1 /*
2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <bcrl@kvack.org>
4 *
5 * Implements an efficient asynchronous io interface.
6 *
7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved.
8 *
9 * See ../COPYING for licensing terms.
10 */
11 #include <linux/kernel.h>
12 #include <linux/init.h>
13 #include <linux/errno.h>
14 #include <linux/time.h>
15 #include <linux/aio_abi.h>
16 #include <linux/module.h>
17 #include <linux/syscalls.h>
18 #include <linux/uio.h>
19
20 #define DEBUG 0
21
22 #include <linux/sched.h>
23 #include <linux/fs.h>
24 #include <linux/file.h>
25 #include <linux/mm.h>
26 #include <linux/mman.h>
27 #include <linux/slab.h>
28 #include <linux/timer.h>
29 #include <linux/aio.h>
30 #include <linux/highmem.h>
31 #include <linux/workqueue.h>
32 #include <linux/security.h>
33 #include <linux/eventfd.h>
34
35 #include <asm/kmap_types.h>
36 #include <asm/uaccess.h>
37 #include <asm/mmu_context.h>
38
39 #if DEBUG > 1
40 #define dprintk printk
41 #else
42 #define dprintk(x...) do { ; } while (0)
43 #endif
44
45 /*------ sysctl variables----*/
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---*/
50
51 static struct kmem_cache *kiocb_cachep;
52 static struct kmem_cache *kioctx_cachep;
53
54 static struct workqueue_struct *aio_wq;
55
56 /* Used for rare fput completion. */
57 static void aio_fput_routine(struct work_struct *);
58 static DECLARE_WORK(fput_work, aio_fput_routine);
59
60 static DEFINE_SPINLOCK(fput_lock);
61 static LIST_HEAD(fput_head);
62
63 static void aio_kick_handler(struct work_struct *);
64 static void aio_queue_work(struct kioctx *);
65
66 /* aio_setup
67 * Creates the slab caches used by the aio routines, panic on
68 * failure as this is done early during the boot sequence.
69 */
70 static int __init aio_setup(void)
71 {
72 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
73 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
74
75 aio_wq = create_workqueue("aio");
76
77 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
78
79 return 0;
80 }
81
82 static void aio_free_ring(struct kioctx *ctx)
83 {
84 struct aio_ring_info *info = &ctx->ring_info;
85 long i;
86
87 for (i=0; i<info->nr_pages; i++)
88 put_page(info->ring_pages[i]);
89
90 if (info->mmap_size) {
91 down_write(&ctx->mm->mmap_sem);
92 do_munmap(ctx->mm, info->mmap_base, info->mmap_size);
93 up_write(&ctx->mm->mmap_sem);
94 }
95
96 if (info->ring_pages && info->ring_pages != info->internal_pages)
97 kfree(info->ring_pages);
98 info->ring_pages = NULL;
99 info->nr = 0;
100 }
101
102 static int aio_setup_ring(struct kioctx *ctx)
103 {
104 struct aio_ring *ring;
105 struct aio_ring_info *info = &ctx->ring_info;
106 unsigned nr_events = ctx->max_reqs;
107 unsigned long size;
108 int nr_pages;
109
110 /* Compensate for the ring buffer's head/tail overlap entry */
111 nr_events += 2; /* 1 is required, 2 for good luck */
112
113 size = sizeof(struct aio_ring);
114 size += sizeof(struct io_event) * nr_events;
115 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
116
117 if (nr_pages < 0)
118 return -EINVAL;
119
120 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
121
122 info->nr = 0;
123 info->ring_pages = info->internal_pages;
124 if (nr_pages > AIO_RING_PAGES) {
125 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
126 if (!info->ring_pages)
127 return -ENOMEM;
128 }
129
130 info->mmap_size = nr_pages * PAGE_SIZE;
131 dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
132 down_write(&ctx->mm->mmap_sem);
133 info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
134 PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE,
135 0);
136 if (IS_ERR((void *)info->mmap_base)) {
137 up_write(&ctx->mm->mmap_sem);
138 info->mmap_size = 0;
139 aio_free_ring(ctx);
140 return -EAGAIN;
141 }
142
143 dprintk("mmap address: 0x%08lx\n", info->mmap_base);
144 info->nr_pages = get_user_pages(current, ctx->mm,
145 info->mmap_base, nr_pages,
146 1, 0, info->ring_pages, NULL);
147 up_write(&ctx->mm->mmap_sem);
148
149 if (unlikely(info->nr_pages != nr_pages)) {
150 aio_free_ring(ctx);
151 return -EAGAIN;
152 }
153
154 ctx->user_id = info->mmap_base;
155
156 info->nr = nr_events; /* trusted copy */
157
158 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
159 ring->nr = nr_events; /* user copy */
160 ring->id = ctx->user_id;
161 ring->head = ring->tail = 0;
162 ring->magic = AIO_RING_MAGIC;
163 ring->compat_features = AIO_RING_COMPAT_FEATURES;
164 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
165 ring->header_length = sizeof(struct aio_ring);
166 kunmap_atomic(ring, KM_USER0);
167
168 return 0;
169 }
170
171
172 /* aio_ring_event: returns a pointer to the event at the given index from
173 * kmap_atomic(, km). Release the pointer with put_aio_ring_event();
174 */
175 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
176 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
177 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
178
179 #define aio_ring_event(info, nr, km) ({ \
180 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
181 struct io_event *__event; \
182 __event = kmap_atomic( \
183 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
184 __event += pos % AIO_EVENTS_PER_PAGE; \
185 __event; \
186 })
187
188 #define put_aio_ring_event(event, km) do { \
189 struct io_event *__event = (event); \
190 (void)__event; \
191 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
192 } while(0)
193
194 /* ioctx_alloc
195 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
196 */
197 static struct kioctx *ioctx_alloc(unsigned nr_events)
198 {
199 struct mm_struct *mm;
200 struct kioctx *ctx;
201
202 /* Prevent overflows */
203 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
204 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
205 pr_debug("ENOMEM: nr_events too high\n");
206 return ERR_PTR(-EINVAL);
207 }
208
209 if ((unsigned long)nr_events > aio_max_nr)
210 return ERR_PTR(-EAGAIN);
211
212 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
213 if (!ctx)
214 return ERR_PTR(-ENOMEM);
215
216 ctx->max_reqs = nr_events;
217 mm = ctx->mm = current->mm;
218 atomic_inc(&mm->mm_count);
219
220 atomic_set(&ctx->users, 1);
221 spin_lock_init(&ctx->ctx_lock);
222 spin_lock_init(&ctx->ring_info.ring_lock);
223 init_waitqueue_head(&ctx->wait);
224
225 INIT_LIST_HEAD(&ctx->active_reqs);
226 INIT_LIST_HEAD(&ctx->run_list);
227 INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
228
229 if (aio_setup_ring(ctx) < 0)
230 goto out_freectx;
231
232 /* limit the number of system wide aios */
233 spin_lock(&aio_nr_lock);
234 if (aio_nr + ctx->max_reqs > aio_max_nr ||
235 aio_nr + ctx->max_reqs < aio_nr)
236 ctx->max_reqs = 0;
237 else
238 aio_nr += ctx->max_reqs;
239 spin_unlock(&aio_nr_lock);
240 if (ctx->max_reqs == 0)
241 goto out_cleanup;
242
243 /* now link into global list. kludge. FIXME */
244 write_lock(&mm->ioctx_list_lock);
245 ctx->next = mm->ioctx_list;
246 mm->ioctx_list = ctx;
247 write_unlock(&mm->ioctx_list_lock);
248
249 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
250 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
251 return ctx;
252
253 out_cleanup:
254 __put_ioctx(ctx);
255 return ERR_PTR(-EAGAIN);
256
257 out_freectx:
258 mmdrop(mm);
259 kmem_cache_free(kioctx_cachep, ctx);
260 ctx = ERR_PTR(-ENOMEM);
261
262 dprintk("aio: error allocating ioctx %p\n", ctx);
263 return ctx;
264 }
265
266 /* aio_cancel_all
267 * Cancels all outstanding aio requests on an aio context. Used
268 * when the processes owning a context have all exited to encourage
269 * the rapid destruction of the kioctx.
