Merge branch 'syscalls' of git://git390.osdl.marist.edu/pub/scm/linux-2.6
[linux-2.6/mini2440.git] / fs / aio.c
blob8fa77e233944474a7f4e850b7aa74f80da915c05
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 static void ctx_rcu_free(struct rcu_head *head)
196 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
197 unsigned nr_events = ctx->max_reqs;
199 kmem_cache_free(kioctx_cachep, ctx);
201 if (nr_events) {
202 spin_lock(&aio_nr_lock);
203 BUG_ON(aio_nr - nr_events > aio_nr);
204 aio_nr -= nr_events;
205 spin_unlock(&aio_nr_lock);
209 /* __put_ioctx
210 * Called when the last user of an aio context has gone away,
211 * and the struct needs to be freed.
213 static void __put_ioctx(struct kioctx *ctx)
215 BUG_ON(ctx->reqs_active);
217 cancel_delayed_work(&ctx->wq);
218 cancel_work_sync(&ctx->wq.work);
219 aio_free_ring(ctx);
220 mmdrop(ctx->mm);
221 ctx->mm = NULL;
222 pr_debug("__put_ioctx: freeing %p\n", ctx);
223 call_rcu(&ctx->rcu_head, ctx_rcu_free);
226 #define get_ioctx(kioctx) do { \
227 BUG_ON(atomic_read(&(kioctx)->users) <= 0); \
228 atomic_inc(&(kioctx)->users); \
229 } while (0)
230 #define put_ioctx(kioctx) do { \
231 BUG_ON(atomic_read(&(kioctx)->users) <= 0); \
232 if (unlikely(atomic_dec_and_test(&(kioctx)->users))) \
233 __put_ioctx(kioctx); \
234 } while (0)
236 /* ioctx_alloc
237 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
239 static struct kioctx *ioctx_alloc(unsigned nr_events)
241 struct mm_struct *mm;
242 struct kioctx *ctx;
243 int did_sync = 0;
245 /* Prevent overflows */
246 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
247 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
248 pr_debug("ENOMEM: nr_events too high\n");
249 return ERR_PTR(-EINVAL);
252 if ((unsigned long)nr_events > aio_max_nr)
253 return ERR_PTR(-EAGAIN);
255 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
256 if (!ctx)
257 return ERR_PTR(-ENOMEM);
259 ctx->max_reqs = nr_events;
260 mm = ctx->mm = current->mm;
261 atomic_inc(&mm->mm_count);
263 atomic_set(&ctx->users, 1);
264 spin_lock_init(&ctx->ctx_lock);
265 spin_lock_init(&ctx->ring_info.ring_lock);
266 init_waitqueue_head(&ctx->wait);
268 INIT_LIST_HEAD(&ctx->active_reqs);
269 INIT_LIST_HEAD(&ctx->run_list);
270 INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
272 if (aio_setup_ring(ctx) < 0)
273 goto out_freectx;
275 /* limit the number of system wide aios */
276 do {
277 spin_lock_bh(&aio_nr_lock);
278 if (aio_nr + nr_events > aio_max_nr ||
279 aio_nr + nr_events < aio_nr)
280 ctx->max_reqs = 0;
281 else
282 aio_nr += ctx->max_reqs;
283 spin_unlock_bh(&aio_nr_lock);
284 if (ctx->max_reqs || did_sync)
285 break;
287 /* wait for rcu callbacks to have completed before giving up */
288 synchronize_rcu();
289 did_sync = 1;
290 ctx->max_reqs = nr_events;
291 } while (1);
293 if (ctx->max_reqs == 0)
294 goto out_cleanup;
296 /* now link into global list. */
297 spin_lock(&mm->ioctx_lock);
298 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
299 spin_unlock(&mm->ioctx_lock);
301 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
302 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
303 return ctx;
305 out_cleanup:
306 __put_ioctx(ctx);
307 return ERR_PTR(-EAGAIN);
309 out_freectx:
310 mmdrop(mm);
311 kmem_cache_free(kioctx_cachep, ctx);
312 ctx = ERR_PTR(-ENOMEM);
314 dprintk("aio: error allocating ioctx %p\n", ctx);
315 return ctx;
318 /* aio_cancel_all
319 * Cancels all outstanding aio requests on an aio context. Used
320 * when the processes owning a context have all exited to encourage
321 * the rapid destruction of the kioctx.
323 static void aio_cancel_all(struct kioctx *ctx)
325 int (*cancel)(struct kiocb *, struct io_event *);
326 struct io_event res;
327 spin_lock_irq(&ctx->ctx_lock);
328 ctx->dead = 1;
329 while (!list_empty(&ctx->active_reqs)) {
330 struct list_head *pos = ctx->active_reqs.next;
331 struct kiocb *iocb = list_kiocb(pos);
332 list_del_init(&iocb->ki_list);
333 cancel = iocb->ki_cancel;
334 kiocbSetCancelled(iocb);
335 if (cancel) {
336 iocb->ki_users++;
337 spin_unlock_irq(&ctx->ctx_lock);
338 cancel(iocb, &res);
339 spin_lock_irq(&ctx->ctx_lock);
342 spin_unlock_irq(&ctx->ctx_lock);
345 static void wait_for_all_aios(struct kioctx *ctx)
347 struct task_struct *tsk = current;
348 DECLARE_WAITQUEUE(wait, tsk);
350 spin_lock_irq(&ctx->ctx_lock);
351 if (!ctx->reqs_active)
352 goto out;
354 add_wait_queue(&ctx->wait, &wait);
355 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
356 while (ctx->reqs_active) {
357 spin_unlock_irq(&ctx->ctx_lock);
358 io_schedule();
359 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
360 spin_lock_irq(&ctx->ctx_lock);
362 __set_task_state(tsk, TASK_RUNNING);
363 remove_wait_queue(&ctx->wait, &wait);
365 out:
366 spin_unlock_irq(&ctx->ctx_lock);
369 /* wait_on_sync_kiocb:
370 * Waits on the given sync kiocb to complete.
