[CRYPTO] twofish: Merge common glue code
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / fs / aio.c
blobf12db415c0f6aa7b2776bdbcec1279a35c17da7f
1 /*
2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <bcrl@kvack.org>
5 * Implements an efficient asynchronous io interface.
7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved.
9 * See ../COPYING for licensing terms.
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>
20 #define DEBUG 0
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>
35 #include <asm/kmap_types.h>
36 #include <asm/uaccess.h>
37 #include <asm/mmu_context.h>
39 #if DEBUG > 1
40 #define dprintk printk
41 #else
42 #define dprintk(x...) do { ; } while (0)
43 #endif
45 /*------ sysctl variables----*/
46 static DEFINE_SPINLOCK(aio_nr_lock);
47 unsigned long aio_nr; /* current system wide number of aio requests */
48 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
49 /*----end sysctl variables---*/
51 static struct kmem_cache *kiocb_cachep;
52 static struct kmem_cache *kioctx_cachep;
54 static struct workqueue_struct *aio_wq;
56 /* Used for rare fput completion. */
57 static void aio_fput_routine(struct work_struct *);
58 static DECLARE_WORK(fput_work, aio_fput_routine);
60 static DEFINE_SPINLOCK(fput_lock);
61 static LIST_HEAD(fput_head);
63 static void aio_kick_handler(struct work_struct *);
64 static void aio_queue_work(struct kioctx *);
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.
70 static int __init aio_setup(void)
72 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
73 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
75 aio_wq = create_workqueue("aio");
77 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
79 return 0;
82 static void aio_free_ring(struct kioctx *ctx)
84 struct aio_ring_info *info = &ctx->ring_info;
85 long i;
87 for (i=0; i<info->nr_pages; i++)
88 put_page(info->ring_pages[i]);
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);
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;
102 static int aio_setup_ring(struct kioctx *ctx)
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;
110 /* Compensate for the ring buffer's head/tail overlap entry */
111 nr_events += 2; /* 1 is required, 2 for good luck */
113 size = sizeof(struct aio_ring);
114 size += sizeof(struct io_event) * nr_events;
115 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
117 if (nr_pages < 0)
118 return -EINVAL;
120 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
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;
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,
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;
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);
149 if (unlikely(info->nr_pages != nr_pages)) {
150 aio_free_ring(ctx);
151 return -EAGAIN;
154 ctx->user_id = info->mmap_base;
156 info->nr = nr_events; /* trusted copy */
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);
168 return 0;
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();
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)
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; \
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)
194 /* ioctx_alloc
195 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
197 static struct kioctx *ioctx_alloc(unsigned nr_events)
199 struct mm_struct *mm;
200 struct kioctx *ctx;
202 /* Prevent overflows */
203 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
204 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
205 pr_debug("ENOMEM: nr_events too high\n");
206 return ERR_PTR(-EINVAL);
209 if ((unsigned long)nr_events > aio_max_nr)
210 return ERR_PTR(-EAGAIN);
212 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
213 if (!ctx)
214 return ERR_PTR(-ENOMEM);
216 ctx->max_reqs = nr_events;
217 mm = ctx->mm = current->mm;
218 atomic_inc(&mm->mm_count);
220 atomic_set(&ctx->users, 1);
221 spin_lock_init(&ctx->ctx_lock);
222 spin_lock_init(&ctx->ring_info.ring_lock);
223 init_waitqueue_head(&ctx->wait);
225 INIT_LIST_HEAD(&ctx->active_reqs);
226 INIT_LIST_HEAD(&ctx->run_list);
227 INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
229 if (aio_setup_ring(ctx) < 0)
230 goto out_freectx;
232 /* limit the number of system wide aios */
233 spin_lock(&aio_nr_lock);
234 if (aio_nr + ctx->max_reqs > aio_max_nr ||
235 aio_nr + ctx->max_reqs < aio_nr)
236 ctx->max_reqs = 0;
237 else
238 aio_nr += ctx->max_reqs;
239 spin_unlock(&aio_nr_lock);
240 if (ctx->max_reqs == 0)
241 goto out_cleanup;
243 /* now link into global list. kludge. FIXME */
244 write_lock(&mm->ioctx_list_lock);
245 ctx->next = mm->ioctx_list;
246 mm->ioctx_list = ctx;
247 write_unlock(&mm->ioctx_list_lock);
249 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
250 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
251 return ctx;
253 out_cleanup:
254 __put_ioctx(ctx);
255 return ERR_PTR(-EAGAIN);
257 out_freectx:
258 mmdrop(mm);
259 kmem_cache_free(kioctx_cachep, ctx);
260 ctx = ERR_PTR(-ENOMEM);
262 dprintk("aio: error allocating ioctx %p\n", ctx);
263 return ctx;
266 /* aio_cancel_all
267 * Cancels all outstanding aio requests on an aio context. Used
268 * when the processes owning a context have all exited to encourage
269 * the rapid destruction of the kioctx.
271 static void aio_cancel_all(struct kioctx *ctx)
273 int (*cancel)(struct kiocb *, struct io_event *);
274 struct io_event res;
275 spin_lock_irq(&ctx->ctx_lock);
276 ctx->dead = 1;
277 while (!list_empty(&ctx->active_reqs)) {
278 struct list_head *pos = ctx->active_reqs.next;
279 struct kiocb *iocb = list_kiocb(pos);
280 list_del_init(&iocb->ki_list);
281 cancel = iocb->ki_cancel;
282 kiocbSetCancelled(iocb);
283 if (cancel) {
284 iocb->ki_users++;
285 spin_unlock_irq(&ctx->ctx_lock);
286 cancel(iocb, &res);
287 spin_lock_irq(&ctx->ctx_lock);
290 spin_unlock_irq(&ctx->ctx_lock);
293 static void wait_for_all_aios(struct kioctx *ctx)
295 struct task_struct *tsk = current;
296 DECLARE_WAITQUEUE(wait, tsk);
298 spin_lock_irq(&ctx->ctx_lock);
299 if (!ctx->reqs_active)
300 goto out;
302 add_wait_queue(&ctx->wait, &wait);
303 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
304 while (ctx->reqs_active) {
305 spin_unlock_irq(&ctx->ctx_lock);
306 io_schedule();
307 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
308 spin_lock_irq(&ctx->ctx_lock);
310 __set_task_state(tsk, TASK_RUNNING);
311 remove_wait_queue(&ctx->wait, &wait);
313 out:
314 spin_unlock_irq(&ctx->ctx_lock);
317 /* wait_on_sync_kiocb:
318 * Waits on the given sync kiocb to complete.
