dmatest: don't use set_freezable_with_signal()
[linux-2.6.git] / fs / aio.c
blob78c514cfd212d66b8e6311d4a25435be6493c26d
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/backing-dev.h>
19 #include <linux/uio.h>
21 #define DEBUG 0
23 #include <linux/sched.h>
24 #include <linux/fs.h>
25 #include <linux/file.h>
26 #include <linux/mm.h>
27 #include <linux/mman.h>
28 #include <linux/mmu_context.h>
29 #include <linux/slab.h>
30 #include <linux/timer.h>
31 #include <linux/aio.h>
32 #include <linux/highmem.h>
33 #include <linux/workqueue.h>
34 #include <linux/security.h>
35 #include <linux/eventfd.h>
36 #include <linux/blkdev.h>
37 #include <linux/compat.h>
39 #include <asm/kmap_types.h>
40 #include <asm/uaccess.h>
42 #if DEBUG > 1
43 #define dprintk printk
44 #else
45 #define dprintk(x...) do { ; } while (0)
46 #endif
48 /*------ sysctl variables----*/
49 static DEFINE_SPINLOCK(aio_nr_lock);
50 unsigned long aio_nr; /* current system wide number of aio requests */
51 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
52 /*----end sysctl variables---*/
54 static struct kmem_cache *kiocb_cachep;
55 static struct kmem_cache *kioctx_cachep;
57 static struct workqueue_struct *aio_wq;
59 /* Used for rare fput completion. */
60 static void aio_fput_routine(struct work_struct *);
61 static DECLARE_WORK(fput_work, aio_fput_routine);
63 static DEFINE_SPINLOCK(fput_lock);
64 static LIST_HEAD(fput_head);
66 static void aio_kick_handler(struct work_struct *);
67 static void aio_queue_work(struct kioctx *);
69 /* aio_setup
70 * Creates the slab caches used by the aio routines, panic on
71 * failure as this is done early during the boot sequence.
73 static int __init aio_setup(void)
75 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
76 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
78 aio_wq = alloc_workqueue("aio", 0, 1); /* used to limit concurrency */
79 BUG_ON(!aio_wq);
81 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
83 return 0;
85 __initcall(aio_setup);
87 static void aio_free_ring(struct kioctx *ctx)
89 struct aio_ring_info *info = &ctx->ring_info;
90 long i;
92 for (i=0; i<info->nr_pages; i++)
93 put_page(info->ring_pages[i]);
95 if (info->mmap_size) {
96 down_write(&ctx->mm->mmap_sem);
97 do_munmap(ctx->mm, info->mmap_base, info->mmap_size);
98 up_write(&ctx->mm->mmap_sem);
101 if (info->ring_pages && info->ring_pages != info->internal_pages)
102 kfree(info->ring_pages);
103 info->ring_pages = NULL;
104 info->nr = 0;
107 static int aio_setup_ring(struct kioctx *ctx)
109 struct aio_ring *ring;
110 struct aio_ring_info *info = &ctx->ring_info;
111 unsigned nr_events = ctx->max_reqs;
112 unsigned long size;
113 int nr_pages;
115 /* Compensate for the ring buffer's head/tail overlap entry */
116 nr_events += 2; /* 1 is required, 2 for good luck */
118 size = sizeof(struct aio_ring);
119 size += sizeof(struct io_event) * nr_events;
120 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
122 if (nr_pages < 0)
123 return -EINVAL;
125 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
127 info->nr = 0;
128 info->ring_pages = info->internal_pages;
129 if (nr_pages > AIO_RING_PAGES) {
130 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
131 if (!info->ring_pages)
132 return -ENOMEM;
135 info->mmap_size = nr_pages * PAGE_SIZE;
136 dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
137 down_write(&ctx->mm->mmap_sem);
138 info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
139 PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE,
141 if (IS_ERR((void *)info->mmap_base)) {
142 up_write(&ctx->mm->mmap_sem);
143 info->mmap_size = 0;
144 aio_free_ring(ctx);
145 return -EAGAIN;
148 dprintk("mmap address: 0x%08lx\n", info->mmap_base);
149 info->nr_pages = get_user_pages(current, ctx->mm,
150 info->mmap_base, nr_pages,
151 1, 0, info->ring_pages, NULL);
152 up_write(&ctx->mm->mmap_sem);
154 if (unlikely(info->nr_pages != nr_pages)) {
155 aio_free_ring(ctx);
156 return -EAGAIN;
159 ctx->user_id = info->mmap_base;
161 info->nr = nr_events; /* trusted copy */
163 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
164 ring->nr = nr_events; /* user copy */
165 ring->id = ctx->user_id;
166 ring->head = ring->tail = 0;
167 ring->magic = AIO_RING_MAGIC;
168 ring->compat_features = AIO_RING_COMPAT_FEATURES;
169 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
170 ring->header_length = sizeof(struct aio_ring);
171 kunmap_atomic(ring, KM_USER0);
173 return 0;
177 /* aio_ring_event: returns a pointer to the event at the given index from
178 * kmap_atomic(, km). Release the pointer with put_aio_ring_event();
180 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
181 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
182 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
184 #define aio_ring_event(info, nr, km) ({ \
185 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
186 struct io_event *__event; \
187 __event = kmap_atomic( \
188 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
189 __event += pos % AIO_EVENTS_PER_PAGE; \
190 __event; \
193 #define put_aio_ring_event(event, km) do { \
194 struct io_event *__event = (event); \
195 (void)__event; \
196 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
197 } while(0)
199 static void ctx_rcu_free(struct rcu_head *head)
201 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
202 unsigned nr_events = ctx->max_reqs;
204 kmem_cache_free(kioctx_cachep, ctx);
206 if (nr_events) {
207 spin_lock(&aio_nr_lock);
208 BUG_ON(aio_nr - nr_events > aio_nr);
209 aio_nr -= nr_events;
210 spin_unlock(&aio_nr_lock);
214 /* __put_ioctx
215 * Called when the last user of an aio context has gone away,
216 * and the struct needs to be freed.