372 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
374 while (iocb->ki_users) {
375 set_current_state(TASK_UNINTERRUPTIBLE);
376 if (!iocb->ki_users)
377 break;
378 io_schedule();
380 __set_current_state(TASK_RUNNING);
381 return iocb->ki_user_data;
384 /* exit_aio: called when the last user of mm goes away. At this point,
385 * there is no way for any new requests to be submited or any of the
386 * io_* syscalls to be called on the context. However, there may be
387 * outstanding requests which hold references to the context; as they
388 * go away, they will call put_ioctx and release any pinned memory
389 * associated with the request (held via struct page * references).
391 void exit_aio(struct mm_struct *mm)
393 struct kioctx *ctx;
395 while (!hlist_empty(&mm->ioctx_list)) {
396 ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
397 hlist_del_rcu(&ctx->list);
399 aio_cancel_all(ctx);
401 wait_for_all_aios(ctx);
403 * Ensure we don't leave the ctx on the aio_wq
405 cancel_work_sync(&ctx->wq.work);
407 if (1 != atomic_read(&ctx->users))
408 printk(KERN_DEBUG
409 "exit_aio:ioctx still alive: %d %d %d\n",
410 atomic_read(&ctx->users), ctx->dead,
411 ctx->reqs_active);
412 put_ioctx(ctx);
416 /* aio_get_req
417 * Allocate a slot for an aio request. Increments the users count
418 * of the kioctx so that the kioctx stays around until all requests are
419 * complete. Returns NULL if no requests are free.
421 * Returns with kiocb->users set to 2. The io submit code path holds
422 * an extra reference while submitting the i/o.
423 * This prevents races between the aio code path referencing the
424 * req (after submitting it) and aio_complete() freeing the req.
426 static struct kiocb *__aio_get_req(struct kioctx *ctx)
428 struct kiocb *req = NULL;
429 struct aio_ring *ring;
430 int okay = 0;
432 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
433 if (unlikely(!req))
434 return NULL;
436 req->ki_flags = 0;
437 req->ki_users = 2;
438 req->ki_key = 0;
439 req->ki_ctx = ctx;
440 req->ki_cancel = NULL;
441 req->ki_retry = NULL;
442 req->ki_dtor = NULL;
443 req->private = NULL;
444 req->ki_iovec = NULL;
445 INIT_LIST_HEAD(&req->ki_run_list);
446 req->ki_eventfd = ERR_PTR(-EINVAL);
448 /* Check if the completion queue has enough free space to
449 * accept an event from this io.
451 spin_lock_irq(&ctx->ctx_lock);
452 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
453 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
454 list_add(&req->ki_list, &ctx->active_reqs);
455 ctx->reqs_active++;
456 okay = 1;
458 kunmap_atomic(ring, KM_USER0);
459 spin_unlock_irq(&ctx->ctx_lock);
461 if (!okay) {
462 kmem_cache_free(kiocb_cachep, req);
463 req = NULL;
466 return req;
469 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
471 struct kiocb *req;
472 /* Handle a potential starvation case -- should be exceedingly rare as
473 * requests will be stuck on fput_head only if the aio_fput_routine is
474 * delayed and the requests were the last user of the struct file.
476 req = __aio_get_req(ctx);
477 if (unlikely(NULL == req)) {
478 aio_fput_routine(NULL);
479 req = __aio_get_req(ctx);
481 return req;
484 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
486 assert_spin_locked(&ctx->ctx_lock);
488 if (!IS_ERR(req->ki_eventfd))
489 fput(req->ki_eventfd);
490 if (req->ki_dtor)
491 req->ki_dtor(req);
492 if (req->ki_iovec != &req->ki_inline_vec)
493 kfree(req->ki_iovec);
494 kmem_cache_free(kiocb_cachep, req);
495 ctx->reqs_active--;
497 if (unlikely(!ctx->reqs_active && ctx->dead))
498 wake_up(&ctx->wait);
501 static void aio_fput_routine(struct work_struct *data)
503 spin_lock_irq(&fput_lock);
504 while (likely(!list_empty(&fput_head))) {
505 struct kiocb *req = list_kiocb(fput_head.next);
506 struct kioctx *ctx = req->ki_ctx;
508 list_del(&req->ki_list);
509 spin_unlock_irq(&fput_lock);
511 /* Complete the fput */
512 __fput(req->ki_filp);
514 /* Link the iocb into the context's free list */
515 spin_lock_irq(&ctx->ctx_lock);
516 really_put_req(ctx, req);
517 spin_unlock_irq(&ctx->ctx_lock);
519 put_ioctx(ctx);
520 spin_lock_irq(&fput_lock);
522 spin_unlock_irq(&fput_lock);
525 /* __aio_put_req
526 * Returns true if this put was the last user of the request.
528 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
530 dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
531 req, atomic_long_read(&req->ki_filp->f_count));
533 assert_spin_locked(&ctx->ctx_lock);
535 req->ki_users --;
536 BUG_ON(req->ki_users < 0);
537 if (likely(req->ki_users))
538 return 0;
539 list_del(&req->ki_list); /* remove from active_reqs */
540 req->ki_cancel = NULL;
541 req->ki_retry = NULL;
543 /* Must be done under the lock to serialise against cancellation.
544 * Call this aio_fput as it duplicates fput via the fput_work.
546 if (unlikely(atomic_long_dec_and_test(&req->ki_filp->f_count))) {
547 get_ioctx(ctx);
548 spin_lock(&fput_lock);
549 list_add(&req->ki_list, &fput_head);
550 spin_unlock(&fput_lock);
551 queue_work(aio_wq, &fput_work);
552 } else
553 really_put_req(ctx, req);
554 return 1;
557 /* aio_put_req
558 * Returns true if this put was the last user of the kiocb,
559 * false if the request is still in use.
561 int aio_put_req(struct kiocb *req)
563 struct kioctx *ctx = req->ki_ctx;
564 int ret;
565 spin_lock_irq(&ctx->ctx_lock);
566 ret = __aio_put_req(ctx, req);
567 spin_unlock_irq(&ctx->ctx_lock);
568 return ret;
571 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
573 struct mm_struct *mm = current->mm;
574 struct kioctx *ctx = NULL;
575 struct hlist_node *n;
577 rcu_read_lock();
579 hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) {
580 if (ctx->user_id == ctx_id && !ctx->dead) {
581 get_ioctx(ctx);
582 break;
586 rcu_read_unlock();
587 return ctx;
591 * use_mm
592 * Makes the calling kernel thread take on the specified
593 * mm context.