320 ssize_t fastcall wait_on_sync_kiocb(struct kiocb *iocb)
322 while (iocb->ki_users) {
323 set_current_state(TASK_UNINTERRUPTIBLE);
324 if (!iocb->ki_users)
325 break;
326 io_schedule();
328 __set_current_state(TASK_RUNNING);
329 return iocb->ki_user_data;
332 /* exit_aio: called when the last user of mm goes away. At this point,
333 * there is no way for any new requests to be submited or any of the
334 * io_* syscalls to be called on the context. However, there may be
335 * outstanding requests which hold references to the context; as they
336 * go away, they will call put_ioctx and release any pinned memory
337 * associated with the request (held via struct page * references).
339 void fastcall exit_aio(struct mm_struct *mm)
341 struct kioctx *ctx = mm->ioctx_list;
342 mm->ioctx_list = NULL;
343 while (ctx) {
344 struct kioctx *next = ctx->next;
345 ctx->next = NULL;
346 aio_cancel_all(ctx);
348 wait_for_all_aios(ctx);
350 * Ensure we don't leave the ctx on the aio_wq
352 cancel_work_sync(&ctx->wq.work);
354 if (1 != atomic_read(&ctx->users))
355 printk(KERN_DEBUG
356 "exit_aio:ioctx still alive: %d %d %d\n",
357 atomic_read(&ctx->users), ctx->dead,
358 ctx->reqs_active);
359 put_ioctx(ctx);
360 ctx = next;
364 /* __put_ioctx
365 * Called when the last user of an aio context has gone away,
366 * and the struct needs to be freed.
368 void fastcall __put_ioctx(struct kioctx *ctx)
370 unsigned nr_events = ctx->max_reqs;
372 BUG_ON(ctx->reqs_active);
374 cancel_delayed_work(&ctx->wq);
375 cancel_work_sync(&ctx->wq.work);
376 aio_free_ring(ctx);
377 mmdrop(ctx->mm);
378 ctx->mm = NULL;
379 pr_debug("__put_ioctx: freeing %p\n", ctx);
380 kmem_cache_free(kioctx_cachep, ctx);
382 if (nr_events) {
383 spin_lock(&aio_nr_lock);
384 BUG_ON(aio_nr - nr_events > aio_nr);
385 aio_nr -= nr_events;
386 spin_unlock(&aio_nr_lock);
390 /* aio_get_req
391 * Allocate a slot for an aio request. Increments the users count
392 * of the kioctx so that the kioctx stays around until all requests are
393 * complete. Returns NULL if no requests are free.
395 * Returns with kiocb->users set to 2. The io submit code path holds
396 * an extra reference while submitting the i/o.
397 * This prevents races between the aio code path referencing the
398 * req (after submitting it) and aio_complete() freeing the req.
400 static struct kiocb *FASTCALL(__aio_get_req(struct kioctx *ctx));
401 static struct kiocb fastcall *__aio_get_req(struct kioctx *ctx)
403 struct kiocb *req = NULL;
404 struct aio_ring *ring;
405 int okay = 0;
407 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
408 if (unlikely(!req))
409 return NULL;
411 req->ki_flags = 0;
412 req->ki_users = 2;
413 req->ki_key = 0;
414 req->ki_ctx = ctx;
415 req->ki_cancel = NULL;
416 req->ki_retry = NULL;
417 req->ki_dtor = NULL;
418 req->private = NULL;
419 req->ki_iovec = NULL;
420 INIT_LIST_HEAD(&req->ki_run_list);
421 req->ki_eventfd = ERR_PTR(-EINVAL);
423 /* Check if the completion queue has enough free space to
424 * accept an event from this io.
426 spin_lock_irq(&ctx->ctx_lock);
427 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
428 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
429 list_add(&req->ki_list, &ctx->active_reqs);
430 ctx->reqs_active++;
431 okay = 1;
433 kunmap_atomic(ring, KM_USER0);
434 spin_unlock_irq(&ctx->ctx_lock);
436 if (!okay) {
437 kmem_cache_free(kiocb_cachep, req);
438 req = NULL;
441 return req;
444 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
446 struct kiocb *req;
447 /* Handle a potential starvation case -- should be exceedingly rare as
448 * requests will be stuck on fput_head only if the aio_fput_routine is
449 * delayed and the requests were the last user of the struct file.
451 req = __aio_get_req(ctx);
452 if (unlikely(NULL == req)) {
453 aio_fput_routine(NULL);
454 req = __aio_get_req(ctx);
456 return req;
459 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
461 assert_spin_locked(&ctx->ctx_lock);
463 if (!IS_ERR(req->ki_eventfd))
464 fput(req->ki_eventfd);
465 if (req->ki_dtor)
466 req->ki_dtor(req);
467 if (req->ki_iovec != &req->ki_inline_vec)
468 kfree(req->ki_iovec);
469 kmem_cache_free(kiocb_cachep, req);
470 ctx->reqs_active--;
472 if (unlikely(!ctx->reqs_active && ctx->dead))
473 wake_up(&ctx->wait);
476 static void aio_fput_routine(struct work_struct *data)
478 spin_lock_irq(&fput_lock);
479 while (likely(!list_empty(&fput_head))) {
480 struct kiocb *req = list_kiocb(fput_head.next);
481 struct kioctx *ctx = req->ki_ctx;
483 list_del(&req->ki_list);
484 spin_unlock_irq(&fput_lock);
486 /* Complete the fput */
487 __fput(req->ki_filp);
489 /* Link the iocb into the context's free list */
490 spin_lock_irq(&ctx->ctx_lock);
491 really_put_req(ctx, req);
492 spin_unlock_irq(&ctx->ctx_lock);
494 put_ioctx(ctx);
495 spin_lock_irq(&fput_lock);
497 spin_unlock_irq(&fput_lock);
500 /* __aio_put_req
501 * Returns true if this put was the last user of the request.