218 static void __put_ioctx(struct kioctx *ctx)
220 BUG_ON(ctx->reqs_active);
222 cancel_delayed_work(&ctx->wq);
223 cancel_work_sync(&ctx->wq.work);
224 aio_free_ring(ctx);
225 mmdrop(ctx->mm);
226 ctx->mm = NULL;
227 pr_debug("__put_ioctx: freeing %p\n", ctx);
228 call_rcu(&ctx->rcu_head, ctx_rcu_free);
231 static inline void get_ioctx(struct kioctx *kioctx)
233 BUG_ON(atomic_read(&kioctx->users) <= 0);
234 atomic_inc(&kioctx->users);
237 static inline int try_get_ioctx(struct kioctx *kioctx)
239 return atomic_inc_not_zero(&kioctx->users);
242 static inline void put_ioctx(struct kioctx *kioctx)
244 BUG_ON(atomic_read(&kioctx->users) <= 0);
245 if (unlikely(atomic_dec_and_test(&kioctx->users)))
246 __put_ioctx(kioctx);
249 /* ioctx_alloc
250 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
252 static struct kioctx *ioctx_alloc(unsigned nr_events)
254 struct mm_struct *mm;
255 struct kioctx *ctx;
256 int did_sync = 0;
258 /* Prevent overflows */
259 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
260 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
261 pr_debug("ENOMEM: nr_events too high\n");
262 return ERR_PTR(-EINVAL);
265 if ((unsigned long)nr_events > aio_max_nr)
266 return ERR_PTR(-EAGAIN);
268 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
269 if (!ctx)
270 return ERR_PTR(-ENOMEM);
272 ctx->max_reqs = nr_events;
273 mm = ctx->mm = current->mm;
274 atomic_inc(&mm->mm_count);
276 atomic_set(&ctx->users, 1);
277 spin_lock_init(&ctx->ctx_lock);
278 spin_lock_init(&ctx->ring_info.ring_lock);
279 init_waitqueue_head(&ctx->wait);
281 INIT_LIST_HEAD(&ctx->active_reqs);
282 INIT_LIST_HEAD(&ctx->run_list);
283 INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
285 if (aio_setup_ring(ctx) < 0)
286 goto out_freectx;
288 /* limit the number of system wide aios */
289 do {
290 spin_lock_bh(&aio_nr_lock);
291 if (aio_nr + nr_events > aio_max_nr ||
292 aio_nr + nr_events < aio_nr)
293 ctx->max_reqs = 0;
294 else
295 aio_nr += ctx->max_reqs;
296 spin_unlock_bh(&aio_nr_lock);
297 if (ctx->max_reqs || did_sync)
298 break;
300 /* wait for rcu callbacks to have completed before giving up */
301 synchronize_rcu();
302 did_sync = 1;
303 ctx->max_reqs = nr_events;
304 } while (1);
306 if (ctx->max_reqs == 0)
307 goto out_cleanup;
309 /* now link into global list. */
310 spin_lock(&mm->ioctx_lock);
311 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
312 spin_unlock(&mm->ioctx_lock);
314 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
315 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
316 return ctx;
318 out_cleanup:
319 __put_ioctx(ctx);
320 return ERR_PTR(-EAGAIN);
322 out_freectx:
323 mmdrop(mm);
324 kmem_cache_free(kioctx_cachep, ctx);
325 ctx = ERR_PTR(-ENOMEM);
327 dprintk("aio: error allocating ioctx %p\n", ctx);
328 return ctx;
331 /* aio_cancel_all
332 * Cancels all outstanding aio requests on an aio context. Used
333 * when the processes owning a context have all exited to encourage
334 * the rapid destruction of the kioctx.
336 static void aio_cancel_all(struct kioctx *ctx)
338 int (*cancel)(struct kiocb *, struct io_event *);
339 struct io_event res;
340 spin_lock_irq(&ctx->ctx_lock);
341 ctx->dead = 1;
342 while (!list_empty(&ctx->active_reqs)) {
343 struct list_head *pos = ctx->active_reqs.next;
344 struct kiocb *iocb = list_kiocb(pos);
345 list_del_init(&iocb->ki_list);
346 cancel = iocb->ki_cancel;
347 kiocbSetCancelled(iocb);
348 if (cancel) {
349 iocb->ki_users++;
350 spin_unlock_irq(&ctx->ctx_lock);
351 cancel(iocb, &res);
352 spin_lock_irq(&ctx->ctx_lock);
355 spin_unlock_irq(&ctx->ctx_lock);
358 static void wait_for_all_aios(struct kioctx *ctx)
360 struct task_struct *tsk = current;
361 DECLARE_WAITQUEUE(wait, tsk);
363 spin_lock_irq(&ctx->ctx_lock);
364 if (!ctx->reqs_active)
365 goto out;
367 add_wait_queue(&ctx->wait, &wait);
368 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
369 while (ctx->reqs_active) {
370 spin_unlock_irq(&ctx->ctx_lock);
371 io_schedule();
372 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
373 spin_lock_irq(&ctx->ctx_lock);
375 __set_task_state(tsk, TASK_RUNNING);
376 remove_wait_queue(&ctx->wait, &wait);
378 out:
379 spin_unlock_irq(&ctx->ctx_lock);
382 /* wait_on_sync_kiocb:
383 * Waits on the given sync kiocb to complete.
385 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
387 while (iocb->ki_users) {
388 set_current_state(TASK_UNINTERRUPTIBLE);
389 if (!iocb->ki_users)
390 break;
391 io_schedule();
393 __set_current_state(TASK_RUNNING);
394 return iocb->ki_user_data;
396 EXPORT_SYMBOL(wait_on_sync_kiocb);
398 /* exit_aio: called when the last user of mm goes away. At this point,
399 * there is no way for any new requests to be submited or any of the
400 * io_* syscalls to be called on the context. However, there may be
401 * outstanding requests which hold references to the context; as they
402 * go away, they will call put_ioctx and release any pinned memory
403 * associated with the request (held via struct page * references).
405 void exit_aio(struct mm_struct *mm)
407 struct kioctx *ctx;
409 while (!hlist_empty(&mm->ioctx_list)) {
410 ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
411 hlist_del_rcu(&ctx->list);
413 aio_cancel_all(ctx);
415 wait_for_all_aios(ctx);
417 * Ensure we don't leave the ctx on the aio_wq
419 cancel_work_sync(&ctx->wq.work);
421 if (1 != atomic_read(&ctx->users))
422 printk(KERN_DEBUG
423 "exit_aio:ioctx still alive: %d %d %d\n",
424 atomic_read(&ctx->users), ctx->dead,
425 ctx->reqs_active);
426 put_ioctx(ctx);
430 /* aio_get_req
431 * Allocate a slot for an aio request. Increments the users count
432 * of the kioctx so that the kioctx stays around until all requests are
433 * complete. Returns NULL if no requests are free.
435 * Returns with kiocb->users set to 2. The io submit code path holds
436 * an extra reference while submitting the i/o.
437 * This prevents races between the aio code path referencing the
438 * req (after submitting it) and aio_complete() freeing the req.
440 static struct kiocb *__aio_get_req(struct kioctx *ctx)
442 struct kiocb *req = NULL;
444 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
445 if (unlikely(!req))
446 return NULL;
448 req->ki_flags = 0;
449 req->ki_users = 2;
450 req->ki_key = 0;
451 req->ki_ctx = ctx;
452 req->ki_cancel = NULL;
453 req->ki_retry = NULL;
454 req->ki_dtor = NULL;
455 req->private = NULL;
456 req->ki_iovec = NULL;
457 INIT_LIST_HEAD(&req->ki_run_list);
458 req->ki_eventfd = NULL;
460 return req;
464 * struct kiocb's are allocated in batches to reduce the number of
465 * times the ctx lock is acquired and released.
