linux/fs/aio.c
<<
>>
Prefs
   1/*
   2 *      An async IO implementation for Linux
   3 *      Written by Benjamin LaHaise <bcrl@kvack.org>
   4 *
   5 *      Implements an efficient asynchronous io interface.
   6 *
   7 *      Copyright 2000, 2001, 2002 Red Hat, Inc.  All Rights Reserved.
   8 *
   9 *      See ../COPYING for licensing terms.
  10 */
  11#include <linux/kernel.h>
  12#include <linux/init.h>
  13#include <linux/errno.h>
  14#include <linux/time.h>
  15#include <linux/aio_abi.h>
  16#include <linux/export.h>
  17#include <linux/syscalls.h>
  18#include <linux/backing-dev.h>
  19#include <linux/uio.h>
  20
  21#define DEBUG 0
  22
  23#include <linux/sched.h>
  24#include <linux/fs.h>
  25#include <linux/file.h>
  26#include <linux/mm.h>
  27#include <linux/mman.h>
  28#include <linux/mmu_context.h>
  29#include <linux/slab.h>
  30#include <linux/timer.h>
  31#include <linux/aio.h>
  32#include <linux/highmem.h>
  33#include <linux/workqueue.h>
  34#include <linux/security.h>
  35#include <linux/eventfd.h>
  36#include <linux/blkdev.h>
  37#include <linux/compat.h>
  38
  39#include <asm/kmap_types.h>
  40#include <asm/uaccess.h>
  41
  42#if DEBUG > 1
  43#define dprintk         printk
  44#else
  45#define dprintk(x...)   do { ; } while (0)
  46#endif
  47
  48/*------ sysctl variables----*/
  49static DEFINE_SPINLOCK(aio_nr_lock);
  50unsigned long aio_nr;           /* current system wide number of aio requests */
  51unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
  52/*----end sysctl variables---*/
  53
  54static struct kmem_cache        *kiocb_cachep;
  55static struct kmem_cache        *kioctx_cachep;
  56
  57static struct workqueue_struct *aio_wq;
  58
  59static void aio_kick_handler(struct work_struct *);
  60static void aio_queue_work(struct kioctx *);
  61
  62/* aio_setup
  63 *      Creates the slab caches used by the aio routines, panic on
  64 *      failure as this is done early during the boot sequence.
  65 */
  66static int __init aio_setup(void)
  67{
  68        kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
  69        kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
  70
  71        aio_wq = alloc_workqueue("aio", 0, 1);  /* used to limit concurrency */
  72        BUG_ON(!aio_wq);
  73
  74        pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
  75
  76        return 0;
  77}
  78__initcall(aio_setup);
  79
  80static void aio_free_ring(struct kioctx *ctx)
  81{
  82        struct aio_ring_info *info = &ctx->ring_info;
  83        long i;
  84
  85        for (i=0; i<info->nr_pages; i++)
  86                put_page(info->ring_pages[i]);
  87
  88        if (info->mmap_size) {
  89                BUG_ON(ctx->mm != current->mm);
  90                vm_munmap(info->mmap_base, info->mmap_size);
  91        }
  92
  93        if (info->ring_pages && info->ring_pages != info->internal_pages)
  94                kfree(info->ring_pages);
  95        info->ring_pages = NULL;
  96        info->nr = 0;
  97}
  98
  99static int aio_setup_ring(struct kioctx *ctx)
 100{
 101        struct aio_ring *ring;
 102        struct aio_ring_info *info = &ctx->ring_info;
 103        unsigned nr_events = ctx->max_reqs;
 104        unsigned long size, populate;
 105        int nr_pages;
 106
 107        /* Compensate for the ring buffer's head/tail overlap entry */
 108        nr_events += 2; /* 1 is required, 2 for good luck */
 109
 110        size = sizeof(struct aio_ring);
 111        size += sizeof(struct io_event) * nr_events;
 112        nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
 113
 114        if (nr_pages < 0)
 115                return -EINVAL;
 116
 117        nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
 118
 119        info->nr = 0;
 120        info->ring_pages = info->internal_pages;
 121        if (nr_pages > AIO_RING_PAGES) {
 122                info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
 123                if (!info->ring_pages)
 124                        return -ENOMEM;
 125        }
 126
 127        info->mmap_size = nr_pages * PAGE_SIZE;
 128        dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
 129        down_write(&ctx->mm->mmap_sem);
 130        info->mmap_base = do_mmap_pgoff(NULL, 0, info->mmap_size, 
 131                                        PROT_READ|PROT_WRITE,
 132                                        MAP_ANONYMOUS|MAP_PRIVATE, 0,
 133                                        &populate);
 134        if (IS_ERR((void *)info->mmap_base)) {
 135                up_write(&ctx->mm->mmap_sem);
 136                info->mmap_size = 0;
 137                aio_free_ring(ctx);
 138                return -EAGAIN;
 139        }
 140
 141        dprintk("mmap address: 0x%08lx\n", info->mmap_base);
 142        info->nr_pages = get_user_pages(current, ctx->mm,
 143                                        info->mmap_base, nr_pages, 
 144                                        1, 0, info->ring_pages, NULL);
 145        up_write(&ctx->mm->mmap_sem);
 146
 147        if (unlikely(info->nr_pages != nr_pages)) {
 148                aio_free_ring(ctx);
 149                return -EAGAIN;
 150        }
 151        if (populate)
 152                mm_populate(info->mmap_base, populate);
 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]);
 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);
 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().  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) ({                                     \
 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]); \
 184        __event += pos % AIO_EVENTS_PER_PAGE;                           \
 185        __event;                                                        \
 186})
 187
 188#define put_aio_ring_event(event) do {          \
 189        struct io_event *__event = (event);     \
 190        (void)__event;                          \
 191        kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK)); \
 192} while(0)
 193
 194static void ctx_rcu_free(struct rcu_head *head)
 195{
 196        struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
 197        kmem_cache_free(kioctx_cachep, ctx);
 198}
 199
 200/* __put_ioctx
 201 *      Called when the last user of an aio context has gone away,
 202 *      and the struct needs to be freed.