270 */
271 static void aio_cancel_all(struct kioctx *ctx)
272 {
273 int (*cancel)(struct kiocb *, struct io_event *);
274 struct io_event res;
275 spin_lock_irq(&ctx->ctx_lock);
276 ctx->dead = 1;
277 while (!list_empty(&ctx->active_reqs)) {
278 struct list_head *pos = ctx->active_reqs.next;
279 struct kiocb *iocb = list_kiocb(pos);
280 list_del_init(&iocb->ki_list);
281 cancel = iocb->ki_cancel;
282 kiocbSetCancelled(iocb);
283 if (cancel) {
284 iocb->ki_users++;
285 spin_unlock_irq(&ctx->ctx_lock);
286 cancel(iocb, &res);
287 spin_lock_irq(&ctx->ctx_lock);
288 }
289 }
290 spin_unlock_irq(&ctx->ctx_lock);
291 }
292
293 static void wait_for_all_aios(struct kioctx *ctx)
294 {
295 struct task_struct *tsk = current;
296 DECLARE_WAITQUEUE(wait, tsk);
297
298 spin_lock_irq(&ctx->ctx_lock);
299 if (!ctx->reqs_active)
300 goto out;
301
302 add_wait_queue(&ctx->wait, &wait);
303 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
304 while (ctx->reqs_active) {
305 spin_unlock_irq(&ctx->ctx_lock);
306 io_schedule();
307 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
308 spin_lock_irq(&ctx->ctx_lock);
309 }
310 __set_task_state(tsk, TASK_RUNNING);
311 remove_wait_queue(&ctx->wait, &wait);
312
313 out:
314 spin_unlock_irq(&ctx->ctx_lock);
315 }
316
317 /* wait_on_sync_kiocb:
318 * Waits on the given sync kiocb to complete.
319 */
320 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
321 {
322 while (iocb->ki_users) {
323 set_current_state(TASK_UNINTERRUPTIBLE);
324 if (!iocb->ki_users)
325 break;
326 io_schedule();
327 }
328 __set_current_state(TASK_RUNNING);
329 return iocb->ki_user_data;
330 }
331
332 /* exit_aio: called when the last user of mm goes away. At this point,
333 * there is no way for any new requests to be submited or any of the
334 * io_* syscalls to be called on the context. However, there may be
335 * outstanding requests which hold references to the context; as they
336 * go away, they will call put_ioctx and release any pinned memory
337 * associated with the request (held via struct page * references).
338 */
339 void exit_aio(struct mm_struct *mm)
340 {
341 struct kioctx *ctx = mm->ioctx_list;
342 mm->ioctx_list = NULL;
343 while (ctx) {
344 struct kioctx *next = ctx->next;
345 ctx->next = NULL;
346 aio_cancel_all(ctx);
347
348 wait_for_all_aios(ctx);
349 /*
350 * Ensure we don't leave the ctx on the aio_wq
351 */
352 cancel_work_sync(&ctx->wq.work);
353
354 if (1 != atomic_read(&ctx->users))
355 printk(KERN_DEBUG
356 "exit_aio:ioctx still alive: %d %d %d\n",
357 atomic_read(&ctx->users), ctx->dead,
358 ctx->reqs_active);
359 put_ioctx(ctx);
360 ctx = next;
361 }
362 }
363
364 /* __put_ioctx
365 * Called when the last user of an aio context has gone away,
366 * and the struct needs to be freed.
367 */
368 void __put_ioctx(struct kioctx *ctx)
369 {
370 unsigned nr_events = ctx->max_reqs;
371
372 BUG_ON(ctx->reqs_active);
373
374 cancel_delayed_work(&ctx->wq);
375 cancel_work_sync(&ctx->wq.work);
376 aio_free_ring(ctx);
377 mmdrop(ctx->mm);
378 ctx->mm = NULL;
379 pr_debug("__put_ioctx: freeing %p\n", ctx);
380 kmem_cache_free(kioctx_cachep, ctx);
381
382 if (nr_events) {
383 spin_lock(&aio_nr_lock);
384 BUG_ON(aio_nr - nr_events > aio_nr);
385 aio_nr -= nr_events;
386 spin_unlock(&aio_nr_lock);
387 }
388 }
389
390 /* aio_get_req
391 * Allocate a slot for an aio request. Increments the users count
392 * of the kioctx so that the kioctx stays around until all requests are
393 * complete. Returns NULL if no requests are free.
394 *
395 * Returns with kiocb->users set to 2. The io submit code path holds
396 * an extra reference while submitting the i/o.
397 * This prevents races between the aio code path referencing the
398 * req (after submitting it) and aio_complete() freeing the req.
399 */
400 static struct kiocb *__aio_get_req(struct kioctx *ctx)
401 {
402 struct kiocb *req = NULL;
403 struct aio_ring *ring;
404 int okay = 0;
405
406 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
407 if (unlikely(!req))
408 return NULL;
409
410 req->ki_flags = 0;
411 req->ki_users = 2;
412 req->ki_key = 0;
413 req->ki_ctx = ctx;
414 req->ki_cancel = NULL;
415 req->ki_retry = NULL;
416 req->ki_dtor = NULL;
417 req->private = NULL;
418 req->ki_iovec = NULL;
419 INIT_LIST_HEAD(&req->ki_run_list);
420 req->ki_eventfd = ERR_PTR(-EINVAL);
421
422 /* Check if the completion queue has enough free space to
423 * accept an event from this io.
424 */
425 spin_lock_irq(&ctx->ctx_lock);
426 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
427 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
428 list_add(&req->ki_list, &ctx->active_reqs);
429 ctx->reqs_active++;
430 okay = 1;
431 }
432 kunmap_atomic(ring, KM_USER0);
433 spin_unlock_irq(&ctx->ctx_lock);
434
435 if (!okay) {
436 kmem_cache_free(kiocb_cachep, req);
437 req = NULL;
438 }
439
440 return req;
441 }
442
443 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
444 {
445 struct kiocb *req;
446 /* Handle a potential starvation case -- should be exceedingly rare as
447 * requests will be stuck on fput_head only if the aio_fput_routine is
448 * delayed and the requests were the last user of the struct file.