467 #define KIOCB_BATCH_SIZE 32L
468 struct kiocb_batch {
469 struct list_head head;
470 long count; /* number of requests left to allocate */
473 static void kiocb_batch_init(struct kiocb_batch *batch, long total)
475 INIT_LIST_HEAD(&batch->head);
476 batch->count = total;
479 static void kiocb_batch_free(struct kiocb_batch *batch)
481 struct kiocb *req, *n;
483 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
484 list_del(&req->ki_batch);
485 kmem_cache_free(kiocb_cachep, req);
490 * Allocate a batch of kiocbs. This avoids taking and dropping the
491 * context lock a lot during setup.
493 static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch)
495 unsigned short allocated, to_alloc;
496 long avail;
497 bool called_fput = false;
498 struct kiocb *req, *n;
499 struct aio_ring *ring;
501 to_alloc = min(batch->count, KIOCB_BATCH_SIZE);
502 for (allocated = 0; allocated < to_alloc; allocated++) {
503 req = __aio_get_req(ctx);
504 if (!req)
505 /* allocation failed, go with what we've got */
506 break;
507 list_add(&req->ki_batch, &batch->head);
510 if (allocated == 0)
511 goto out;
513 retry:
514 spin_lock_irq(&ctx->ctx_lock);
515 ring = kmap_atomic(ctx->ring_info.ring_pages[0]);
517 avail = aio_ring_avail(&ctx->ring_info, ring) - ctx->reqs_active;
518 BUG_ON(avail < 0);
519 if (avail == 0 && !called_fput) {
521 * Handle a potential starvation case. It is possible that
522 * we hold the last reference on a struct file, causing us
523 * to delay the final fput to non-irq context. In this case,
524 * ctx->reqs_active is artificially high. Calling the fput
525 * routine here may free up a slot in the event completion
526 * ring, allowing this allocation to succeed.
528 kunmap_atomic(ring);
529 spin_unlock_irq(&ctx->ctx_lock);
530 aio_fput_routine(NULL);
531 called_fput = true;
532 goto retry;
535 if (avail < allocated) {
536 /* Trim back the number of requests. */
537 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
538 list_del(&req->ki_batch);
539 kmem_cache_free(kiocb_cachep, req);
540 if (--allocated <= avail)
541 break;
545 batch->count -= allocated;
546 list_for_each_entry(req, &batch->head, ki_batch) {
547 list_add(&req->ki_list, &ctx->active_reqs);
548 ctx->reqs_active++;
551 kunmap_atomic(ring);
552 spin_unlock_irq(&ctx->ctx_lock);
554 out:
555 return allocated;
558 static inline struct kiocb *aio_get_req(struct kioctx *ctx,
559 struct kiocb_batch *batch)
561 struct kiocb *req;
563 if (list_empty(&batch->head))
564 if (kiocb_batch_refill(ctx, batch) == 0)
565 return NULL;
566 req = list_first_entry(&batch->head, struct kiocb, ki_batch);
567 list_del(&req->ki_batch);
568 return req;
571 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
573 assert_spin_locked(&ctx->ctx_lock);
575 if (req->ki_eventfd != NULL)
576 eventfd_ctx_put(req->ki_eventfd);
577 if (req->ki_dtor)
578 req->ki_dtor(req);
579 if (req->ki_iovec != &req->ki_inline_vec)
580 kfree(req->ki_iovec);
581 kmem_cache_free(kiocb_cachep, req);
582 ctx->reqs_active--;
584 if (unlikely(!ctx->reqs_active && ctx->dead))
585 wake_up_all(&ctx->wait);
588 static void aio_fput_routine(struct work_struct *data)
590 spin_lock_irq(&fput_lock);
591 while (likely(!list_empty(&fput_head))) {
592 struct kiocb *req = list_kiocb(fput_head.next);
593 struct kioctx *ctx = req->ki_ctx;
595 list_del(&req->ki_list);
596 spin_unlock_irq(&fput_lock);
598 /* Complete the fput(s) */
599 if (req->ki_filp != NULL)
600 fput(req->ki_filp);
602 /* Link the iocb into the context's free list */
603 spin_lock_irq(&ctx->ctx_lock);
604 really_put_req(ctx, req);
605 spin_unlock_irq(&ctx->ctx_lock);
607 put_ioctx(ctx);
608 spin_lock_irq(&fput_lock);
610 spin_unlock_irq(&fput_lock);
613 /* __aio_put_req
614 * Returns true if this put was the last user of the request.
616 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
618 dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
619 req, atomic_long_read(&req->ki_filp->f_count));
621 assert_spin_locked(&ctx->ctx_lock);
623 req->ki_users--;
624 BUG_ON(req->ki_users < 0);
625 if (likely(req->ki_users))
626 return 0;
627 list_del(&req->ki_list); /* remove from active_reqs */
628 req->ki_cancel = NULL;
629 req->ki_retry = NULL;
632 * Try to optimize the aio and eventfd file* puts, by avoiding to
633 * schedule work in case it is not final fput() time. In normal cases,
634 * we would not be holding the last reference to the file*, so
635 * this function will be executed w/out any aio kthread wakeup.
637 if (unlikely(!fput_atomic(req->ki_filp))) {
638 get_ioctx(ctx);
639 spin_lock(&fput_lock);
640 list_add(&req->ki_list, &fput_head);
641 spin_unlock(&fput_lock);
642 schedule_work(&fput_work);
643 } else {
644 req->ki_filp = NULL;
645 really_put_req(ctx, req);
647 return 1;
650 /* aio_put_req
651 * Returns true if this put was the last user of the kiocb,
652 * false if the request is still in use.
654 int aio_put_req(struct kiocb *req)
656 struct kioctx *ctx = req->ki_ctx;
657 int ret;
658 spin_lock_irq(&ctx->ctx_lock);
659 ret = __aio_put_req(ctx, req);
660 spin_unlock_irq(&ctx->ctx_lock);
661 return ret;
663 EXPORT_SYMBOL(aio_put_req);
665 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
667 struct mm_struct *mm = current->mm;
668 struct kioctx *ctx, *ret = NULL;
669 struct hlist_node *n;
671 rcu_read_lock();
673 hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) {
675 * RCU protects us against accessing freed memory but
676 * we have to be careful not to get a reference when the
677 * reference count already dropped to 0 (ctx->dead test
678 * is unreliable because of races).
680 if (ctx->user_id == ctx_id && !ctx->dead && try_get_ioctx(ctx)){
681 ret = ctx;
682 break;
686 rcu_read_unlock();
687 return ret;
691 * Queue up a kiocb to be retried. Assumes that the kiocb
692 * has already been marked as kicked, and places it on
693 * the retry run list for the corresponding ioctx, if it
694 * isn't already queued. Returns 1 if it actually queued
695 * the kiocb (to tell the caller to activate the work
696 * queue to process it), or 0, if it found that it was
697 * already queued.