 203 */
 204static void __put_ioctx(struct kioctx *ctx)
 205{
 206        unsigned nr_events = ctx->max_reqs;
 207        BUG_ON(ctx->reqs_active);
 208
 209        cancel_delayed_work_sync(&ctx->wq);
 210        aio_free_ring(ctx);
 211        mmdrop(ctx->mm);
 212        ctx->mm = NULL;
 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        pr_debug("__put_ioctx: freeing %p\n", ctx);
 220        call_rcu(&ctx->rcu_head, ctx_rcu_free);
 221}
 222
 223static inline int try_get_ioctx(struct kioctx *kioctx)
 224{
 225        return atomic_inc_not_zero(&kioctx->users);
 226}
 227
 228static inline void put_ioctx(struct kioctx *kioctx)
 229{
 230        BUG_ON(atomic_read(&kioctx->users) <= 0);
 231        if (unlikely(atomic_dec_and_test(&kioctx->users)))
 232                __put_ioctx(kioctx);
 233}
 234
 235/* ioctx_alloc
 236 *      Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
 237 */
 238static struct kioctx *ioctx_alloc(unsigned nr_events)
 239{
 240        struct mm_struct *mm;
 241        struct kioctx *ctx;
 242        int err = -ENOMEM;
 243
 244        /* Prevent overflows */
 245        if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
 246            (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
 247                pr_debug("ENOMEM: nr_events too high\n");
 248                return ERR_PTR(-EINVAL);
 249        }
 250
 251        if (!nr_events || (unsigned long)nr_events > aio_max_nr)
 252                return ERR_PTR(-EAGAIN);
 253
 254        ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
 255        if (!ctx)
 256                return ERR_PTR(-ENOMEM);
 257
 258        ctx->max_reqs = nr_events;
 259        mm = ctx->mm = current->mm;
 260        atomic_inc(&mm->mm_count);
 261
 262        atomic_set(&ctx->users, 2);
 263        spin_lock_init(&ctx->ctx_lock);
 264        spin_lock_init(&ctx->ring_info.ring_lock);
 265        init_waitqueue_head(&ctx->wait);
 266
 267        INIT_LIST_HEAD(&ctx->active_reqs);
 268        INIT_LIST_HEAD(&ctx->run_list);
 269        INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
 270
 271        if (aio_setup_ring(ctx) < 0)
 272                goto out_freectx;
 273
 274        /* limit the number of system wide aios */
 275        spin_lock(&aio_nr_lock);
 276        if (aio_nr + nr_events > aio_max_nr ||
 277            aio_nr + nr_events < aio_nr) {
 278                spin_unlock(&aio_nr_lock);
 279                goto out_cleanup;
 280        }
 281        aio_nr += ctx->max_reqs;
 282        spin_unlock(&aio_nr_lock);
 283
 284        /* now link into global list. */
 285        spin_lock(&mm->ioctx_lock);
 286        hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
 287        spin_unlock(&mm->ioctx_lock);
 288
 289        dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
 290                ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
 291        return ctx;
 292
 293out_cleanup:
 294        err = -EAGAIN;
 295        aio_free_ring(ctx);
 296out_freectx:
 297        mmdrop(mm);
 298        kmem_cache_free(kioctx_cachep, ctx);
 299        dprintk("aio: error allocating ioctx %d\n", err);
 300        return ERR_PTR(err);
 301}
 302
 303/* kill_ctx
 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 kill_ctx(struct kioctx *ctx)
 309{
 310        int (*cancel)(struct kiocb *, struct io_event *);
 311        struct task_struct *tsk = current;
 312        DECLARE_WAITQUEUE(wait, tsk);
 313        struct io_event res;
 314
 315        spin_lock_irq(&ctx->ctx_lock);
 316        ctx->dead = 1;
 317        while (!list_empty(&ctx->active_reqs)) {
 318                struct list_head *pos = ctx->active_reqs.next;
 319                struct kiocb *iocb = list_kiocb(pos);
 320                list_del_init(&iocb->ki_list);
 321                cancel = iocb->ki_cancel;
 322                kiocbSetCancelled(iocb);
 323                if (cancel) {
 324                        iocb->ki_users++;
 325                        spin_unlock_irq(&ctx->ctx_lock);
 326                        cancel(iocb, &res);
 327                        spin_lock_irq(&ctx->ctx_lock);
 328                }
 329        }
 330
 331        if (!ctx->reqs_active)
 332                goto out;
 333
 334        add_wait_queue(&ctx->wait, &wait);
 335        set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 336        while (ctx->reqs_active) {
 337                spin_unlock_irq(&ctx->ctx_lock);
 338                io_schedule();
 339                set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 340                spin_lock_irq(&ctx->ctx_lock);
 341        }
 342        __set_task_state(tsk, TASK_RUNNING);
 343        remove_wait_queue(&ctx->wait, &wait);
 344
 345out:
 346        spin_unlock_irq(&ctx->ctx_lock);
 347}
 348
 349/* wait_on_sync_kiocb:
 350 *      Waits on the given sync kiocb to complete.
 351 */
 352ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
 353{
 354        while (iocb->ki_users) {
 355                set_current_state(TASK_UNINTERRUPTIBLE);
 356                if (!iocb->ki_users)
 357                        break;
 358                io_schedule();
 359        }
 360        __set_current_state(TASK_RUNNING);
 361        return iocb->ki_user_data;
 362}
 363EXPORT_SYMBOL(wait_on_sync_kiocb);
 364
 365/* exit_aio: called when the last user of mm goes away.  At this point, 
 366 * there is no way for any new requests to be submited or any of the 
 367 * io_* syscalls to be called on the context.  However, there may be 
 368 * outstanding requests which hold references to the context; as they 
 369 * go away, they will call put_ioctx and release any pinned memory
 370 * associated with the request (held via struct page * references).
 371 */
 372void exit_aio(struct mm_struct *mm)
 373{
 374        struct kioctx *ctx;
 375
 376        while (!hlist_empty(&mm->ioctx_list)) {
 377                ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
 378                hlist_del_rcu(&ctx->list);
 379
 380                kill_ctx(ctx);
 381
 382                if (1 != atomic_read(&ctx->users))
 383                        printk(KERN_DEBUG
 384                                "exit_aio:ioctx still alive: %d %d %d\n",
 385                                atomic_read(&ctx->users), ctx->dead,
 386                                ctx->reqs_active);
 387                /*
 388                 * We don't need to bother with munmap() here -
 389                 * exit_mmap(mm) is coming and it'll unmap everything.
 390                 * Since aio_free_ring() uses non-zero ->mmap_size
 391                 * as indicator that it needs to unmap the area,
 392                 * just set it to 0; aio_free_ring() is the only
 393                 * place that uses ->mmap_size, so it's safe.
 394                 * That way we get all munmap done to current->mm -
 395                 * all other callers have ctx->mm == current->mm.
 396                 */
 397                ctx->ring_info.mmap_size = 0;
 398                put_ioctx(ctx);
 399        }
 400}
 401
 402/* aio_get_req
 403 *      Allocate a slot for an aio request.  Increments the users count
 404 * of the kioctx so that the kioctx stays around until all requests are
 405 * complete.  Returns NULL if no requests are free.
 406 *
 407 * Returns with kiocb->users set to 2.  The io submit code path holds
 408 * an extra reference while submitting the i/o.
 409 * This prevents races between the aio code path referencing the
 410 * req (after submitting it) and aio_complete() freeing the req.
 411 */
 412static struct kiocb *__aio_get_req(struct kioctx *ctx)
 413{
 414        struct kiocb *req = NULL;
 415
 416        req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
 417        if (unlikely(!req))
 418                return NULL;
 419
 420        req->ki_flags = 0;
 421        req->ki_users = 2;
 422        req->ki_key = 0;
 423        req->ki_ctx = ctx;
 424        req->ki_cancel = NULL;
 425        req->ki_retry = NULL;
 426        req->ki_dtor = NULL;
 427        req->private = NULL;
 428        req->ki_iovec = NULL;
 429        INIT_LIST_HEAD(&req->ki_run_list);
 430        req->ki_eventfd = NULL;
 431
 432        return req;
 433}
 434
 435/*
 436 * struct kiocb's are allocated in batches to reduce the number of
 437 * times the ctx lock is acquired and released.
 438 */
 439#define KIOCB_BATCH_SIZE        32L
 440struct kiocb_batch {
 441        struct list_head head;
 442        long count; /* number of requests left to allocate */
 443};
 444
 445static void kiocb_batch_init(struct kiocb_batch *batch, long total)
 446{
 447        INIT_LIST_HEAD(&batch->head);
 448        batch->count = total;
 449}
 450
 451static void kiocb_batch_free(struct kioctx *ctx, struct kiocb_batch *batch)
 452{
 453        struct kiocb *req, *n;
 454
 455        if (list_empty(&batch->head))
 456                return;
 457
 458        spin_lock_irq(&ctx->ctx_lock);
 459        list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
 460                list_del(&req->ki_batch);
 461                list_del(&req->ki_list);
 462                kmem_cache_free(kiocb_cachep, req);
 463                ctx->reqs_active--;
 464        }
 465        if (unlikely(!ctx->reqs_active && ctx->dead))
 466                wake_up_all(&ctx->wait);
 467        spin_unlock_irq(&ctx->ctx_lock);
 468}
 469
 470/*
 471 * Allocate a batch of kiocbs.  This avoids taking and dropping the
 472 * context lock a lot during setup.
