144                do {
1145                        set_task_state(tsk, TASK_INTERRUPTIBLE);
1146                        ret = aio_read_evt(ctx, &ent);
1147                        if (ret)
1148                                break;
1149                        if (min_nr <= i)
1150                                break;
1151                        if (unlikely(ctx->dead)) {
1152                                ret = -EINVAL;
1153                                break;
1154                        }
1155                        if (to.timed_out)       /* Only check after read evt */
1156                                break;
1157                        /* Try to only show up in io wait if there are ops
1158                         *  in flight */
1159                        if (ctx->reqs_active)
1160                                io_schedule();
1161                        else
1162                                schedule();
1163                        if (signal_pending(tsk)) {
1164                                ret = -EINTR;
1165                                break;
1166                        }
1167                        /*ret = aio_read_evt(ctx, &ent);*/
1168                } while (1) ;
1169
1170                set_task_state(tsk, TASK_RUNNING);
1171                remove_wait_queue(&ctx->wait, &wait);
1172
1173                if (unlikely(ret <= 0))
1174                        break;
1175
1176                ret = -EFAULT;
1177                if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1178                        dprintk("aio: lost an event due to EFAULT.\n");
1179                        break;
1180                }
1181
1182                /* Good, event copied to userland, update counts. */
1183                event ++;
1184                i ++;
1185        }
1186
1187        if (timeout)
1188                clear_timeout(&to);
1189out:
1190        destroy_timer_on_stack(&to.timer);
1191        return i ? i : ret;
1192}
1193
1194/* Take an ioctx and remove it from the list of ioctx's.  Protects 
1195 * against races with itself via ->dead.
1196 */
1197static void io_destroy(struct kioctx *ioctx)
1198{
1199        struct mm_struct *mm = current->mm;
1200        int was_dead;
1201
1202        /* delete the entry from the list is someone else hasn't already */
1203        spin_lock(&mm->ioctx_lock);
1204        was_dead = ioctx->dead;
1205        ioctx->dead = 1;
1206        hlist_del_rcu(&ioctx->list);
1207        spin_unlock(&mm->ioctx_lock);
1208
1209        dprintk("aio_release(%p)\n", ioctx);
1210        if (likely(!was_dead))
1211                put_ioctx(ioctx);       /* twice for the list */
1212
1213        aio_cancel_all(ioctx);
1214        wait_for_all_aios(ioctx);
1215
1216        /*
1217         * Wake up any waiters.  The setting of ctx->dead must be seen
1218         * by other CPUs at this point.  Right now, we rely on the
1219         * locking done by the above calls to ensure this consistency.
1220         */
1221        wake_up_all(&ioctx->wait);
1222        put_ioctx(ioctx);       /* once for the lookup */
1223}
1224
1225/* sys_io_setup:
1226 *      Create an aio_context capable of receiving at least nr_events.
1227 *      ctxp must not point to an aio_context that already exists, and
1228 *      must be initialized to 0 prior to the call.  On successful
1229 *      creation of the aio_context, *ctxp is filled in with the resulting 
1230 *      handle.  May fail with -EINVAL if *ctxp is not initialized,
1231 *      if the specified nr_events exceeds internal limits.  May fail 
1232 *      with -EAGAIN if the specified nr_events exceeds the user's limit 
1233 *      of available events.  May fail with -ENOMEM if insufficient kernel
1234 *      resources are available.  May fail with -EFAULT if an invalid
1235 *      pointer is passed for ctxp.  Will fail with -ENOSYS if not
1236 *      implemented.
1237 */
1238SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1239{
1240        struct kioctx *ioctx = NULL;
1241        unsigned long ctx;
1242        long ret;
1243
1244        ret = get_user(ctx, ctxp);
1245        if (unlikely(ret))
1246                goto out;
1247
1248        ret = -EINVAL;
1249        if (unlikely(ctx || nr_events == 0)) {
1250                pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1251                         ctx, nr_events);
1252                goto out;
1253        }
1254
1255        ioctx = ioctx_alloc(nr_events);
1256        ret = PTR_ERR(ioctx);
1257        if (!IS_ERR(ioctx)) {
1258                ret = put_user(ioctx->user_id, ctxp);
1259                if (!ret)
1260                        return 0;
1261
1262                get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1263                io_destroy(ioctx);
1264        }
1265
1266out:
1267        return ret;
1268}
1269
1270/* sys_io_destroy:
1271 *      Destroy the aio_context specified.  May cancel any outstanding 
1272 *      AIOs and block on completion.  Will fail with -ENOSYS if not
1273 *      implemented.  May fail with -EINVAL if the context pointed to
1274 *      is invalid.