699 static inline int __queue_kicked_iocb(struct kiocb *iocb)
701 struct kioctx *ctx = iocb->ki_ctx;
703 assert_spin_locked(&ctx->ctx_lock);
705 if (list_empty(&iocb->ki_run_list)) {
706 list_add_tail(&iocb->ki_run_list,
707 &ctx->run_list);
708 return 1;
710 return 0;
713 /* aio_run_iocb
714 * This is the core aio execution routine. It is
715 * invoked both for initial i/o submission and
716 * subsequent retries via the aio_kick_handler.
717 * Expects to be invoked with iocb->ki_ctx->lock
718 * already held. The lock is released and reacquired
719 * as needed during processing.
721 * Calls the iocb retry method (already setup for the
722 * iocb on initial submission) for operation specific
723 * handling, but takes care of most of common retry
724 * execution details for a given iocb. The retry method
725 * needs to be non-blocking as far as possible, to avoid
726 * holding up other iocbs waiting to be serviced by the
727 * retry kernel thread.
729 * The trickier parts in this code have to do with
730 * ensuring that only one retry instance is in progress
731 * for a given iocb at any time. Providing that guarantee
732 * simplifies the coding of individual aio operations as
733 * it avoids various potential races.
735 static ssize_t aio_run_iocb(struct kiocb *iocb)
737 struct kioctx *ctx = iocb->ki_ctx;
738 ssize_t (*retry)(struct kiocb *);
739 ssize_t ret;
741 if (!(retry = iocb->ki_retry)) {
742 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
743 return 0;
747 * We don't want the next retry iteration for this
748 * operation to start until this one has returned and
749 * updated the iocb state. However, wait_queue functions
750 * can trigger a kick_iocb from interrupt context in the
751 * meantime, indicating that data is available for the next
752 * iteration. We want to remember that and enable the
753 * next retry iteration _after_ we are through with
754 * this one.
756 * So, in order to be able to register a "kick", but
757 * prevent it from being queued now, we clear the kick
758 * flag, but make the kick code *think* that the iocb is
759 * still on the run list until we are actually done.
760 * When we are done with this iteration, we check if
761 * the iocb was kicked in the meantime and if so, queue
762 * it up afresh.
765 kiocbClearKicked(iocb);
768 * This is so that aio_complete knows it doesn't need to
769 * pull the iocb off the run list (We can't just call
770 * INIT_LIST_HEAD because we don't want a kick_iocb to
771 * queue this on the run list yet)
773 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
774 spin_unlock_irq(&ctx->ctx_lock);
776 /* Quit retrying if the i/o has been cancelled */
777 if (kiocbIsCancelled(iocb)) {
778 ret = -EINTR;
779 aio_complete(iocb, ret, 0);
780 /* must not access the iocb after this */
781 goto out;
785 * Now we are all set to call the retry method in async
786 * context.
788 ret = retry(iocb);
790 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
792 * There's no easy way to restart the syscall since other AIO's
793 * may be already running. Just fail this IO with EINTR.
795 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
796 ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
797 ret = -EINTR;
798 aio_complete(iocb, ret, 0);
800 out:
801 spin_lock_irq(&ctx->ctx_lock);
803 if (-EIOCBRETRY == ret) {
805 * OK, now that we are done with this iteration
806 * and know that there is more left to go,
807 * this is where we let go so that a subsequent
808 * "kick" can start the next iteration
811 /* will make __queue_kicked_iocb succeed from here on */
812 INIT_LIST_HEAD(&iocb->ki_run_list);
813 /* we must queue the next iteration ourselves, if it
814 * has already been kicked */
815 if (kiocbIsKicked(iocb)) {
816 __queue_kicked_iocb(iocb);
819 * __queue_kicked_iocb will always return 1 here, because
820 * iocb->ki_run_list is empty at this point so it should
821 * be safe to unconditionally queue the context into the
822 * work queue.
824 aio_queue_work(ctx);
827 return ret;
831 * __aio_run_iocbs:
832 * Process all pending retries queued on the ioctx
833 * run list.
834 * Assumes it is operating within the aio issuer's mm
835 * context.
837 static int __aio_run_iocbs(struct kioctx *ctx)
839 struct kiocb *iocb;
840 struct list_head run_list;
842 assert_spin_locked(&ctx->ctx_lock);
844 list_replace_init(&ctx->run_list, &run_list);
845 while (!list_empty(&run_list)) {
846 iocb = list_entry(run_list.next, struct kiocb,
847 ki_run_list);
848 list_del(&iocb->ki_run_list);
850 * Hold an extra reference while retrying i/o.
852 iocb->ki_users++; /* grab extra reference */
853 aio_run_iocb(iocb);
854 __aio_put_req(ctx, iocb);
856 if (!list_empty(&ctx->run_list))
857 return 1;
858 return 0;
861 static void aio_queue_work(struct kioctx * ctx)
863 unsigned long timeout;
865 * if someone is waiting, get the work started right
866 * away, otherwise, use a longer delay
868 smp_mb();
869 if (waitqueue_active(&ctx->wait))
870 timeout = 1;
871 else
872 timeout = HZ/10;
873 queue_delayed_work(aio_wq, &ctx->wq, timeout);
877 * aio_run_all_iocbs:
878 * Process all pending retries queued on the ioctx
879 * run list, and keep running them until the list
880 * stays empty.
881 * Assumes it is operating within the aio issuer's mm context.
883 static inline void aio_run_all_iocbs(struct kioctx *ctx)
885 spin_lock_irq(&ctx->ctx_lock);
886 while (__aio_run_iocbs(ctx))
888 spin_unlock_irq(&ctx->ctx_lock);
892 * aio_kick_handler:
893 * Work queue handler triggered to process pending
894 * retries on an ioctx. Takes on the aio issuer's
895 * mm context before running the iocbs, so that
896 * copy_xxx_user operates on the issuer's address
897 * space.
898 * Run on aiod's context.
900 static void aio_kick_handler(struct work_struct *work)
902 struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
903 mm_segment_t oldfs = get_fs();
904 struct mm_struct *mm;
905 int requeue;
907 set_fs(USER_DS);
908 use_mm(ctx->mm);
909 spin_lock_irq(&ctx->ctx_lock);
910 requeue =__aio_run_iocbs(ctx);
911 mm = ctx->mm;
912 spin_unlock_irq(&ctx->ctx_lock);
913 unuse_mm(mm);
914 set_fs(oldfs);
916 * we're in a worker thread already, don't use queue_delayed_work,
918 if (requeue)
919 queue_delayed_work(aio_wq, &ctx->wq, 0);
924 * Called by kick_iocb to queue the kiocb for retry
925 * and if required activate the aio work queue to process
926 * it
928 static void try_queue_kicked_iocb(struct kiocb *iocb)
930 struct kioctx *ctx = iocb->ki_ctx;
931 unsigned long flags;
932 int run = 0;
934 spin_lock_irqsave(&ctx->ctx_lock, flags);
935 /* set this inside the lock so that we can't race with aio_run_iocb()
936 * testing it and putting the iocb on the run list under the lock */
937 if (!kiocbTryKick(iocb))
938 run = __queue_kicked_iocb(iocb);
939 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
940 if (run)
941 aio_queue_work(ctx);
945 * kick_iocb:
946 * Called typically from a wait queue callback context
947 * to trigger a retry of the iocb.