 473 */
 474static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch)
 475{
 476        unsigned short allocated, to_alloc;
 477        long avail;
 478        struct kiocb *req, *n;
 479        struct aio_ring *ring;
 480
 481        to_alloc = min(batch->count, KIOCB_BATCH_SIZE);
 482        for (allocated = 0; allocated < to_alloc; allocated++) {
 483                req = __aio_get_req(ctx);
 484                if (!req)
 485                        /* allocation failed, go with what we've got */
 486                        break;
 487                list_add(&req->ki_batch, &batch->head);
 488        }
 489
 490        if (allocated == 0)
 491                goto out;
 492
 493        spin_lock_irq(&ctx->ctx_lock);
 494        ring = kmap_atomic(ctx->ring_info.ring_pages[0]);
 495
 496        avail = aio_ring_avail(&ctx->ring_info, ring) - ctx->reqs_active;
 497        BUG_ON(avail < 0);
 498        if (avail < allocated) {
 499                /* Trim back the number of requests. */
 500                list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
 501                        list_del(&req->ki_batch);
 502                        kmem_cache_free(kiocb_cachep, req);
 503                        if (--allocated <= avail)
 504                                break;
 505                }
 506        }
 507
 508        batch->count -= allocated;
 509        list_for_each_entry(req, &batch->head, ki_batch) {
 510                list_add(&req->ki_list, &ctx->active_reqs);
 511                ctx->reqs_active++;
 512        }
 513
 514        kunmap_atomic(ring);
 515        spin_unlock_irq(&ctx->ctx_lock);
 516
 517out:
 518        return allocated;
 519}
 520
 521static inline struct kiocb *aio_get_req(struct kioctx *ctx,
 522                                        struct kiocb_batch *batch)
 523{
 524        struct kiocb *req;
 525
 526        if (list_empty(&batch->head))
 527                if (kiocb_batch_refill(ctx, batch) == 0)
 528                        return NULL;
 529        req = list_first_entry(&batch->head, struct kiocb, ki_batch);
 530        list_del(&req->ki_batch);
 531        return req;
 532}
 533
 534static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
 535{
 536        assert_spin_locked(&ctx->ctx_lock);
 537
 538        if (req->ki_eventfd != NULL)
 539                eventfd_ctx_put(req->ki_eventfd);
 540        if (req->ki_dtor)
 541                req->ki_dtor(req);
 542        if (req->ki_iovec != &req->ki_inline_vec)
 543                kfree(req->ki_iovec);
 544        kmem_cache_free(kiocb_cachep, req);
 545        ctx->reqs_active--;
 546
 547        if (unlikely(!ctx->reqs_active && ctx->dead))
 548                wake_up_all(&ctx->wait);
 549}
 550
 551/* __aio_put_req
 552 *      Returns true if this put was the last user of the request.
 553 */
 554static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
 555{
 556        dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
 557                req, atomic_long_read(&req->ki_filp->f_count));
 558
 559        assert_spin_locked(&ctx->ctx_lock);
 560
 561        req->ki_users--;
 562        BUG_ON(req->ki_users < 0);
 563        if (likely(req->ki_users))
 564                return 0;
 565        list_del(&req->ki_list);                /* remove from active_reqs */
 566        req->ki_cancel = NULL;
 567        req->ki_retry = NULL;
 568
 569        fput(req->ki_filp);
 570        req->ki_filp = NULL;
 571        really_put_req(ctx, req);
 572        return 1;
 573}
 574
 575/* aio_put_req
 576 *      Returns true if this put was the last user of the kiocb,
 577 *      false if the request is still in use.
 578 */
 579int aio_put_req(struct kiocb *req)
 580{
 581        struct kioctx *ctx = req->ki_ctx;
 582        int ret;
 583        spin_lock_irq(&ctx->ctx_lock);
 584        ret = __aio_put_req(ctx, req);
 585        spin_unlock_irq(&ctx->ctx_lock);
 586        return ret;
 587}
 588EXPORT_SYMBOL(aio_put_req);
 589
 590static struct kioctx *lookup_ioctx(unsigned long ctx_id)
 591{
 592        struct mm_struct *mm = current->mm;
 593        struct kioctx *ctx, *ret = NULL;
 594
 595        rcu_read_lock();
 596
 597        hlist_for_each_entry_rcu(ctx, &mm->ioctx_list, list) {
 598                /*
 599                 * RCU protects us against accessing freed memory but
 600                 * we have to be careful not to get a reference when the
 601                 * reference count already dropped to 0 (ctx->dead test
 602                 * is unreliable because of races).
 603                 */
 604                if (ctx->user_id == ctx_id && !ctx->dead && try_get_ioctx(ctx)){
 605                        ret = ctx;
 606                        break;
 607                }
 608        }
 609
 610        rcu_read_unlock();
 611        return ret;
 612}
 613
 614/*
 615 * Queue up a kiocb to be retried. Assumes that the kiocb
 616 * has already been marked as kicked, and places it on
 617 * the retry run list for the corresponding ioctx, if it
 618 * isn't already queued. Returns 1 if it actually queued
 619 * the kiocb (to tell the caller to activate the work
 620 * queue to process it), or 0, if it found that it was
 621 * already queued.
 622 */
 623static inline int __queue_kicked_iocb(struct kiocb *iocb)
 624{
 625        struct kioctx *ctx = iocb->ki_ctx;
 626
 627        assert_spin_locked(&ctx->ctx_lock);
 628
 629        if (list_empty(&iocb->ki_run_list)) {
 630                list_add_tail(&iocb->ki_run_list,
 631                        &ctx->run_list);
 632                return 1;
 633        }
 634        return 0;
 635}
 636
 637/* aio_run_iocb
 638 *      This is the core aio execution routine. It is
 639 *      invoked both for initial i/o submission and
 640 *      subsequent retries via the aio_kick_handler.
 641 *      Expects to be invoked with iocb->ki_ctx->lock
 642 *      already held. The lock is released and reacquired
 643 *      as needed during processing.
 644 *
 645 * Calls the iocb retry method (already setup for the
 646 * iocb on initial submission) for operation specific
 647 * handling, but takes care of most of common retry
 648 * execution details for a given iocb. The retry method
 649 * needs to be non-blocking as far as possible, to avoid
 650 * holding up other iocbs waiting to be serviced by the
 651 * retry kernel thread.
 652 *
 653 * The trickier parts in this code have to do with
 654 * ensuring that only one retry instance is in progress
 655 * for a given iocb at any time. Providing that guarantee
 656 * simplifies the coding of individual aio operations as
 657 * it avoids various potential races.
 658 */
 659static ssize_t aio_run_iocb(struct kiocb *iocb)
 660{
 661        struct kioctx   *ctx = iocb->ki_ctx;
 662        ssize_t (*retry)(struct kiocb *);
 663        ssize_t ret;
 664
 665        if (!(retry = iocb->ki_retry)) {
 666                printk("aio_run_iocb: iocb->ki_retry = NULL\n");
 667                return 0;
 668        }
 669
 670        /*
 671         * We don't want the next retry iteration for this
 672         * operation to start until this one has returned and
 673         * updated the iocb state. However, wait_queue functions
 674         * can trigger a kick_iocb from interrupt context in the
 675         * meantime, indicating that data is available for the next
 676         * iteration. We want to remember that and enable the
 677         * next retry iteration _after_ we are through with
 678         * this one.
