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/uio.h>
  19
  20#define DEBUG 0
  21
  22#include <linux/sched.h>
  23#include <linux/fs.h>
  24#include <linux/file.h>
  25#include <linux/mm.h>
  26#include <linux/mman.h>
  27#include <linux/slab.h>
  28#include <linux/timer.h>
  29#include <linux/aio.h>
  30#include <linux/highmem.h>
  31#include <linux/workqueue.h>
  32#include <linux/security.h>
  33#include <linux/eventfd.h>
  34
  35#include <asm/kmap_types.h>
  36#include <asm/uaccess.h>
  37#include <asm/mmu_context.h>
  38
  39#if DEBUG > 1
  40#define dprintk         printk
  41#else
  42#define dprintk(x...)   do { ; } while (0)
  43#endif
  44
  45/*------ sysctl variables----*/
  46static DEFINE_SPINLOCK(aio_nr_lock);
  47unsigned long aio_nr;           /* current system wide number of aio requests */
  48unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
  49/*----end sysctl variables---*/
  50
  51static struct kmem_cache        *kiocb_cachep;
  52static struct kmem_cache        *kioctx_cachep;
  53
  54static struct workqueue_struct *aio_wq;
  55
  56/* Used for rare fput completion. */
  57static void aio_fput_routine(struct work_struct *);
  58static DECLARE_WORK(fput_work, aio_fput_routine);
  59
  60static DEFINE_SPINLOCK(fput_lock);
  61static LIST_HEAD(fput_head);
  62
  63static void aio_kick_handler(struct work_struct *);
  64static void aio_queue_work(struct kioctx *);
  65
  66/* aio_setup
  67 *      Creates the slab caches used by the aio routines, panic on
  68 *      failure as this is done early during the boot sequence.
  69 */
  70static int __init aio_setup(void)
  71{
  72        kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
  73        kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
  74
  75        aio_wq = create_workqueue("aio");
  76
  77        pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
  78
  79        return 0;
  80}
  81
  82static void aio_free_ring(struct kioctx *ctx)
  83{
  84        struct aio_ring_info *info = &ctx->ring_info;
  85        long i;
  86
  87        for (i=0; i<info->nr_pages; i++)
  88                put_page(info->ring_pages[i]);
  89
  90        if (info->mmap_size) {
  91                down_write(&ctx->mm->mmap_sem);
  92                do_munmap(ctx->mm, info->mmap_base, info->mmap_size);
  93                up_write(&ctx->mm->mmap_sem);
  94        }
  95
  96        if (info->ring_pages && info->ring_pages != info->internal_pages)
  97                kfree(info->ring_pages);
  98        info->ring_pages = NULL;
  99        info->nr = 0;
 100}
 101
 102static int aio_setup_ring(struct kioctx *ctx)
 103{
 104        struct aio_ring *ring;
 105        struct aio_ring_info *info = &ctx->ring_info;
 106        unsigned nr_events = ctx->max_reqs;
 107        unsigned long size;
 108        int nr_pages;
 109
 110        /* Compensate for the ring buffer's head/tail overlap entry */
 111        nr_events += 2; /* 1 is required, 2 for good luck */
 112
 113        size = sizeof(struct aio_ring);
 114        size += sizeof(struct io_event) * nr_events;
 115        nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
 116
 117        if (nr_pages < 0)
 118                return -EINVAL;
 119
 120        nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
 121
 122        info->nr = 0;
 123        info->ring_pages = info->internal_pages;
 124        if (nr_pages > AIO_RING_PAGES) {
 125                info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
 126                if (!info->ring_pages)
 127                        return -ENOMEM;
 128        }
 129
 130        info->mmap_size = nr_pages * PAGE_SIZE;
 131        dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
 132        down_write(&ctx->mm->mmap_sem);
 133        info->mmap_base = do_mmap(NULL, 0, info->mmap_size, 
 134                                  PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE,
 135                                  0);
 136        if (IS_ERR((void *)info->mmap_base)) {
 137                up_write(&ctx->mm->mmap_sem);
 138                info->mmap_size = 0;
 139                aio_free_ring(ctx);
 140                return -EAGAIN;
 141        }
 142
 143        dprintk("mmap address: 0x%08lx\n", info->mmap_base);
 144        info->nr_pages = get_user_pages(current, ctx->mm,
 145                                        info->mmap_base, nr_pages, 
 146                                        1, 0, info->ring_pages, NULL);
 147        up_write(&ctx->mm->mmap_sem);
 148
 149        if (unlikely(info->nr_pages != nr_pages)) {
 150                aio_free_ring(ctx);
 151                return -EAGAIN;
 152        }
 153
 154        ctx->user_id = info->mmap_base;
 155
 156        info->nr = nr_events;           /* trusted copy */
 157
 158        ring = kmap_atomic(info->ring_pages[0], KM_USER0);
 159        ring->nr = nr_events;   /* user copy */
 160        ring->id = ctx->user_id;
 161        ring->head = ring->tail = 0;
 162        ring->magic = AIO_RING_MAGIC;
 163        ring->compat_features = AIO_RING_COMPAT_FEATURES;
 164        ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
 165        ring->header_length = sizeof(struct aio_ring);
 166        kunmap_atomic(ring, KM_USER0);
 167
 168        return 0;
 169}
 170
 171
 172/* aio_ring_event: returns a pointer to the event at the given index from
 173 * kmap_atomic(, km).  Release the pointer with put_aio_ring_event();
 174 */
 175#define AIO_EVENTS_PER_PAGE     (PAGE_SIZE / sizeof(struct io_event))
 176#define AIO_EVENTS_FIRST_PAGE   ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
 177#define AIO_EVENTS_OFFSET       (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
 178
 179#define aio_ring_event(info, nr, km) ({                                 \
 180        unsigned pos = (nr) + AIO_EVENTS_OFFSET;                        \
 181        struct io_event *__event;                                       \
 182        __event = kmap_atomic(                                          \
 183                        (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
 184        __event += pos % AIO_EVENTS_PER_PAGE;                           \
 185        __event;                                                        \
 186})
 187
 188#define put_aio_ring_event(event, km) do {      \
 189        struct io_event *__event = (event);     \
 190        (void)__event;                          \
 191        kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
 192} while(0)
 193
 194
 195/* __put_ioctx
 196 *      Called when the last user of an aio context has gone away,
 197 *      and the struct needs to be freed.