594 * Called by the retry thread execute retries within the
595 * iocb issuer's mm context, so that copy_from/to_user
596 * operations work seamlessly for aio.
597 * (Note: this routine is intended to be called only
598 * from a kernel thread context)
600 static void use_mm(struct mm_struct *mm)
602 struct mm_struct *active_mm;
603 struct task_struct *tsk = current;
605 task_lock(tsk);
606 active_mm = tsk->active_mm;
607 atomic_inc(&mm->mm_count);
608 tsk->mm = mm;
609 tsk->active_mm = mm;
610 switch_mm(active_mm, mm, tsk);
611 task_unlock(tsk);
613 mmdrop(active_mm);
617 * unuse_mm
618 * Reverses the effect of use_mm, i.e. releases the
619 * specified mm context which was earlier taken on
620 * by the calling kernel thread
621 * (Note: this routine is intended to be called only
622 * from a kernel thread context)
624 static void unuse_mm(struct mm_struct *mm)
626 struct task_struct *tsk = current;
628 task_lock(tsk);
629 tsk->mm = NULL;
630 /* active_mm is still 'mm' */
631 enter_lazy_tlb(mm, tsk);
632 task_unlock(tsk);
636 * Queue up a kiocb to be retried. Assumes that the kiocb
637 * has already been marked as kicked, and places it on
638 * the retry run list for the corresponding ioctx, if it
639 * isn't already queued. Returns 1 if it actually queued
640 * the kiocb (to tell the caller to activate the work
641 * queue to process it), or 0, if it found that it was
642 * already queued.
644 static inline int __queue_kicked_iocb(struct kiocb *iocb)
646 struct kioctx *ctx = iocb->ki_ctx;
648 assert_spin_locked(&ctx->ctx_lock);
650 if (list_empty(&iocb->ki_run_list)) {
651 list_add_tail(&iocb->ki_run_list,
652 &ctx->run_list);
653 return 1;
655 return 0;
658 /* aio_run_iocb
659 * This is the core aio execution routine. It is
660 * invoked both for initial i/o submission and
661 * subsequent retries via the aio_kick_handler.
662 * Expects to be invoked with iocb->ki_ctx->lock
663 * already held. The lock is released and reacquired
664 * as needed during processing.
666 * Calls the iocb retry method (already setup for the
667 * iocb on initial submission) for operation specific
668 * handling, but takes care of most of common retry
669 * execution details for a given iocb. The retry method
670 * needs to be non-blocking as far as possible, to avoid
671 * holding up other iocbs waiting to be serviced by the
672 * retry kernel thread.
674 * The trickier parts in this code have to do with
675 * ensuring that only one retry instance is in progress
676 * for a given iocb at any time. Providing that guarantee
677 * simplifies the coding of individual aio operations as
678 * it avoids various potential races.
680 static ssize_t aio_run_iocb(struct kiocb *iocb)
682 struct kioctx *ctx = iocb->ki_ctx;
683 ssize_t (*retry)(struct kiocb *);
684 ssize_t ret;
686 if (!(retry = iocb->ki_retry)) {
687 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
688 return 0;
692 * We don't want the next retry iteration for this
693 * operation to start until this one has returned and
694 * updated the iocb state. However, wait_queue functions
695 * can trigger a kick_iocb from interrupt context in the
696 * meantime, indicating that data is available for the next
697 * iteration. We want to remember that and enable the
698 * next retry iteration _after_ we are through with
699 * this one.
701 * So, in order to be able to register a "kick", but
702 * prevent it from being queued now, we clear the kick
703 * flag, but make the kick code *think* that the iocb is
704 * still on the run list until we are actually done.
705 * When we are done with this iteration, we check if
706 * the iocb was kicked in the meantime and if so, queue
707 * it up afresh.
710 kiocbClearKicked(iocb);
713 * This is so that aio_complete knows it doesn't need to
714 * pull the iocb off the run list (We can't just call
715 * INIT_LIST_HEAD because we don't want a kick_iocb to
716 * queue this on the run list yet)
718 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
719 spin_unlock_irq(&ctx->ctx_lock);
721 /* Quit retrying if the i/o has been cancelled */
722 if (kiocbIsCancelled(iocb)) {
723 ret = -EINTR;
724 aio_complete(iocb, ret, 0);
725 /* must not access the iocb after this */
726 goto out;
730 * Now we are all set to call the retry method in async
731 * context.
733 ret = retry(iocb);
735 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
736 BUG_ON(!list_empty(&iocb->ki_wait.task_list));
737 aio_complete(iocb, ret, 0);
739 out:
740 spin_lock_irq(&ctx->ctx_lock);
742 if (-EIOCBRETRY == ret) {
744 * OK, now that we are done with this iteration
745 * and know that there is more left to go,
746 * this is where we let go so that a subsequent
747 * "kick" can start the next iteration
750 /* will make __queue_kicked_iocb succeed from here on */
751 INIT_LIST_HEAD(&iocb->ki_run_list);
752 /* we must queue the next iteration ourselves, if it
753 * has already been kicked */
754 if (kiocbIsKicked(iocb)) {
755 __queue_kicked_iocb(iocb);
758 * __queue_kicked_iocb will always return 1 here, because
759 * iocb->ki_run_list is empty at this point so it should
760 * be safe to unconditionally queue the context into the
761 * work queue.
763 aio_queue_work(ctx);
766 return ret;
770 * __aio_run_iocbs:
771 * Process all pending retries queued on the ioctx
772 * run list.
773 * Assumes it is operating within the aio issuer's mm
774 * context.
776 static int __aio_run_iocbs(struct kioctx *ctx)
778 struct kiocb *iocb;
779 struct list_head run_list;
781 assert_spin_locked(&ctx->ctx_lock);
783 list_replace_init(&ctx->run_list, &run_list);
784 while (!list_empty(&run_list)) {
785 iocb = list_entry(run_list.next, struct kiocb,
786 ki_run_list);
787 list_del(&iocb->ki_run_list);
789 * Hold an extra reference while retrying i/o.