 679         *
 680         * So, in order to be able to register a "kick", but
 681         * prevent it from being queued now, we clear the kick
 682         * flag, but make the kick code *think* that the iocb is
 683         * still on the run list until we are actually done.
 684         * When we are done with this iteration, we check if
 685         * the iocb was kicked in the meantime and if so, queue
 686         * it up afresh.
 687         */
 688
 689        kiocbClearKicked(iocb);
 690
 691        /*
 692         * This is so that aio_complete knows it doesn't need to
 693         * pull the iocb off the run list (We can't just call
 694         * INIT_LIST_HEAD because we don't want a kick_iocb to
 695         * queue this on the run list yet)
 696         */
 697        iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
 698        spin_unlock_irq(&ctx->ctx_lock);
 699
 700        /* Quit retrying if the i/o has been cancelled */
 701        if (kiocbIsCancelled(iocb)) {
 702                ret = -EINTR;
 703                aio_complete(iocb, ret, 0);
 704                /* must not access the iocb after this */
 705                goto out;
 706        }
 707
 708        /*
 709         * Now we are all set to call the retry method in async
 710         * context.
 711         */
 712        ret = retry(iocb);
 713
 714        if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
 715                /*
 716                 * There's no easy way to restart the syscall since other AIO's
 717                 * may be already running. Just fail this IO with EINTR.
 718                 */
 719                if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
 720                             ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
 721                        ret = -EINTR;
 722                aio_complete(iocb, ret, 0);
 723        }
 724out:
 725        spin_lock_irq(&ctx->ctx_lock);
 726
 727        if (-EIOCBRETRY == ret) {
 728                /*
 729                 * OK, now that we are done with this iteration
 730                 * and know that there is more left to go,
 731                 * this is where we let go so that a subsequent
 732                 * "kick" can start the next iteration
 733                 */
 734
 735                /* will make __queue_kicked_iocb succeed from here on */
 736                INIT_LIST_HEAD(&iocb->ki_run_list);
 737                /* we must queue the next iteration ourselves, if it
 738                 * has already been kicked */
 739                if (kiocbIsKicked(iocb)) {
 740                        __queue_kicked_iocb(iocb);
 741
 742                        /*
 743                         * __queue_kicked_iocb will always return 1 here, because
 744                         * iocb->ki_run_list is empty at this point so it should
 745                         * be safe to unconditionally queue the context into the
 746                         * work queue.
 747                         */
 748                        aio_queue_work(ctx);
 749                }
 750        }
 751        return ret;
 752}
 753
 754/*
 755 * __aio_run_iocbs:
 756 *      Process all pending retries queued on the ioctx
 757 *      run list.
 758 * Assumes it is operating within the aio issuer's mm
 759 * context.
 760 */
 761static int __aio_run_iocbs(struct kioctx *ctx)
 762{
 763        struct kiocb *iocb;
 764        struct list_head run_list;
 765
 766        assert_spin_locked(&ctx->ctx_lock);
 767
 768        list_replace_init(&ctx->run_list, &run_list);
 769        while (!list_empty(&run_list)) {
 770                iocb = list_entry(run_list.next, struct kiocb,
 771                        ki_run_list);
 772                list_del(&iocb->ki_run_list);
 773                /*
 774                 * Hold an extra reference while retrying i/o.
 775                 */
 776                iocb->ki_users++;       /* grab extra reference */
 777                aio_run_iocb(iocb);
 778                __aio_put_req(ctx, iocb);
 779        }
 780        if (!list_empty(&ctx->run_list))
 781                return 1;
 782        return 0;
 783}
 784
 785static void aio_queue_work(struct kioctx * ctx)
 786{
 787        unsigned long timeout;
 788        /*
 789         * if someone is waiting, get the work started right
 790         * away, otherwise, use a longer delay
 791         */
 792        smp_mb();
 793        if (waitqueue_active(&ctx->wait))
 794                timeout = 1;
 795        else
 796                timeout = HZ/10;
 797        queue_delayed_work(aio_wq, &ctx->wq, timeout);
 798}
 799
 800/*
 801 * aio_run_all_iocbs:
 802 *      Process all pending retries queued on the ioctx
 803 *      run list, and keep running them until the list
 804 *      stays empty.
 805 * Assumes it is operating within the aio issuer's mm context.
 806 */
 807static inline void aio_run_all_iocbs(struct kioctx *ctx)
 808{
 809        spin_lock_irq(&ctx->ctx_lock);
 810        while (__aio_run_iocbs(ctx))
 811                ;
 812        spin_unlock_irq(&ctx->ctx_lock);
 813}
 814
 815/*
 816 * aio_kick_handler:
 817 *      Work queue handler triggered to process pending
 818 *      retries on an ioctx. Takes on the aio issuer's
 819 *      mm context before running the iocbs, so that
 820 *      copy_xxx_user operates on the issuer's address
 821 *      space.
 822 * Run on aiod's context.
 823 */
 824static void aio_kick_handler(struct work_struct *work)
 825{
 826        struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
 827        mm_segment_t oldfs = get_fs();
 828        struct mm_struct *mm;
 829        int requeue;
 830
 831        set_fs(USER_DS);
 832        use_mm(ctx->mm);
 833        spin_lock_irq(&ctx->ctx_lock);
 834        requeue =__aio_run_iocbs(ctx);
 835        mm = ctx->mm;
 836        spin_unlock_irq(&ctx->ctx_lock);
 837        unuse_mm(mm);
 838        set_fs(oldfs);
 839        /*
 840         * we're in a worker thread already; no point using non-zero delay
 841         */
 842        if (requeue)
 843                queue_delayed_work(aio_wq, &ctx->wq, 0);
 844}
 845
 846
 847/*
 848 * Called by kick_iocb to queue the kiocb for retry
 849 * and if required activate the aio work queue to process
 850 * it
 851 */
 852static void try_queue_kicked_iocb(struct kiocb *iocb)
 853{
 854        struct kioctx   *ctx = iocb->ki_ctx;
 855        unsigned long flags;
 856        int run = 0;
 857
 858        spin_lock_irqsave(&ctx->ctx_lock, flags);
 859        /* set this inside the lock so that we can't race with aio_run_iocb()
 860         * testing it and putting the iocb on the run list under the lock */
 861        if (!kiocbTryKick(iocb))
 862                run = __queue_kicked_iocb(iocb);
 863        spin_unlock_irqrestore(&ctx->ctx_lock, flags);
 864        if (run)
 865                aio_queue_work(ctx);
 866}
 867
 868/*
 869 * kick_iocb:
 870 *      Called typically from a wait queue callback context
 871 *      to trigger a retry of the iocb.
 872 *      The retry is usually executed by aio workqueue
 873 *      threads (See aio_kick_handler).
 874 */
 875void kick_iocb(struct kiocb *iocb)
 876{
 877        /* sync iocbs are easy: they can only ever be executing from a 
 878         * single context. */
 879        if (is_sync_kiocb(iocb)) {
 880                kiocbSetKicked(iocb);
 881                wake_up_process(iocb->ki_obj.tsk);
 882                return;
 883        }
 884
 885        try_queue_kicked_iocb(iocb);
 886}
 887EXPORT_SYMBOL(kick_iocb);
 888
 889/* aio_complete
 890 *      Called when the io request on the given iocb is complete.
 891 *      Returns true if this is the last user of the request.  The 
 892 *      only other user of the request can be the cancellation code.