503 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
505 dprintk(KERN_DEBUG "aio_put(%p): f_count=%d\n",
506 req, atomic_read(&req->ki_filp->f_count));
508 assert_spin_locked(&ctx->ctx_lock);
510 req->ki_users --;
511 BUG_ON(req->ki_users < 0);
512 if (likely(req->ki_users))
513 return 0;
514 list_del(&req->ki_list); /* remove from active_reqs */
515 req->ki_cancel = NULL;
516 req->ki_retry = NULL;
518 /* Must be done under the lock to serialise against cancellation.
519 * Call this aio_fput as it duplicates fput via the fput_work.
521 if (unlikely(atomic_dec_and_test(&req->ki_filp->f_count))) {
522 get_ioctx(ctx);
523 spin_lock(&fput_lock);
524 list_add(&req->ki_list, &fput_head);
525 spin_unlock(&fput_lock);
526 queue_work(aio_wq, &fput_work);
527 } else
528 really_put_req(ctx, req);
529 return 1;
532 /* aio_put_req
533 * Returns true if this put was the last user of the kiocb,
534 * false if the request is still in use.
536 int fastcall aio_put_req(struct kiocb *req)
538 struct kioctx *ctx = req->ki_ctx;
539 int ret;
540 spin_lock_irq(&ctx->ctx_lock);
541 ret = __aio_put_req(ctx, req);
542 spin_unlock_irq(&ctx->ctx_lock);
543 return ret;
546 /* Lookup an ioctx id. ioctx_list is lockless for reads.
547 * FIXME: this is O(n) and is only suitable for development.
549 struct kioctx *lookup_ioctx(unsigned long ctx_id)
551 struct kioctx *ioctx;
552 struct mm_struct *mm;
554 mm = current->mm;
555 read_lock(&mm->ioctx_list_lock);
556 for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next)
557 if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) {
558 get_ioctx(ioctx);
559 break;
561 read_unlock(&mm->ioctx_list_lock);
563 return ioctx;
567 * use_mm
568 * Makes the calling kernel thread take on the specified
569 * mm context.
570 * Called by the retry thread execute retries within the
571 * iocb issuer's mm context, so that copy_from/to_user
572 * operations work seamlessly for aio.
573 * (Note: this routine is intended to be called only
574 * from a kernel thread context)
576 static void use_mm(struct mm_struct *mm)
578 struct mm_struct *active_mm;
579 struct task_struct *tsk = current;
581 task_lock(tsk);
582 tsk->flags |= PF_BORROWED_MM;
583 active_mm = tsk->active_mm;
584 atomic_inc(&mm->mm_count);
585 tsk->mm = mm;
586 tsk->active_mm = mm;
588 * Note that on UML this *requires* PF_BORROWED_MM to be set, otherwise
589 * it won't work. Update it accordingly if you change it here
591 switch_mm(active_mm, mm, tsk);
592 task_unlock(tsk);
594 mmdrop(active_mm);
598 * unuse_mm
599 * Reverses the effect of use_mm, i.e. releases the
600 * specified mm context which was earlier taken on
601 * by the calling kernel thread
602 * (Note: this routine is intended to be called only
603 * from a kernel thread context)
605 static void unuse_mm(struct mm_struct *mm)
607 struct task_struct *tsk = current;
609 task_lock(tsk);
610 tsk->flags &= ~PF_BORROWED_MM;
611 tsk->mm = NULL;
612 /* active_mm is still 'mm' */
613 enter_lazy_tlb(mm, tsk);
614 task_unlock(tsk);
618 * Queue up a kiocb to be retried. Assumes that the kiocb
619 * has already been marked as kicked, and places it on
620 * the retry run list for the corresponding ioctx, if it
621 * isn't already queued. Returns 1 if it actually queued
622 * the kiocb (to tell the caller to activate the work
623 * queue to process it), or 0, if it found that it was
624 * already queued.
626 static inline int __queue_kicked_iocb(struct kiocb *iocb)
628 struct kioctx *ctx = iocb->ki_ctx;
630 assert_spin_locked(&ctx->ctx_lock);
632 if (list_empty(&iocb->ki_run_list)) {
633 list_add_tail(&iocb->ki_run_list,
634 &ctx->run_list);
635 return 1;
637 return 0;
640 /* aio_run_iocb
641 * This is the core aio execution routine. It is
642 * invoked both for initial i/o submission and
643 * subsequent retries via the aio_kick_handler.
644 * Expects to be invoked with iocb->ki_ctx->lock
645 * already held. The lock is released and reacquired
646 * as needed during processing.
648 * Calls the iocb retry method (already setup for the
649 * iocb on initial submission) for operation specific
650 * handling, but takes care of most of common retry
651 * execution details for a given iocb. The retry method
652 * needs to be non-blocking as far as possible, to avoid
653 * holding up other iocbs waiting to be serviced by the
654 * retry kernel thread.
656 * The trickier parts in this code have to do with
657 * ensuring that only one retry instance is in progress
658 * for a given iocb at any time. Providing that guarantee
659 * simplifies the coding of individual aio operations as
660 * it avoids various potential races.
662 static ssize_t aio_run_iocb(struct kiocb *iocb)
664 struct kioctx *ctx = iocb->ki_ctx;
665 ssize_t (*retry)(struct kiocb *);
666 ssize_t ret;
668 if (!(retry = iocb->ki_retry)) {
669 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
670 return 0;
674 * We don't want the next retry iteration for this
675 * operation to start until this one has returned and
676 * updated the iocb state. However, wait_queue functions
677 * can trigger a kick_iocb from interrupt context in the
678 * meantime, indicating that data is available for the next
679 * iteration. We want to remember that and enable the
680 * next retry iteration _after_ we are through with
681 * this one.
683 * So, in order to be able to register a "kick", but
684 * prevent it from being queued now, we clear the kick
685 * flag, but make the kick code *think* that the iocb is
686 * still on the run list until we are actually done.
687 * When we are done with this iteration, we check if
688 * the iocb was kicked in the meantime and if so, queue
689 * it up afresh.
692 kiocbClearKicked(iocb);
695 * This is so that aio_complete knows it doesn't need to
696 * pull the iocb off the run list (We can't just call
697 * INIT_LIST_HEAD because we don't want a kick_iocb to
698 * queue this on the run list yet)
700 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
701 spin_unlock_irq(&ctx->ctx_lock);
703 /* Quit retrying if the i/o has been cancelled */
704 if (kiocbIsCancelled(iocb)) {
705 ret = -EINTR;
706 aio_complete(iocb, ret, 0);
707 /* must not access the iocb after this */
708 goto out;
712 * Now we are all set to call the retry method in async
713 * context.