791 iocb->ki_users++; /* grab extra reference */
792 aio_run_iocb(iocb);
793 __aio_put_req(ctx, iocb);
795 if (!list_empty(&ctx->run_list))
796 return 1;
797 return 0;
800 static void aio_queue_work(struct kioctx * ctx)
802 unsigned long timeout;
804 * if someone is waiting, get the work started right
805 * away, otherwise, use a longer delay
807 smp_mb();
808 if (waitqueue_active(&ctx->wait))
809 timeout = 1;
810 else
811 timeout = HZ/10;
812 queue_delayed_work(aio_wq, &ctx->wq, timeout);
817 * aio_run_iocbs:
818 * Process all pending retries queued on the ioctx
819 * run list.
820 * Assumes it is operating within the aio issuer's mm
821 * context.
823 static inline void aio_run_iocbs(struct kioctx *ctx)
825 int requeue;
827 spin_lock_irq(&ctx->ctx_lock);
829 requeue = __aio_run_iocbs(ctx);
830 spin_unlock_irq(&ctx->ctx_lock);
831 if (requeue)
832 aio_queue_work(ctx);
836 * just like aio_run_iocbs, but keeps running them until
837 * the list stays empty
839 static inline void aio_run_all_iocbs(struct kioctx *ctx)
841 spin_lock_irq(&ctx->ctx_lock);
842 while (__aio_run_iocbs(ctx))
844 spin_unlock_irq(&ctx->ctx_lock);
848 * aio_kick_handler:
849 * Work queue handler triggered to process pending
850 * retries on an ioctx. Takes on the aio issuer's
851 * mm context before running the iocbs, so that
852 * copy_xxx_user operates on the issuer's address
853 * space.
854 * Run on aiod's context.
856 static void aio_kick_handler(struct work_struct *work)
858 struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
859 mm_segment_t oldfs = get_fs();
860 struct mm_struct *mm;
861 int requeue;
863 set_fs(USER_DS);
864 use_mm(ctx->mm);
865 spin_lock_irq(&ctx->ctx_lock);
866 requeue =__aio_run_iocbs(ctx);
867 mm = ctx->mm;
868 spin_unlock_irq(&ctx->ctx_lock);
869 unuse_mm(mm);
870 set_fs(oldfs);
872 * we're in a worker thread already, don't use queue_delayed_work,
874 if (requeue)
875 queue_delayed_work(aio_wq, &ctx->wq, 0);
880 * Called by kick_iocb to queue the kiocb for retry
881 * and if required activate the aio work queue to process
882 * it
884 static void try_queue_kicked_iocb(struct kiocb *iocb)
886 struct kioctx *ctx = iocb->ki_ctx;
887 unsigned long flags;
888 int run = 0;
890 /* We're supposed to be the only path putting the iocb back on the run
891 * list. If we find that the iocb is *back* on a wait queue already
892 * than retry has happened before we could queue the iocb. This also
893 * means that the retry could have completed and freed our iocb, no
894 * good. */
895 BUG_ON((!list_empty(&iocb->ki_wait.task_list)));
897 spin_lock_irqsave(&ctx->ctx_lock, flags);
898 /* set this inside the lock so that we can't race with aio_run_iocb()
899 * testing it and putting the iocb on the run list under the lock */
900 if (!kiocbTryKick(iocb))
901 run = __queue_kicked_iocb(iocb);
902 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
903 if (run)
904 aio_queue_work(ctx);
908 * kick_iocb:
909 * Called typically from a wait queue callback context
910 * (aio_wake_function) to trigger a retry of the iocb.
911 * The retry is usually executed by aio workqueue
912 * threads (See aio_kick_handler).
914 void kick_iocb(struct kiocb *iocb)
916 /* sync iocbs are easy: they can only ever be executing from a
917 * single context. */
918 if (is_sync_kiocb(iocb)) {
919 kiocbSetKicked(iocb);
920 wake_up_process(iocb->ki_obj.tsk);
921 return;
924 try_queue_kicked_iocb(iocb);
926 EXPORT_SYMBOL(kick_iocb);
928 /* aio_complete
929 * Called when the io request on the given iocb is complete.
930 * Returns true if this is the last user of the request. The
931 * only other user of the request can be the cancellation code.
933 int aio_complete(struct kiocb *iocb, long res, long res2)
935 struct kioctx *ctx = iocb->ki_ctx;
936 struct aio_ring_info *info;
937 struct aio_ring *ring;
938 struct io_event *event;
939 unsigned long flags;
940 unsigned long tail;
941 int ret;
944 * Special case handling for sync iocbs:
945 * - events go directly into the iocb for fast handling
946 * - the sync task with the iocb in its stack holds the single iocb
947 * ref, no other paths have a way to get another ref
948 * - the sync task helpfully left a reference to itself in the iocb
950 if (is_sync_kiocb(iocb)) {
951 BUG_ON(iocb->ki_users != 1);
952 iocb->ki_user_data = res;
953 iocb->ki_users = 0;
954 wake_up_process(iocb->ki_obj.tsk);
955 return 1;
958 info = &ctx->ring_info;
960 /* add a completion event to the ring buffer.
961 * must be done holding ctx->ctx_lock to prevent
962 * other code from messing with the tail
963 * pointer since we might be called from irq
964 * context.
966 spin_lock_irqsave(&ctx->ctx_lock, flags);
968 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
969 list_del_init(&iocb->ki_run_list);
972 * cancelled requests don't get events, userland was given one
973 * when the event got cancelled.
975 if (kiocbIsCancelled(iocb))
976 goto put_rq;
978 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
980 tail = info->tail;
981 event = aio_ring_event(info, tail, KM_IRQ0);
982 if (++tail >= info->nr)
983 tail = 0;
985 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
986 event->data = iocb->ki_user_data;
987 event->res = res;
988 event->res2 = res2;
990 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
991 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
992 res, res2);
994 /* after flagging the request as done, we
995 * must never even look at it again
997 smp_wmb(); /* make event visible before updating tail */
999 info->tail = tail;
1000 ring->tail = tail;
1002 put_aio_ring_event(event, KM_IRQ0);
1003 kunmap_atomic(ring, KM_IRQ1);
1005 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
1008 * Check if the user asked us to deliver the result through an
1009 * eventfd. The eventfd_signal() function is safe to be called
1010 * from IRQ context.