 893 */
 894int aio_complete(struct kiocb *iocb, long res, long res2)
 895{
 896        struct kioctx   *ctx = iocb->ki_ctx;
 897        struct aio_ring_info    *info;
 898        struct aio_ring *ring;
 899        struct io_event *event;
 900        unsigned long   flags;
 901        unsigned long   tail;
 902        int             ret;
 903
 904        /*
 905         * Special case handling for sync iocbs:
 906         *  - events go directly into the iocb for fast handling
 907         *  - the sync task with the iocb in its stack holds the single iocb
 908         *    ref, no other paths have a way to get another ref
 909         *  - the sync task helpfully left a reference to itself in the iocb
 910         */
 911        if (is_sync_kiocb(iocb)) {
 912                BUG_ON(iocb->ki_users != 1);
 913                iocb->ki_user_data = res;
 914                iocb->ki_users = 0;
 915                wake_up_process(iocb->ki_obj.tsk);
 916                return 1;
 917        }
 918
 919        info = &ctx->ring_info;
 920
 921        /* add a completion event to the ring buffer.
 922         * must be done holding ctx->ctx_lock to prevent
 923         * other code from messing with the tail
 924         * pointer since we might be called from irq
 925         * context.
 926         */
 927        spin_lock_irqsave(&ctx->ctx_lock, flags);
 928
 929        if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
 930                list_del_init(&iocb->ki_run_list);
 931
 932        /*
 933         * cancelled requests don't get events, userland was given one
 934         * when the event got cancelled.
 935         */
 936        if (kiocbIsCancelled(iocb))
 937                goto put_rq;
 938
 939        ring = kmap_atomic(info->ring_pages[0]);
 940
 941        tail = info->tail;
 942        event = aio_ring_event(info, tail);
 943        if (++tail >= info->nr)
 944                tail = 0;
 945
 946        event->obj = (u64)(unsigned long)iocb->ki_obj.user;
 947        event->data = iocb->ki_user_data;
 948        event->res = res;
 949        event->res2 = res2;
 950
 951        dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
 952                ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
 953                res, res2);
 954
 955        /* after flagging the request as done, we
 956         * must never even look at it again
 957         */
 958        smp_wmb();      /* make event visible before updating tail */
 959
 960        info->tail = tail;
 961        ring->tail = tail;
 962
 963        put_aio_ring_event(event);
 964        kunmap_atomic(ring);
 965
 966        pr_debug("added to ring %p at [%lu]\n", iocb, tail);
 967
 968        /*
 969         * Check if the user asked us to deliver the result through an
 970         * eventfd. The eventfd_signal() function is safe to be called
 971         * from IRQ context.
 972         */
 973        if (iocb->ki_eventfd != NULL)
 974                eventfd_signal(iocb->ki_eventfd, 1);
 975
 976put_rq:
 977        /* everything turned out well, dispose of the aiocb. */
 978        ret = __aio_put_req(ctx, iocb);
 979
 980        /*
 981         * We have to order our ring_info tail store above and test
 982         * of the wait list below outside the wait lock.  This is
 983         * like in wake_up_bit() where clearing a bit has to be
 984         * ordered with the unlocked test.
 985         */
 986        smp_mb();
 987
 988        if (waitqueue_active(&ctx->wait))
 989                wake_up(&ctx->wait);
 990
 991        spin_unlock_irqrestore(&ctx->ctx_lock, flags);
 992        return ret;
 993}
 994EXPORT_SYMBOL(aio_complete);
 995
 996/* aio_read_evt
 997 *      Pull an event off of the ioctx's event ring.  Returns the number of 
 998 *      events fetched (0 or 1 ;-)
 999 *      FIXME: make this use cmpxchg.
1000 *      TODO: make the ringbuffer user mmap()able (requires FIXME).
1001 */
1002static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1003{
1004        struct aio_ring_info *info = &ioctx->ring_info;
1005        struct aio_ring *ring;
1006        unsigned long head;
1007        int ret = 0;
1008
1009        ring = kmap_atomic(info->ring_pages[0]);
1010        dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1011                 (unsigned long)ring->head, (unsigned long)ring->tail,
1012                 (unsigned long)ring->nr);
1013
1014        if (ring->head == ring->tail)
1015                goto out;
1016
1017        spin_lock(&info->ring_lock);
1018
1019        head = ring->head % info->nr;
1020        if (head != ring->tail) {
1021                struct io_event *evp = aio_ring_event(info, head);
1022                *ent = *evp;
1023                head = (head + 1) % info->nr;
1024                smp_mb(); /* finish reading the event before updatng the head */
1025                ring->head = head;
1026                ret = 1;
1027                put_aio_ring_event(evp);
1028        }
1029        spin_unlock(&info->ring_lock);
1030
1031out:
1032        dprintk("leaving aio_read_evt: %d  h%lu t%lu\n", ret,
1033                 (unsigned long)ring->head, (unsigned long)ring->tail);
1034        kunmap_atomic(ring);
1035        return ret;
1036}
1037
1038struct aio_timeout {
1039        struct timer_list       timer;
1040        int                     timed_out;
1041        struct task_struct      *p;
1042};
1043
1044static void timeout_func(unsigned long data)
1045{
1046        struct aio_timeout *to = (struct aio_timeout *)data;
1047
1048        to->timed_out = 1;
1049        wake_up_process(to->p);
1050}
1051
1052static inline void init_timeout(struct aio_timeout *to)
1053{
1054        setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
1055        to->timed_out = 0;
1056        to->p = current;
1057}
1058
1059static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1060                               const struct timespec *ts)
1061{
1062        to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1063        if (time_after(to->timer.expires, jiffies))
1064                add_timer(&to->timer);
1065        else
1066                to->timed_out = 1;
1067}
1068
1069static inline void clear_timeout(struct aio_timeout *to)
1070{
1071        del_singleshot_timer_sync(&to->timer);
1072}
1073
1074static int read_events(struct kioctx *ctx,
1075                        long min_nr, long nr,
1076                        struct io_event __user *event,
1077                        struct timespec __user *timeout)
1078{
1079        long                    start_jiffies = jiffies;
1080        struct task_struct      *tsk = current;
1081        DECLARE_WAITQUEUE(wait, tsk);
1082        int                     ret;
1083        int                     i = 0;
1084        struct io_event         ent;
1085        struct aio_timeout      to;
1086        int                     retry = 0;
1087
1088        /* needed to zero any padding within an entry (there shouldn't be 
1089         * any, but C is fun!