 198 */
 199static void __put_ioctx(struct kioctx *ctx)
 200{
 201        unsigned nr_events = ctx->max_reqs;
 202
 203        BUG_ON(ctx->reqs_active);
 204
 205        cancel_delayed_work(&ctx->wq);
 206        cancel_work_sync(&ctx->wq.work);
 207        aio_free_ring(ctx);
 208        mmdrop(ctx->mm);
 209        ctx->mm = NULL;
 210        pr_debug("__put_ioctx: freeing %p\n", ctx);
 211        kmem_cache_free(kioctx_cachep, ctx);
 212
 213        if (nr_events) {
 214                spin_lock(&aio_nr_lock);
 215                BUG_ON(aio_nr - nr_events > aio_nr);
 216                aio_nr -= nr_events;
 217                spin_unlock(&aio_nr_lock);
 218        }
 219}
 220
 221#define get_ioctx(kioctx) do {                                          \
 222        BUG_ON(atomic_read(&(kioctx)->users) <= 0);                     \
 223        atomic_inc(&(kioctx)->users);                                   \
 224} while (0)
 225#define put_ioctx(kioctx) do {                                          \
 226        BUG_ON(atomic_read(&(kioctx)->users) <= 0);                     \
 227        if (unlikely(atomic_dec_and_test(&(kioctx)->users)))            \
 228                __put_ioctx(kioctx);                                    \
 229} while (0)
 230
 231/* ioctx_alloc
 232 *      Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
 233 */
 234static struct kioctx *ioctx_alloc(unsigned nr_events)
 235{
 236        struct mm_struct *mm;
 237        struct kioctx *ctx;
 238
 239        /* Prevent overflows */
 240        if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
 241            (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
 242                pr_debug("ENOMEM: nr_events too high\n");
 243                return ERR_PTR(-EINVAL);
 244        }
 245
 246        if ((unsigned long)nr_events > aio_max_nr)
 247                return ERR_PTR(-EAGAIN);
 248
 249        ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
 250        if (!ctx)
 251                return ERR_PTR(-ENOMEM);
 252
 253        ctx->max_reqs = nr_events;
 254        mm = ctx->mm = current->mm;
 255        atomic_inc(&mm->mm_count);
 256
 257        atomic_set(&ctx->users, 1);
 258        spin_lock_init(&ctx->ctx_lock);
 259        spin_lock_init(&ctx->ring_info.ring_lock);
 260        init_waitqueue_head(&ctx->wait);
 261
 262        INIT_LIST_HEAD(&ctx->active_reqs);
 263        INIT_LIST_HEAD(&ctx->run_list);
 264        INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
 265
 266        if (aio_setup_ring(ctx) < 0)
 267                goto out_freectx;
 268
 269        /* limit the number of system wide aios */
 270        spin_lock(&aio_nr_lock);
 271        if (aio_nr + ctx->max_reqs > aio_max_nr ||
 272            aio_nr + ctx->max_reqs < aio_nr)
 273                ctx->max_reqs = 0;
 274        else
 275                aio_nr += ctx->max_reqs;
 276        spin_unlock(&aio_nr_lock);
 277        if (ctx->max_reqs == 0)
 278                goto out_cleanup;
 279
 280        /* now link into global list. */
 281        write_lock(&mm->ioctx_list_lock);
 282        ctx->next = mm->ioctx_list;
 283        mm->ioctx_list = ctx;
 284        write_unlock(&mm->ioctx_list_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        __put_ioctx(ctx);
 292        return ERR_PTR(-EAGAIN);
 293
 294out_freectx:
 295        mmdrop(mm);
 296        kmem_cache_free(kioctx_cachep, ctx);
 297        ctx = ERR_PTR(-ENOMEM);
 298
 299        dprintk("aio: error allocating ioctx %p\n", ctx);
 300        return ctx;
 301}
 302
 303/* aio_cancel_all
 304 *      Cancels all outstanding aio requests on an aio context.  Used 
 305 *      when the processes owning a context have all exited to encourage 
 306 *      the rapid destruction of the kioctx.
 307 */
 308static void aio_cancel_all(struct kioctx *ctx)
 309{
 310        int (*cancel)(struct kiocb *, struct io_event *);
 311        struct io_event res;
 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        spin_unlock_irq(&ctx->ctx_lock);
 328}
 329
 330static void wait_for_all_aios(struct kioctx *ctx)
 331{
 332        struct task_struct *tsk = current;
 333        DECLARE_WAITQUEUE(wait, tsk);
 334
 335        spin_lock_irq(&ctx->ctx_lock);
 336        if (!ctx->reqs_active)
 337                goto out;
 338
 339        add_wait_queue(&ctx->wait, &wait);
 340        set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 341        while (ctx->reqs_active) {
 342                spin_unlock_irq(&ctx->ctx_lock);
 343                io_schedule();
 344                set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 345                spin_lock_irq(&ctx->ctx_lock);
 346        }
 347        __set_task_state(tsk, TASK_RUNNING);
 348        remove_wait_queue(&ctx->wait, &wait);
 349
 350out:
 351        spin_unlock_irq(&ctx->ctx_lock);
 352}
 353
 354/* wait_on_sync_kiocb:
 355 *      Waits on the given sync kiocb to complete.
 356 */
 357ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
 358{
 359        while (iocb->ki_users) {
 360                set_current_state(TASK_UNINTERRUPTIBLE);
 361                if (!iocb->ki_users)
 362                        break;
 363                io_schedule();
 364        }
 365        __set_current_state(TASK_RUNNING);
 366        return iocb->ki_user_data;
 367}
 368
 369/* exit_aio: called when the last user of mm goes away.  At this point, 
 370 * there is no way for any new requests to be submited or any of the 
 371 * io_* syscalls to be called on the context.  However, there may be 
 372 * outstanding requests which hold references to the context; as they 
 373 * go away, they will call put_ioctx and release any pinned memory
 374 * associated with the request (held via struct page * references).
 375 */
 376void exit_aio(struct mm_struct *mm)
 377{
 378        struct kioctx *ctx = mm->ioctx_list;
 379        mm->ioctx_list = NULL;
 380        while (ctx) {
 381                struct kioctx *next = ctx->next;
 382                ctx->next = NULL;
 383                aio_cancel_all(ctx);
 384
 385                wait_for_all_aios(ctx);
 386                /*
 387                 * Ensure we don't leave the ctx on the aio_wq
 388                 */
 389                cancel_work_sync(&ctx->wq.work);
 390
 391                if (1 != atomic_read(&ctx->users))
 392                        printk(KERN_DEBUG
 393                                "exit_aio:ioctx still alive: %d %d %d\n",
 394                                atomic_read(&ctx->users), ctx->dead,
 395                                ctx->reqs_active);
 396                put_ioctx(ctx);
 397                ctx = next;
 398        }
 399}
 400
 401/* aio_get_req
 402 *      Allocate a slot for an aio request.  Increments the users count
 403 * of the kioctx so that the kioctx stays around until all requests are
 404 * complete.  Returns NULL if no requests are free.
 405 *
 406 * Returns with kiocb->users set to 2.  The io submit code path holds
 407 * an extra reference while submitting the i/o.
 408 * This prevents races between the aio code path referencing the
 409 * req (after submitting it) and aio_complete() freeing the req.
 410 */
 411static struct kiocb *__aio_get_req(struct kioctx *ctx)
 412{
 413        struct kiocb *req = NULL;
 414        struct aio_ring *ring;
 415        int okay = 0;
 416
 417        req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
 418        if (unlikely(!req))
 419                return NULL;
 420
 421        req->ki_flags = 0;
 422        req->ki_users = 2;
 423        req->ki_key = 0;
 424        req->ki_ctx = ctx;
 425        req->ki_cancel = NULL;
 426        req->ki_retry = NULL;
 427        req->ki_dtor = NULL;
 428        req->private = NULL;
 429        req->ki_iovec = NULL;
 430        INIT_LIST_HEAD(&req->ki_run_list);
 431        req->ki_eventfd = ERR_PTR(-EINVAL);
 432
 433        /* Check if the completion queue has enough free space to
 434         * accept an event from this io.