1275 */
1276SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1277{
1278        struct kioctx *ioctx = lookup_ioctx(ctx);
1279        if (likely(NULL != ioctx)) {
1280                io_destroy(ioctx);
1281                return 0;
1282        }
1283        pr_debug("EINVAL: io_destroy: invalid context id\n");
1284        return -EINVAL;
1285}
1286
1287static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1288{
1289        struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1290
1291        BUG_ON(ret <= 0);
1292
1293        while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1294                ssize_t this = min((ssize_t)iov->iov_len, ret);
1295                iov->iov_base += this;
1296                iov->iov_len -= this;
1297                iocb->ki_left -= this;
1298                ret -= this;
1299                if (iov->iov_len == 0) {
1300                        iocb->ki_cur_seg++;
1301                        iov++;
1302                }
1303        }
1304
1305        /* the caller should not have done more io than what fit in
1306         * the remaining iovecs */
1307        BUG_ON(ret > 0 && iocb->ki_left == 0);
1308}
1309
1310static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1311{
1312        struct file *file = iocb->ki_filp;
1313        struct address_space *mapping = file->f_mapping;
1314        struct inode *inode = mapping->host;
1315        ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1316                         unsigned long, loff_t);
1317        ssize_t ret = 0;
1318        unsigned short opcode;
1319
1320        if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1321                (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1322                rw_op = file->f_op->aio_read;
1323                opcode = IOCB_CMD_PREADV;
1324        } else {
1325                rw_op = file->f_op->aio_write;
1326                opcode = IOCB_CMD_PWRITEV;
1327        }
1328
1329        /* This matches the pread()/pwrite() logic */
1330        if (iocb->ki_pos < 0)
1331                return -EINVAL;
1332
1333        do {
1334                ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1335                            iocb->ki_nr_segs - iocb->ki_cur_seg,
1336                            iocb->ki_pos);
1337                if (ret > 0)
1338                        aio_advance_iovec(iocb, ret);
1339
1340        /* retry all partial writes.  retry partial reads as long as its a
1341         * regular file. */
1342        } while (ret > 0 && iocb->ki_left > 0 &&
1343                 (opcode == IOCB_CMD_PWRITEV ||
1344                  (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1345
1346        /* This means we must have transferred all that we could */
1347        /* No need to retry anymore */
1348        if ((ret == 0) || (iocb->ki_left == 0))
1349                ret = iocb->ki_nbytes - iocb->ki_left;
1350
1351        /* If we managed to write some out we return that, rather than
1352         * the eventual error. */
1353        if (opcode == IOCB_CMD_PWRITEV
1354            && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
1355            && iocb->ki_nbytes - iocb->ki_left)
1356                ret = iocb->ki_nbytes - iocb->ki_left;
1357
1358        return ret;
1359}
1360
1361static ssize_t aio_fdsync(struct kiocb *iocb)
1362{
1363        struct file *file = iocb->ki_filp;
1364        ssize_t ret = -EINVAL;
1365
1366        if (file->f_op->aio_fsync)
1367                ret = file->f_op->aio_fsync(iocb, 1);
1368        return ret;
1369}
1370
1371static ssize_t aio_fsync(struct kiocb *iocb)
1372{
1373        struct file *file = iocb->ki_filp;
1374        ssize_t ret = -EINVAL;
1375
1376        if (file->f_op->aio_fsync)
1377                ret = file->f_op->aio_fsync(iocb, 0);
1378        return ret;
1379}
1380
1381static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
1382{
1383        ssize_t ret;
1384
1385#ifdef CONFIG_COMPAT
1386        if (compat)
1387                ret = compat_rw_copy_check_uvector(type,
1388                                (struct compat_iovec __user *)kiocb->ki_buf,
1389                                kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1390                                &kiocb->ki_iovec);
1391        else
1392#endif
1393                ret = rw_copy_check_uvector(type,
1394                                (struct iovec __user *)kiocb->ki_buf,
1395                                kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1396                                &kiocb->ki_iovec);
1397        if (ret < 0)
1398                goto out;
1399
1400        kiocb->ki_nr_segs = kiocb->ki_nbytes;
1401        kiocb->ki_cur_seg = 0;
1402        /* ki_nbytes/left now reflect bytes instead of segs */
1403        kiocb->ki_nbytes = ret;
1404        kiocb->ki_left = ret;
1405
1406        ret = 0;
1407out:
1408        return ret;
1409}
1410
1411static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
1412{
1413        kiocb->ki_iovec = &kiocb->ki_inline_vec;
1414        kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1415        kiocb->ki_iovec->iov_len = kiocb->ki_left;
1416        kiocb->ki_nr_segs = 1;
1417        kiocb->ki_cur_seg = 0;
1418        return 0;
1419}
1420
1421/*
1422 * aio_setup_iocb:
1423 *      Performs the initial checks and aio retry method
1424 *      setup for the kiocb at the time of io submission.