1012 if (!IS_ERR(iocb->ki_eventfd))
1013 eventfd_signal(iocb->ki_eventfd, 1);
1015 put_rq:
1016 /* everything turned out well, dispose of the aiocb. */
1017 ret = __aio_put_req(ctx, iocb);
1020 * We have to order our ring_info tail store above and test
1021 * of the wait list below outside the wait lock. This is
1022 * like in wake_up_bit() where clearing a bit has to be
1023 * ordered with the unlocked test.
1025 smp_mb();
1027 if (waitqueue_active(&ctx->wait))
1028 wake_up(&ctx->wait);
1030 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1031 return ret;
1034 /* aio_read_evt
1035 * Pull an event off of the ioctx's event ring. Returns the number of
1036 * events fetched (0 or 1 ;-)
1037 * FIXME: make this use cmpxchg.
1038 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1040 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1042 struct aio_ring_info *info = &ioctx->ring_info;
1043 struct aio_ring *ring;
1044 unsigned long head;
1045 int ret = 0;
1047 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1048 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1049 (unsigned long)ring->head, (unsigned long)ring->tail,
1050 (unsigned long)ring->nr);
1052 if (ring->head == ring->tail)
1053 goto out;
1055 spin_lock(&info->ring_lock);
1057 head = ring->head % info->nr;
1058 if (head != ring->tail) {
1059 struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1060 *ent = *evp;
1061 head = (head + 1) % info->nr;
1062 smp_mb(); /* finish reading the event before updatng the head */
1063 ring->head = head;
1064 ret = 1;
1065 put_aio_ring_event(evp, KM_USER1);
1067 spin_unlock(&info->ring_lock);
1069 out:
1070 kunmap_atomic(ring, KM_USER0);
1071 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1072 (unsigned long)ring->head, (unsigned long)ring->tail);
1073 return ret;
1076 struct aio_timeout {
1077 struct timer_list timer;
1078 int timed_out;
1079 struct task_struct *p;
1082 static void timeout_func(unsigned long data)
1084 struct aio_timeout *to = (struct aio_timeout *)data;
1086 to->timed_out = 1;
1087 wake_up_process(to->p);
1090 static inline void init_timeout(struct aio_timeout *to)
1092 setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
1093 to->timed_out = 0;
1094 to->p = current;
1097 static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1098 const struct timespec *ts)
1100 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1101 if (time_after(to->timer.expires, jiffies))
1102 add_timer(&to->timer);
1103 else
1104 to->timed_out = 1;
1107 static inline void clear_timeout(struct aio_timeout *to)
1109 del_singleshot_timer_sync(&to->timer);
1112 static int read_events(struct kioctx *ctx,
1113 long min_nr, long nr,
1114 struct io_event __user *event,
1115 struct timespec __user *timeout)
1117 long start_jiffies = jiffies;
1118 struct task_struct *tsk = current;
1119 DECLARE_WAITQUEUE(wait, tsk);
1120 int ret;
1121 int i = 0;
1122 struct io_event ent;
1123 struct aio_timeout to;
1124 int retry = 0;
1126 /* needed to zero any padding within an entry (there shouldn't be
1127 * any, but C is fun!
1129 memset(&ent, 0, sizeof(ent));
1130 retry:
1131 ret = 0;
1132 while (likely(i < nr)) {
1133 ret = aio_read_evt(ctx, &ent);
1134 if (unlikely(ret <= 0))
1135 break;
1137 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1138 ent.data, ent.obj, ent.res, ent.res2);
1140 /* Could we split the check in two? */
1141 ret = -EFAULT;
1142 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1143 dprintk("aio: lost an event due to EFAULT.\n");
1144 break;
1146 ret = 0;
1148 /* Good, event copied to userland, update counts. */
1149 event ++;
1150 i ++;
1153 if (min_nr <= i)
1154 return i;
1155 if (ret)
1156 return ret;
1158 /* End fast path */
1160 /* racey check, but it gets redone */
1161 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1162 retry = 1;
1163 aio_run_all_iocbs(ctx);
1164 goto retry;
1167 init_timeout(&to);
1168 if (timeout) {
1169 struct timespec ts;
1170 ret = -EFAULT;
1171 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1172 goto out;
1174 set_timeout(start_jiffies, &to, &ts);
1177 while (likely(i < nr)) {
1178 add_wait_queue_exclusive(&ctx->wait, &wait);
1179 do {
1180 set_task_state(tsk, TASK_INTERRUPTIBLE);
1181 ret = aio_read_evt(ctx, &ent);
1182 if (ret)
1183 break;
1184 if (min_nr <= i)
1185 break;
1186 if (unlikely(ctx->dead)) {
1187 ret = -EINVAL;
1188 break;
1190 if (to.timed_out) /* Only check after read evt */
1191 break;
1192 /* Try to only show up in io wait if there are ops
1193 * in flight */
1194 if (ctx->reqs_active)
1195 io_schedule();
1196 else
1197 schedule();
1198 if (signal_pending(tsk)) {
1199 ret = -EINTR;
1200 break;
1202 /*ret = aio_read_evt(ctx, &ent);*/
1203 } while (1) ;
1205 set_task_state(tsk, TASK_RUNNING);
1206 remove_wait_queue(&ctx->wait, &wait);
1208 if (unlikely(ret <= 0))
1209 break;
1211 ret = -EFAULT;
1212 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1213 dprintk("aio: lost an event due to EFAULT.\n");
1214 break;
1217 /* Good, event copied to userland, update counts. */
1218 event ++;
1219 i ++;
1222 if (timeout)
1223 clear_timeout(&to);
1224 out:
1225 destroy_timer_on_stack(&to.timer);
1226 return i ? i : ret;
1229 /* Take an ioctx and remove it from the list of ioctx's. Protects
1230 * against races with itself via ->dead.