1090         */
1091        memset(&ent, 0, sizeof(ent));
1092retry:
1093        ret = 0;
1094        while (likely(i < nr)) {
1095                ret = aio_read_evt(ctx, &ent);
1096                if (unlikely(ret <= 0))
1097                        break;
1098
1099                dprintk("read event: %Lx %Lx %Lx %Lx\n",
1100                        ent.data, ent.obj, ent.res, ent.res2);
1101
1102                /* Could we split the check in two? */
1103                ret = -EFAULT;
1104                if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1105                        dprintk("aio: lost an event due to EFAULT.\n");
1106                        break;
1107                }
1108                ret = 0;
1109
1110                /* Good, event copied to userland, update counts. */
1111                event ++;
1112                i ++;
1113        }
1114
1115        if (min_nr <= i)
1116                return i;
1117        if (ret)
1118                return ret;
1119
1120        /* End fast path */
1121
1122        /* racey check, but it gets redone */
1123        if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1124                retry = 1;
1125                aio_run_all_iocbs(ctx);
1126                goto retry;
1127        }
1128
1129        init_timeout(&to);
1130        if (timeout) {
1131                struct timespec ts;
1132                ret = -EFAULT;
1133                if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1134                        goto out;
1135
1136                set_timeout(start_jiffies, &to, &ts);
1137        }
1138
1139        while (likely(i < nr)) {
1140                add_wait_queue_exclusive(&ctx->wait, &wait);
1141                do {
1142                        set_task_state(tsk, TASK_INTERRUPTIBLE);
1143                        ret = aio_read_evt(ctx, &ent);
1144                        if (ret)
1145                                break;
1146                        if (min_nr <= i)
1147                                break;
1148                        if (unlikely(ctx->dead)) {
1149                                ret = -EINVAL;
1150                                break;
1151                        }
1152                        if (to.timed_out)       /* Only check after read evt */
1153                                break;
1154                        /* Try to only show up in io wait if there are ops
1155                         *  in flight */
1156                        if (ctx->reqs_active)
1157                                io_schedule();
1158                        else
1159                                schedule();
1160                        if (signal_pending(tsk)) {
1161                                ret = -EINTR;
1162                                break;
1163                        }
1164                        /*ret = aio_read_evt(ctx, &ent);*/
1165                } while (1) ;
1166
1167                set_task_state(tsk, TASK_RUNNING);
1168                remove_wait_queue(&ctx->wait, &wait);
1169
1170                if (unlikely(ret <= 0))
1171                        break;
1172
1173                ret = -EFAULT;
1174                if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1175                        dprintk("aio: lost an event due to EFAULT.\n");
1176                        break;
1177                }
1178
1179                /* Good, event copied to userland, update counts. */
1180                event ++;
1181                i ++;
1182        }
1183
1184        if (timeout)
1185                clear_timeout(&to);
1186out:
1187        destroy_timer_on_stack(&to.timer);
1188        return i ? i : ret;
1189}
1190
1191/* Take an ioctx and remove it from the list of ioctx's.  Protects 
1192 * against races with itself via ->dead.
1193 */
1194static void io_destroy(struct kioctx *ioctx)
1195{
1196        struct mm_struct *mm = current->mm;
1197        int was_dead;
1198
1199        /* delete the entry from the list is someone else hasn't already */
1200        spin_lock(&mm->ioctx_lock);
1201        was_dead = ioctx->dead;
1202        ioctx->dead = 1;
1203        hlist_del_rcu(&ioctx->list);
1204        spin_unlock(&mm->ioctx_lock);
1205
1206        dprintk("aio_release(%p)\n", ioctx);
1207        if (likely(!was_dead))
1208                put_ioctx(ioctx);       /* twice for the list */
1209
1210        kill_ctx(ioctx);
1211
1212        /*
1213         * Wake up any waiters.  The setting of ctx->dead must be seen
1214         * by other CPUs at this point.  Right now, we rely on the
1215         * locking done by the above calls to ensure this consistency.
1216         */
1217        wake_up_all(&ioctx->wait);
1218}
1219
1220/* sys_io_setup:
1221 *      Create an aio_context capable of receiving at least nr_events.
1222 *      ctxp must not point to an aio_context that already exists, and
1223 *      must be initialized to 0 prior to the call.  On successful
1224 *      creation of the aio_context, *ctxp is filled in with the resulting 
1225 *      handle.  May fail with -EINVAL if *ctxp is not initialized,
1226 *      if the specified nr_events exceeds internal limits.  May fail 
1227 *      with -EAGAIN if the specified nr_events exceeds the user's limit 
1228 *      of available events.  May fail with -ENOMEM if insufficient kernel
1229 *      resources are available.  May fail with -EFAULT if an invalid
1230 *      pointer is passed for ctxp.  Will fail with -ENOSYS if not
1231 *      implemented.
1232 */
1233SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1234{
1235        struct kioctx *ioctx = NULL;
1236        unsigned long ctx;
1237        long ret;
1238
1239        ret = get_user(ctx, ctxp);
1240        if (unlikely(ret))
1241                goto out;
1242
1243        ret = -EINVAL;
1244        if (unlikely(ctx || nr_events == 0)) {
1245                pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1246                         ctx, nr_events);
1247                goto out;
1248        }
1249
1250        ioctx = ioctx_alloc(nr_events);
1251        ret = PTR_ERR(ioctx);
1252        if (!IS_ERR(ioctx)) {
1253                ret = put_user(ioctx->user_id, ctxp);
1254                if (ret)
1255                        io_destroy(ioctx);
1256                put_ioctx(ioctx);
1257        }
1258
1259out:
1260        return ret;
1261}
1262
1263/* sys_io_destroy:
1264 *      Destroy the aio_context specified.  May cancel any outstanding 
1265 *      AIOs and block on completion.  Will fail with -ENOSYS if not
1266 *      implemented.  May fail with -EINVAL if the context pointed to
1267 *      is invalid.
1268 */
1269SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1270{
1271        struct kioctx *ioctx = lookup_ioctx(ctx);
1272        if (likely(NULL != ioctx)) {
1273                io_destroy(ioctx);
1274                put_ioctx(ioctx);
1275                return 0;
1276        }
1277        pr_debug("EINVAL: io_destroy: invalid context id\n");
1278        return -EINVAL;
1279}
1280
1281static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1282{
1283        struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1284
1285        BUG_ON(ret <= 0);
1286
1287        while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1288                ssize_t this = min((ssize_t)iov->iov_len, ret);
1289                iov->iov_base += this;
1290                iov->iov_len -= this;
1291                iocb->ki_left -= this;
1292                ret -= this;
1293                if (iov->iov_len == 0) {
1294                        iocb->ki_cur_seg++;
1295                        iov++;
1296                }
1297        }
1298
1299        /* the caller should not have done more io than what fit in
1300         * the remaining iovecs */
1301        BUG_ON(ret > 0 && iocb->ki_left == 0);
1302}
1303
1304static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1305{
1306        struct file *file = iocb->ki_filp;
1307        struct address_space *mapping = file->f_mapping;
1308        struct inode *inode = mapping->host;
1309        ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1310                         unsigned long, loff_t);
1311        ssize_t ret = 0;
1312        unsigned short opcode;
1313
1314        if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1315                (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1316                rw_op = file->f_op->aio_read;
1317                opcode = IOCB_CMD_PREADV;
1318        } else {
1319                rw_op = file->f_op->aio_write;
1320                opcode = IOCB_CMD_PWRITEV;
1321        }
1322
1323        /* This matches the pread()/pwrite() logic */
1324        if (iocb->ki_pos < 0)
1325                return -EINVAL;
1326
1327        do {
1328                ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1329                            iocb->ki_nr_segs - iocb->ki_cur_seg,
1330                            iocb->ki_pos);
1331                if (ret > 0)
1332                        aio_advance_iovec(iocb, ret);
1333
1334        /* retry all partial writes.  retry partial reads as long as its a
1335         * regular file. */
1336        } while (ret > 0 && iocb->ki_left > 0 &&
1337                 (opcode == IOCB_CMD_PWRITEV ||
1338                  (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1339
1340        /* This means we must have transferred all that we could */
1341        /* No need to retry anymore */
1342        if ((ret == 0) || (iocb->ki_left == 0))