715 ret = retry(iocb);
717 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
718 BUG_ON(!list_empty(&iocb->ki_wait.task_list));
719 aio_complete(iocb, ret, 0);
721 out:
722 spin_lock_irq(&ctx->ctx_lock);
724 if (-EIOCBRETRY == ret) {
726 * OK, now that we are done with this iteration
727 * and know that there is more left to go,
728 * this is where we let go so that a subsequent
729 * "kick" can start the next iteration
732 /* will make __queue_kicked_iocb succeed from here on */
733 INIT_LIST_HEAD(&iocb->ki_run_list);
734 /* we must queue the next iteration ourselves, if it
735 * has already been kicked */
736 if (kiocbIsKicked(iocb)) {
737 __queue_kicked_iocb(iocb);
740 * __queue_kicked_iocb will always return 1 here, because
741 * iocb->ki_run_list is empty at this point so it should
742 * be safe to unconditionally queue the context into the
743 * work queue.
745 aio_queue_work(ctx);
748 return ret;
752 * __aio_run_iocbs:
753 * Process all pending retries queued on the ioctx
754 * run list.
755 * Assumes it is operating within the aio issuer's mm
756 * context.
758 static int __aio_run_iocbs(struct kioctx *ctx)
760 struct kiocb *iocb;
761 struct list_head run_list;
763 assert_spin_locked(&ctx->ctx_lock);
765 list_replace_init(&ctx->run_list, &run_list);
766 while (!list_empty(&run_list)) {
767 iocb = list_entry(run_list.next, struct kiocb,
768 ki_run_list);
769 list_del(&iocb->ki_run_list);
771 * Hold an extra reference while retrying i/o.
773 iocb->ki_users++; /* grab extra reference */
774 aio_run_iocb(iocb);
775 __aio_put_req(ctx, iocb);
777 if (!list_empty(&ctx->run_list))
778 return 1;
779 return 0;
782 static void aio_queue_work(struct kioctx * ctx)
784 unsigned long timeout;
786 * if someone is waiting, get the work started right
787 * away, otherwise, use a longer delay
789 smp_mb();
790 if (waitqueue_active(&ctx->wait))
791 timeout = 1;
792 else
793 timeout = HZ/10;
794 queue_delayed_work(aio_wq, &ctx->wq, timeout);
799 * aio_run_iocbs:
800 * Process all pending retries queued on the ioctx
801 * run list.
802 * Assumes it is operating within the aio issuer's mm
803 * context.
805 static inline void aio_run_iocbs(struct kioctx *ctx)
807 int requeue;
809 spin_lock_irq(&ctx->ctx_lock);
811 requeue = __aio_run_iocbs(ctx);
812 spin_unlock_irq(&ctx->ctx_lock);
813 if (requeue)
814 aio_queue_work(ctx);
818 * just like aio_run_iocbs, but keeps running them until
819 * the list stays empty
821 static inline void aio_run_all_iocbs(struct kioctx *ctx)
823 spin_lock_irq(&ctx->ctx_lock);
824 while (__aio_run_iocbs(ctx))
826 spin_unlock_irq(&ctx->ctx_lock);
830 * aio_kick_handler:
831 * Work queue handler triggered to process pending
832 * retries on an ioctx. Takes on the aio issuer's
833 * mm context before running the iocbs, so that
834 * copy_xxx_user operates on the issuer's address
835 * space.
836 * Run on aiod's context.
838 static void aio_kick_handler(struct work_struct *work)
840 struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
841 mm_segment_t oldfs = get_fs();
842 struct mm_struct *mm;
843 int requeue;
845 set_fs(USER_DS);
846 use_mm(ctx->mm);
847 spin_lock_irq(&ctx->ctx_lock);
848 requeue =__aio_run_iocbs(ctx);
849 mm = ctx->mm;
850 spin_unlock_irq(&ctx->ctx_lock);
851 unuse_mm(mm);
852 set_fs(oldfs);
854 * we're in a worker thread already, don't use queue_delayed_work,
856 if (requeue)
857 queue_delayed_work(aio_wq, &ctx->wq, 0);
862 * Called by kick_iocb to queue the kiocb for retry
863 * and if required activate the aio work queue to process
864 * it
866 static void try_queue_kicked_iocb(struct kiocb *iocb)
868 struct kioctx *ctx = iocb->ki_ctx;
869 unsigned long flags;
870 int run = 0;
872 /* We're supposed to be the only path putting the iocb back on the run
873 * list. If we find that the iocb is *back* on a wait queue already
874 * than retry has happened before we could queue the iocb. This also
875 * means that the retry could have completed and freed our iocb, no
876 * good. */
877 BUG_ON((!list_empty(&iocb->ki_wait.task_list)));
879 spin_lock_irqsave(&ctx->ctx_lock, flags);
880 /* set this inside the lock so that we can't race with aio_run_iocb()
881 * testing it and putting the iocb on the run list under the lock */
882 if (!kiocbTryKick(iocb))
883 run = __queue_kicked_iocb(iocb);
884 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
885 if (run)
886 aio_queue_work(ctx);
890 * kick_iocb:
891 * Called typically from a wait queue callback context
892 * (aio_wake_function) to trigger a retry of the iocb.
893 * The retry is usually executed by aio workqueue
894 * threads (See aio_kick_handler).
896 void fastcall kick_iocb(struct kiocb *iocb)
898 /* sync iocbs are easy: they can only ever be executing from a
899 * single context. */
900 if (is_sync_kiocb(iocb)) {
901 kiocbSetKicked(iocb);
902 wake_up_process(iocb->ki_obj.tsk);
903 return;
906 try_queue_kicked_iocb(iocb);
908 EXPORT_SYMBOL(kick_iocb);
910 /* aio_complete
911 * Called when the io request on the given iocb is complete.
912 * Returns true if this is the last user of the request. The
913 * only other user of the request can be the cancellation code.