1232 static void io_destroy(struct kioctx *ioctx)
1234 struct mm_struct *mm = current->mm;
1235 int was_dead;
1237 /* delete the entry from the list is someone else hasn't already */
1238 spin_lock(&mm->ioctx_lock);
1239 was_dead = ioctx->dead;
1240 ioctx->dead = 1;
1241 hlist_del_rcu(&ioctx->list);
1242 spin_unlock(&mm->ioctx_lock);
1244 dprintk("aio_release(%p)\n", ioctx);
1245 if (likely(!was_dead))
1246 put_ioctx(ioctx); /* twice for the list */
1248 aio_cancel_all(ioctx);
1249 wait_for_all_aios(ioctx);
1252 * Wake up any waiters. The setting of ctx->dead must be seen
1253 * by other CPUs at this point. Right now, we rely on the
1254 * locking done by the above calls to ensure this consistency.
1256 wake_up(&ioctx->wait);
1257 put_ioctx(ioctx); /* once for the lookup */
1260 /* sys_io_setup:
1261 * Create an aio_context capable of receiving at least nr_events.
1262 * ctxp must not point to an aio_context that already exists, and
1263 * must be initialized to 0 prior to the call. On successful
1264 * creation of the aio_context, *ctxp is filled in with the resulting
1265 * handle. May fail with -EINVAL if *ctxp is not initialized,
1266 * if the specified nr_events exceeds internal limits. May fail
1267 * with -EAGAIN if the specified nr_events exceeds the user's limit
1268 * of available events. May fail with -ENOMEM if insufficient kernel
1269 * resources are available. May fail with -EFAULT if an invalid
1270 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1271 * implemented.
1273 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1275 struct kioctx *ioctx = NULL;
1276 unsigned long ctx;
1277 long ret;
1279 ret = get_user(ctx, ctxp);
1280 if (unlikely(ret))
1281 goto out;
1283 ret = -EINVAL;
1284 if (unlikely(ctx || nr_events == 0)) {
1285 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1286 ctx, nr_events);
1287 goto out;
1290 ioctx = ioctx_alloc(nr_events);
1291 ret = PTR_ERR(ioctx);
1292 if (!IS_ERR(ioctx)) {
1293 ret = put_user(ioctx->user_id, ctxp);
1294 if (!ret)
1295 return 0;
1297 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1298 io_destroy(ioctx);
1301 out:
1302 return ret;
1305 /* sys_io_destroy:
1306 * Destroy the aio_context specified. May cancel any outstanding
1307 * AIOs and block on completion. Will fail with -ENOSYS if not
1308 * implemented. May fail with -EFAULT if the context pointed to
1309 * is invalid.
1311 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1313 struct kioctx *ioctx = lookup_ioctx(ctx);
1314 if (likely(NULL != ioctx)) {
1315 io_destroy(ioctx);
1316 return 0;
1318 pr_debug("EINVAL: io_destroy: invalid context id\n");
1319 return -EINVAL;
1322 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1324 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1326 BUG_ON(ret <= 0);
1328 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1329 ssize_t this = min((ssize_t)iov->iov_len, ret);
1330 iov->iov_base += this;
1331 iov->iov_len -= this;
1332 iocb->ki_left -= this;
1333 ret -= this;
1334 if (iov->iov_len == 0) {
1335 iocb->ki_cur_seg++;
1336 iov++;
1340 /* the caller should not have done more io than what fit in
1341 * the remaining iovecs */
1342 BUG_ON(ret > 0 && iocb->ki_left == 0);
1345 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1347 struct file *file = iocb->ki_filp;
1348 struct address_space *mapping = file->f_mapping;
1349 struct inode *inode = mapping->host;
1350 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1351 unsigned long, loff_t);
1352 ssize_t ret = 0;
1353 unsigned short opcode;
1355 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1356 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1357 rw_op = file->f_op->aio_read;
1358 opcode = IOCB_CMD_PREADV;
1359 } else {
1360 rw_op = file->f_op->aio_write;
1361 opcode = IOCB_CMD_PWRITEV;
1364 /* This matches the pread()/pwrite() logic */
1365 if (iocb->ki_pos < 0)
1366 return -EINVAL;
1368 do {
1369 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1370 iocb->ki_nr_segs - iocb->ki_cur_seg,
1371 iocb->ki_pos);
1372 if (ret > 0)
1373 aio_advance_iovec(iocb, ret);
1375 /* retry all partial writes. retry partial reads as long as its a
1376 * regular file. */
1377 } while (ret > 0 && iocb->ki_left > 0 &&
1378 (opcode == IOCB_CMD_PWRITEV ||
1379 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1381 /* This means we must have transferred all that we could */
1382 /* No need to retry anymore */
1383 if ((ret == 0) || (iocb->ki_left == 0))
1384 ret = iocb->ki_nbytes - iocb->ki_left;
1386 /* If we managed to write some out we return that, rather than
1387 * the eventual error. */
1388 if (opcode == IOCB_CMD_PWRITEV
1389 && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
1390 && iocb->ki_nbytes - iocb->ki_left)
1391 ret = iocb->ki_nbytes - iocb->ki_left;
1393 return ret;
1396 static ssize_t aio_fdsync(struct kiocb *iocb)
1398 struct file *file = iocb->ki_filp;
1399 ssize_t ret = -EINVAL;
1401 if (file->f_op->aio_fsync)
1402 ret = file->f_op->aio_fsync(iocb, 1);
1403 return ret;
1406 static ssize_t aio_fsync(struct kiocb *iocb)
1408 struct file *file = iocb->ki_filp;
1409 ssize_t ret = -EINVAL;
1411 if (file->f_op->aio_fsync)
1412 ret = file->f_op->aio_fsync(iocb, 0);
1413 return ret;
1416 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb)
1418 ssize_t ret;
1420 ret = rw_copy_check_uvector(type, (struct iovec __user *)kiocb->ki_buf,
1421 kiocb->ki_nbytes, 1,
1422 &kiocb->ki_inline_vec, &kiocb->ki_iovec);
1423 if (ret < 0)
1424 goto out;
1426 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1427 kiocb->ki_cur_seg = 0;
1428 /* ki_nbytes/left now reflect bytes instead of segs */
1429 kiocb->ki_nbytes = ret;
1430 kiocb->ki_left = ret;
1432 ret = 0;
1433 out:
1434 return ret;
1437 static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
1439 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1440 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1441 kiocb->ki_iovec->iov_len = kiocb->ki_left;
1442 kiocb->ki_nr_segs = 1;
1443 kiocb->ki_cur_seg = 0;
1444 return 0;
1448 * aio_setup_iocb:
1449 * Performs the initial checks and aio retry method
1450 * setup for the kiocb at the time of io submission.