1343                ret = iocb->ki_nbytes - iocb->ki_left;
1344
1345        /* If we managed to write some out we return that, rather than
1346         * the eventual error. */
1347        if (opcode == IOCB_CMD_PWRITEV
1348            && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
1349            && iocb->ki_nbytes - iocb->ki_left)
1350                ret = iocb->ki_nbytes - iocb->ki_left;
1351
1352        return ret;
1353}
1354
1355static ssize_t aio_fdsync(struct kiocb *iocb)
1356{
1357        struct file *file = iocb->ki_filp;
1358        ssize_t ret = -EINVAL;
1359
1360        if (file->f_op->aio_fsync)
1361                ret = file->f_op->aio_fsync(iocb, 1);
1362        return ret;
1363}
1364
1365static ssize_t aio_fsync(struct kiocb *iocb)
1366{
1367        struct file *file = iocb->ki_filp;
1368        ssize_t ret = -EINVAL;
1369
1370        if (file->f_op->aio_fsync)
1371                ret = file->f_op->aio_fsync(iocb, 0);
1372        return ret;
1373}
1374
1375static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
1376{
1377        ssize_t ret;
1378
1379#ifdef CONFIG_COMPAT
1380        if (compat)
1381                ret = compat_rw_copy_check_uvector(type,
1382                                (struct compat_iovec __user *)kiocb->ki_buf,
1383                                kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1384                                &kiocb->ki_iovec);
1385        else
1386#endif
1387                ret = rw_copy_check_uvector(type,
1388                                (struct iovec __user *)kiocb->ki_buf,
1389                                kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1390                                &kiocb->ki_iovec);
1391        if (ret < 0)
1392                goto out;
1393
1394        ret = rw_verify_area(type, kiocb->ki_filp, &kiocb->ki_pos, ret);
1395        if (ret < 0)
1396                goto out;
1397
1398        kiocb->ki_nr_segs = kiocb->ki_nbytes;
1399        kiocb->ki_cur_seg = 0;
1400        /* ki_nbytes/left now reflect bytes instead of segs */
1401        kiocb->ki_nbytes = ret;
1402        kiocb->ki_left = ret;
1403
1404        ret = 0;
1405out:
1406        return ret;
1407}
1408
1409static ssize_t aio_setup_single_vector(int type, struct file * file, struct kiocb *kiocb)
1410{
1411        int bytes;
1412
1413        bytes = rw_verify_area(type, file, &kiocb->ki_pos, kiocb->ki_left);
1414        if (bytes < 0)
1415                return bytes;
1416
1417        kiocb->ki_iovec = &kiocb->ki_inline_vec;
1418        kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1419        kiocb->ki_iovec->iov_len = bytes;
1420        kiocb->ki_nr_segs = 1;
1421        kiocb->ki_cur_seg = 0;
1422        return 0;
1423}
1424
1425/*
1426 * aio_setup_iocb:
1427 *      Performs the initial checks and aio retry method
1428 *      setup for the kiocb at the time of io submission.
1429 */
1430static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
1431{
1432        struct file *file = kiocb->ki_filp;
1433        ssize_t ret = 0;
1434
1435        switch (kiocb->ki_opcode) {
1436        case IOCB_CMD_PREAD:
1437                ret = -EBADF;
1438                if (unlikely(!(file->f_mode & FMODE_READ)))
1439                        break;
1440                ret = -EFAULT;
1441                if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1442                        kiocb->ki_left)))
1443                        break;
1444                ret = aio_setup_single_vector(READ, file, kiocb);
1445                if (ret)
1446                        break;
1447                ret = -EINVAL;
1448                if (file->f_op->aio_read)
1449                        kiocb->ki_retry = aio_rw_vect_retry;
1450                break;
1451        case IOCB_CMD_PWRITE:
1452                ret = -EBADF;
1453                if (unlikely(!(file->f_mode & FMODE_WRITE)))
1454                        break;
1455                ret = -EFAULT;
1456                if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1457                        kiocb->ki_left)))
1458                        break;
1459                ret = aio_setup_single_vector(WRITE, file, kiocb);
1460                if (ret)
1461                        break;
1462                ret = -EINVAL;
1463                if (file->f_op->aio_write)
1464                        kiocb->ki_retry = aio_rw_vect_retry;
1465                break;
1466        case IOCB_CMD_PREADV:
1467                ret = -EBADF;
1468                if (unlikely(!(file->f_mode & FMODE_READ)))
1469                        break;
1470                ret = aio_setup_vectored_rw(READ, kiocb, compat);
1471                if (ret)
1472                        break;
1473                ret = -EINVAL;
1474                if (file->f_op->aio_read)
1475                        kiocb->ki_retry = aio_rw_vect_retry;
1476                break;
1477        case IOCB_CMD_PWRITEV:
1478                ret = -EBADF;
1479                if (unlikely(!(file->f_mode & FMODE_WRITE)))
1480                        break;
1481                ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1482                if (ret)
1483                        break;
1484                ret = -EINVAL;
1485                if (file->f_op->aio_write)
1486                        kiocb->ki_retry = aio_rw_vect_retry;
1487                break;
1488        case IOCB_CMD_FDSYNC:
1489                ret = -EINVAL;
1490                if (file->f_op->aio_fsync)
1491                        kiocb->ki_retry = aio_fdsync;
1492                break;
1493        case IOCB_CMD_FSYNC:
1494                ret = -EINVAL;
1495                if (file->f_op->aio_fsync)
1496                        kiocb->ki_retry = aio_fsync;
1497                break;
1498        default:
1499                dprintk("EINVAL: io_submit: no operation provided\n");
1500                ret = -EINVAL;
1501        }
1502
1503        if (!kiocb->ki_retry)
1504                return ret;
1505
1506        return 0;
1507}
1508
1509static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1510                         struct iocb *iocb, struct kiocb_batch *batch,
1511                         bool compat)
1512{
1513        struct kiocb *req;
1514        struct file *file;
1515        ssize_t ret;
1516
1517        /* enforce forwards compatibility on users */
1518        if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1519                pr_debug("EINVAL: io_submit: reserve field set\n");
1520                return -EINVAL;
1521        }
1522
1523        /* prevent overflows */
1524        if (unlikely(
1525            (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1526            (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1527            ((ssize_t)iocb->aio_nbytes < 0)
1528           )) {
1529                pr_debug("EINVAL: io_submit: overflow check\n");
1530                return -EINVAL;
1531        }
1532
1533        file = fget(iocb->aio_fildes);
1534        if (unlikely(!file))
1535                return -EBADF;
1536
1537        req = aio_get_req(ctx, batch);  /* returns with 2 references to req */
1538        if (unlikely(!req)) {
1539                fput(file);
1540                return -EAGAIN;
1541        }
1542        req->ki_filp = file;
1543        if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1544                /*
1545                 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1546                 * instance of the file* now. The file descriptor must be
1547                 * an eventfd() fd, and will be signaled for each completed
1548                 * event using the eventfd_signal() function.
1549                 */
1550                req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1551                if (IS_ERR(req->ki_eventfd)) {
1552                        ret = PTR_ERR(req->ki_eventfd);
1553                        req->ki_eventfd = NULL;
1554                        goto out_put_req;
1555                }
1556        }
1557
1558        ret = put_user(req->ki_key, &user_iocb->aio_key);
1559        if (unlikely(ret)) {
1560                dprintk("EFAULT: aio_key\n");
1561                goto out_put_req;
1562        }
1563
1564        req->ki_obj.user = user_iocb;
1565        req->ki_user_data = iocb->aio_data;
1566        req->ki_pos = iocb->aio_offset;
1567
1568        req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1569        req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1570        req->ki_opcode = iocb->aio_lio_opcode;
1571
1572        ret = aio_setup_iocb(req, compat);
1573
1574        if (ret)
1575                goto out_put_req;
1576
1577        spin_lock_irq(&ctx->ctx_lock);
1578        /*
1579         * We could have raced with io_destroy() and are currently holding a
1580         * reference to ctx which should be destroyed. We cannot submit IO
1581         * since ctx gets freed as soon as io_submit() puts its reference.  The
1582         * check here is reliable: io_destroy() sets ctx->dead before waiting
1583         * for outstanding IO and the barrier between these two is realized by
1584         * unlock of mm->ioctx_lock and lock of ctx->ctx_lock.  Analogously we
1585         * increment ctx->reqs_active before checking for ctx->dead and the
1586         * barrier is realized by unlock and lock of ctx->ctx_lock. Thus if we
1587         * don't see ctx->dead set here, io_destroy() waits for our IO to
1588         * finish.