948 * The retry is usually executed by aio workqueue
949 * threads (See aio_kick_handler).
951 void kick_iocb(struct kiocb *iocb)
953 /* sync iocbs are easy: they can only ever be executing from a
954 * single context. */
955 if (is_sync_kiocb(iocb)) {
956 kiocbSetKicked(iocb);
957 wake_up_process(iocb->ki_obj.tsk);
958 return;
961 try_queue_kicked_iocb(iocb);
963 EXPORT_SYMBOL(kick_iocb);
965 /* aio_complete
966 * Called when the io request on the given iocb is complete.
967 * Returns true if this is the last user of the request. The
968 * only other user of the request can be the cancellation code.
970 int aio_complete(struct kiocb *iocb, long res, long res2)
972 struct kioctx *ctx = iocb->ki_ctx;
973 struct aio_ring_info *info;
974 struct aio_ring *ring;
975 struct io_event *event;
976 unsigned long flags;
977 unsigned long tail;
978 int ret;
981 * Special case handling for sync iocbs:
982 * - events go directly into the iocb for fast handling
983 * - the sync task with the iocb in its stack holds the single iocb
984 * ref, no other paths have a way to get another ref
985 * - the sync task helpfully left a reference to itself in the iocb
987 if (is_sync_kiocb(iocb)) {
988 BUG_ON(iocb->ki_users != 1);
989 iocb->ki_user_data = res;
990 iocb->ki_users = 0;
991 wake_up_process(iocb->ki_obj.tsk);
992 return 1;
995 info = &ctx->ring_info;
997 /* add a completion event to the ring buffer.
998 * must be done holding ctx->ctx_lock to prevent
999 * other code from messing with the tail
1000 * pointer since we might be called from irq
1001 * context.
1003 spin_lock_irqsave(&ctx->ctx_lock, flags);
1005 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
1006 list_del_init(&iocb->ki_run_list);
1009 * cancelled requests don't get events, userland was given one
1010 * when the event got cancelled.
1012 if (kiocbIsCancelled(iocb))
1013 goto put_rq;
1015 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
1017 tail = info->tail;
1018 event = aio_ring_event(info, tail, KM_IRQ0);
1019 if (++tail >= info->nr)
1020 tail = 0;
1022 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
1023 event->data = iocb->ki_user_data;
1024 event->res = res;
1025 event->res2 = res2;
1027 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
1028 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
1029 res, res2);
1031 /* after flagging the request as done, we
1032 * must never even look at it again
1034 smp_wmb(); /* make event visible before updating tail */
1036 info->tail = tail;
1037 ring->tail = tail;
1039 put_aio_ring_event(event, KM_IRQ0);
1040 kunmap_atomic(ring, KM_IRQ1);
1042 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
1045 * Check if the user asked us to deliver the result through an
1046 * eventfd. The eventfd_signal() function is safe to be called
1047 * from IRQ context.
1049 if (iocb->ki_eventfd != NULL)
1050 eventfd_signal(iocb->ki_eventfd, 1);
1052 put_rq:
1053 /* everything turned out well, dispose of the aiocb. */
1054 ret = __aio_put_req(ctx, iocb);
1057 * We have to order our ring_info tail store above and test
1058 * of the wait list below outside the wait lock. This is
1059 * like in wake_up_bit() where clearing a bit has to be
1060 * ordered with the unlocked test.
1062 smp_mb();
1064 if (waitqueue_active(&ctx->wait))
1065 wake_up(&ctx->wait);
1067 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1068 return ret;
1070 EXPORT_SYMBOL(aio_complete);
1072 /* aio_read_evt
1073 * Pull an event off of the ioctx's event ring. Returns the number of
1074 * events fetched (0 or 1 ;-)
1075 * FIXME: make this use cmpxchg.
1076 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1078 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1080 struct aio_ring_info *info = &ioctx->ring_info;
1081 struct aio_ring *ring;
1082 unsigned long head;
1083 int ret = 0;
1085 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1086 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1087 (unsigned long)ring->head, (unsigned long)ring->tail,
1088 (unsigned long)ring->nr);
1090 if (ring->head == ring->tail)
1091 goto out;
1093 spin_lock(&info->ring_lock);
1095 head = ring->head % info->nr;
1096 if (head != ring->tail) {
1097 struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1098 *ent = *evp;
1099 head = (head + 1) % info->nr;
1100 smp_mb(); /* finish reading the event before updatng the head */
1101 ring->head = head;
1102 ret = 1;
1103 put_aio_ring_event(evp, KM_USER1);
1105 spin_unlock(&info->ring_lock);
1107 out:
1108 kunmap_atomic(ring, KM_USER0);
1109 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1110 (unsigned long)ring->head, (unsigned long)ring->tail);
1111 return ret;
1114 struct aio_timeout {
1115 struct timer_list timer;
1116 int timed_out;
1117 struct task_struct *p;
1120 static void timeout_func(unsigned long data)
1122 struct aio_timeout *to = (struct aio_timeout *)data;
1124 to->timed_out = 1;
1125 wake_up_process(to->p);
1128 static inline void init_timeout(struct aio_timeout *to)
1130 setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
1131 to->timed_out = 0;
1132 to->p = current;
1135 static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1136 const struct timespec *ts)
1138 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1139 if (time_after(to->timer.expires, jiffies))
1140 add_timer(&to->timer);
1141 else
1142 to->timed_out = 1;
1145 static inline void clear_timeout(struct aio_timeout *to)
1147 del_singleshot_timer_sync(&to->timer);
1150 static int read_events(struct kioctx *ctx,
1151 long min_nr, long nr,
1152 struct io_event __user *event,
1153 struct timespec __user *timeout)
1155 long start_jiffies = jiffies;
1156 struct task_struct *tsk = current;
1157 DECLARE_WAITQUEUE(wait, tsk);
1158 int ret;
1159 int i = 0;
1160 struct io_event ent;
1161 struct aio_timeout to;
1162 int retry = 0;
1164 /* needed to zero any padding within an entry (there shouldn't be
1165 * any, but C is fun!