1452 static ssize_t aio_setup_iocb(struct kiocb *kiocb)
1454 struct file *file = kiocb->ki_filp;
1455 ssize_t ret = 0;
1457 switch (kiocb->ki_opcode) {
1458 case IOCB_CMD_PREAD:
1459 ret = -EBADF;
1460 if (unlikely(!(file->f_mode & FMODE_READ)))
1461 break;
1462 ret = -EFAULT;
1463 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1464 kiocb->ki_left)))
1465 break;
1466 ret = security_file_permission(file, MAY_READ);
1467 if (unlikely(ret))
1468 break;
1469 ret = aio_setup_single_vector(kiocb);
1470 if (ret)
1471 break;
1472 ret = -EINVAL;
1473 if (file->f_op->aio_read)
1474 kiocb->ki_retry = aio_rw_vect_retry;
1475 break;
1476 case IOCB_CMD_PWRITE:
1477 ret = -EBADF;
1478 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1479 break;
1480 ret = -EFAULT;
1481 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1482 kiocb->ki_left)))
1483 break;
1484 ret = security_file_permission(file, MAY_WRITE);
1485 if (unlikely(ret))
1486 break;
1487 ret = aio_setup_single_vector(kiocb);
1488 if (ret)
1489 break;
1490 ret = -EINVAL;
1491 if (file->f_op->aio_write)
1492 kiocb->ki_retry = aio_rw_vect_retry;
1493 break;
1494 case IOCB_CMD_PREADV:
1495 ret = -EBADF;
1496 if (unlikely(!(file->f_mode & FMODE_READ)))
1497 break;
1498 ret = security_file_permission(file, MAY_READ);
1499 if (unlikely(ret))
1500 break;
1501 ret = aio_setup_vectored_rw(READ, kiocb);
1502 if (ret)
1503 break;
1504 ret = -EINVAL;
1505 if (file->f_op->aio_read)
1506 kiocb->ki_retry = aio_rw_vect_retry;
1507 break;
1508 case IOCB_CMD_PWRITEV:
1509 ret = -EBADF;
1510 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1511 break;
1512 ret = security_file_permission(file, MAY_WRITE);
1513 if (unlikely(ret))
1514 break;
1515 ret = aio_setup_vectored_rw(WRITE, kiocb);
1516 if (ret)
1517 break;
1518 ret = -EINVAL;
1519 if (file->f_op->aio_write)
1520 kiocb->ki_retry = aio_rw_vect_retry;
1521 break;
1522 case IOCB_CMD_FDSYNC:
1523 ret = -EINVAL;
1524 if (file->f_op->aio_fsync)
1525 kiocb->ki_retry = aio_fdsync;
1526 break;
1527 case IOCB_CMD_FSYNC:
1528 ret = -EINVAL;
1529 if (file->f_op->aio_fsync)
1530 kiocb->ki_retry = aio_fsync;
1531 break;
1532 default:
1533 dprintk("EINVAL: io_submit: no operation provided\n");
1534 ret = -EINVAL;
1537 if (!kiocb->ki_retry)
1538 return ret;
1540 return 0;
1544 * aio_wake_function:
1545 * wait queue callback function for aio notification,
1546 * Simply triggers a retry of the operation via kick_iocb.
1548 * This callback is specified in the wait queue entry in
1549 * a kiocb.
1551 * Note:
1552 * This routine is executed with the wait queue lock held.
1553 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests
1554 * the ioctx lock inside the wait queue lock. This is safe
1555 * because this callback isn't used for wait queues which
1556 * are nested inside ioctx lock (i.e. ctx->wait)
1558 static int aio_wake_function(wait_queue_t *wait, unsigned mode,
1559 int sync, void *key)
1561 struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait);
1563 list_del_init(&wait->task_list);
1564 kick_iocb(iocb);
1565 return 1;
1568 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1569 struct iocb *iocb)
1571 struct kiocb *req;
1572 struct file *file;
1573 ssize_t ret;
1575 /* enforce forwards compatibility on users */
1576 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1577 pr_debug("EINVAL: io_submit: reserve field set\n");
1578 return -EINVAL;
1581 /* prevent overflows */
1582 if (unlikely(
1583 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1584 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1585 ((ssize_t)iocb->aio_nbytes < 0)
1586 )) {
1587 pr_debug("EINVAL: io_submit: overflow check\n");
1588 return -EINVAL;
1591 file = fget(iocb->aio_fildes);
1592 if (unlikely(!file))
1593 return -EBADF;
1595 req = aio_get_req(ctx); /* returns with 2 references to req */
1596 if (unlikely(!req)) {
1597 fput(file);
1598 return -EAGAIN;
1600 req->ki_filp = file;
1601 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1603 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1604 * instance of the file* now. The file descriptor must be
1605 * an eventfd() fd, and will be signaled for each completed
1606 * event using the eventfd_signal() function.
