blob: 5a28b69ad223b243121c8eb65495f9a519891d33 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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
19#define DEBUG 0
20
21#include <linux/sched.h>
22#include <linux/fs.h>
23#include <linux/file.h>
24#include <linux/mm.h>
25#include <linux/mman.h>
26#include <linux/slab.h>
27#include <linux/timer.h>
28#include <linux/aio.h>
29#include <linux/highmem.h>
30#include <linux/workqueue.h>
31#include <linux/security.h>
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070032#include <linux/rcuref.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070033
34#include <asm/kmap_types.h>
35#include <asm/uaccess.h>
36#include <asm/mmu_context.h>
37
38#if DEBUG > 1
39#define dprintk printk
40#else
41#define dprintk(x...) do { ; } while (0)
42#endif
43
Linus Torvalds1da177e2005-04-16 15:20:36 -070044/*------ sysctl variables----*/
Zach Brownd55b5fd2005-11-07 00:59:31 -080045static DEFINE_SPINLOCK(aio_nr_lock);
46unsigned long aio_nr; /* current system wide number of aio requests */
47unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
Linus Torvalds1da177e2005-04-16 15:20:36 -070048/*----end sysctl variables---*/
49
50static kmem_cache_t *kiocb_cachep;
51static kmem_cache_t *kioctx_cachep;
52
53static struct workqueue_struct *aio_wq;
54
55/* Used for rare fput completion. */
56static void aio_fput_routine(void *);
57static DECLARE_WORK(fput_work, aio_fput_routine, NULL);
58
59static DEFINE_SPINLOCK(fput_lock);
Adrian Bunk25ee7e32005-04-25 08:18:14 -070060static LIST_HEAD(fput_head);
Linus Torvalds1da177e2005-04-16 15:20:36 -070061
62static void aio_kick_handler(void *);
Sébastien Duguc016e222005-06-28 20:44:59 -070063static void aio_queue_work(struct kioctx *);
Linus Torvalds1da177e2005-04-16 15:20:36 -070064
65/* aio_setup
66 * Creates the slab caches used by the aio routines, panic on
67 * failure as this is done early during the boot sequence.
68 */
69static int __init aio_setup(void)
70{
71 kiocb_cachep = kmem_cache_create("kiocb", sizeof(struct kiocb),
72 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
73 kioctx_cachep = kmem_cache_create("kioctx", sizeof(struct kioctx),
74 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
75
76 aio_wq = create_workqueue("aio");
77
78 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
79
80 return 0;
81}
82
83static void aio_free_ring(struct kioctx *ctx)
84{
85 struct aio_ring_info *info = &ctx->ring_info;
86 long i;
87
88 for (i=0; i<info->nr_pages; i++)
89 put_page(info->ring_pages[i]);
90
91 if (info->mmap_size) {
92 down_write(&ctx->mm->mmap_sem);
93 do_munmap(ctx->mm, info->mmap_base, info->mmap_size);
94 up_write(&ctx->mm->mmap_sem);
95 }
96
97 if (info->ring_pages && info->ring_pages != info->internal_pages)
98 kfree(info->ring_pages);
99 info->ring_pages = NULL;
100 info->nr = 0;
101}
102
103static int aio_setup_ring(struct kioctx *ctx)
104{
105 struct aio_ring *ring;
106 struct aio_ring_info *info = &ctx->ring_info;
107 unsigned nr_events = ctx->max_reqs;
108 unsigned long size;
109 int nr_pages;
110
111 /* Compensate for the ring buffer's head/tail overlap entry */
112 nr_events += 2; /* 1 is required, 2 for good luck */
113
114 size = sizeof(struct aio_ring);
115 size += sizeof(struct io_event) * nr_events;
116 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
117
118 if (nr_pages < 0)
119 return -EINVAL;
120
121 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
122
123 info->nr = 0;
124 info->ring_pages = info->internal_pages;
125 if (nr_pages > AIO_RING_PAGES) {
126 info->ring_pages = kmalloc(sizeof(struct page *) * nr_pages, GFP_KERNEL);
127 if (!info->ring_pages)
128 return -ENOMEM;
129 memset(info->ring_pages, 0, sizeof(struct page *) * nr_pages);
130 }
131
132 info->mmap_size = nr_pages * PAGE_SIZE;
133 dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
134 down_write(&ctx->mm->mmap_sem);
135 info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
136 PROT_READ|PROT_WRITE, MAP_ANON|MAP_PRIVATE,
137 0);
138 if (IS_ERR((void *)info->mmap_base)) {
139 up_write(&ctx->mm->mmap_sem);
140 printk("mmap err: %ld\n", -info->mmap_base);
141 info->mmap_size = 0;
142 aio_free_ring(ctx);
143 return -EAGAIN;
144 }
145
146 dprintk("mmap address: 0x%08lx\n", info->mmap_base);
147 info->nr_pages = get_user_pages(current, ctx->mm,
148 info->mmap_base, nr_pages,
149 1, 0, info->ring_pages, NULL);
150 up_write(&ctx->mm->mmap_sem);
151
152 if (unlikely(info->nr_pages != nr_pages)) {
153 aio_free_ring(ctx);
154 return -EAGAIN;
155 }
156
157 ctx->user_id = info->mmap_base;
158
159 info->nr = nr_events; /* trusted copy */
160