1167 memset(&ent, 0, sizeof(ent));
1168 retry:
1169 ret = 0;
1170 while (likely(i < nr)) {
1171 ret = aio_read_evt(ctx, &ent);
1172 if (unlikely(ret <= 0))
1173 break;
1175 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1176 ent.data, ent.obj, ent.res, ent.res2);
1178 /* Could we split the check in two? */
1179 ret = -EFAULT;
1180 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1181 dprintk("aio: lost an event due to EFAULT.\n");
1182 break;
1184 ret = 0;
1186 /* Good, event copied to userland, update counts. */
1187 event ++;
1188 i ++;
1191 if (min_nr <= i)
1192 return i;
1193 if (ret)
1194 return ret;
1196 /* End fast path */
1198 /* racey check, but it gets redone */
1199 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1200 retry = 1;
1201 aio_run_all_iocbs(ctx);
1202 goto retry;
1205 init_timeout(&to);
1206 if (timeout) {
1207 struct timespec ts;
1208 ret = -EFAULT;
1209 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1210 goto out;
1212 set_timeout(start_jiffies, &to, &ts);
1215 while (likely(i < nr)) {
1216 add_wait_queue_exclusive(&ctx->wait, &wait);
1217 do {
1218 set_task_state(tsk, TASK_INTERRUPTIBLE);
1219 ret = aio_read_evt(ctx, &ent);
1220 if (ret)
1221 break;
1222 if (min_nr <= i)
1223 break;
1224 if (unlikely(ctx->dead)) {
1225 ret = -EINVAL;
1226 break;
1228 if (to.timed_out) /* Only check after read evt */
1229 break;
1230 /* Try to only show up in io wait if there are ops
1231 * in flight */
1232 if (ctx->reqs_active)
1233 io_schedule();
1234 else
1235 schedule();
1236 if (signal_pending(tsk)) {
1237 ret = -EINTR;
1238 break;
1240 /*ret = aio_read_evt(ctx, &ent);*/
1241 } while (1) ;
1243 set_task_state(tsk, TASK_RUNNING);
1244 remove_wait_queue(&ctx->wait, &wait);
1246 if (unlikely(ret <= 0))
1247 break;
1249 ret = -EFAULT;
1250 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1251 dprintk("aio: lost an event due to EFAULT.\n");
1252 break;
1255 /* Good, event copied to userland, update counts. */
1256 event ++;
1257 i ++;
1260 if (timeout)
1261 clear_timeout(&to);
1262 out:
1263 destroy_timer_on_stack(&to.timer);
1264 return i ? i : ret;
1267 /* Take an ioctx and remove it from the list of ioctx's. Protects
1268 * against races with itself via ->dead.
1270 static void io_destroy(struct kioctx *ioctx)
1272 struct mm_struct *mm = current->mm;
1273 int was_dead;
1275 /* delete the entry from the list is someone else hasn't already */
1276 spin_lock(&mm->ioctx_lock);
1277 was_dead = ioctx->dead;
1278 ioctx->dead = 1;
1279 hlist_del_rcu(&ioctx->list);
1280 spin_unlock(&mm->ioctx_lock);
1282 dprintk("aio_release(%p)\n", ioctx);
1283 if (likely(!was_dead))
1284 put_ioctx(ioctx); /* twice for the list */
1286 aio_cancel_all(ioctx);
1287 wait_for_all_aios(ioctx);
1290 * Wake up any waiters. The setting of ctx->dead must be seen
1291 * by other CPUs at this point. Right now, we rely on the
1292 * locking done by the above calls to ensure this consistency.
1294 wake_up_all(&ioctx->wait);
1295 put_ioctx(ioctx); /* once for the lookup */
1298 /* sys_io_setup:
1299 * Create an aio_context capable of receiving at least nr_events.
1300 * ctxp must not point to an aio_context that already exists, and
1301 * must be initialized to 0 prior to the call. On successful
1302 * creation of the aio_context, *ctxp is filled in with the resulting
1303 * handle. May fail with -EINVAL if *ctxp is not initialized,
1304 * if the specified nr_events exceeds internal limits. May fail
1305 * with -EAGAIN if the specified nr_events exceeds the user's limit
1306 * of available events. May fail with -ENOMEM if insufficient kernel
1307 * resources are available. May fail with -EFAULT if an invalid
1308 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1309 * implemented.
1311 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1313 struct kioctx *ioctx = NULL;
1314 unsigned long ctx;
1315 long ret;
1317 ret = get_user(ctx, ctxp);
1318 if (unlikely(ret))
1319 goto out;
1321 ret = -EINVAL;
1322 if (unlikely(ctx || nr_events == 0)) {
1323 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1324 ctx, nr_events);
1325 goto out;
1328 ioctx = ioctx_alloc(nr_events);
1329 ret = PTR_ERR(ioctx);
1330 if (!IS_ERR(ioctx)) {
1331 ret = put_user(ioctx->user_id, ctxp);
1332 if (!ret)
1333 return 0;
1335 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1336 io_destroy(ioctx);
1339 out:
1340 return ret;
1343 /* sys_io_destroy:
1344 * Destroy the aio_context specified. May cancel any outstanding
1345 * AIOs and block on completion. Will fail with -ENOSYS if not
1346 * implemented. May fail with -EINVAL if the context pointed to
1347 * is invalid.
1349 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1351 struct kioctx *ioctx = lookup_ioctx(ctx);
1352 if (likely(NULL != ioctx)) {
1353 io_destroy(ioctx);
1354 return 0;
1356 pr_debug("EINVAL: io_destroy: invalid context id\n");
1357 return -EINVAL;
1360 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1362 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1364 BUG_ON(ret <= 0);
1366 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1367 ssize_t this = min((ssize_t)iov->iov_len, ret);
1368 iov->iov_base += this;
1369 iov->iov_len -= this;
1370 iocb->ki_left -= this;
1371 ret -= this;
1372 if (iov->iov_len == 0) {
1373 iocb->ki_cur_seg++;
1374 iov++;
1378 /* the caller should not have done more io than what fit in
1379 * the remaining iovecs */
1380 BUG_ON(ret > 0 && iocb->ki_left == 0);
1383 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1385 struct file *file = iocb->ki_filp;
1386 struct address_space *mapping = file->f_mapping;
1387 struct inode *inode = mapping->host;
1388 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1389 unsigned long, loff_t);
1390 ssize_t ret = 0;
1391 unsigned short opcode;
1393 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1394 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1395 rw_op = file->f_op->aio_read;
1396 opcode = IOCB_CMD_PREADV;
1397 } else {
1398 rw_op = file->f_op->aio_write;
1399 opcode = IOCB_CMD_PWRITEV;
1402 /* This matches the pread()/pwrite() logic */
1403 if (iocb->ki_pos < 0)
1404 return -EINVAL;
1406 do {
1407 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1408 iocb->ki_nr_segs - iocb->ki_cur_seg,
1409 iocb->ki_pos);
1410 if (ret > 0)
1411 aio_advance_iovec(iocb, ret);
1413 /* retry all partial writes. retry partial reads as long as its a
1414 * regular file. */
1415 } while (ret > 0 && iocb->ki_left > 0 &&
1416 (opcode == IOCB_CMD_PWRITEV ||
1417 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1419 /* This means we must have transferred all that we could */
1420 /* No need to retry anymore */
1421 if ((ret == 0) || (iocb->ki_left == 0))
1422 ret = iocb->ki_nbytes - iocb->ki_left;
1424 /* If we managed to write some out we return that, rather than
1425 * the eventual error. */
1426 if (opcode == IOCB_CMD_PWRITEV
1427 && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
1428 && iocb->ki_nbytes - iocb->ki_left)
1429 ret = iocb->ki_nbytes - iocb->ki_left;
1431 return ret;
1434 static ssize_t aio_fdsync(struct kiocb *iocb)
1436 struct file *file = iocb->ki_filp;
1437 ssize_t ret = -EINVAL;
1439 if (file->f_op->aio_fsync)
1440 ret = file->f_op->aio_fsync(iocb, 1);
1441 return ret;
1444 static ssize_t aio_fsync(struct kiocb *iocb)
1446 struct file *file = iocb->ki_filp;
1447 ssize_t ret = -EINVAL;
1449 if (file->f_op->aio_fsync)
1450 ret = file->f_op->aio_fsync(iocb, 0);
1451 return ret;
1454 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
1456 ssize_t ret;
1458 #ifdef CONFIG_COMPAT
1459 if (compat)
1460 ret = compat_rw_copy_check_uvector(type,
1461 (struct compat_iovec __user *)kiocb->ki_buf,
1462 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1463 &kiocb->ki_iovec, 1);
1464 else
1465 #endif
1466 ret = rw_copy_check_uvector(type,
1467 (struct iovec __user *)kiocb->ki_buf,
1468 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1469 &kiocb->ki_iovec, 1);
1470 if (ret < 0)
1471 goto out;
1473 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1474 kiocb->ki_cur_seg = 0;
1475 /* ki_nbytes/left now reflect bytes instead of segs */
1476 kiocb->ki_nbytes = ret;
1477 kiocb->ki_left = ret;
1479 ret = 0;
1480 out:
1481 return ret;
1484 static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
1486 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1487 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1488 kiocb->ki_iovec->iov_len = kiocb->ki_left;
1489 kiocb->ki_nr_segs = 1;
1490 kiocb->ki_cur_seg = 0;
1491 return 0;
1495 * aio_setup_iocb:
1496 * Performs the initial checks and aio retry method
1497 * setup for the kiocb at the time of io submission.