1425 */
1426static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
1427{
1428        struct file *file = kiocb->ki_filp;
1429        ssize_t ret = 0;
1430
1431        switch (kiocb->ki_opcode) {
1432        case IOCB_CMD_PREAD:
1433                ret = -EBADF;
1434                if (unlikely(!(file->f_mode & FMODE_READ)))
1435                        break;
1436                ret = -EFAULT;
1437                if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1438                        kiocb->ki_left)))
1439                        break;
1440                ret = security_file_permission(file, MAY_READ);
1441                if (unlikely(ret))
1442                        break;
1443                ret = aio_setup_single_vector(kiocb);
1444                if (ret)
1445                        break;
1446                ret = -EINVAL;
1447                if (file->f_op->aio_read)
1448                        kiocb->ki_retry = aio_rw_vect_retry;
1449                break;
1450        case IOCB_CMD_PWRITE:
1451                ret = -EBADF;
1452                if (unlikely(!(file->f_mode & FMODE_WRITE)))
1453                        break;
1454                ret = -EFAULT;
1455                if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1456                        kiocb->ki_left)))
1457                        break;
1458                ret = security_file_permission(file, MAY_WRITE);
1459                if (unlikely(ret))
1460                        break;
1461                ret = aio_setup_single_vector(kiocb);
1462                if (ret)
1463                        break;
1464                ret = -EINVAL;
1465                if (file->f_op->aio_write)
1466                        kiocb->ki_retry = aio_rw_vect_retry;
1467                break;
1468        case IOCB_CMD_PREADV:
1469                ret = -EBADF;
1470                if (unlikely(!(file->f_mode & FMODE_READ)))
1471                        break;
1472                ret = security_file_permission(file, MAY_READ);
1473                if (unlikely(ret))
1474                        break;
1475                ret = aio_setup_vectored_rw(READ, kiocb, compat);
1476                if (ret)
1477                        break;
1478                ret = -EINVAL;
1479                if (file->f_op->aio_read)
1480                        kiocb->ki_retry = aio_rw_vect_retry;
1481                break;
1482        case IOCB_CMD_PWRITEV:
1483                ret = -EBADF;
1484                if (unlikely(!(file->f_mode & FMODE_WRITE)))
1485                        break;
1486                ret = security_file_permission(file, MAY_WRITE);
1487                if (unlikely(ret))
1488                        break;
1489                ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1490                if (ret)
1491                        break;
1492                ret = -EINVAL;
1493                if (file->f_op->aio_write)
1494                        kiocb->ki_retry = aio_rw_vect_retry;
1495                break;
1496        case IOCB_CMD_FDSYNC:
1497                ret = -EINVAL;
1498                if (file->f_op->aio_fsync)
1499                        kiocb->ki_retry = aio_fdsync;
1500                break;
1501        case IOCB_CMD_FSYNC:
1502                ret = -EINVAL;
1503                if (file->f_op->aio_fsync)
1504                        kiocb->ki_retry = aio_fsync;
1505                break;
1506        default:
1507                dprintk("EINVAL: io_submit: no operation provided\n");
1508                ret = -EINVAL;
1509        }
1510
1511        if (!kiocb->ki_retry)
1512                return ret;
1513
1514        return 0;
1515}
1516
1517static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1518                         struct iocb *iocb, bool compat)
1519{
1520        struct kiocb *req;
1521        struct file *file;
1522        ssize_t ret;
1523
1524        /* enforce forwards compatibility on users */
1525        if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1526                pr_debug("EINVAL: io_submit: reserve field set\n");
1527                return -EINVAL;
1528        }
1529
1530        /* prevent overflows */
1531        if (unlikely(
1532            (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1533            (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1534            ((ssize_t)iocb->aio_nbytes < 0)
1535           )) {
1536                pr_debug("EINVAL: io_submit: overflow check\n");
1537                return -EINVAL;
1538        }
1539
1540        file = fget(iocb->aio_fildes);
1541        if (unlikely(!file))
1542                return -EBADF;
1543
1544        req = aio_get_req(ctx);         /* returns with 2 references to req */
1545        if (unlikely(!req)) {
1546                fput(file);
1547                return -EAGAIN;
1548        }
1549        req->ki_filp = file;
1550        if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1551                /*
1552                 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1553                 * instance of the file* now. The file descriptor must be
1554                 * an eventfd() fd, and will be signaled for each completed
1555                 * event using the eventfd_signal() function.