 435         */
 436        spin_lock_irq(&ctx->ctx_lock);
 437        ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
 438        if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
 439                list_add(&req->ki_list, &ctx->active_reqs);
 440                ctx->reqs_active++;
 441                okay = 1;
 442        }
 443        kunmap_atomic(ring, KM_USER0);
 444        spin_unlock_irq(&ctx->ctx_lock);
 445
 446        if (!okay) {
 447                kmem_cache_free(kiocb_cachep, req);
 448                req = NULL;
 449        }
 450
 451        return req;
 452}
 453
 454static inline struct kiocb *aio_get_req(struct kioctx *ctx)
 455{
 456        struct kiocb *req;
 457        /* Handle a potential starvation case -- should be exceedingly rare as 
 458         * requests will be stuck on fput_head only if the aio_fput_routine is 
 459         * delayed and the requests were the last user of the struct file.
 460         */
 461        req = __aio_get_req(ctx);
 462        if (unlikely(NULL == req)) {
 463                aio_fput_routine(NULL);
 464                req = __aio_get_req(ctx);
 465        }
 466        return req;
 467}
 468
 469static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
 470{
 471        assert_spin_locked(&ctx->ctx_lock);
 472
 473        if (!IS_ERR(req->ki_eventfd))
 474                fput(req->ki_eventfd);
 475        if (req->ki_dtor)
 476                req->ki_dtor(req);
 477        if (req->ki_iovec != &req->ki_inline_vec)
 478                kfree(req->ki_iovec);
 479        kmem_cache_free(kiocb_cachep, req);
 480        ctx->reqs_active--;
 481
 482        if (unlikely(!ctx->reqs_active && ctx->dead))
 483                wake_up(&ctx->wait);
 484}
 485
 486static void aio_fput_routine(struct work_struct *data)
 487{
 488        spin_lock_irq(&fput_lock);
 489        while (likely(!list_empty(&fput_head))) {
 490                struct kiocb *req = list_kiocb(fput_head.next);
 491                struct kioctx *ctx = req->ki_ctx;
 492
 493                list_del(&req->ki_list);
 494                spin_unlock_irq(&fput_lock);
 495
 496                /* Complete the fput */
 497                __fput(req->ki_filp);
 498
 499                /* Link the iocb into the context's free list */
 500                spin_lock_irq(&ctx->ctx_lock);
 501                really_put_req(ctx, req);
 502                spin_unlock_irq(&ctx->ctx_lock);
 503
 504                put_ioctx(ctx);
 505                spin_lock_irq(&fput_lock);
 506        }
 507        spin_unlock_irq(&fput_lock);
 508}
 509
 510/* __aio_put_req
 511 *      Returns true if this put was the last user of the request.
 512 */
 513static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
 514{
 515        dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
 516                req, atomic_long_read(&req->ki_filp->f_count));
 517
 518        assert_spin_locked(&ctx->ctx_lock);
 519
 520        req->ki_users --;
 521        BUG_ON(req->ki_users < 0);
 522        if (likely(req->ki_users))
 523                return 0;
 524        list_del(&req->ki_list);                /* remove from active_reqs */
 525        req->ki_cancel = NULL;
 526        req->ki_retry = NULL;
 527
 528        /* Must be done under the lock to serialise against cancellation.
 529         * Call this aio_fput as it duplicates fput via the fput_work.
 530         */
 531        if (unlikely(atomic_long_dec_and_test(&req->ki_filp->f_count))) {
 532                get_ioctx(ctx);
 533                spin_lock(&fput_lock);
 534                list_add(&req->ki_list, &fput_head);
 535                spin_unlock(&fput_lock);
 536                queue_work(aio_wq, &fput_work);
 537        } else
 538                really_put_req(ctx, req);
 539        return 1;
 540}
 541
 542/* aio_put_req
 543 *      Returns true if this put was the last user of the kiocb,
 544 *      false if the request is still in use.
 545 */
 546int aio_put_req(struct kiocb *req)
 547{
 548        struct kioctx *ctx = req->ki_ctx;
 549        int ret;
 550        spin_lock_irq(&ctx->ctx_lock);
 551        ret = __aio_put_req(ctx, req);
 552        spin_unlock_irq(&ctx->ctx_lock);
 553        return ret;
 554}
 555
 556static struct kioctx *lookup_ioctx(unsigned long ctx_id)
 557{
 558        struct kioctx *ioctx;
 559        struct mm_struct *mm;
 560
 561        mm = current->mm;
 562        read_lock(&mm->ioctx_list_lock);
 563        for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next)
 564                if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) {
 565                        get_ioctx(ioctx);
 566                        break;
 567                }
 568        read_unlock(&mm->ioctx_list_lock);
 569
 570        return ioctx;
 571}
 572
 573/*
 574 * use_mm
 575 *      Makes the calling kernel thread take on the specified
 576 *      mm context.
 577 *      Called by the retry thread execute retries within the
 578 *      iocb issuer's mm context, so that copy_from/to_user
 579 *      operations work seamlessly for aio.
 580 *      (Note: this routine is intended to be called only
 581 *      from a kernel thread context)
 582 */
 583static void use_mm(struct mm_struct *mm)
 584{
 585        struct mm_struct *active_mm;
 586        struct task_struct *tsk = current;
 587
 588        task_lock(tsk);
 589        active_mm = tsk->active_mm;
 590        atomic_inc(&mm->mm_count);
 591        tsk->mm = mm;
 592        tsk->active_mm = mm;
 593        switch_mm(active_mm, mm, tsk);
 594        task_unlock(tsk);
 595
 596        mmdrop(active_mm);
 597}
 598
 599/*
 600 * unuse_mm
 601 *      Reverses the effect of use_mm, i.e. releases the
 602 *      specified mm context which was earlier taken on
 603 *      by the calling kernel thread
 604 *      (Note: this routine is intended to be called only
 605 *      from a kernel thread context)
 606 */
 607static void unuse_mm(struct mm_struct *mm)
 608{
 609        struct task_struct *tsk = current;
 610
 611        task_lock(tsk);
 612        tsk->mm = NULL;
 613        /* active_mm is still 'mm' */
 614        enter_lazy_tlb(mm, tsk);
 615        task_unlock(tsk);
 616}
 617
 618/*
 619 * Queue up a kiocb to be retried. Assumes that the kiocb
 620 * has already been marked as kicked, and places it on
 621 * the retry run list for the corresponding ioctx, if it
 622 * isn't already queued. Returns 1 if it actually queued
 623 * the kiocb (to tell the caller to activate the work
 624 * queue to process it), or 0, if it found that it was
 625 * already queued.
 626 */
 627static inline int __queue_kicked_iocb(struct kiocb *iocb)
 628{
 629        struct kioctx *ctx = iocb->ki_ctx;
 630
 631        assert_spin_locked(&ctx->ctx_lock);
 632
 633        if (list_empty(&iocb->ki_run_list)) {
 634                list_add_tail(&iocb->ki_run_list,
 635                        &ctx->run_list);
 636                return 1;
 637        }
 638        return 0;
 639}
 640
 641/* aio_run_iocb
 642 *      This is the core aio execution routine. It is
 643 *      invoked both for initial i/o submission and
 644 *      subsequent retries via the aio_kick_handler.
 645 *      Expects to be invoked with iocb->ki_ctx->lock
 646 *      already held. The lock is released and reacquired
 647 *      as needed during processing.
 648 *
 649 * Calls the iocb retry method (already setup for the
 650 * iocb on initial submission) for operation specific
 651 * handling, but takes care of most of common retry
 652 * execution details for a given iocb. The retry method
 653 * needs to be non-blocking as far as possible, to avoid
 654 * holding up other iocbs waiting to be serviced by the
 655 * retry kernel thread.