1499 static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
1501 struct file *file = kiocb->ki_filp;
1502 ssize_t ret = 0;
1504 switch (kiocb->ki_opcode) {
1505 case IOCB_CMD_PREAD:
1506 ret = -EBADF;
1507 if (unlikely(!(file->f_mode & FMODE_READ)))
1508 break;
1509 ret = -EFAULT;
1510 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1511 kiocb->ki_left)))
1512 break;
1513 ret = security_file_permission(file, MAY_READ);
1514 if (unlikely(ret))
1515 break;
1516 ret = aio_setup_single_vector(kiocb);
1517 if (ret)
1518 break;
1519 ret = -EINVAL;
1520 if (file->f_op->aio_read)
1521 kiocb->ki_retry = aio_rw_vect_retry;
1522 break;
1523 case IOCB_CMD_PWRITE:
1524 ret = -EBADF;
1525 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1526 break;
1527 ret = -EFAULT;
1528 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1529 kiocb->ki_left)))
1530 break;
1531 ret = security_file_permission(file, MAY_WRITE);
1532 if (unlikely(ret))
1533 break;
1534 ret = aio_setup_single_vector(kiocb);
1535 if (ret)
1536 break;
1537 ret = -EINVAL;
1538 if (file->f_op->aio_write)
1539 kiocb->ki_retry = aio_rw_vect_retry;
1540 break;
1541 case IOCB_CMD_PREADV:
1542 ret = -EBADF;
1543 if (unlikely(!(file->f_mode & FMODE_READ)))
1544 break;
1545 ret = security_file_permission(file, MAY_READ);
1546 if (unlikely(ret))
1547 break;
1548 ret = aio_setup_vectored_rw(READ, kiocb, compat);
1549 if (ret)
1550 break;
1551 ret = -EINVAL;
1552 if (file->f_op->aio_read)
1553 kiocb->ki_retry = aio_rw_vect_retry;
1554 break;
1555 case IOCB_CMD_PWRITEV:
1556 ret = -EBADF;
1557 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1558 break;
1559 ret = security_file_permission(file, MAY_WRITE);
1560 if (unlikely(ret))
1561 break;
1562 ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1563 if (ret)
1564 break;
1565 ret = -EINVAL;
1566 if (file->f_op->aio_write)
1567 kiocb->ki_retry = aio_rw_vect_retry;
1568 break;
1569 case IOCB_CMD_FDSYNC:
1570 ret = -EINVAL;
1571 if (file->f_op->aio_fsync)
1572 kiocb->ki_retry = aio_fdsync;
1573 break;
1574 case IOCB_CMD_FSYNC:
1575 ret = -EINVAL;
1576 if (file->f_op->aio_fsync)
1577 kiocb->ki_retry = aio_fsync;
1578 break;
1579 default:
1580 dprintk("EINVAL: io_submit: no operation provided\n");
1581 ret = -EINVAL;
1584 if (!kiocb->ki_retry)
1585 return ret;
1587 return 0;
1590 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1591 struct iocb *iocb, struct kiocb_batch *batch,
1592 bool compat)
1594 struct kiocb *req;
1595 struct file *file;
1596 ssize_t ret;
1598 /* enforce forwards compatibility on users */
1599 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1600 pr_debug("EINVAL: io_submit: reserve field set\n");
1601 return -EINVAL;
1604 /* prevent overflows */
1605 if (unlikely(
1606 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1607 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1608 ((ssize_t)iocb->aio_nbytes < 0)
1609 )) {
1610 pr_debug("EINVAL: io_submit: overflow check\n");
1611 return -EINVAL;
1614 file = fget(iocb->aio_fildes);
1615 if (unlikely(!file))
1616 return -EBADF;
1618 req = aio_get_req(ctx, batch); /* returns with 2 references to req */
1619 if (unlikely(!req)) {
1620 fput(file);
1621 return -EAGAIN;
1623 req->ki_filp = file;
1624 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1626 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1627 * instance of the file* now. The file descriptor must be
1628 * an eventfd() fd, and will be signaled for each completed
1629 * event using the eventfd_signal() function.
1631 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1632 if (IS_ERR(req->ki_eventfd)) {
1633 ret = PTR_ERR(req->ki_eventfd);
1634 req->ki_eventfd = NULL;
1635 goto out_put_req;
1639 ret = put_user(req->ki_key, &user_iocb->aio_key);
1640 if (unlikely(ret)) {
1641 dprintk("EFAULT: aio_key\n");
1642 goto out_put_req;
1645 req->ki_obj.user = user_iocb;
1646 req->ki_user_data = iocb->aio_data;
1647 req->ki_pos = iocb->aio_offset;
1649 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1650 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1651 req->ki_opcode = iocb->aio_lio_opcode;
1653 ret = aio_setup_iocb(req, compat);
1655 if (ret)
1656 goto out_put_req;
1658 spin_lock_irq(&ctx->ctx_lock);
1660 * We could have raced with io_destroy() and are currently holding a
1661 * reference to ctx which should be destroyed. We cannot submit IO
1662 * since ctx gets freed as soon as io_submit() puts its reference. The
1663 * check here is reliable: io_destroy() sets ctx->dead before waiting
1664 * for outstanding IO and the barrier between these two is realized by
1665 * unlock of mm->ioctx_lock and lock of ctx->ctx_lock. Analogously we
1666 * increment ctx->reqs_active before checking for ctx->dead and the
1667 * barrier is realized by unlock and lock of ctx->ctx_lock. Thus if we
1668 * don't see ctx->dead set here, io_destroy() waits for our IO to
1669 * finish.