915 int fastcall aio_complete(struct kiocb *iocb, long res, long res2)
917 struct kioctx *ctx = iocb->ki_ctx;
918 struct aio_ring_info *info;
919 struct aio_ring *ring;
920 struct io_event *event;
921 unsigned long flags;
922 unsigned long tail;
923 int ret;
926 * Special case handling for sync iocbs:
927 * - events go directly into the iocb for fast handling
928 * - the sync task with the iocb in its stack holds the single iocb
929 * ref, no other paths have a way to get another ref
930 * - the sync task helpfully left a reference to itself in the iocb
932 if (is_sync_kiocb(iocb)) {
933 BUG_ON(iocb->ki_users != 1);
934 iocb->ki_user_data = res;
935 iocb->ki_users = 0;
936 wake_up_process(iocb->ki_obj.tsk);
937 return 1;
941 * Check if the user asked us to deliver the result through an
942 * eventfd. The eventfd_signal() function is safe to be called
943 * from IRQ context.
945 if (!IS_ERR(iocb->ki_eventfd))
946 eventfd_signal(iocb->ki_eventfd, 1);
948 info = &ctx->ring_info;
950 /* add a completion event to the ring buffer.
951 * must be done holding ctx->ctx_lock to prevent
952 * other code from messing with the tail
953 * pointer since we might be called from irq
954 * context.
956 spin_lock_irqsave(&ctx->ctx_lock, flags);
958 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
959 list_del_init(&iocb->ki_run_list);
962 * cancelled requests don't get events, userland was given one
963 * when the event got cancelled.
965 if (kiocbIsCancelled(iocb))
966 goto put_rq;
968 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
970 tail = info->tail;
971 event = aio_ring_event(info, tail, KM_IRQ0);
972 if (++tail >= info->nr)
973 tail = 0;
975 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
976 event->data = iocb->ki_user_data;
977 event->res = res;
978 event->res2 = res2;
980 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
981 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
982 res, res2);
984 /* after flagging the request as done, we
985 * must never even look at it again
987 smp_wmb(); /* make event visible before updating tail */
989 info->tail = tail;
990 ring->tail = tail;
992 put_aio_ring_event(event, KM_IRQ0);
993 kunmap_atomic(ring, KM_IRQ1);
995 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
996 put_rq:
997 /* everything turned out well, dispose of the aiocb. */
998 ret = __aio_put_req(ctx, iocb);
1000 if (waitqueue_active(&ctx->wait))
1001 wake_up(&ctx->wait);
1003 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1004 return ret;
1007 /* aio_read_evt
1008 * Pull an event off of the ioctx's event ring. Returns the number of
1009 * events fetched (0 or 1 ;-)
1010 * FIXME: make this use cmpxchg.
1011 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1013 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1015 struct aio_ring_info *info = &ioctx->ring_info;
1016 struct aio_ring *ring;
1017 unsigned long head;
1018 int ret = 0;
1020 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1021 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1022 (unsigned long)ring->head, (unsigned long)ring->tail,
1023 (unsigned long)ring->nr);
1025 if (ring->head == ring->tail)
1026 goto out;
1028 spin_lock(&info->ring_lock);
1030 head = ring->head % info->nr;
1031 if (head != ring->tail) {
1032 struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1033 *ent = *evp;
1034 head = (head + 1) % info->nr;
1035 smp_mb(); /* finish reading the event before updatng the head */
1036 ring->head = head;
1037 ret = 1;
1038 put_aio_ring_event(evp, KM_USER1);
1040 spin_unlock(&info->ring_lock);
1042 out:
1043 kunmap_atomic(ring, KM_USER0);
1044 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1045 (unsigned long)ring->head, (unsigned long)ring->tail);
1046 return ret;
1049 struct aio_timeout {
1050 struct timer_list timer;
1051 int timed_out;
1052 struct task_struct *p;
1055 static void timeout_func(unsigned long data)
1057 struct aio_timeout *to = (struct aio_timeout *)data;
1059 to->timed_out = 1;
1060 wake_up_process(to->p);
1063 static inline void init_timeout(struct aio_timeout *to)
1065 init_timer(&to->timer);
1066 to->timer.data = (unsigned long)to;
1067 to->timer.function = timeout_func;
1068 to->timed_out = 0;
1069 to->p = current;
1072 static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1073 const struct timespec *ts)
1075 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1076 if (time_after(to->timer.expires, jiffies))
1077 add_timer(&to->timer);
1078 else
1079 to->timed_out = 1;
1082 static inline void clear_timeout(struct aio_timeout *to)
1084 del_singleshot_timer_sync(&to->timer);
1087 static int read_events(struct kioctx *ctx,
1088 long min_nr, long nr,
1089 struct io_event __user *event,
1090 struct timespec __user *timeout)
1092 long start_jiffies = jiffies;
1093 struct task_struct *tsk = current;
1094 DECLARE_WAITQUEUE(wait, tsk);
1095 int ret;
1096 int i = 0;
1097 struct io_event ent;
1098 struct aio_timeout to;
1099 int retry = 0;
1101 /* needed to zero any padding within an entry (there shouldn't be
1102 * any, but C is fun!
1104 memset(&ent, 0, sizeof(ent));
1105 retry:
1106 ret = 0;
1107 while (likely(i < nr)) {
1108 ret = aio_read_evt(ctx, &ent);
1109 if (unlikely(ret <= 0))
1110 break;
1112 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1113 ent.data, ent.obj, ent.res, ent.res2);
1115 /* Could we split the check in two? */
1116 ret = -EFAULT;
1117 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1118 dprintk("aio: lost an event due to EFAULT.\n");
1119 break;
1121 ret = 0;
1123 /* Good, event copied to userland, update counts. */
1124 event ++;
1125 i ++;
1128 if (min_nr <= i)
1129 return i;
1130 if (ret)
1131 return ret;
1133 /* End fast path */
1135 /* racey check, but it gets redone */
1136 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1137 retry = 1;
1138 aio_run_all_iocbs(ctx);
1139 goto retry;
1142 init_timeout(&to);
1143 if (timeout) {
1144 struct timespec ts;
1145 ret = -EFAULT;
1146 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1147 goto out;
1149 set_timeout(start_jiffies, &to, &ts);
1152 while (likely(i < nr)) {
1153 add_wait_queue_exclusive(&ctx->wait, &wait);
1154 do {
1155 set_task_state(tsk, TASK_INTERRUPTIBLE);
1156 ret = aio_read_evt(ctx, &ent);
1157 if (ret)
1158 break;
1159 if (min_nr <= i)
1160 break;
1161 ret = 0;
1162 if (to.timed_out) /* Only check after read evt */
1163 break;
1164 io_schedule();
1165 if (signal_pending(tsk)) {
1166 ret = -EINTR;
1167 break;
1169 /*ret = aio_read_evt(ctx, &ent);*/
1170 } while (1) ;
1172 set_task_state(tsk, TASK_RUNNING);
1173 remove_wait_queue(&ctx->wait, &wait);
1175 if (unlikely(ret <= 0))
1176 break;
1178 ret = -EFAULT;
1179 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1180 dprintk("aio: lost an event due to EFAULT.\n");
1181 break;
1184 /* Good, event copied to userland, update counts. */
1185 event ++;
1186 i ++;
1189 if (timeout)
1190 clear_timeout(&to);
1191 out:
1192 return i ? i : ret;
1195 /* Take an ioctx and remove it from the list of ioctx's. Protects
1196 * against races with itself via ->dead.