 656 *
 657 * The trickier parts in this code have to do with
 658 * ensuring that only one retry instance is in progress
 659 * for a given iocb at any time. Providing that guarantee
 660 * simplifies the coding of individual aio operations as
 661 * it avoids various potential races.
 662 */
 663static ssize_t aio_run_iocb(struct kiocb *iocb)
 664{
 665        struct kioctx   *ctx = iocb->ki_ctx;
 666        ssize_t (*retry)(struct kiocb *);
 667        ssize_t ret;
 668
 669        if (!(retry = iocb->ki_retry)) {
 670                printk("aio_run_iocb: iocb->ki_retry = NULL\n");
 671                return 0;
 672        }
 673
 674        /*
 675         * We don't want the next retry iteration for this
 676         * operation to start until this one has returned and
 677         * updated the iocb state. However, wait_queue functions
 678         * can trigger a kick_iocb from interrupt context in the
 679         * meantime, indicating that data is available for the next
 680         * iteration. We want to remember that and enable the
 681         * next retry iteration _after_ we are through with
 682         * this one.
 683         *
 684         * So, in order to be able to register a "kick", but
 685         * prevent it from being queued now, we clear the kick
 686         * flag, but make the kick code *think* that the iocb is
 687         * still on the run list until we are actually done.
 688         * When we are done with this iteration, we check if
 689         * the iocb was kicked in the meantime and if so, queue
 690         * it up afresh.
 691         */
 692
 693        kiocbClearKicked(iocb);
 694
 695        /*
 696         * This is so that aio_complete knows it doesn't need to
 697         * pull the iocb off the run list (We can't just call
 698         * INIT_LIST_HEAD because we don't want a kick_iocb to
 699         * queue this on the run list yet)
 700         */
 701        iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
 702        spin_unlock_irq(&ctx->ctx_lock);
 703
 704        /* Quit retrying if the i/o has been cancelled */
 705        if (kiocbIsCancelled(iocb)) {
 706                ret = -EINTR;
 707                aio_complete(iocb, ret, 0);
 708                /* must not access the iocb after this */
 709                goto out;
 710        }
 711
 712        /*
 713         * Now we are all set to call the retry method in async
 714         * context.
 715         */
 716        ret = retry(iocb);
 717
 718        if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
 719                BUG_ON(!list_empty(&iocb->ki_wait.task_list));
 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/*
 800 * aio_run_iocbs:
 801 *      Process all pending retries queued on the ioctx
 802 *      run list.
 803 * Assumes it is operating within the aio issuer's mm
 804 * context.
 805 */
 806static inline void aio_run_iocbs(struct kioctx *ctx)
 807{
 808        int requeue;
 809
 810        spin_lock_irq(&ctx->ctx_lock);
 811
 812        requeue = __aio_run_iocbs(ctx);
 813        spin_unlock_irq(&ctx->ctx_lock);
 814        if (requeue)
 815                aio_queue_work(ctx);
 816}
 817
 818/*
 819 * just like aio_run_iocbs, but keeps running them until
 820 * the list stays empty
 821 */
 822static inline void aio_run_all_iocbs(struct kioctx *ctx)
 823{
 824        spin_lock_irq(&ctx->ctx_lock);
 825        while (__aio_run_iocbs(ctx))
 826                ;
 827        spin_unlock_irq(&ctx->ctx_lock);
 828}
 829
 830/*
 831 * aio_kick_handler:
 832 *      Work queue handler triggered to process pending
 833 *      retries on an ioctx. Takes on the aio issuer's
 834 *      mm context before running the iocbs, so that
 835 *      copy_xxx_user operates on the issuer's address
 836 *      space.
 837 * Run on aiod's context.
 838 */
 839static void aio_kick_handler(struct work_struct *work)
 840{
 841        struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
 842        mm_segment_t oldfs = get_fs();
 843        struct mm_struct *mm;
 844        int requeue;
 845
 846        set_fs(USER_DS);
 847        use_mm(ctx->mm);
 848        spin_lock_irq(&ctx->ctx_lock);
 849        requeue =__aio_run_iocbs(ctx);
 850        mm = ctx->mm;
 851        spin_unlock_irq(&ctx->ctx_lock);
 852        unuse_mm(mm);
 853        set_fs(oldfs);
 854        /*
 855         * we're in a worker thread already, don't use queue_delayed_work,
 856         */
 857        if (requeue)
 858                queue_delayed_work(aio_wq, &ctx->wq, 0);
 859}
 860
 861
 862/*
 863 * Called by kick_iocb to queue the kiocb for retry
 864 * and if required activate the aio work queue to process
 865 * it
 866 */
 867static void try_queue_kicked_iocb(struct kiocb *iocb)
 868{
 869        struct kioctx   *ctx = iocb->ki_ctx;
 870        unsigned long flags;
 871        int run = 0;
 872
 873        /* We're supposed to be the only path putting the iocb back on the run
 874         * list.  If we find that the iocb is *back* on a wait queue already
 875         * than retry has happened before we could queue the iocb.  This also
 876         * means that the retry could have completed and freed our iocb, no
 877         * good. */
 878        BUG_ON((!list_empty(&iocb->ki_wait.task_list)));
 879
 880        spin_lock_irqsave(&ctx->ctx_lock, flags);
 881        /* set this inside the lock so that we can't race with aio_run_iocb()
 882         * testing it and putting the iocb on the run list under the lock */
 883        if (!kiocbTryKick(iocb))
 884                run = __queue_kicked_iocb(iocb);
 885        spin_unlock_irqrestore(&ctx->ctx_lock, flags);
 886        if (run)
 887                aio_queue_work(ctx);
 888}
 889
 890/*
 891 * kick_iocb:
 892 *      Called typically from a wait queue callback context
 893 *      (aio_wake_function) to trigger a retry of the iocb.
 894 *      The retry is usually executed by aio workqueue
 895 *      threads (See aio_kick_handler).
 896 */
 897void kick_iocb(struct kiocb *iocb)
 898{
 899        /* sync iocbs are easy: they can only ever be executing from a 
 900         * single context. */
 901        if (is_sync_kiocb(iocb)) {
 902                kiocbSetKicked(iocb);
 903                wake_up_process(iocb->ki_obj.tsk);
 904                return;
 905        }
 906
 907        try_queue_kicked_iocb(iocb);
 908}
 909EXPORT_SYMBOL(kick_iocb);
 910
 911/* aio_complete
 912 *      Called when the io request on the given iocb is complete.
 913 *      Returns true if this is the last user of the request.  The 
 914 *      only other user of the request can be the cancellation code.
 915 */
 916int aio_complete(struct kiocb *iocb, long res, long res2)
 917{
 918        struct kioctx   *ctx = iocb->ki_ctx;
 919        struct aio_ring_info    *info;
 920        struct aio_ring *ring;
 921        struct io_event *event;
 922        unsigned long   flags;
 923        unsigned long   tail;
 924        int             ret;
 925
 926        /*
 927         * Special case handling for sync iocbs:
 928         *  - events go directly into the iocb for fast handling
 929         *  - the sync task with the iocb in its stack holds the single iocb
 930         *    ref, no other paths have a way to get another ref
 931         *  - the sync task helpfully left a reference to itself in the iocb
 932         */
 933        if (is_sync_kiocb(iocb)) {
 934                BUG_ON(iocb->ki_users != 1);
 935                iocb->ki_user_data = res;
 936                iocb->ki_users = 0;
 937                wake_up_process(iocb->ki_obj.tsk);
 938                return 1;
 939        }
 940
 941        info = &ctx->ring_info;
 942
 943        /* add a completion event to the ring buffer.