449 */
450 req = __aio_get_req(ctx);
451 if (unlikely(NULL == req)) {
452 aio_fput_routine(NULL);
453 req = __aio_get_req(ctx);
454 }
455 return req;
456 }
457
458 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
459 {
460 assert_spin_locked(&ctx->ctx_lock);
461
462 if (!IS_ERR(req->ki_eventfd))
463 fput(req->ki_eventfd);
464 if (req->ki_dtor)
465 req->ki_dtor(req);
466 if (req->ki_iovec != &req->ki_inline_vec)
467 kfree(req->ki_iovec);
468 kmem_cache_free(kiocb_cachep, req);
469 ctx->reqs_active--;
470
471 if (unlikely(!ctx->reqs_active && ctx->dead))
472 wake_up(&ctx->wait);
473 }
474
475 static void aio_fput_routine(struct work_struct *data)
476 {
477 spin_lock_irq(&fput_lock);
478 while (likely(!list_empty(&fput_head))) {
479 struct kiocb *req = list_kiocb(fput_head.next);
480 struct kioctx *ctx = req->ki_ctx;
481
482 list_del(&req->ki_list);
483 spin_unlock_irq(&fput_lock);
484
485 /* Complete the fput */
486 __fput(req->ki_filp);
487
488 /* Link the iocb into the context's free list */
489 spin_lock_irq(&ctx->ctx_lock);
490 really_put_req(ctx, req);
491 spin_unlock_irq(&ctx->ctx_lock);
492
493 put_ioctx(ctx);
494 spin_lock_irq(&fput_lock);
495 }
496 spin_unlock_irq(&fput_lock);
497 }
498
499 /* __aio_put_req
500 * Returns true if this put was the last user of the request.
501 */
502 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
503 {
504 dprintk(KERN_DEBUG "aio_put(%p): f_count=%d\n",
505 req, atomic_read(&req->ki_filp->f_count));
506
507 assert_spin_locked(&ctx->ctx_lock);
508
509 req->ki_users --;
510 BUG_ON(req->ki_users < 0);
511 if (likely(req->ki_users))
512 return 0;
513 list_del(&req->ki_list); /* remove from active_reqs */
514 req->ki_cancel = NULL;
515 req->ki_retry = NULL;
516
517 /* Must be done under the lock to serialise against cancellation.
518 * Call this aio_fput as it duplicates fput via the fput_work.
519 */
520 if (unlikely(atomic_dec_and_test(&req->ki_filp->f_count))) {
521 get_ioctx(ctx);
522 spin_lock(&fput_lock);
523 list_add(&req->ki_list, &fput_head);
524 spin_unlock(&fput_lock);
525 queue_work(aio_wq, &fput_work);
526 } else
527 really_put_req(ctx, req);
528 return 1;
529 }
530
531 /* aio_put_req
532 * Returns true if this put was the last user of the kiocb,
533 * false if the request is still in use.
534 */
535 int aio_put_req(struct kiocb *req)
536 {
537 struct kioctx *ctx = req->ki_ctx;
538 int ret;
539 spin_lock_irq(&ctx->ctx_lock);
540 ret = __aio_put_req(ctx, req);
541 spin_unlock_irq(&ctx->ctx_lock);
542 return ret;
543 }
544
545 /* Lookup an ioctx id. ioctx_list is lockless for reads.
546 * FIXME: this is O(n) and is only suitable for development.
547 */
548 struct kioctx *lookup_ioctx(unsigned long ctx_id)
549 {
550 struct kioctx *ioctx;
551 struct mm_struct *mm;
552
553 mm = current->mm;
554 read_lock(&mm->ioctx_list_lock);
555 for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next)
556 if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) {
557 get_ioctx(ioctx);
558 break;
559 }
560 read_unlock(&mm->ioctx_list_lock);
561
562 return ioctx;
563 }
564
565 /*
566 * use_mm
567 * Makes the calling kernel thread take on the specified
568 * mm context.
569 * Called by the retry thread execute retries within the
570 * iocb issuer's mm context, so that copy_from/to_user
571 * operations work seamlessly for aio.
572 * (Note: this routine is intended to be called only
573 * from a kernel thread context)
574 */
575 static void use_mm(struct mm_struct *mm)
576 {
577 struct mm_struct *active_mm;
578 struct task_struct *tsk = current;
579
580 task_lock(tsk);
581 tsk->flags |= PF_BORROWED_MM;
582 active_mm = tsk->active_mm;
583 atomic_inc(&mm->mm_count);
584 local_irq_disable(); // FIXME
585 /*
586 * Note that on UML this *requires* PF_BORROWED_MM to be set, otherwise
587 * it won't work. Update it accordingly if you change it here
588 */
589 switch_mm(active_mm, mm, tsk);
590 tsk->mm = mm;
591 tsk->active_mm = mm;
592 local_irq_enable();
593 task_unlock(tsk);
594
595 mmdrop(active_mm);
596 }
597
598 /*
599 * unuse_mm
600 * Reverses the effect of use_mm, i.e. releases the
601 * specified mm context which was earlier taken on
602 * by the calling kernel thread
603 * (Note: this routine is intended to be called only
604 * from a kernel thread context)
605 */
606 static void unuse_mm(struct mm_struct *mm)
607 {
608 struct task_struct *tsk = current;
609
610 task_lock(tsk);
611 tsk->flags &= ~PF_BORROWED_MM;
612 tsk->mm = NULL;
613 /* active_mm is still 'mm' */
614 enter_lazy_tlb(mm, tsk);
615 task_unlock(tsk);
616 }
617
618 /*
619 * Queue up a kiocb to be retried. Assumes that the kiocb
620 * has already been marked as kicked, and places it on
621 * the retry run list for the corresponding ioctx, if it
622 * isn't already queued. Returns 1 if it actually queued
623 * the kiocb (to tell the caller to activate the work
624 * queue to process it), or 0, if it found that it was
625 * already queued.
626 */
627 static inline int __queue_kicked_iocb(struct kiocb *iocb)
628 {
629 struct kioctx *ctx = iocb->ki_ctx;
630
631 assert_spin_locked(&ctx->ctx_lock);
632
633 if (list_empty(&iocb->ki_run_list)) {
634 list_add_tail(&iocb->ki_run_list,
635 &ctx->run_list);
636 return 1;
637 }
638 return 0;
639 }
640
641 /* aio_run_iocb
642 * This is the core aio execution routine. It is
643 * invoked both for initial i/o submission and
644 * subsequent retries via the aio_kick_handler.
645 * Expects to be invoked with iocb->ki_ctx->lock
646 * already held. The lock is released and reacquired
647 * as needed during processing.
648 *
649 * Calls the iocb retry method (already setup for the
650 * iocb on initial submission) for operation specific
651 * handling, but takes care of most of common retry
652 * execution details for a given iocb. The retry method
653 * needs to be non-blocking as far as possible, to avoid
654 * holding up other iocbs waiting to be serviced by the
655 * retry kernel thread.
656 *
657 * The trickier parts in this code have to do with
658 * ensuring that only one retry instance is in progress
659 * for a given iocb at any time. Providing that guarantee
660 * simplifies the coding of individual aio operations as
661 * it avoids various potential races.
662 */
663 static ssize_t aio_run_iocb(struct kiocb *iocb)
664 {
665 struct kioctx *ctx = iocb->ki_ctx;
666 ssize_t (*retry)(struct kiocb *);
667 ssize_t ret;
668
669 if (!(retry = iocb->ki_retry)) {
670 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
671 return 0;
672 }
673
674 /*
675 * We don't want the next retry iteration for this
676 * operation to start until this one has returned and
677 * updated the iocb state. However, wait_queue functions
678 * can trigger a kick_iocb from interrupt context in the
679 * meantime, indicating that data is available for the next
680 * iteration. We want to remember that and enable the
681 * next retry iteration _after_ we are through with
682 * this one.