1198 static void io_destroy(struct kioctx *ioctx)
1200 struct mm_struct *mm = current->mm;
1201 struct kioctx **tmp;
1202 int was_dead;
1204 /* delete the entry from the list is someone else hasn't already */
1205 write_lock(&mm->ioctx_list_lock);
1206 was_dead = ioctx->dead;
1207 ioctx->dead = 1;
1208 for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx;
1209 tmp = &(*tmp)->next)
1211 if (*tmp)
1212 *tmp = ioctx->next;
1213 write_unlock(&mm->ioctx_list_lock);
1215 dprintk("aio_release(%p)\n", ioctx);
1216 if (likely(!was_dead))
1217 put_ioctx(ioctx); /* twice for the list */
1219 aio_cancel_all(ioctx);
1220 wait_for_all_aios(ioctx);
1221 put_ioctx(ioctx); /* once for the lookup */
1224 /* sys_io_setup:
1225 * Create an aio_context capable of receiving at least nr_events.
1226 * ctxp must not point to an aio_context that already exists, and
1227 * must be initialized to 0 prior to the call. On successful
1228 * creation of the aio_context, *ctxp is filled in with the resulting
1229 * handle. May fail with -EINVAL if *ctxp is not initialized,
1230 * if the specified nr_events exceeds internal limits. May fail
1231 * with -EAGAIN if the specified nr_events exceeds the user's limit
1232 * of available events. May fail with -ENOMEM if insufficient kernel
1233 * resources are available. May fail with -EFAULT if an invalid
1234 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1235 * implemented.
1237 asmlinkage long sys_io_setup(unsigned nr_events, aio_context_t __user *ctxp)
1239 struct kioctx *ioctx = NULL;
1240 unsigned long ctx;
1241 long ret;
1243 ret = get_user(ctx, ctxp);
1244 if (unlikely(ret))
1245 goto out;
1247 ret = -EINVAL;
1248 if (unlikely(ctx || nr_events == 0)) {
1249 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1250 ctx, nr_events);
1251 goto out;
1254 ioctx = ioctx_alloc(nr_events);
1255 ret = PTR_ERR(ioctx);
1256 if (!IS_ERR(ioctx)) {
1257 ret = put_user(ioctx->user_id, ctxp);
1258 if (!ret)
1259 return 0;
1261 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1262 io_destroy(ioctx);
1265 out:
1266 return ret;
1269 /* sys_io_destroy:
1270 * Destroy the aio_context specified. May cancel any outstanding
1271 * AIOs and block on completion. Will fail with -ENOSYS if not
1272 * implemented. May fail with -EFAULT if the context pointed to
1273 * is invalid.
1275 asmlinkage long sys_io_destroy(aio_context_t ctx)
1277 struct kioctx *ioctx = lookup_ioctx(ctx);
1278 if (likely(NULL != ioctx)) {
1279 io_destroy(ioctx);
1280 return 0;
1282 pr_debug("EINVAL: io_destroy: invalid context id\n");
1283 return -EINVAL;
1286 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1288 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1290 BUG_ON(ret <= 0);
1292 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1293 ssize_t this = min((ssize_t)iov->iov_len, ret);
1294 iov->iov_base += this;
1295 iov->iov_len -= this;
1296 iocb->ki_left -= this;
1297 ret -= this;
1298 if (iov->iov_len == 0) {
1299 iocb->ki_cur_seg++;
1300 iov++;
1304 /* the caller should not have done more io than what fit in
1305 * the remaining iovecs */
1306 BUG_ON(ret > 0 && iocb->ki_left == 0);
1309 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1311 struct file *file = iocb->ki_filp;
1312 struct address_space *mapping = file->f_mapping;
1313 struct inode *inode = mapping->host;
1314 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1315 unsigned long, loff_t);
1316 ssize_t ret = 0;
1317 unsigned short opcode;
1319 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1320 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1321 rw_op = file->f_op->aio_read;
1322 opcode = IOCB_CMD_PREADV;
1323 } else {
1324 rw_op = file->f_op->aio_write;
1325 opcode = IOCB_CMD_PWRITEV;
1328 do {
1329 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1330 iocb->ki_nr_segs - iocb->ki_cur_seg,
1331 iocb->ki_pos);
1332 if (ret > 0)
1333 aio_advance_iovec(iocb, ret);
1335 /* retry all partial writes. retry partial reads as long as its a
1336 * regular file. */
1337 } while (ret > 0 && iocb->ki_left > 0 &&
1338 (opcode == IOCB_CMD_PWRITEV ||
1339 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1341 /* This means we must have transferred all that we could */
1342 /* No need to retry anymore */
1343 if ((ret == 0) || (iocb->ki_left == 0))
1344 ret = iocb->ki_nbytes - iocb->ki_left;
1346 return ret;
1349 static ssize_t aio_fdsync(struct kiocb *iocb)
1351 struct file *file = iocb->ki_filp;
1352 ssize_t ret = -EINVAL;
1354 if (file->f_op->aio_fsync)
1355 ret = file->f_op->aio_fsync(iocb, 1);
1356 return ret;
1359 static ssize_t aio_fsync(struct kiocb *iocb)
1361 struct file *file = iocb->ki_filp;
1362 ssize_t ret = -EINVAL;
1364 if (file->f_op->aio_fsync)
1365 ret = file->f_op->aio_fsync(iocb, 0);
1366 return ret;
1369 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb)
1371 ssize_t ret;
1373 ret = rw_copy_check_uvector(type, (struct iovec __user *)kiocb->ki_buf,
1374 kiocb->ki_nbytes, 1,
1375 &kiocb->ki_inline_vec, &kiocb->ki_iovec);
1376 if (ret < 0)
1377 goto out;
1379 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1380 kiocb->ki_cur_seg = 0;
1381 /* ki_nbytes/left now reflect bytes instead of segs */
1382 kiocb->ki_nbytes = ret;
1383 kiocb->ki_left = ret;
1385 ret = 0;
1386 out:
1387 return ret;
1390 static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
1392 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1393 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1394 kiocb->ki_iovec->iov_len = kiocb->ki_left;
1395 kiocb->ki_nr_segs = 1;
1396 kiocb->ki_cur_seg = 0;
1397 return 0;
1401 * aio_setup_iocb:
1402 * Performs the initial checks and aio retry method
1403 * setup for the kiocb at the time of io submission.