 944         * must be done holding ctx->ctx_lock to prevent
 945         * other code from messing with the tail
 946         * pointer since we might be called from irq
 947         * context.
 948         */
 949        spin_lock_irqsave(&ctx->ctx_lock, flags);
 950
 951        if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
 952                list_del_init(&iocb->ki_run_list);
 953
 954        /*
 955         * cancelled requests don't get events, userland was given one
 956         * when the event got cancelled.
 957         */
 958        if (kiocbIsCancelled(iocb))
 959                goto put_rq;
 960
 961        ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
 962
 963        tail = info->tail;
 964        event = aio_ring_event(info, tail, KM_IRQ0);
 965        if (++tail >= info->nr)
 966                tail = 0;
 967
 968        event->obj = (u64)(unsigned long)iocb->ki_obj.user;
 969        event->data = iocb->ki_user_data;
 970        event->res = res;
 971        event->res2 = res2;
 972
 973        dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
 974                ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
 975                res, res2);
 976
 977        /* after flagging the request as done, we
 978         * must never even look at it again
 979         */
 980        smp_wmb();      /* make event visible before updating tail */
 981
 982        info->tail = tail;
 983        ring->tail = tail;
 984
 985        put_aio_ring_event(event, KM_IRQ0);
 986        kunmap_atomic(ring, KM_IRQ1);
 987
 988        pr_debug("added to ring %p at [%lu]\n", iocb, tail);
 989
 990        /*
 991         * Check if the user asked us to deliver the result through an
 992         * eventfd. The eventfd_signal() function is safe to be called
 993         * from IRQ context.
 994         */
 995        if (!IS_ERR(iocb->ki_eventfd))
 996                eventfd_signal(iocb->ki_eventfd, 1);
 997
 998put_rq:
 999        /* everything turned out well, dispose of the aiocb. */
1000        ret = __aio_put_req(ctx, iocb);
1001
1002        /*
1003         * We have to order our ring_info tail store above and test
1004         * of the wait list below outside the wait lock.  This is
1005         * like in wake_up_bit() where clearing a bit has to be
1006         * ordered with the unlocked test.
1007         */
1008        smp_mb();
1009
1010        if (waitqueue_active(&ctx->wait))
1011                wake_up(&ctx->wait);
1012
1013        spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1014        return ret;
1015}
1016
1017/* aio_read_evt
1018 *      Pull an event off of the ioctx's event ring.  Returns the number of 
1019 *      events fetched (0 or 1 ;-)
1020 *      FIXME: make this use cmpxchg.
1021 *      TODO: make the ringbuffer user mmap()able (requires FIXME).
1022 */
1023static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1024{
1025        struct aio_ring_info *info = &ioctx->ring_info;
1026        struct aio_ring *ring;
1027        unsigned long head;
1028        int ret = 0;
1029
1030        ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1031        dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1032                 (unsigned long)ring->head, (unsigned long)ring->tail,
1033                 (unsigned long)ring->nr);
1034
1035        if (ring->head == ring->tail)
1036                goto out;
1037
1038        spin_lock(&info->ring_lock);
1039
1040        head = ring->head % info->nr;
1041        if (head != ring->tail) {
1042                struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1043                *ent = *evp;
1044                head = (head + 1) % info->nr;
1045                smp_mb(); /* finish reading the event before updatng the head */
1046                ring->head = head;
1047                ret = 1;
1048                put_aio_ring_event(evp, KM_USER1);
1049        }
1050        spin_unlock(&info->ring_lock);
1051
1052out:
1053        kunmap_atomic(ring, KM_USER0);
1054        dprintk("leaving aio_read_evt: %d  h%lu t%lu\n", ret,
1055                 (unsigned long)ring->head, (unsigned long)ring->tail);
1056        return ret;
1057}
1058
1059struct aio_timeout {
1060        struct timer_list       timer;
1061        int                     timed_out;
1062        struct task_struct      *p;
1063};
1064
1065static void timeout_func(unsigned long data)
1066{
1067        struct aio_timeout *to = (struct aio_timeout *)data;
1068
1069        to->timed_out = 1;
1070        wake_up_process(to->p);
1071}
1072
1073static inline void init_timeout(struct aio_timeout *to)
1074{
1075        setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
1076        to->timed_out = 0;
1077        to->p = current;
1078}
1079
1080static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1081                               const struct timespec *ts)
1082{
1083        to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1084        if (time_after(to->timer.expires, jiffies))
1085                add_timer(&to->timer);
1086        else
1087                to->timed_out = 1;
1088}
1089
1090static inline void clear_timeout(struct aio_timeout *to)
1091{
1092        del_singleshot_timer_sync(&to->timer);
1093}
1094
1095static int read_events(struct kioctx *ctx,
1096                        long min_nr, long nr,
1097                        struct io_event __user *event,
1098                        struct timespec __user *timeout)
1099{
1100        long                    start_jiffies = jiffies;
1101        struct task_struct      *tsk = current;
1102        DECLARE_WAITQUEUE(wait, tsk);
1103        int                     ret;
1104        int                     i = 0;
1105        struct io_event         ent;
1106        struct aio_timeout      to;
1107        int                     retry = 0;
1108
1109        /* needed to zero any padding within an entry (there shouldn't be 
1110         * any, but C is fun!
1111         */
1112        memset(&ent, 0, sizeof(ent));
1113retry:
1114        ret = 0;
1115        while (likely(i < nr)) {
1116                ret = aio_read_evt(ctx, &ent);
1117                if (unlikely(ret <= 0))
1118                        break;
1119
1120                dprintk("read event: %Lx %Lx %Lx %Lx\n",
1121                        ent.data, ent.obj, ent.res, ent.res2);
1122
1123                /* Could we split the check in two? */
1124                ret = -EFAULT;
1125                if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1126                        dprintk("aio: lost an event due to EFAULT.\n");
1127                        break;
1128                }
1129                ret = 0;
1130
1131                /* Good, event copied to userland, update counts. */
1132                event ++;
1133                i ++;
1134        }
1135
1136        if (min_nr <= i)
1137                return i;
1138        if (ret)
1139                return ret;
1140
1141        /* End fast path */
1142
1143        /* racey check, but it gets redone */
1144        if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1145                retry = 1;
1146                aio_run_all_iocbs(ctx);
1147                goto retry;
1148        }
1149
1150        init_timeout(&to);
1151        if (timeout) {
1152                struct timespec ts;
1153                ret = -EFAULT;
1154                if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1155                        goto out;
1156
1157                set_timeout(start_jiffies, &to, &ts);
1158        }
1159
1160        while (likely(i < nr)) {
1161                add_wait_queue_exclusive(&ctx->wait, &wait);
1162                do {
1163                        set_task_state(tsk, TASK_INTERRUPTIBLE);
1164                        ret = aio_read_evt(ctx, &ent);
1165                        if (ret)
1166                                break;
1167                        if (min_nr <= i)
1168                                break;
1169                        if (unlikely(ctx->dead)) {
1170                                ret = -EINVAL;
1171                                break;
1172                        }
1173                        if (to.timed_out)       /* Only check after read evt */
1174                                break;
1175                        /* Try to only show up in io wait if there are ops
1176                         *  in flight */
1177                        if (ctx->reqs_active)
1178                                io_schedule();
1179                        else
1180                                schedule();
1181                        if (signal_pending(tsk)) {
1182                                ret = -EINTR;
1183                                break;
1184                        }
1185                        /*ret = aio_read_evt(ctx, &ent);*/
1186                } while (1) ;
1187
1188                set_task_state(tsk, TASK_RUNNING);
1189                remove_wait_queue(&ctx->wait, &wait);
1190
1191                if (unlikely(ret <= 0))
1192                        break;
1193
1194                ret = -EFAULT;
1195                if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1196                        dprintk("aio: lost an event due to EFAULT.\n");
1197                        break;
1198                }
1199
1200                /* Good, event copied to userland, update counts. */
1201                event ++;
1202                i ++;
1203        }
1204
1205        if (timeout)
1206                clear_timeout(&to);
1207out:
1208        destroy_timer_on_stack(&to.timer);
1209        return i ? i : ret;
1210}
1211
1212/* Take an ioctx and remove it from the list of ioctx's.  Protects 
1213 * against races with itself via ->dead.