1671 if (ctx->dead) {
1672 spin_unlock_irq(&ctx->ctx_lock);
1673 ret = -EINVAL;
1674 goto out_put_req;
1676 aio_run_iocb(req);
1677 if (!list_empty(&ctx->run_list)) {
1678 /* drain the run list */
1679 while (__aio_run_iocbs(ctx))
1682 spin_unlock_irq(&ctx->ctx_lock);
1684 aio_put_req(req); /* drop extra ref to req */
1685 return 0;
1687 out_put_req:
1688 aio_put_req(req); /* drop extra ref to req */
1689 aio_put_req(req); /* drop i/o ref to req */
1690 return ret;
1693 long do_io_submit(aio_context_t ctx_id, long nr,
1694 struct iocb __user *__user *iocbpp, bool compat)
1696 struct kioctx *ctx;
1697 long ret = 0;
1698 int i = 0;
1699 struct blk_plug plug;
1700 struct kiocb_batch batch;
1702 if (unlikely(nr < 0))
1703 return -EINVAL;
1705 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1706 nr = LONG_MAX/sizeof(*iocbpp);
1708 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1709 return -EFAULT;
1711 ctx = lookup_ioctx(ctx_id);
1712 if (unlikely(!ctx)) {
1713 pr_debug("EINVAL: io_submit: invalid context id\n");
1714 return -EINVAL;
1717 kiocb_batch_init(&batch, nr);
1719 blk_start_plug(&plug);
1722 * AKPM: should this return a partial result if some of the IOs were
1723 * successfully submitted?
1725 for (i=0; i<nr; i++) {
1726 struct iocb __user *user_iocb;
1727 struct iocb tmp;
1729 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1730 ret = -EFAULT;
1731 break;
1734 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1735 ret = -EFAULT;
1736 break;
1739 ret = io_submit_one(ctx, user_iocb, &tmp, &batch, compat);
1740 if (ret)
1741 break;
1743 blk_finish_plug(&plug);
1745 kiocb_batch_free(&batch);
1746 put_ioctx(ctx);
1747 return i ? i : ret;
1750 /* sys_io_submit:
1751 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1752 * the number of iocbs queued. May return -EINVAL if the aio_context
1753 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1754 * *iocbpp[0] is not properly initialized, if the operation specified
1755 * is invalid for the file descriptor in the iocb. May fail with
1756 * -EFAULT if any of the data structures point to invalid data. May
1757 * fail with -EBADF if the file descriptor specified in the first
1758 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1759 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1760 * fail with -ENOSYS if not implemented.
1762 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1763 struct iocb __user * __user *, iocbpp)
1765 return do_io_submit(ctx_id, nr, iocbpp, 0);
1768 /* lookup_kiocb
1769 * Finds a given iocb for cancellation.
1771 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1772 u32 key)
1774 struct list_head *pos;
1776 assert_spin_locked(&ctx->ctx_lock);
1778 /* TODO: use a hash or array, this sucks. */
1779 list_for_each(pos, &ctx->active_reqs) {
1780 struct kiocb *kiocb = list_kiocb(pos);
1781 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1782 return kiocb;
1784 return NULL;
1787 /* sys_io_cancel:
1788 * Attempts to cancel an iocb previously passed to io_submit. If
1789 * the operation is successfully cancelled, the resulting event is
1790 * copied into the memory pointed to by result without being placed
1791 * into the completion queue and 0 is returned. May fail with
1792 * -EFAULT if any of the data structures pointed to are invalid.
1793 * May fail with -EINVAL if aio_context specified by ctx_id is
1794 * invalid. May fail with -EAGAIN if the iocb specified was not
1795 * cancelled. Will fail with -ENOSYS if not implemented.
1797 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1798 struct io_event __user *, result)
1800 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1801 struct kioctx *ctx;
1802 struct kiocb *kiocb;
1803 u32 key;
1804 int ret;
1806 ret = get_user(key, &iocb->aio_key);
1807 if (unlikely(ret))
1808 return -EFAULT;
1810 ctx = lookup_ioctx(ctx_id);
1811 if (unlikely(!ctx))
1812 return -EINVAL;
1814 spin_lock_irq(&ctx->ctx_lock);
1815 ret = -EAGAIN;
1816 kiocb = lookup_kiocb(ctx, iocb, key);
1817 if (kiocb && kiocb->ki_cancel) {
1818 cancel = kiocb->ki_cancel;
1819 kiocb->ki_users ++;
1820 kiocbSetCancelled(kiocb);
1821 } else
1822 cancel = NULL;
1823 spin_unlock_irq(&ctx->ctx_lock);
1825 if (NULL != cancel) {
1826 struct io_event tmp;
1827 pr_debug("calling cancel\n");
1828 memset(&tmp, 0, sizeof(tmp));
1829 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1830 tmp.data = kiocb->ki_user_data;
1831 ret = cancel(kiocb, &tmp);
1832 if (!ret) {
1833 /* Cancellation succeeded -- copy the result
1834 * into the user's buffer.
1836 if (copy_to_user(result, &tmp, sizeof(tmp)))
1837 ret = -EFAULT;
1839 } else
1840 ret = -EINVAL;
1842 put_ioctx(ctx);
1844 return ret;
1847 /* io_getevents:
1848 * Attempts to read at least min_nr events and up to nr events from
1849 * the completion queue for the aio_context specified by ctx_id. If
1850 * it succeeds, the number of read events is returned. May fail with
1851 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1852 * out of range, if timeout is out of range. May fail with -EFAULT
1853 * if any of the memory specified is invalid. May return 0 or
1854 * < min_nr if the timeout specified by timeout has elapsed
1855 * before sufficient events are available, where timeout == NULL
1856 * specifies an infinite timeout. Note that the timeout pointed to by
1857 * timeout is relative and will be updated if not NULL and the
1858 * operation blocks. Will fail with -ENOSYS if not implemented.
1860 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1861 long, min_nr,
1862 long, nr,
1863 struct io_event __user *, events,
1864 struct timespec __user *, timeout)
1866 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1867 long ret = -EINVAL;
1869 if (likely(ioctx)) {
1870 if (likely(min_nr <= nr && min_nr >= 0))
1871 ret = read_events(ioctx, min_nr, nr, events, timeout);
1872 put_ioctx(ioctx);
1875 asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);
1876 return ret;