1405 static ssize_t aio_setup_iocb(struct kiocb *kiocb)
1407 struct file *file = kiocb->ki_filp;
1408 ssize_t ret = 0;
1410 switch (kiocb->ki_opcode) {
1411 case IOCB_CMD_PREAD:
1412 ret = -EBADF;
1413 if (unlikely(!(file->f_mode & FMODE_READ)))
1414 break;
1415 ret = -EFAULT;
1416 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1417 kiocb->ki_left)))
1418 break;
1419 ret = security_file_permission(file, MAY_READ);
1420 if (unlikely(ret))
1421 break;
1422 ret = aio_setup_single_vector(kiocb);
1423 if (ret)
1424 break;
1425 ret = -EINVAL;
1426 if (file->f_op->aio_read)
1427 kiocb->ki_retry = aio_rw_vect_retry;
1428 break;
1429 case IOCB_CMD_PWRITE:
1430 ret = -EBADF;
1431 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1432 break;
1433 ret = -EFAULT;
1434 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1435 kiocb->ki_left)))
1436 break;
1437 ret = security_file_permission(file, MAY_WRITE);
1438 if (unlikely(ret))
1439 break;
1440 ret = aio_setup_single_vector(kiocb);
1441 if (ret)
1442 break;
1443 ret = -EINVAL;
1444 if (file->f_op->aio_write)
1445 kiocb->ki_retry = aio_rw_vect_retry;
1446 break;
1447 case IOCB_CMD_PREADV:
1448 ret = -EBADF;
1449 if (unlikely(!(file->f_mode & FMODE_READ)))
1450 break;
1451 ret = security_file_permission(file, MAY_READ);
1452 if (unlikely(ret))
1453 break;
1454 ret = aio_setup_vectored_rw(READ, kiocb);
1455 if (ret)
1456 break;
1457 ret = -EINVAL;
1458 if (file->f_op->aio_read)
1459 kiocb->ki_retry = aio_rw_vect_retry;
1460 break;
1461 case IOCB_CMD_PWRITEV:
1462 ret = -EBADF;
1463 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1464 break;
1465 ret = security_file_permission(file, MAY_WRITE);
1466 if (unlikely(ret))
1467 break;
1468 ret = aio_setup_vectored_rw(WRITE, kiocb);
1469 if (ret)
1470 break;
1471 ret = -EINVAL;
1472 if (file->f_op->aio_write)
1473 kiocb->ki_retry = aio_rw_vect_retry;
1474 break;
1475 case IOCB_CMD_FDSYNC:
1476 ret = -EINVAL;
1477 if (file->f_op->aio_fsync)
1478 kiocb->ki_retry = aio_fdsync;
1479 break;
1480 case IOCB_CMD_FSYNC:
1481 ret = -EINVAL;
1482 if (file->f_op->aio_fsync)
1483 kiocb->ki_retry = aio_fsync;
1484 break;
1485 default:
1486 dprintk("EINVAL: io_submit: no operation provided\n");
1487 ret = -EINVAL;
1490 if (!kiocb->ki_retry)
1491 return ret;
1493 return 0;
1497 * aio_wake_function:
1498 * wait queue callback function for aio notification,
1499 * Simply triggers a retry of the operation via kick_iocb.
1501 * This callback is specified in the wait queue entry in
1502 * a kiocb.
1504 * Note:
1505 * This routine is executed with the wait queue lock held.
1506 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests
1507 * the ioctx lock inside the wait queue lock. This is safe
1508 * because this callback isn't used for wait queues which
1509 * are nested inside ioctx lock (i.e. ctx->wait)
1511 static int aio_wake_function(wait_queue_t *wait, unsigned mode,
1512 int sync, void *key)
1514 struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait);
1516 list_del_init(&wait->task_list);
1517 kick_iocb(iocb);
1518 return 1;
1521 int fastcall io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1522 struct iocb *iocb)
1524 struct kiocb *req;
1525 struct file *file;
1526 ssize_t ret;
1528 /* enforce forwards compatibility on users */
1529 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1530 pr_debug("EINVAL: io_submit: reserve field set\n");
1531 return -EINVAL;
1534 /* prevent overflows */
1535 if (unlikely(
1536 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1537 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1538 ((ssize_t)iocb->aio_nbytes < 0)
1539 )) {
1540 pr_debug("EINVAL: io_submit: overflow check\n");
1541 return -EINVAL;
1544 file = fget(iocb->aio_fildes);
1545 if (unlikely(!file))
1546 return -EBADF;
1548 req = aio_get_req(ctx); /* returns with 2 references to req */
1549 if (unlikely(!req)) {
1550 fput(file);
1551 return -EAGAIN;
1553 req->ki_filp = file;
1554 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1556 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1557 * instance of the file* now. The file descriptor must be
1558 * an eventfd() fd, and will be signaled for each completed
1559 * event using the eventfd_signal() function.