1214 */
1215static void io_destroy(struct kioctx *ioctx)
1216{
1217        struct mm_struct *mm = current->mm;
1218        struct kioctx **tmp;
1219        int was_dead;
1220
1221        /* delete the entry from the list is someone else hasn't already */
1222        write_lock(&mm->ioctx_list_lock);
1223        was_dead = ioctx->dead;
1224        ioctx->dead = 1;
1225        for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx;
1226             tmp = &(*tmp)->next)
1227                ;
1228        if (*tmp)
1229                *tmp = ioctx->next;
1230        write_unlock(&mm->ioctx_list_lock);
1231
1232        dprintk("aio_release(%p)\n", ioctx);
1233        if (likely(!was_dead))
1234                put_ioctx(ioctx);       /* twice for the list */
1235
1236        aio_cancel_all(ioctx);
1237        wait_for_all_aios(ioctx);
1238
1239        /*
1240         * Wake up any waiters.  The setting of ctx->dead must be seen
1241         * by other CPUs at this point.  Right now, we rely on the
1242         * locking done by the above calls to ensure this consistency.
1243         */
1244        wake_up(&ioctx->wait);
1245        put_ioctx(ioctx);       /* once for the lookup */
1246}
1247
1248/* sys_io_setup:
1249 *      Create an aio_context capable of receiving at least nr_events.
1250 *      ctxp must not point to an aio_context that already exists, and
1251 *      must be initialized to 0 prior to the call.  On successful
1252 *      creation of the aio_context, *ctxp is filled in with the resulting 
1253 *      handle.  May fail with -EINVAL if *ctxp is not initialized,
1254 *      if the specified nr_events exceeds internal limits.  May fail 
1255 *      with -EAGAIN if the specified nr_events exceeds the user's limit 
1256 *      of available events.  May fail with -ENOMEM if insufficient kernel
1257 *      resources are available.  May fail with -EFAULT if an invalid
1258 *      pointer is passed for ctxp.  Will fail with -ENOSYS if not
1259 *      implemented.
1260 */
1261SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1262{
1263        struct kioctx *ioctx = NULL;
1264        unsigned long ctx;
1265        long ret;
1266
1267        ret = get_user(ctx, ctxp);
1268        if (unlikely(ret))
1269                goto out;
1270
1271        ret = -EINVAL;
1272        if (unlikely(ctx || nr_events == 0)) {
1273                pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1274                         ctx, nr_events);
1275                goto out;
1276        }
1277
1278        ioctx = ioctx_alloc(nr_events);
1279        ret = PTR_ERR(ioctx);
1280        if (!IS_ERR(ioctx)) {
1281                ret = put_user(ioctx->user_id, ctxp);
1282                if (!ret)
1283                        return 0;
1284
1285                get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1286                io_destroy(ioctx);
1287        }
1288
1289out:
1290        return ret;
1291}
1292
1293/* sys_io_destroy:
1294 *      Destroy the aio_context specified.  May cancel any outstanding 
1295 *      AIOs and block on completion.  Will fail with -ENOSYS if not
1296 *      implemented.  May fail with -EFAULT if the context pointed to
1297 *      is invalid.
1298 */
1299SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1300{
1301        struct kioctx *ioctx = lookup_ioctx(ctx);
1302        if (likely(NULL != ioctx)) {
1303                io_destroy(ioctx);
1304                return 0;
1305        }
1306        pr_debug("EINVAL: io_destroy: invalid context id\n");
1307        return -EINVAL;
1308}
1309
1310static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1311{
1312        struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1313
1314        BUG_ON(ret <= 0);
1315
1316        while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1317                ssize_t this = min((ssize_t)iov->iov_len, ret);
1318                iov->iov_base += this;
1319                iov->iov_len -= this;
1320                iocb->ki_left -= this;
1321                ret -= this;
1322                if (iov->iov_len == 0) {
1323                        iocb->ki_cur_seg++;
1324                        iov++;
1325                }
1326        }
1327
1328        /* the caller should not have done more io than what fit in
1329         * the remaining iovecs */
1330        BUG_ON(ret > 0 && iocb->ki_left == 0);
1331}
1332
1333static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1334{
1335        struct file *file = iocb->ki_filp;
1336        struct address_space *mapping = file->f_mapping;
1337        struct inode *inode = mapping->host;
1338        ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1339                         unsigned long, loff_t);
1340        ssize_t ret = 0;
1341        unsigned short opcode;
1342
1343        if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1344                (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1345                rw_op = file->f_op->aio_read;
1346                opcode = IOCB_CMD_PREADV;
1347        } else {
1348                rw_op = file->f_op->aio_write;
1349                opcode = IOCB_CMD_PWRITEV;
1350        }
1351
1352        /* This matches the pread()/pwrite() logic */
1353        if (iocb->ki_pos < 0)
1354                return -EINVAL;
1355
1356        do {
1357                ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1358                            iocb->ki_nr_segs - iocb->ki_cur_seg,
1359                            iocb->ki_pos);
1360                if (ret > 0)
1361                        aio_advance_iovec(iocb, ret);
1362
1363        /* retry all partial writes.  retry partial reads as long as its a
1364         * regular file. */
1365        } while (ret > 0 && iocb->ki_left > 0 &&
1366                 (opcode == IOCB_CMD_PWRITEV ||
1367                  (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1368
1369        /* This means we must have transferred all that we could */
1370        /* No need to retry anymore */
1371        if ((ret == 0) || (iocb->ki_left == 0))
1372                ret = iocb->ki_nbytes - iocb->ki_left;
1373
1374        /* If we managed to write some out we return that, rather than
1375         * the eventual error. */
1376        if (opcode == IOCB_CMD_PWRITEV
1377            && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
1378            && iocb->ki_nbytes - iocb->ki_left)
1379                ret = iocb->ki_nbytes - iocb->ki_left;
1380
1381        return ret;
1382}
1383
1384static ssize_t aio_fdsync(struct kiocb *iocb)
1385{
1386        struct file *file = iocb->ki_filp;
1387        ssize_t ret = -EINVAL;
1388
1389        if (file->f_op->aio_fsync)
1390                ret = file->f_op->aio_fsync(iocb, 1);
1391        return ret;
1392}
1393
1394static ssize_t aio_fsync(struct kiocb *iocb)
1395{
1396        struct file *file = iocb->ki_filp;
1397        ssize_t ret = -EINVAL;
1398
1399        if (file->f_op->aio_fsync)
1400                ret = file->f_op->aio_fsync(iocb, 0);
1401        return ret;
1402}
1403
1404static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb)
1405{
1406        ssize_t ret;
1407
1408        ret = rw_copy_check_uvector(type, (struct iovec __user *)kiocb->ki_buf,
1409                                    kiocb->ki_nbytes, 1,
1410                                    &kiocb->ki_inline_vec, &kiocb->ki_iovec);
1411        if (ret < 0)
1412                goto out;
1413
1414        kiocb->ki_nr_segs = kiocb->ki_nbytes;
1415        kiocb->ki_cur_seg = 0;
1416        /* ki_nbytes/left now reflect bytes instead of segs */
1417        kiocb->ki_nbytes = ret;
1418        kiocb->ki_left = ret;
1419
1420        ret = 0;
1421out:
1422        return ret;
1423}
1424
1425static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
1426{
1427        kiocb->ki_iovec = &kiocb->ki_inline_vec;
1428        kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1429        kiocb->ki_iovec->iov_len = kiocb->ki_left;
1430        kiocb->ki_nr_segs = 1;
1431        kiocb->ki_cur_seg = 0;
1432        return 0;
1433}
1434
1435/*
1436 * aio_setup_iocb:
1437 *      Performs the initial checks and aio retry method
1438 *      setup for the kiocb at the time of io submission.