1556                 */
1557                req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1558                if (IS_ERR(req->ki_eventfd)) {
1559                        ret = PTR_ERR(req->ki_eventfd);
1560                        req->ki_eventfd = NULL;
1561                        goto out_put_req;
1562                }
1563        }
1564
1565        ret = put_user(req->ki_key, &user_iocb->aio_key);
1566        if (unlikely(ret)) {
1567                dprintk("EFAULT: aio_key\n");
1568                goto out_put_req;
1569        }
1570
1571        req->ki_obj.user = user_iocb;
1572        req->ki_user_data = iocb->aio_data;
1573        req->ki_pos = iocb->aio_offset;
1574
1575        req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1576        req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1577        req->ki_opcode = iocb->aio_lio_opcode;
1578
1579        ret = aio_setup_iocb(req, compat);
1580
1581        if (ret)
1582                goto out_put_req;
1583
1584        spin_lock_irq(&ctx->ctx_lock);
1585        /*
1586         * We could have raced with io_destroy() and are currently holding a
1587         * reference to ctx which should be destroyed. We cannot submit IO
1588         * since ctx gets freed as soon as io_submit() puts its reference.  The
1589         * check here is reliable: io_destroy() sets ctx->dead before waiting
1590         * for outstanding IO and the barrier between these two is realized by
1591         * unlock of mm->ioctx_lock and lock of ctx->ctx_lock.  Analogously we
1592         * increment ctx->reqs_active before checking for ctx->dead and the
1593         * barrier is realized by unlock and lock of ctx->ctx_lock. Thus if we
1594         * don't see ctx->dead set here, io_destroy() waits for our IO to
1595         * finish.
1596         */
1597        if (ctx->dead) {
1598                spin_unlock_irq(&ctx->ctx_lock);
1599                ret = -EINVAL;
1600                goto out_put_req;
1601        }
1602        aio_run_iocb(req);
1603        if (!list_empty(&ctx->run_list)) {
1604                /* drain the run list */
1605                while (__aio_run_iocbs(ctx))
1606                        ;
1607        }
1608        spin_unlock_irq(&ctx->ctx_lock);
1609
1610        aio_put_req(req);       /* drop extra ref to req */
1611        return 0;
1612
1613out_put_req:
1614        aio_put_req(req);       /* drop extra ref to req */
1615        aio_put_req(req);       /* drop i/o ref to req */
1616        return ret;
1617}
1618
1619long do_io_submit(aio_context_t ctx_id, long nr,
1620                  struct iocb __user *__user *iocbpp, bool compat)
1621{
1622        struct kioctx *ctx;
1623        long ret = 0;
1624        int i;
1625        struct blk_plug plug;
1626
1627        if (unlikely(nr < 0))
1628                return -EINVAL;
1629
1630        if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1631                nr = LONG_MAX/sizeof(*iocbpp);
1632
1633        if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1634                return -EFAULT;
1635
1636        ctx = lookup_ioctx(ctx_id);
1637        if (unlikely(!ctx)) {
1638                pr_debug("EINVAL: io_submit: invalid context id\n");
1639                return -EINVAL;
1640        }
1641
1642        blk_start_plug(&plug);
1643
1644        /*
1645         * AKPM: should this return a partial result if some of the IOs were
1646         * successfully submitted?
1647         */
1648        for (i=0; i<nr; i++) {
1649                struct iocb __user *user_iocb;
1650                struct iocb tmp;
1651
1652                if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1653                        ret = -EFAULT;
1654                        break;
1655                }
1656
1657                if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1658                        ret = -EFAULT;
1659                        break;
1660                }
1661
1662                ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1663                if (ret)
1664                        break;
1665        }
1666        blk_finish_plug(&plug);
1667
1668        put_ioctx(ctx);
1669        return i ? i : ret;
1670}
1671
1672/* sys_io_submit:
1673 *      Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1674 *      the number of iocbs queued.  May return -EINVAL if the aio_context
1675 *      specified by ctx_id is invalid, if nr is < 0, if the iocb at
1676 *      *iocbpp[0] is not properly initialized, if the operation specified
1677 *      is invalid for the file descriptor in the iocb.  May fail with
1678 *      -EFAULT if any of the data structures point to invalid data.  May
1679 *      fail with -EBADF if the file descriptor specified in the first
1680 *      iocb is invalid.  May fail with -EAGAIN if insufficient resources
1681 *      are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1682 *      fail with -ENOSYS if not implemented.