683 *
684 * So, in order to be able to register a "kick", but
685 * prevent it from being queued now, we clear the kick
686 * flag, but make the kick code *think* that the iocb is
687 * still on the run list until we are actually done.
688 * When we are done with this iteration, we check if
689 * the iocb was kicked in the meantime and if so, queue
690 * it up afresh.
691 */
692
693 kiocbClearKicked(iocb);
694
695 /*
696 * This is so that aio_complete knows it doesn't need to
697 * pull the iocb off the run list (We can't just call
698 * INIT_LIST_HEAD because we don't want a kick_iocb to
699 * queue this on the run list yet)
700 */
701 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
702 spin_unlock_irq(&ctx->ctx_lock);
703
704 /* Quit retrying if the i/o has been cancelled */
705 if (kiocbIsCancelled(iocb)) {
706 ret = -EINTR;
707 aio_complete(iocb, ret, 0);
708 /* must not access the iocb after this */
709 goto out;
710 }
711
712 /*
713 * Now we are all set to call the retry method in async
714 * context.
715 */
716 ret = retry(iocb);
717
718 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
719 BUG_ON(!list_empty(&iocb->ki_wait.task_list));
720 aio_complete(iocb, ret, 0);
721 }
722 out:
723 spin_lock_irq(&ctx->ctx_lock);
724
725 if (-EIOCBRETRY == ret) {
726 /*
727 * OK, now that we are done with this iteration
728 * and know that there is more left to go,
729 * this is where we let go so that a subsequent
730 * "kick" can start the next iteration
731 */
732
733 /* will make __queue_kicked_iocb succeed from here on */
734 INIT_LIST_HEAD(&iocb->ki_run_list);
735 /* we must queue the next iteration ourselves, if it
736 * has already been kicked */
737 if (kiocbIsKicked(iocb)) {
738 __queue_kicked_iocb(iocb);
739
740 /*
741 * __queue_kicked_iocb will always return 1 here, because
742 * iocb->ki_run_list is empty at this point so it should
743 * be safe to unconditionally queue the context into the
744 * work queue.
745 */
746 aio_queue_work(ctx);
747 }
748 }
749 return ret;
750 }
751
752 /*
753 * __aio_run_iocbs:
754 * Process all pending retries queued on the ioctx
755 * run list.
756 * Assumes it is operating within the aio issuer's mm
757 * context.
758 */
759 static int __aio_run_iocbs(struct kioctx *ctx)
760 {
761 struct kiocb *iocb;
762 struct list_head run_list;
763
764 assert_spin_locked(&ctx->ctx_lock);
765
766 list_replace_init(&ctx->run_list, &run_list);
767 while (!list_empty(&run_list)) {
768 iocb = list_entry(run_list.next, struct kiocb,
769 ki_run_list);
770 list_del(&iocb->ki_run_list);
771 /*
772 * Hold an extra reference while retrying i/o.
773 */
774 iocb->ki_users++; /* grab extra reference */
775 aio_run_iocb(iocb);
776 __aio_put_req(ctx, iocb);
777 }
778 if (!list_empty(&ctx->run_list))
779 return 1;
780 return 0;
781 }
782
783 static void aio_queue_work(struct kioctx * ctx)
784 {
785 unsigned long timeout;
786 /*
787 * if someone is waiting, get the work started right
788 * away, otherwise, use a longer delay
789 */
790 smp_mb();
791 if (waitqueue_active(&ctx->wait))
792 timeout = 1;
793 else
794 timeout = HZ/10;
795 queue_delayed_work(aio_wq, &ctx->wq, timeout);
796 }
797
798
799 /*
800 * aio_run_iocbs:
801 * Process all pending retries queued on the ioctx
802 * run list.
803 * Assumes it is operating within the aio issuer's mm
804 * context.
805 */
806 static inline void aio_run_iocbs(struct kioctx *ctx)
807 {
808 int requeue;
809
810 spin_lock_irq(&ctx->ctx_lock);
811
812 requeue = __aio_run_iocbs(ctx);
813 spin_unlock_irq(&ctx->ctx_lock);
814 if (requeue)
815 aio_queue_work(ctx);
816 }
817
818 /*
819 * just like aio_run_iocbs, but keeps running them until
820 * the list stays empty
821 */
822 static inline void aio_run_all_iocbs(struct kioctx *ctx)
823 {
824 spin_lock_irq(&ctx->ctx_lock);
825 while (__aio_run_iocbs(ctx))
826 ;
827 spin_unlock_irq(&ctx->ctx_lock);
828 }
829
830 /*
831 * aio_kick_handler:
832 * Work queue handler triggered to process pending
833 * retries on an ioctx. Takes on the aio issuer's
834 * mm context before running the iocbs, so that
835 * copy_xxx_user operates on the issuer's address
836 * space.
837 * Run on aiod's context.
838 */
839 static void aio_kick_handler(struct work_struct *work)
840 {
841 struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
842 mm_segment_t oldfs = get_fs();
843 struct mm_struct *mm;
844 int requeue;
845
846 set_fs(USER_DS);
847 use_mm(ctx->mm);
848 spin_lock_irq(&ctx->ctx_lock);
849 requeue =__aio_run_iocbs(ctx);
850 mm = ctx->mm;
851 spin_unlock_irq(&ctx->ctx_lock);
852 unuse_mm(mm);
853 set_fs(oldfs);
854 /*
855 * we're in a worker thread already, don't use queue_delayed_work,
856 */
857 if (requeue)
858 queue_delayed_work(aio_wq, &ctx->wq, 0);
859 }
860
861
862 /*
863 * Called by kick_iocb to queue the kiocb for retry
864 * and if required activate the aio work queue to process
865 * it
866 */
867 static void try_queue_kicked_iocb(struct kiocb *iocb)
868 {
869 struct kioctx *ctx = iocb->ki_ctx;
870 unsigned long flags;
871 int run = 0;
872
873 /* We're supposed to be the only path putting the iocb back on the run
874 * list. If we find that the iocb is *back* on a wait queue already
875 * than retry has happened before we could queue the iocb. This also
876 * means that the retry could have completed and freed our iocb, no
877 * good. */
878 BUG_ON((!list_empty(&iocb->ki_wait.task_list)));
879
880 spin_lock_irqsave(&ctx->ctx_lock, flags);
881 /* set this inside the lock so that we can't race with aio_run_iocb()
882 * testing it and putting the iocb on the run list under the lock */
883 if (!kiocbTryKick(iocb))
884 run = __queue_kicked_iocb(iocb);
885 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
886 if (run)
887 aio_queue_work(ctx);
888 }
889
890 /*
891 * kick_iocb:
892 * Called typically from a wait queue callback context
893 * (aio_wake_function) to trigger a retry of the iocb.
894 * The retry is usually executed by aio workqueue
895 * threads (See aio_kick_handler).