1608 req->ki_eventfd = eventfd_fget((int) iocb->aio_resfd);
1609 if (IS_ERR(req->ki_eventfd)) {
1610 ret = PTR_ERR(req->ki_eventfd);
1611 goto out_put_req;
1615 ret = put_user(req->ki_key, &user_iocb->aio_key);
1616 if (unlikely(ret)) {
1617 dprintk("EFAULT: aio_key\n");
1618 goto out_put_req;
1621 req->ki_obj.user = user_iocb;
1622 req->ki_user_data = iocb->aio_data;
1623 req->ki_pos = iocb->aio_offset;
1625 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1626 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1627 req->ki_opcode = iocb->aio_lio_opcode;
1628 init_waitqueue_func_entry(&req->ki_wait, aio_wake_function);
1629 INIT_LIST_HEAD(&req->ki_wait.task_list);
1631 ret = aio_setup_iocb(req);
1633 if (ret)
1634 goto out_put_req;
1636 spin_lock_irq(&ctx->ctx_lock);
1637 aio_run_iocb(req);
1638 if (!list_empty(&ctx->run_list)) {
1639 /* drain the run list */
1640 while (__aio_run_iocbs(ctx))
1643 spin_unlock_irq(&ctx->ctx_lock);
1644 aio_put_req(req); /* drop extra ref to req */
1645 return 0;
1647 out_put_req:
1648 aio_put_req(req); /* drop extra ref to req */
1649 aio_put_req(req); /* drop i/o ref to req */
1650 return ret;
1653 /* sys_io_submit:
1654 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1655 * the number of iocbs queued. May return -EINVAL if the aio_context
1656 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1657 * *iocbpp[0] is not properly initialized, if the operation specified
1658 * is invalid for the file descriptor in the iocb. May fail with
1659 * -EFAULT if any of the data structures point to invalid data. May
1660 * fail with -EBADF if the file descriptor specified in the first
1661 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1662 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1663 * fail with -ENOSYS if not implemented.
1665 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1666 struct iocb __user * __user *, iocbpp)
1668 struct kioctx *ctx;
1669 long ret = 0;
1670 int i;
1672 if (unlikely(nr < 0))
1673 return -EINVAL;
1675 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1676 return -EFAULT;
1678 ctx = lookup_ioctx(ctx_id);
1679 if (unlikely(!ctx)) {
1680 pr_debug("EINVAL: io_submit: invalid context id\n");
1681 return -EINVAL;
1685 * AKPM: should this return a partial result if some of the IOs were
1686 * successfully submitted?
1688 for (i=0; i<nr; i++) {
1689 struct iocb __user *user_iocb;
1690 struct iocb tmp;
1692 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1693 ret = -EFAULT;
1694 break;
1697 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1698 ret = -EFAULT;
1699 break;
1702 ret = io_submit_one(ctx, user_iocb, &tmp);
1703 if (ret)
1704 break;
1707 put_ioctx(ctx);
1708 return i ? i : ret;
1711 /* lookup_kiocb
1712 * Finds a given iocb for cancellation.
1714 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1715 u32 key)
1717 struct list_head *pos;
1719 assert_spin_locked(&ctx->ctx_lock);
1721 /* TODO: use a hash or array, this sucks. */
1722 list_for_each(pos, &ctx->active_reqs) {
1723 struct kiocb *kiocb = list_kiocb(pos);
1724 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1725 return kiocb;
1727 return NULL;
1730 /* sys_io_cancel:
1731 * Attempts to cancel an iocb previously passed to io_submit. If
1732 * the operation is successfully cancelled, the resulting event is
1733 * copied into the memory pointed to by result without being placed
1734 * into the completion queue and 0 is returned. May fail with
1735 * -EFAULT if any of the data structures pointed to are invalid.
1736 * May fail with -EINVAL if aio_context specified by ctx_id is
1737 * invalid. May fail with -EAGAIN if the iocb specified was not
1738 * cancelled. Will fail with -ENOSYS if not implemented.
1740 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1741 struct io_event __user *, result)
1743 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1744 struct kioctx *ctx;
1745 struct kiocb *kiocb;
1746 u32 key;
1747 int ret;
1749 ret = get_user(key, &iocb->aio_key);
1750 if (unlikely(ret))
1751 return -EFAULT;
1753 ctx = lookup_ioctx(ctx_id);
1754 if (unlikely(!ctx))
1755 return -EINVAL;
1757 spin_lock_irq(&ctx->ctx_lock);
1758 ret = -EAGAIN;
1759 kiocb = lookup_kiocb(ctx, iocb, key);
1760 if (kiocb && kiocb->ki_cancel) {
1761 cancel = kiocb->ki_cancel;
1762 kiocb->ki_users ++;
1763 kiocbSetCancelled(kiocb);
1764 } else
1765 cancel = NULL;
1766 spin_unlock_irq(&ctx->ctx_lock);
1768 if (NULL != cancel) {
1769 struct io_event tmp;
1770 pr_debug("calling cancel\n");
1771 memset(&tmp, 0, sizeof(tmp));
1772 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1773 tmp.data = kiocb->ki_user_data;
1774 ret = cancel(kiocb, &tmp);
1775 if (!ret) {
1776 /* Cancellation succeeded -- copy the result
1777 * into the user's buffer.
1779 if (copy_to_user(result, &tmp, sizeof(tmp)))
1780 ret = -EFAULT;
1782 } else
1783 ret = -EINVAL;
1785 put_ioctx(ctx);
1787 return ret;
1790 /* io_getevents:
1791 * Attempts to read at least min_nr events and up to nr events from
1792 * the completion queue for the aio_context specified by ctx_id. May
1793 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range,
1794 * if nr is out of range, if when is out of range. May fail with
1795 * -EFAULT if any of the memory specified to is invalid. May return
1796 * 0 or < min_nr if no events are available and the timeout specified
1797 * by when has elapsed, where when == NULL specifies an infinite
1798 * timeout. Note that the timeout pointed to by when is relative and
1799 * will be updated if not NULL and the operation blocks. Will fail
1800 * with -ENOSYS if not implemented.
1802 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1803 long, min_nr,
1804 long, nr,
1805 struct io_event __user *, events,
1806 struct timespec __user *, timeout)
1808 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1809 long ret = -EINVAL;
1811 if (likely(ioctx)) {
1812 if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0))
1813 ret = read_events(ioctx, min_nr, nr, events, timeout);
1814 put_ioctx(ioctx);
1817 asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);
1818 return ret;
1821 __initcall(aio_setup);
1823 EXPORT_SYMBOL(aio_complete);
1824 EXPORT_SYMBOL(aio_put_req);
1825 EXPORT_SYMBOL(wait_on_sync_kiocb);