1589         */
1590        if (ctx->dead) {
1591                spin_unlock_irq(&ctx->ctx_lock);
1592                ret = -EINVAL;
1593                goto out_put_req;
1594        }
1595        aio_run_iocb(req);
1596        if (!list_empty(&ctx->run_list)) {
1597                /* drain the run list */
1598                while (__aio_run_iocbs(ctx))
1599                        ;
1600        }
1601        spin_unlock_irq(&ctx->ctx_lock);
1602
1603        aio_put_req(req);       /* drop extra ref to req */
1604        return 0;
1605
1606out_put_req:
1607        aio_put_req(req);       /* drop extra ref to req */
1608        aio_put_req(req);       /* drop i/o ref to req */
1609        return ret;
1610}
1611
1612long do_io_submit(aio_context_t ctx_id, long nr,
1613                  struct iocb __user *__user *iocbpp, bool compat)
1614{
1615        struct kioctx *ctx;
1616        long ret = 0;
1617        int i = 0;
1618        struct blk_plug plug;
1619        struct kiocb_batch batch;
1620
1621        if (unlikely(nr < 0))
1622                return -EINVAL;
1623
1624        if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1625                nr = LONG_MAX/sizeof(*iocbpp);
1626
1627        if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1628                return -EFAULT;
1629
1630        ctx = lookup_ioctx(ctx_id);
1631        if (unlikely(!ctx)) {
1632                pr_debug("EINVAL: io_submit: invalid context id\n");
1633                return -EINVAL;
1634        }
1635
1636        kiocb_batch_init(&batch, nr);
1637
1638        blk_start_plug(&plug);
1639
1640        /*
1641         * AKPM: should this return a partial result if some of the IOs were
1642         * successfully submitted?
1643         */
1644        for (i=0; i<nr; i++) {
1645                struct iocb __user *user_iocb;
1646                struct iocb tmp;
1647
1648                if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1649                        ret = -EFAULT;
1650                        break;
1651                }
1652
1653                if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1654                        ret = -EFAULT;
1655                        break;
1656                }
1657
1658                ret = io_submit_one(ctx, user_iocb, &tmp, &batch, compat);
1659                if (ret)
1660                        break;
1661        }
1662        blk_finish_plug(&plug);
1663
1664        kiocb_batch_free(ctx, &batch);
1665        put_ioctx(ctx);
1666        return i ? i : ret;
1667}
1668
1669/* sys_io_submit:
1670 *      Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1671 *      the number of iocbs queued.  May return -EINVAL if the aio_context
1672 *      specified by ctx_id is invalid, if nr is < 0, if the iocb at
1673 *      *iocbpp[0] is not properly initialized, if the operation specified
1674 *      is invalid for the file descriptor in the iocb.  May fail with
1675 *      -EFAULT if any of the data structures point to invalid data.  May
1676 *      fail with -EBADF if the file descriptor specified in the first
1677 *      iocb is invalid.  May fail with -EAGAIN if insufficient resources
1678 *      are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1679 *      fail with -ENOSYS if not implemented.
1680 */
1681SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1682                struct iocb __user * __user *, iocbpp)
1683{
1684        return do_io_submit(ctx_id, nr, iocbpp, 0);
1685}
1686
1687/* lookup_kiocb
1688 *      Finds a given iocb for cancellation.
1689 */
1690static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1691                                  u32 key)
1692{
1693        struct list_head *pos;
1694
1695        assert_spin_locked(&ctx->ctx_lock);
1696
1697        /* TODO: use a hash or array, this sucks. */
1698        list_for_each(pos, &ctx->active_reqs) {
1699                struct kiocb *kiocb = list_kiocb(pos);
1700                if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1701                        return kiocb;
1702        }
1703        return NULL;
1704}
1705
1706/* sys_io_cancel:
1707 *      Attempts to cancel an iocb previously passed to io_submit.  If
1708 *      the operation is successfully cancelled, the resulting event is
1709 *      copied into the memory pointed to by result without being placed
1710 *      into the completion queue and 0 is returned.  May fail with
1711 *      -EFAULT if any of the data structures pointed to are invalid.
1712 *      May fail with -EINVAL if aio_context specified by ctx_id is
1713 *      invalid.  May fail with -EAGAIN if the iocb specified was not
1714 *      cancelled.  Will fail with -ENOSYS if not implemented.
1715 */
1716SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1717                struct io_event __user *, result)
1718{
1719        int (*cancel)(struct kiocb *iocb, struct io_event *res);
1720        struct kioctx *ctx;
1721        struct kiocb *kiocb;
1722        u32 key;
1723        int ret;
1724
1725        ret = get_user(key, &iocb->aio_key);
1726        if (unlikely(ret))
1727                return -EFAULT;
1728
1729        ctx = lookup_ioctx(ctx_id);
1730        if (unlikely(!ctx))
1731                return -EINVAL;
1732
1733        spin_lock_irq(&ctx->ctx_lock);
1734        ret = -EAGAIN;
1735        kiocb = lookup_kiocb(ctx, iocb, key);
1736        if (kiocb && kiocb->ki_cancel) {
1737                cancel = kiocb->ki_cancel;
1738                kiocb->ki_users ++;
1739                kiocbSetCancelled(kiocb);
1740        } else
1741                cancel = NULL;
1742        spin_unlock_irq(&ctx->ctx_lock);
1743
1744        if (NULL != cancel) {
1745                struct io_event tmp;
1746                pr_debug("calling cancel\n");
1747                memset(&tmp, 0, sizeof(tmp));
1748                tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1749                tmp.data = kiocb->ki_user_data;
1750                ret = cancel(kiocb, &tmp);
1751                if (!ret) {
1752                        /* Cancellation succeeded -- copy the result
1753                         * into the user's buffer.
1754                         */
1755                        if (copy_to_user(result, &tmp, sizeof(tmp)))
1756                                ret = -EFAULT;
1757                }
1758        } else
1759                ret = -EINVAL;
1760
1761        put_ioctx(ctx);
1762
1763        return ret;
1764}
1765
1766/* io_getevents:
1767 *      Attempts to read at least min_nr events and up to nr events from
1768 *      the completion queue for the aio_context specified by ctx_id. If
1769 *      it succeeds, the number of read events is returned. May fail with
1770 *      -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1771 *      out of range, if timeout is out of range.  May fail with -EFAULT
1772 *      if any of the memory specified is invalid.  May return 0 or
1773 *      < min_nr if the timeout specified by timeout has elapsed
1774 *      before sufficient events are available, where timeout == NULL
1775 *      specifies an infinite timeout. Note that the timeout pointed to by
1776 *      timeout is relative and will be updated if not NULL and the
1777 *      operation blocks. Will fail with -ENOSYS if not implemented.
1778 */
1779SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1780                long, min_nr,
1781                long, nr,
1782                struct io_event __user *, events,
1783                struct timespec __user *, timeout)
1784{
1785        struct kioctx *ioctx = lookup_ioctx(ctx_id);
1786        long ret = -EINVAL;
1787
1788        if (likely(ioctx)) {
1789                if (likely(min_nr <= nr && min_nr >= 0))
1790                        ret = read_events(ioctx, min_nr, nr, events, timeout);
1791                put_ioctx(ioctx);
1792        }
1793
1794        asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);
1795        return ret;
1796}
1797
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.