896 */
897 void kick_iocb(struct kiocb *iocb)
898 {
899 /* sync iocbs are easy: they can only ever be executing from a
900 * single context. */
901 if (is_sync_kiocb(iocb)) {
902 kiocbSetKicked(iocb);
903 wake_up_process(iocb->ki_obj.tsk);
904 return;
905 }
906
907 try_queue_kicked_iocb(iocb);
908 }
909 EXPORT_SYMBOL(kick_iocb);
910
911 /* aio_complete
912 * Called when the io request on the given iocb is complete.
913 * Returns true if this is the last user of the request. The
914 * only other user of the request can be the cancellation code.
915 */
916 int aio_complete(struct kiocb *iocb, long res, long res2)
917 {
918 struct kioctx *ctx = iocb->ki_ctx;
919 struct aio_ring_info *info;
920 struct aio_ring *ring;
921 struct io_event *event;
922 unsigned long flags;
923 unsigned long tail;
924 int ret;
925
926 /*
927 * Special case handling for sync iocbs:
928 * - events go directly into the iocb for fast handling
929 * - the sync task with the iocb in its stack holds the single iocb
930 * ref, no other paths have a way to get another ref
931 * - the sync task helpfully left a reference to itself in the iocb
932 */
933 if (is_sync_kiocb(iocb)) {
934 BUG_ON(iocb->ki_users != 1);
935 iocb->ki_user_data = res;
936 iocb->ki_users = 0;
937 wake_up_process(iocb->ki_obj.tsk);
938 return 1;
939 }
940
941 info = &ctx->ring_info;
942
943 /* add a completion event to the ring buffer.
944 * must be done holding ctx->ctx_lock to prevent
945 * other code from messing with the tail
946 * pointer since we might be called from irq
947 * context.
948 */
949 spin_lock_irqsave(&ctx->ctx_lock, flags);
950
951 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
952 list_del_init(&iocb->ki_run_list);
953
954 /*
955 * cancelled requests don't get events, userland was given one
956 * when the event got cancelled.
957 */
958 if (kiocbIsCancelled(iocb))
959 goto put_rq;
960
961 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
962
963 tail = info->tail;
964 event = aio_ring_event(info, tail, KM_IRQ0);
965 if (++tail >= info->nr)
966 tail = 0;
967
968 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
969 event->data = iocb->ki_user_data;
970 event->res = res;
971 event->res2 = res2;
972
973 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
974 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
975 res, res2);
976
977 /* after flagging the request as done, we
978 * must never even look at it again
979 */
980 smp_wmb(); /* make event visible before updating tail */
981
982 info->tail = tail;
983 ring->tail = tail;
984
985 put_aio_ring_event(event, KM_IRQ0);
986 kunmap_atomic(ring, KM_IRQ1);
987
988 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
989
990 /*
991 * Check if the user asked us to deliver the result through an
992 * eventfd. The eventfd_signal() function is safe to be called
993 * from IRQ context.
994 */
995 if (!IS_ERR(iocb->ki_eventfd))
996 eventfd_signal(iocb->ki_eventfd, 1);
997
998 put_rq:
999 /* everything turned out well, dispose of the aiocb. */
1000 ret = __aio_put_req(ctx, iocb);
1001
1002 /*
1003 * We have to order our ring_info tail store above and test
1004 * of the wait list below outside the wait lock. This is
1005 * like in wake_up_bit() where clearing a bit has to be
1006 * ordered with the unlocked test.
1007 */
1008 smp_mb();
1009
1010 if (waitqueue_active(&ctx->wait))
1011 wake_up(&ctx->wait);
1012
1013 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1014 return ret;
1015 }
1016
1017 /* aio_read_evt
1018 * Pull an event off of the ioctx's event ring. Returns the number of
1019 * events fetched (0 or 1 ;-)
1020 * FIXME: make this use cmpxchg.
1021 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1022 */
1023 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1024 {
1025 struct aio_ring_info *info = &ioctx->ring_info;
1026 struct aio_ring *ring;
1027 unsigned long head;
1028 int ret = 0;
1029
1030 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1031 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1032 (unsigned long)ring->head, (unsigned long)ring->tail,
1033 (unsigned long)ring->nr);
1034
1035 if (ring->head == ring->tail)
1036 goto out;
1037
1038 spin_lock(&info->ring_lock);
1039
1040 head = ring->head % info->nr;
1041 if (head != ring->tail) {
1042 struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1043 *ent = *evp;
1044 head = (head + 1) % info->nr;
1045 smp_mb(); /* finish reading the event before updatng the head */
1046 ring->head = head;
1047 ret = 1;
1048 put_aio_ring_event(evp, KM_USER1);
1049 }
1050 spin_unlock(&info->ring_lock);
1051
1052 out:
1053 kunmap_atomic(ring, KM_USER0);
1054 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1055 (unsigned long)ring->head, (unsigned long)ring->tail);
1056 return ret;
1057 }
1058
1059 struct aio_timeout {
1060 struct timer_list timer;
1061 int timed_out;
1062 struct task_struct *p;
1063 };
1064
1065 static void timeout_func(unsigned long data)
1066 {
1067 struct aio_timeout *to = (struct aio_timeout *)data;
1068
1069 to->timed_out = 1;
1070 wake_up_process(to->p);
1071 }
1072
1073 static inline void init_timeout(struct aio_timeout *to)
1074 {
1075 init_timer(&to->timer);
1076 to->timer.data = (unsigned long)to;
1077 to->timer.function = timeout_func;
1078 to->timed_out = 0;
1079 to->p = current;
1080 }
1081
1082 static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1083 const struct timespec *ts)
1084 {
1085 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1086 if (time_after(to->timer.expires, jiffies))
1087 add_timer(&to->timer);
1088 else
1089 to->timed_out = 1;
1090 }
1091
1092 static inline void clear_timeout(struct aio_timeout *to)
1093 {
1094 del_singleshot_timer_sync(&to->timer);
1095 }
1096
1097 static int read_events(struct kioctx *ctx,
1098 long min_nr, long nr,
1099 struct io_event __user *event,
1100 struct timespec __user *timeout)
1101 {
1102 long start_jiffies = jiffies;
1103 struct task_struct *tsk = current;
1104 DECLARE_WAITQUEUE(wait, tsk);
1105 int ret;
1106 int i = 0;
1107 struct io_event ent;
1108 struct aio_timeout to;
1109 int retry = 0;
1110
1111 /* needed to zero any padding within an entry (there shouldn't be
1112 * any, but C is fun!