1683 */
1684SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1685                struct iocb __user * __user *, iocbpp)
1686{
1687        return do_io_submit(ctx_id, nr, iocbpp, 0);
1688}
1689
1690/* lookup_kiocb
1691 *      Finds a given iocb for cancellation.
1692 */
1693static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1694                                  u32 key)
1695{
1696        struct list_head *pos;
1697
1698        assert_spin_locked(&ctx->ctx_lock);
1699
1700        /* TODO: use a hash or array, this sucks. */
1701        list_for_each(pos, &ctx->active_reqs) {
1702                struct kiocb *kiocb = list_kiocb(pos);
1703                if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1704                        return kiocb;
1705        }
1706        return NULL;
1707}
1708
1709/* sys_io_cancel:
1710 *      Attempts to cancel an iocb previously passed to io_submit.  If
1711 *      the operation is successfully cancelled, the resulting event is
1712 *      copied into the memory pointed to by result without being placed
1713 *      into the completion queue and 0 is returned.  May fail with
1714 *      -EFAULT if any of the data structures pointed to are invalid.
1715 *      May fail with -EINVAL if aio_context specified by ctx_id is
1716 *      invalid.  May fail with -EAGAIN if the iocb specified was not
1717 *      cancelled.  Will fail with -ENOSYS if not implemented.
1718 */
1719SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1720                struct io_event __user *, result)
1721{
1722        int (*cancel)(struct kiocb *iocb, struct io_event *res);
1723        struct kioctx *ctx;
1724        struct kiocb *kiocb;
1725        u32 key;
1726        int ret;
1727
1728        ret = get_user(key, &iocb->aio_key);
1729        if (unlikely(ret))
1730                return -EFAULT;
1731
1732        ctx = lookup_ioctx(ctx_id);
1733        if (unlikely(!ctx))
1734                return -EINVAL;
1735
1736        spin_lock_irq(&ctx->ctx_lock);
1737        ret = -EAGAIN;
1738        kiocb = lookup_kiocb(ctx, iocb, key);
1739        if (kiocb && kiocb->ki_cancel) {
1740                cancel = kiocb->ki_cancel;
1741                kiocb->ki_users ++;
1742                kiocbSetCancelled(kiocb);
1743        } else
1744                cancel = NULL;
1745        spin_unlock_irq(&ctx->ctx_lock);
1746
1747        if (NULL != cancel) {
1748                struct io_event tmp;
1749                pr_debug("calling cancel\n");
1750                memset(&tmp, 0, sizeof(tmp));
1751                tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1752                tmp.data = kiocb->ki_user_data;
1753                ret = cancel(kiocb, &tmp);
1754                if (!ret) {
1755                        /* Cancellation succeeded -- copy the result
1756                         * into the user's buffer.
1757                         */
1758                        if (copy_to_user(result, &tmp, sizeof(tmp)))
1759                                ret = -EFAULT;
1760                }
1761        } else
1762                ret = -EINVAL;
1763
1764        put_ioctx(ctx);
1765
1766        return ret;
1767}
1768
1769/* io_getevents:
1770 *      Attempts to read at least min_nr events and up to nr events from
1771 *      the completion queue for the aio_context specified by ctx_id. If
1772 *      it succeeds, the number of read events is returned. May fail with
1773 *      -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1774 *      out of range, if timeout is out of range.  May fail with -EFAULT
1775 *      if any of the memory specified is invalid.  May return 0 or
1776 *      < min_nr if the timeout specified by timeout has elapsed
1777 *      before sufficient events are available, where timeout == NULL
1778 *      specifies an infinite timeout. Note that the timeout pointed to by
1779 *      timeout is relative and will be updated if not NULL and the
1780 *      operation blocks. Will fail with -ENOSYS if not implemented.
1781 */
1782SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1783                long, min_nr,
1784                long, nr,
1785                struct io_event __user *, events,
1786                struct timespec __user *, timeout)
1787{
1788        struct kioctx *ioctx = lookup_ioctx(ctx_id);
1789        long ret = -EINVAL;
1790
1791        if (likely(ioctx)) {
1792                if (likely(min_nr <= nr && min_nr >= 0))
1793                        ret = read_events(ioctx, min_nr, nr, events, timeout);
1794                put_ioctx(ioctx);
1795        }
1796
1797        asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);
1798        return ret;
1799}
1800
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.