1439 */
1440static ssize_t aio_setup_iocb(struct kiocb *kiocb)
1441{
1442        struct file *file = kiocb->ki_filp;
1443        ssize_t ret = 0;
1444
1445        switch (kiocb->ki_opcode) {
1446        case IOCB_CMD_PREAD:
1447                ret = -EBADF;
1448                if (unlikely(!(file->f_mode & FMODE_READ)))
1449                        break;
1450                ret = -EFAULT;
1451                if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1452                        kiocb->ki_left)))
1453                        break;
1454                ret = security_file_permission(file, MAY_READ);
1455                if (unlikely(ret))
1456                        break;
1457                ret = aio_setup_single_vector(kiocb);
1458                if (ret)
1459                        break;
1460                ret = -EINVAL;
1461                if (file->f_op->aio_read)
1462                        kiocb->ki_retry = aio_rw_vect_retry;
1463                break;
1464        case IOCB_CMD_PWRITE:
1465                ret = -EBADF;
1466                if (unlikely(!(file->f_mode & FMODE_WRITE)))
1467                        break;
1468                ret = -EFAULT;
1469                if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1470                        kiocb->ki_left)))
1471                        break;
1472                ret = security_file_permission(file, MAY_WRITE);
1473                if (unlikely(ret))
1474                        break;
1475                ret = aio_setup_single_vector(kiocb);
1476                if (ret)
1477                        break;
1478                ret = -EINVAL;
1479                if (file->f_op->aio_write)
1480                        kiocb->ki_retry = aio_rw_vect_retry;
1481                break;
1482        case IOCB_CMD_PREADV:
1483                ret = -EBADF;
1484                if (unlikely(!(file->f_mode & FMODE_READ)))
1485                        break;
1486                ret = security_file_permission(file, MAY_READ);
1487                if (unlikely(ret))
1488                        break;
1489                ret = aio_setup_vectored_rw(READ, kiocb);
1490                if (ret)
1491                        break;
1492                ret = -EINVAL;
1493                if (file->f_op->aio_read)
1494                        kiocb->ki_retry = aio_rw_vect_retry;
1495                break;
1496        case IOCB_CMD_PWRITEV:
1497                ret = -EBADF;
1498                if (unlikely(!(file->f_mode & FMODE_WRITE)))
1499                        break;
1500                ret = security_file_permission(file, MAY_WRITE);
1501                if (unlikely(ret))
1502                        break;
1503                ret = aio_setup_vectored_rw(WRITE, kiocb);
1504                if (ret)
1505                        break;
1506                ret = -EINVAL;
1507                if (file->f_op->aio_write)
1508                        kiocb->ki_retry = aio_rw_vect_retry;
1509                break;
1510        case IOCB_CMD_FDSYNC:
1511                ret = -EINVAL;
1512                if (file->f_op->aio_fsync)
1513                        kiocb->ki_retry = aio_fdsync;
1514                break;
1515        case IOCB_CMD_FSYNC:
1516                ret = -EINVAL;
1517                if (file->f_op->aio_fsync)
1518                        kiocb->ki_retry = aio_fsync;
1519                break;
1520        default:
1521                dprintk("EINVAL: io_submit: no operation provided\n");
1522                ret = -EINVAL;
1523        }
1524
1525        if (!kiocb->ki_retry)
1526                return ret;
1527
1528        return 0;
1529}
1530
1531/*
1532 * aio_wake_function:
1533 *      wait queue callback function for aio notification,
1534 *      Simply triggers a retry of the operation via kick_iocb.
1535 *
1536 *      This callback is specified in the wait queue entry in
1537 *      a kiocb.
1538 *
1539 * Note:
1540 * This routine is executed with the wait queue lock held.
1541 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests
1542 * the ioctx lock inside the wait queue lock. This is safe
1543 * because this callback isn't used for wait queues which
1544 * are nested inside ioctx lock (i.e. ctx->wait)
1545 */
1546static int aio_wake_function(wait_queue_t *wait, unsigned mode,
1547                             int sync, void *key)
1548{
1549        struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait);
1550
1551        list_del_init(&wait->task_list);
1552        kick_iocb(iocb);
1553        return 1;
1554}
1555
1556static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1557                         struct iocb *iocb)
1558{
1559        struct kiocb *req;
1560        struct file *file;
1561        ssize_t ret;
1562
1563        /* enforce forwards compatibility on users */
1564        if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1565                pr_debug("EINVAL: io_submit: reserve field set\n");
1566                return -EINVAL;
1567        }
1568
1569        /* prevent overflows */
1570        if (unlikely(
1571            (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1572            (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1573            ((ssize_t)iocb->aio_nbytes < 0)
1574           )) {
1575                pr_debug("EINVAL: io_submit: overflow check\n");
1576                return -EINVAL;
1577        }
1578
1579        file = fget(iocb->aio_fildes);
1580        if (unlikely(!file))
1581                return -EBADF;
1582
1583        req = aio_get_req(ctx);         /* returns with 2 references to req */
1584        if (unlikely(!req)) {
1585                fput(file);
1586                return -EAGAIN;
1587        }
1588        req->ki_filp = file;
1589        if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1590                /*
1591                 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1592                 * instance of the file* now. The file descriptor must be
1593                 * an eventfd() fd, and will be signaled for each completed
1594                 * event using the eventfd_signal() function.