1113 */
1114 memset(&ent, 0, sizeof(ent));
1115 retry:
1116 ret = 0;
1117 while (likely(i < nr)) {
1118 ret = aio_read_evt(ctx, &ent);
1119 if (unlikely(ret <= 0))
1120 break;
1121
1122 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1123 ent.data, ent.obj, ent.res, ent.res2);
1124
1125 /* Could we split the check in two? */
1126 ret = -EFAULT;
1127 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1128 dprintk("aio: lost an event due to EFAULT.\n");
1129 break;
1130 }
1131 ret = 0;
1132
1133 /* Good, event copied to userland, update counts. */
1134 event ++;
1135 i ++;
1136 }
1137
1138 if (min_nr <= i)
1139 return i;
1140 if (ret)
1141 return ret;
1142
1143 /* End fast path */
1144
1145 /* racey check, but it gets redone */
1146 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1147 retry = 1;
1148 aio_run_all_iocbs(ctx);
1149 goto retry;
1150 }
1151
1152 init_timeout(&to);
1153 if (timeout) {
1154 struct timespec ts;
1155 ret = -EFAULT;
1156 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1157 goto out;
1158
1159 set_timeout(start_jiffies, &to, &ts);
1160 }
1161
1162 while (likely(i < nr)) {
1163 add_wait_queue_exclusive(&ctx->wait, &wait);
1164 do {
1165 set_task_state(tsk, TASK_INTERRUPTIBLE);
1166 ret = aio_read_evt(ctx, &ent);
1167 if (ret)
1168 break;
1169 if (min_nr <= i)
1170 break;
1171 if (unlikely(ctx->dead)) {
1172 ret = -EINVAL;
1173 break;
1174 }
1175 if (to.timed_out) /* Only check after read evt */
1176 break;
1177 /* Try to only show up in io wait if there are ops
1178 * in flight */
1179 if (ctx->reqs_active)
1180 io_schedule();
1181 else
1182 schedule();
1183 if (signal_pending(tsk)) {
1184 ret = -EINTR;
1185 break;
1186 }
1187 /*ret = aio_read_evt(ctx, &ent);*/
1188 } while (1) ;
1189
1190 set_task_state(tsk, TASK_RUNNING);
1191 remove_wait_queue(&ctx->wait, &wait);
1192
1193 if (unlikely(ret <= 0))
1194 break;
1195
1196 ret = -EFAULT;
1197 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1198 dprintk("aio: lost an event due to EFAULT.\n");
1199 break;
1200 }
1201
1202 /* Good, event copied to userland, update counts. */
1203 event ++;
1204 i ++;
1205 }
1206
1207 if (timeout)
1208 clear_timeout(&to);
1209 out:
1210 return i ? i : ret;
1211 }
1212
1213 /* Take an ioctx and remove it from the list of ioctx's. Protects
1214 * against races with itself via ->dead.
1215 */
1216 static void io_destroy(struct kioctx *ioctx)
1217 {
1218 struct mm_struct *mm = current->mm;
1219 struct kioctx **tmp;
1220 int was_dead;
1221
1222 /* delete the entry from the list is someone else hasn't already */
1223 write_lock(&mm->ioctx_list_lock);
1224 was_dead = ioctx->dead;
1225 ioctx->dead = 1;
1226 for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx;
1227 tmp = &(*tmp)->next)
1228 ;
1229 if (*tmp)
1230 *tmp = ioctx->next;
1231 write_unlock(&mm->ioctx_list_lock);
1232
1233 dprintk("aio_release(%p)\n", ioctx);
1234 if (likely(!was_dead))
1235 put_ioctx(ioctx); /* twice for the list */
1236
1237 aio_cancel_all(ioctx);
1238 wait_for_all_aios(ioctx);
1239
1240 /*
1241 * Wake up any waiters. The setting of ctx->dead must be seen
1242 * by other CPUs at this point. Right now, we rely on the
1243 * locking done by the above calls to ensure this consistency.
1244 */
1245 wake_up(&ioctx->wait);
1246 put_ioctx(ioctx); /* once for the lookup */
1247 }
1248
1249 /* sys_io_setup:
1250 * Create an aio_context capable of receiving at least nr_events.
1251 * ctxp must not point to an aio_context that already exists, and
1252 * must be initialized to 0 prior to the call. On successful
1253 * creation of the aio_context, *ctxp is filled in with the resulting
1254 * handle. May fail with -EINVAL if *ctxp is not initialized,
1255 * if the specified nr_events exceeds internal limits. May fail
1256 * with -EAGAIN if the specified nr_events exceeds the user's limit
1257 * of available events. May fail with -ENOMEM if insufficient kernel
1258 * resources are available. May fail with -EFAULT if an invalid
1259 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1260 * implemented.
1261 */
1262 asmlinkage long sys_io_setup(unsigned nr_events, aio_context_t __user *ctxp)
1263 {
1264 struct kioctx *ioctx = NULL;
1265 unsigned long ctx;
1266 long ret;
1267
1268 ret = get_user(ctx, ctxp);
1269 if (unlikely(ret))
1270 goto out;
1271
1272 ret = -EINVAL;
1273 if (unlikely(ctx || nr_events == 0)) {
1274 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1275 ctx, nr_events);
1276 goto out;
1277 }
1278
1279 ioctx = ioctx_alloc(nr_events);
1280 ret = PTR_ERR(ioctx);
1281 if (!IS_ERR(ioctx)) {
1282 ret = put_user(ioctx->user_id, ctxp);
1283 if (!ret)
1284 return 0;
1285
1286 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1287 io_destroy(ioctx);
1288 }
1289
1290 out:
1291 return ret;
1292 }
1293
1294 /* sys_io_destroy:
1295 * Destroy the aio_context specified. May cancel any outstanding
1296 * AIOs and block on completion. Will fail with -ENOSYS if not
1297 * implemented. May fail with -EFAULT if the context pointed to
1298 * is invalid.
1299 */
1300 asmlinkage long sys_io_destroy(aio_context_t ctx)
1301 {
1302 struct kioctx *ioctx = lookup_ioctx(ctx);
1303 if (likely(NULL != ioctx)) {
1304 io_destroy(ioctx);
1305 return 0;
1306 }
1307 pr_debug("EINVAL: io_destroy: invalid context id\n");
1308 return -EINVAL;
1309 }
1310
1311 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1312 {
1313 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1314
1315 BUG_ON(ret <= 0);
1316
1317 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1318 ssize_t this = min((ssize_t)iov->iov_len, ret);
1319 iov->iov_base += this;
1320 iov->iov_len -= this;
1321 iocb->ki_left -= this;
1322 ret -= this;
1323 if (iov->iov_len == 0) {
1324 iocb->ki_cur_seg++;
1325 iov++;
1326 }
1327 }
1328
1329 /* the caller should not have done more io than what fit in
1330 * the remaining iovecs */
1331 BUG_ON(ret > 0 && iocb->ki_left == 0);
1332 }
1333
1334 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1335 {
1336 struct file *file = iocb->ki_filp;
1337 struct address_space *mapping = file->f_mapping;
1338 struct inode *inode = mapping->host;
1339 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1340 unsigned long, loff_t);
1341 ssize_t ret = 0;
1342 unsigned short opcode;
1343
1344 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1345 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1346 rw_op = file->f_op->aio_read;
1347 opcode = IOCB_CMD_PREADV;
1348 } else {
1349 rw_op = file->f_op->aio_write;
1350 opcode = IOCB_CMD_PWRITEV;
1351 }
1352
1353 /* This matches the pread()/pwrite() logic */
1354 if (iocb->ki_pos < 0)
1355 return -EINVAL;
1356
1357 do {
1358 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1359 iocb->ki_nr_segs - iocb->ki_cur_seg,
1360 iocb->ki_pos);
1361 if (ret > 0)
1362 aio_advance_iovec(iocb, ret);
1363
1364 /* retry all partial writes. retry partial reads as long as its a
1365 * regular file. */
1366 } while (ret > 0 && iocb->ki_left > 0 &&
1367 (opcode == IOCB_CMD_PWRITEV ||
1368 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1369
1370 /* This means we must have transferred all that we could */
1371 /* No need to retry anymore */
1372 if ((ret == 0) || (iocb->ki_left == 0))
1373 ret = iocb->ki_nbytes - iocb->ki_left;
1374
1375 /* If we managed to write some out we return that, rather than
1376 * the eventual error. */
1377 if (opcode == IOCB_CMD_PWRITEV
1378 && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
1379 && iocb->ki_nbytes - iocb->ki_left)
1380 ret = iocb->ki_nbytes - iocb->ki_left;
1381
1382 return ret;
1383 }
1384
1385 static ssize_t aio_fdsync(struct kiocb *iocb)
1386 {
1387 struct file *file = iocb->ki_filp;
1388 ssize_t ret = -EINVAL;
1389
1390 if (file->f_op->aio_fsync)
1391 ret = file->f_op->aio_fsync(iocb, 1);
1392 return ret;
1393 }
1394
1395 static ssize_t aio_fsync(struct kiocb *iocb)
1396 {
1397 struct file *file = iocb->ki_filp;
1398 ssize_t ret = -EINVAL;
1399
1400 if (file->f_op->aio_fsync)
1401 ret = file->f_op->aio_fsync(iocb, 0);
1402 return ret;
1403 }
1404
1405 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb)
1406 {
1407 ssize_t ret;
1408
1409 ret = rw_copy_check_uvector(type, (struct iovec __user *)kiocb->ki_buf,
1410 kiocb->ki_nbytes, 1,
1411 &kiocb->ki_inline_vec, &kiocb->ki_iovec);
1412 if (ret < 0)
1413 goto out;
1414
1415 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1416 kiocb->ki_cur_seg = 0;
1417 /* ki_nbytes/left now reflect bytes instead of segs */
1418 kiocb->ki_nbytes = ret;
1419 kiocb->ki_left = ret;
1420
1421 ret = 0;
1422 out:
1423 return ret;
1424 }
1425
1426 static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
1427 {
1428 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1429 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1430 kiocb->ki_iovec->iov_len = kiocb->ki_left;
1431 kiocb->ki_nr_segs = 1;
1432 kiocb->ki_cur_seg = 0;
1433 return 0;
1434 }
1435
1436 /*
1437 * aio_setup_iocb:
1438 * Performs the initial checks and aio retry method
1439 * setup for the kiocb at the time of io submission.