1561 req->ki_eventfd = eventfd_fget((int) iocb->aio_resfd);
1562 if (unlikely(IS_ERR(req->ki_eventfd))) {
1563 ret = PTR_ERR(req->ki_eventfd);
1564 goto out_put_req;
1568 ret = put_user(req->ki_key, &user_iocb->aio_key);
1569 if (unlikely(ret)) {
1570 dprintk("EFAULT: aio_key\n");
1571 goto out_put_req;
1574 req->ki_obj.user = user_iocb;
1575 req->ki_user_data = iocb->aio_data;
1576 req->ki_pos = iocb->aio_offset;
1578 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1579 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1580 req->ki_opcode = iocb->aio_lio_opcode;
1581 init_waitqueue_func_entry(&req->ki_wait, aio_wake_function);
1582 INIT_LIST_HEAD(&req->ki_wait.task_list);
1584 ret = aio_setup_iocb(req);
1586 if (ret)
1587 goto out_put_req;
1589 spin_lock_irq(&ctx->ctx_lock);
1590 aio_run_iocb(req);
1591 if (!list_empty(&ctx->run_list)) {
1592 /* drain the run list */
1593 while (__aio_run_iocbs(ctx))
1596 spin_unlock_irq(&ctx->ctx_lock);
1597 aio_put_req(req); /* drop extra ref to req */
1598 return 0;
1600 out_put_req:
1601 aio_put_req(req); /* drop extra ref to req */
1602 aio_put_req(req); /* drop i/o ref to req */
1603 return ret;
1606 /* sys_io_submit:
1607 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1608 * the number of iocbs queued. May return -EINVAL if the aio_context
1609 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1610 * *iocbpp[0] is not properly initialized, if the operation specified
1611 * is invalid for the file descriptor in the iocb. May fail with
1612 * -EFAULT if any of the data structures point to invalid data. May
1613 * fail with -EBADF if the file descriptor specified in the first
1614 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1615 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1616 * fail with -ENOSYS if not implemented.
1618 asmlinkage long sys_io_submit(aio_context_t ctx_id, long nr,
1619 struct iocb __user * __user *iocbpp)
1621 struct kioctx *ctx;
1622 long ret = 0;
1623 int i;
1625 if (unlikely(nr < 0))
1626 return -EINVAL;
1628 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1629 return -EFAULT;
1631 ctx = lookup_ioctx(ctx_id);
1632 if (unlikely(!ctx)) {
1633 pr_debug("EINVAL: io_submit: invalid context id\n");
1634 return -EINVAL;
1638 * AKPM: should this return a partial result if some of the IOs were
1639 * successfully submitted?
1641 for (i=0; i<nr; i++) {
1642 struct iocb __user *user_iocb;
1643 struct iocb tmp;
1645 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1646 ret = -EFAULT;
1647 break;
1650 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1651 ret = -EFAULT;
1652 break;
1655 ret = io_submit_one(ctx, user_iocb, &tmp);
1656 if (ret)
1657 break;
1660 put_ioctx(ctx);
1661 return i ? i : ret;
1664 /* lookup_kiocb
1665 * Finds a given iocb for cancellation.
1667 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1668 u32 key)
1670 struct list_head *pos;
1672 assert_spin_locked(&ctx->ctx_lock);
1674 /* TODO: use a hash or array, this sucks. */
1675 list_for_each(pos, &ctx->active_reqs) {
1676 struct kiocb *kiocb = list_kiocb(pos);
1677 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1678 return kiocb;
1680 return NULL;
1683 /* sys_io_cancel:
1684 * Attempts to cancel an iocb previously passed to io_submit. If
1685 * the operation is successfully cancelled, the resulting event is
1686 * copied into the memory pointed to by result without being placed
1687 * into the completion queue and 0 is returned. May fail with
1688 * -EFAULT if any of the data structures pointed to are invalid.
1689 * May fail with -EINVAL if aio_context specified by ctx_id is
1690 * invalid. May fail with -EAGAIN if the iocb specified was not
1691 * cancelled. Will fail with -ENOSYS if not implemented.
1693 asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb,
1694 struct io_event __user *result)
1696 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1697 struct kioctx *ctx;
1698 struct kiocb *kiocb;
1699 u32 key;
1700 int ret;
1702 ret = get_user(key, &iocb->aio_key);
1703 if (unlikely(ret))
1704 return -EFAULT;
1706 ctx = lookup_ioctx(ctx_id);
1707 if (unlikely(!ctx))
1708 return -EINVAL;
1710 spin_lock_irq(&ctx->ctx_lock);
1711 ret = -EAGAIN;
1712 kiocb = lookup_kiocb(ctx, iocb, key);
1713 if (kiocb && kiocb->ki_cancel) {
1714 cancel = kiocb->ki_cancel;
1715 kiocb->ki_users ++;
1716 kiocbSetCancelled(kiocb);
1717 } else
1718 cancel = NULL;
1719 spin_unlock_irq(&ctx->ctx_lock);
1721 if (NULL != cancel) {
1722 struct io_event tmp;
1723 pr_debug("calling cancel\n");
1724 memset(&tmp, 0, sizeof(tmp));
1725 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1726 tmp.data = kiocb->ki_user_data;
1727 ret = cancel(kiocb, &tmp);
1728 if (!ret) {
1729 /* Cancellation succeeded -- copy the result
1730 * into the user's buffer.
1732 if (copy_to_user(result, &tmp, sizeof(tmp)))
1733 ret = -EFAULT;
1735 } else
1736 ret = -EINVAL;
1738 put_ioctx(ctx);
1740 return ret;
1743 /* io_getevents:
1744 * Attempts to read at least min_nr events and up to nr events from
1745 * the completion queue for the aio_context specified by ctx_id. May
1746 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range,
1747 * if nr is out of range, if when is out of range. May fail with
1748 * -EFAULT if any of the memory specified to is invalid. May return
1749 * 0 or < min_nr if no events are available and the timeout specified
1750 * by when has elapsed, where when == NULL specifies an infinite
1751 * timeout. Note that the timeout pointed to by when is relative and
1752 * will be updated if not NULL and the operation blocks. Will fail
1753 * with -ENOSYS if not implemented.
1755 asmlinkage long sys_io_getevents(aio_context_t ctx_id,
1756 long min_nr,
1757 long nr,
1758 struct io_event __user *events,
1759 struct timespec __user *timeout)
1761 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1762 long ret = -EINVAL;
1764 if (likely(ioctx)) {
1765 if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0))
1766 ret = read_events(ioctx, min_nr, nr, events, timeout);
1767 put_ioctx(ioctx);
1770 return ret;
1773 __initcall(aio_setup);
1775 EXPORT_SYMBOL(aio_complete);
1776 EXPORT_SYMBOL(aio_put_req);
1777 EXPORT_SYMBOL(wait_on_sync_kiocb);