1595                 */
1596                req->ki_eventfd = eventfd_fget((int) iocb->aio_resfd);
1597                if (IS_ERR(req->ki_eventfd)) {
1598                        ret = PTR_ERR(req->ki_eventfd);
1599                        goto out_put_req;
1600                }
1601        }
1602
1603        ret = put_user(req->ki_key, &user_iocb->aio_key);
1604        if (unlikely(ret)) {
1605                dprintk("EFAULT: aio_key\n");
1606                goto out_put_req;
1607        }
1608
1609        req->ki_obj.user = user_iocb;
1610        req->ki_user_data = iocb->aio_data;
1611        req->ki_pos = iocb->aio_offset;
1612
1613        req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1614        req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1615        req->ki_opcode = iocb->aio_lio_opcode;
1616        init_waitqueue_func_entry(&req->ki_wait, aio_wake_function);
1617        INIT_LIST_HEAD(&req->ki_wait.task_list);
1618
1619        ret = aio_setup_iocb(req);
1620
1621        if (ret)
1622                goto out_put_req;
1623
1624        spin_lock_irq(&ctx->ctx_lock);
1625        aio_run_iocb(req);
1626        if (!list_empty(&ctx->run_list)) {
1627                /* drain the run list */
1628                while (__aio_run_iocbs(ctx))
1629                        ;
1630        }
1631        spin_unlock_irq(&ctx->ctx_lock);
1632        aio_put_req(req);       /* drop extra ref to req */
1633        return 0;
1634
1635out_put_req:
1636        aio_put_req(req);       /* drop extra ref to req */
1637        aio_put_req(req);       /* drop i/o ref to req */
1638        return ret;
1639}
1640
1641/* sys_io_submit:
1642 *      Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1643 *      the number of iocbs queued.  May return -EINVAL if the aio_context
1644 *      specified by ctx_id is invalid, if nr is < 0, if the iocb at
1645 *      *iocbpp[0] is not properly initialized, if the operation specified
1646 *      is invalid for the file descriptor in the iocb.  May fail with
1647 *      -EFAULT if any of the data structures point to invalid data.  May
1648 *      fail with -EBADF if the file descriptor specified in the first
1649 *      iocb is invalid.  May fail with -EAGAIN if insufficient resources
1650 *      are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1651 *      fail with -ENOSYS if not implemented.
1652 */
1653SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1654                struct iocb __user * __user *, iocbpp)
1655{
1656        struct kioctx *ctx;
1657        long ret = 0;
1658        int i;
1659
1660        if (unlikely(nr < 0))
1661                return -EINVAL;
1662
1663        if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1664                return -EFAULT;
1665
1666        ctx = lookup_ioctx(ctx_id);
1667        if (unlikely(!ctx)) {
1668                pr_debug("EINVAL: io_submit: invalid context id\n");
1669                return -EINVAL;
1670        }
1671
1672        /*
1673         * AKPM: should this return a partial result if some of the IOs were
1674         * successfully submitted?
1675         */
1676        for (i=0; i<nr; i++) {
1677                struct iocb __user *user_iocb;
1678                struct iocb tmp;
1679
1680                if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1681                        ret = -EFAULT;
1682                        break;
1683                }
1684
1685                if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1686                        ret = -EFAULT;
1687                        break;
1688                }
1689
1690                ret = io_submit_one(ctx, user_iocb, &tmp);
1691                if (ret)
1692                        break;
1693        }
1694
1695        put_ioctx(ctx);
1696        return i ? i : ret;
1697}
1698
1699/* lookup_kiocb
1700 *      Finds a given iocb for cancellation.
1701 */
1702static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1703                                  u32 key)
1704{
1705        struct list_head *pos;
1706
1707        assert_spin_locked(&ctx->ctx_lock);
1708
1709        /* TODO: use a hash or array, this sucks. */
1710        list_for_each(pos, &ctx->active_reqs) {
1711                struct kiocb *kiocb = list_kiocb(pos);
1712                if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1713                        return kiocb;
1714        }
1715        return NULL;
1716}
1717
1718/* sys_io_cancel:
1719 *      Attempts to cancel an iocb previously passed to io_submit.  If
1720 *      the operation is successfully cancelled, the resulting event is
1721 *      copied into the memory pointed to by result without being placed
1722 *      into the completion queue and 0 is returned.  May fail with
1723 *      -EFAULT if any of the data structures pointed to are invalid.
1724 *      May fail with -EINVAL if aio_context specified by ctx_id is
1725 *      invalid.  May fail with -EAGAIN if the iocb specified was not
1726 *      cancelled.  Will fail with -ENOSYS if not implemented.
1727 */
1728SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1729                struct io_event __user *, result)
1730{
1731        int (*cancel)(struct kiocb *iocb, struct io_event *res);
1732        struct kioctx *ctx;
1733        struct kiocb *kiocb;
1734        u32 key;
1735        int ret;
1736
1737        ret = get_user(key, &iocb->aio_key);
1738        if (unlikely(ret))
1739                return -EFAULT;
1740
1741        ctx = lookup_ioctx(ctx_id);
1742        if (unlikely(!ctx))
1743                return -EINVAL;
1744
1745        spin_lock_irq(&ctx->ctx_lock);
1746        ret = -EAGAIN;
1747        kiocb = lookup_kiocb(ctx, iocb, key);
1748        if (kiocb && kiocb->ki_cancel) {
1749                cancel = kiocb->ki_cancel;
1750                kiocb->ki_users ++;
1751                kiocbSetCancelled(kiocb);
1752        } else
1753                cancel = NULL;
1754        spin_unlock_irq(&ctx->ctx_lock);
1755
1756        if (NULL != cancel) {
1757                struct io_event tmp;
1758                pr_debug("calling cancel\n");
1759                memset(&tmp, 0, sizeof(tmp));
1760                tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1761                tmp.data = kiocb->ki_user_data;
1762                ret = cancel(kiocb, &tmp);
1763                if (!ret) {
1764                        /* Cancellation succeeded -- copy the result
1765                         * into the user's buffer.
1766                         */
1767                        if (copy_to_user(result, &tmp, sizeof(tmp)))
1768                                ret = -EFAULT;
1769                }
1770        } else
1771                ret = -EINVAL;
1772
1773        put_ioctx(ctx);
1774
1775        return ret;
1776}
1777
1778/* io_getevents:
1779 *      Attempts to read at least min_nr events and up to nr events from
1780 *      the completion queue for the aio_context specified by ctx_id.  May
1781 *      fail with -EINVAL if ctx_id is invalid, if min_nr is out of range,
1782 *      if nr is out of range, if when is out of range.  May fail with
1783 *      -EFAULT if any of the memory specified to is invalid.  May return
1784 *      0 or < min_nr if no events are available and the timeout specified
1785 *      by when has elapsed, where when == NULL specifies an infinite
1786 *      timeout.  Note that the timeout pointed to by when is relative and
1787 *      will be updated if not NULL and the operation blocks.  Will fail
1788 *      with -ENOSYS if not implemented.
1789 */
1790SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1791                long, min_nr,
1792                long, nr,
1793                struct io_event __user *, events,
1794                struct timespec __user *, timeout)
1795{
1796        struct kioctx *ioctx = lookup_ioctx(ctx_id);
1797        long ret = -EINVAL;
1798
1799        if (likely(ioctx)) {
1800                if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0))
1801                        ret = read_events(ioctx, min_nr, nr, events, timeout);
1802                put_ioctx(ioctx);
1803        }
1804
1805        asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);
1806        return ret;
1807}
1808
1809__initcall(aio_setup);
1810
1811EXPORT_SYMBOL(aio_complete);
1812EXPORT_SYMBOL(aio_put_req);
1813EXPORT_SYMBOL(wait_on_sync_kiocb);
1814
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.