1440 */
1441 static ssize_t aio_setup_iocb(struct kiocb *kiocb)
1442 {
1443 struct file *file = kiocb->ki_filp;
1444 ssize_t ret = 0;
1445
1446 switch (kiocb->ki_opcode) {
1447 case IOCB_CMD_PREAD:
1448 ret = -EBADF;
1449 if (unlikely(!(file->f_mode & FMODE_READ)))
1450 break;
1451 ret = -EFAULT;
1452 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1453 kiocb->ki_left)))
1454 break;
1455 ret = security_file_permission(file, MAY_READ);
1456 if (unlikely(ret))
1457 break;
1458 ret = aio_setup_single_vector(kiocb);
1459 if (ret)
1460 break;
1461 ret = -EINVAL;
1462 if (file->f_op->aio_read)
1463 kiocb->ki_retry = aio_rw_vect_retry;
1464 break;
1465 case IOCB_CMD_PWRITE:
1466 ret = -EBADF;
1467 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1468 break;
1469 ret = -EFAULT;
1470 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1471 kiocb->ki_left)))
1472 break;
1473 ret = security_file_permission(file, MAY_WRITE);
1474 if (unlikely(ret))
1475 break;
1476 ret = aio_setup_single_vector(kiocb);
1477 if (ret)
1478 break;
1479 ret = -EINVAL;
1480 if (file->f_op->aio_write)
1481 kiocb->ki_retry = aio_rw_vect_retry;
1482 break;
1483 case IOCB_CMD_PREADV:
1484 ret = -EBADF;
1485 if (unlikely(!(file->f_mode & FMODE_READ)))
1486 break;
1487 ret = security_file_permission(file, MAY_READ);
1488 if (unlikely(ret))
1489 break;
1490 ret = aio_setup_vectored_rw(READ, kiocb);
1491 if (ret)
1492 break;
1493 ret = -EINVAL;
1494 if (file->f_op->aio_read)
1495 kiocb->ki_retry = aio_rw_vect_retry;
1496 break;
1497 case IOCB_CMD_PWRITEV:
1498 ret = -EBADF;
1499 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1500 break;
1501 ret = security_file_permission(file, MAY_WRITE);
1502 if (unlikely(ret))
1503 break;
1504 ret = aio_setup_vectored_rw(WRITE, kiocb);
1505 if (ret)
1506 break;
1507 ret = -EINVAL;
1508 if (file->f_op->aio_write)
1509 kiocb->ki_retry = aio_rw_vect_retry;
1510 break;
1511 case IOCB_CMD_FDSYNC:
1512 ret = -EINVAL;
1513 if (file->f_op->aio_fsync)
1514 kiocb->ki_retry = aio_fdsync;
1515 break;
1516 case IOCB_CMD_FSYNC:
1517 ret = -EINVAL;
1518 if (file->f_op->aio_fsync)
1519 kiocb->ki_retry = aio_fsync;
1520 break;
1521 default:
1522 dprintk("EINVAL: io_submit: no operation provided\n");
1523 ret = -EINVAL;
1524 }
1525
1526 if (!kiocb->ki_retry)
1527 return ret;
1528
1529 return 0;
1530 }
1531
1532 /*
1533 * aio_wake_function:
1534 * wait queue callback function for aio notification,
1535 * Simply triggers a retry of the operation via kick_iocb.
1536 *
1537 * This callback is specified in the wait queue entry in
1538 * a kiocb.
1539 *
1540 * Note:
1541 * This routine is executed with the wait queue lock held.
1542 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests
1543 * the ioctx lock inside the wait queue lock. This is safe
1544 * because this callback isn't used for wait queues which
1545 * are nested inside ioctx lock (i.e. ctx->wait)
1546 */
1547 static int aio_wake_function(wait_queue_t *wait, unsigned mode,
1548 int sync, void *key)
1549 {
1550 struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait);
1551
1552 list_del_init(&wait->task_list);
1553 kick_iocb(iocb);
1554 return 1;
1555 }
1556
1557 int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1558 struct iocb *iocb)
1559 {
1560 struct kiocb *req;
1561 struct file *file;
1562 ssize_t ret;
1563
1564 /* enforce forwards compatibility on users */
1565 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1566 pr_debug("EINVAL: io_submit: reserve field set\n");
1567 return -EINVAL;
1568 }
1569
1570 /* prevent overflows */
1571 if (unlikely(
1572 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1573 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1574 ((ssize_t)iocb->aio_nbytes < 0)
1575 )) {
1576 pr_debug("EINVAL: io_submit: overflow check\n");
1577 return -EINVAL;
1578 }
1579
1580 file = fget(iocb->aio_fildes);
1581 if (unlikely(!file))
1582 return -EBADF;
1583
1584 req = aio_get_req(ctx); /* returns with 2 references to req */
1585 if (unlikely(!req)) {
1586 fput(file);
1587 return -EAGAIN;
1588 }
1589 req->ki_filp = file;
1590 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1591 /*
1592 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1593 * instance of the file* now. The file descriptor must be
1594 * an eventfd() fd, and will be signaled for each completed
1595 * event using the eventfd_signal() function.
1596 */
1597 req->ki_eventfd = eventfd_fget((int) iocb->aio_resfd);
1598 if (unlikely(IS_ERR(req->ki_eventfd))) {
1599 ret = PTR_ERR(req->ki_eventfd);
1600 goto out_put_req;
1601 }
1602 }
1603
1604 ret = put_user(req->ki_key, &user_iocb->aio_key);
1605 if (unlikely(ret)) {
1606 dprintk("EFAULT: aio_key\n");
1607 goto out_put_req;
1608 }
1609
1610 req->ki_obj.user = user_iocb;
1611 req->ki_user_data = iocb->aio_data;
1612 req->ki_pos = iocb->aio_offset;
1613
1614 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1615 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1616 req->ki_opcode = iocb->aio_lio_opcode;
1617 init_waitqueue_func_entry(&req->ki_wait, aio_wake_function);
1618 INIT_LIST_HEAD(&req->ki_wait.task_list);
1619
1620 ret = aio_setup_iocb(req);
1621
1622 if (ret)
1623 goto out_put_req;
1624
1625 spin_lock_irq(&ctx->ctx_lock);
1626 aio_run_iocb(req);
1627 if (!list_empty(&ctx->run_list)) {
1628 /* drain the run list */
1629 while (__aio_run_iocbs(ctx))
1630 ;
1631 }
1632 spin_unlock_irq(&ctx->ctx_lock);
1633 aio_put_req(req); /* drop extra ref to req */
1634 return 0;
1635
1636 out_put_req:
1637 aio_put_req(req); /* drop extra ref to req */
1638 aio_put_req(req); /* drop i/o ref to req */
1639 return ret;
1640 }
1641
1642 /* sys_io_submit:
1643 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1644 * the number of iocbs queued. May return -EINVAL if the aio_context
1645 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1646 * *iocbpp[0] is not properly initialized, if the operation specified
1647 * is invalid for the file descriptor in the iocb. May fail with
1648 * -EFAULT if any of the data structures point to invalid data. May
1649 * fail with -EBADF if the file descriptor specified in the first
1650 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1651 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1652 * fail with -ENOSYS if not implemented.
1653 */
1654 asmlinkage long sys_io_submit(aio_context_t ctx_id, long nr,
1655 struct iocb __user * __user *iocbpp)
1656 {
1657 struct kioctx *ctx;
1658 long ret = 0;
1659 int i;
1660
1661 if (unlikely(nr < 0))
1662 return -EINVAL;
1663
1664 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1665 return -EFAULT;
1666
1667 ctx = lookup_ioctx(ctx_id);
1668 if (unlikely(!ctx)) {
1669 pr_debug("EINVAL: io_submit: invalid context id\n");
1670 return -EINVAL;
1671 }
1672
1673 /*
1674 * AKPM: should this return a partial result if some of the IOs were
1675 * successfully submitted?
1676 */
1677 for (i=0; i<nr; i++) {
1678 struct iocb __user *user_iocb;
1679 struct iocb tmp;
1680
1681 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1682 ret = -EFAULT;
1683 break;
1684 }
1685
1686 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1687 ret = -EFAULT;
1688 break;
1689 }
1690
1691 ret = io_submit_one(ctx, user_iocb, &tmp);
1692 if (ret)
1693 break;
1694 }
1695
1696 put_ioctx(ctx);
1697 return i ? i : ret;
1698 }
1699
1700 /* lookup_kiocb
1701 * Finds a given iocb for cancellation.
1702 */
1703 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1704 u32 key)
1705 {
1706 struct list_head *pos;
1707
1708 assert_spin_locked(&ctx->ctx_lock);
1709
1710 /* TODO: use a hash or array, this sucks. */
1711 list_for_each(pos, &ctx->active_reqs) {
1712 struct kiocb *kiocb = list_kiocb(pos);
1713 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1714 return kiocb;
1715 }
1716 return NULL;
1717 }
1718
1719 /* sys_io_cancel:
1720 * Attempts to cancel an iocb previously passed to io_submit. If
1721 * the operation is successfully cancelled, the resulting event is
1722 * copied into the memory pointed to by result without being placed
1723 * into the completion queue and 0 is returned. May fail with
1724 * -EFAULT if any of the data structures pointed to are invalid.
1725 * May fail with -EINVAL if aio_context specified by ctx_id is
1726 * invalid. May fail with -EAGAIN if the iocb specified was not
1727 * cancelled. Will fail with -ENOSYS if not implemented.
1728 */
1729 asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb,
1730 struct io_event __user *result)
1731 {
1732 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1733 struct kioctx *ctx;
1734 struct kiocb *kiocb;
1735 u32 key;
1736 int ret;
1737
1738 ret = get_user(key, &iocb->aio_key);
1739 if (unlikely(ret))
1740 return -EFAULT;
1741
1742 ctx = lookup_ioctx(ctx_id);
1743 if (unlikely(!ctx))
1744 return -EINVAL;
1745
1746 spin_lock_irq(&ctx->ctx_lock);
1747 ret = -EAGAIN;
1748 kiocb = lookup_kiocb(ctx, iocb, key);
1749 if (kiocb && kiocb->ki_cancel) {
1750 cancel = kiocb->ki_cancel;
1751 kiocb->ki_users ++;
1752 kiocbSetCancelled(kiocb);
1753 } else
1754 cancel = NULL;
1755 spin_unlock_irq(&ctx->ctx_lock);
1756
1757 if (NULL != cancel) {
1758 struct io_event tmp;
1759 pr_debug("calling cancel\n");
1760 memset(&tmp, 0, sizeof(tmp));
1761 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1762 tmp.data = kiocb->ki_user_data;
1763 ret = cancel(kiocb, &tmp);
1764 if (!ret) {
1765 /* Cancellation succeeded -- copy the result
1766 * into the user's buffer.
1767 */
1768 if (copy_to_user(result, &tmp, sizeof(tmp)))
1769 ret = -EFAULT;
1770 }
1771 } else
1772 ret = -EINVAL;
1773
1774 put_ioctx(ctx);
1775
1776 return ret;
1777 }
1778
1779 /* io_getevents:
1780 * Attempts to read at least min_nr events and up to nr events from
1781 * the completion queue for the aio_context specified by ctx_id. May
1782 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range,
1783 * if nr is out of range, if when is out of range. May fail with
1784 * -EFAULT if any of the memory specified to is invalid. May return
1785 * 0 or < min_nr if no events are available and the timeout specified
1786 * by when has elapsed, where when == NULL specifies an infinite
1787 * timeout. Note that the timeout pointed to by when is relative and
1788 * will be updated if not NULL and the operation blocks. Will fail
1789 * with -ENOSYS if not implemented.
1790 */
1791 asmlinkage long sys_io_getevents(aio_context_t ctx_id,
1792 long min_nr,
1793 long nr,
1794 struct io_event __user *events,
1795 struct timespec __user *timeout)
1796 {
1797 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1798 long ret = -EINVAL;
1799
1800 if (likely(ioctx)) {
1801 if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0))
1802 ret = read_events(ioctx, min_nr, nr, events, timeout);
1803 put_ioctx(ioctx);
1804 }
1805
1806 asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);
1807 return ret;
1808 }
1809
1810 __initcall(aio_setup);
1811
1812 EXPORT_SYMBOL(aio_complete);
1813 EXPORT_SYMBOL(aio_put_req);
1814 EXPORT_SYMBOL(wait_on_sync_kiocb);
1815
|
This page was automatically generated by the
LXR engine.
|