161 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
162 ring->nr = nr_events; /* user copy */
163 ring->id = ctx->user_id;
164 ring->head = ring->tail = 0;
165 ring->magic = AIO_RING_MAGIC;
166 ring->compat_features = AIO_RING_COMPAT_FEATURES;
167 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
168 ring->header_length = sizeof(struct aio_ring);
169 kunmap_atomic(ring, KM_USER0);
170
171 return 0;
172}
173
174
175/* aio_ring_event: returns a pointer to the event at the given index from
176 * kmap_atomic(, km). Release the pointer with put_aio_ring_event();
177 */
178#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
179#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
180#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
181
182#define aio_ring_event(info, nr, km) ({ \
183 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
184 struct io_event *__event; \
185 __event = kmap_atomic( \
186 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
187 __event += pos % AIO_EVENTS_PER_PAGE; \
188 __event; \
189})
190
191#define put_aio_ring_event(event, km) do { \
192 struct io_event *__event = (event); \
193 (void)__event; \
194 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
195} while(0)
196
197/* ioctx_alloc
198 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
199 */
200static struct kioctx *ioctx_alloc(unsigned nr_events)
201{
202 struct mm_struct *mm;
203 struct kioctx *ctx;
204
205 /* Prevent overflows */
206 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
207 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
208 pr_debug("ENOMEM: nr_events too high\n");
209 return ERR_PTR(-EINVAL);
210 }
211
Zach Brownd55b5fd2005-11-07 00:59:31 -0800212 if ((unsigned long)nr_events > aio_max_nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213 return ERR_PTR(-EAGAIN);
214
215 ctx = kmem_cache_alloc(kioctx_cachep, GFP_KERNEL);
216 if (!ctx)
217 return ERR_PTR(-ENOMEM);
218
219 memset(ctx, 0, sizeof(*ctx));
220 ctx->max_reqs = nr_events;
221 mm = ctx->mm = current->mm;
222 atomic_inc(&mm->mm_count);
223
224 atomic_set(&ctx->users, 1);
225 spin_lock_init(&ctx->ctx_lock);
226 spin_lock_init(&ctx->ring_info.ring_lock);
227 init_waitqueue_head(&ctx->wait);
228
229 INIT_LIST_HEAD(&ctx->active_reqs);
230 INIT_LIST_HEAD(&ctx->run_list);
231 INIT_WORK(&ctx->wq, aio_kick_handler, ctx);
232
233 if (aio_setup_ring(ctx) < 0)
234 goto out_freectx;
235
236 /* limit the number of system wide aios */
Zach Brownd55b5fd2005-11-07 00:59:31 -0800237 spin_lock(&aio_nr_lock);
238 if (aio_nr + ctx->max_reqs > aio_max_nr ||
239 aio_nr + ctx->max_reqs < aio_nr)
240 ctx->max_reqs = 0;
241 else
242 aio_nr += ctx->max_reqs;
243 spin_unlock(&aio_nr_lock);
244 if (ctx->max_reqs == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700245 goto out_cleanup;
246
247 /* now link into global list. kludge. FIXME */
248 write_lock(&mm->ioctx_list_lock);
249 ctx->next = mm->ioctx_list;
250 mm->ioctx_list = ctx;
251 write_unlock(&mm->ioctx_list_lock);
252
253 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
254 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
255 return ctx;
256
257out_cleanup:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258 __put_ioctx(ctx);
259 return ERR_PTR(-EAGAIN);
260
261out_freectx:
262 mmdrop(mm);
263 kmem_cache_free(kioctx_cachep, ctx);
264 ctx = ERR_PTR(-ENOMEM);
265
266 dprintk("aio: error allocating ioctx %p\n", ctx);
267 return ctx;
268}
269
270/* aio_cancel_all
271 * Cancels all outstanding aio requests on an aio context. Used
272 * when the processes owning a context have all exited to encourage
273 * the rapid destruction of the kioctx.
274 */
275static void aio_cancel_all(struct kioctx *ctx)
276{
277 int (*cancel)(struct kiocb *, struct io_event *);
278 struct io_event res;
279 spin_lock_irq(&ctx->ctx_lock);
280 ctx->dead = 1;
281 while (!list_empty(&ctx->active_reqs)) {
282 struct list_head *pos = ctx->active_reqs.next;
283 struct kiocb *iocb = list_kiocb(pos);
284 list_del_init(&iocb->ki_list);
285 cancel = iocb->ki_cancel;
286 kiocbSetCancelled(iocb);
287 if (cancel) {
288 iocb->ki_users++;
289 spin_unlock_irq(&ctx->ctx_lock);
290 cancel(iocb, &res);
291 spin_lock_irq(&ctx->ctx_lock);
292 }
293 }
294 spin_unlock_irq(&ctx->ctx_lock);
295}
296
Adrian Bunk25ee7e32005-04-25 08:18:14 -0700297static void wait_for_all_aios(struct kioctx *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298{
299 struct task_struct *tsk = current;
300 DECLARE_WAITQUEUE(wait, tsk);
301
302 if (!ctx->reqs_active)
303 return;
304
305 add_wait_queue(&ctx->wait, &wait);
306 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
307 while (ctx->reqs_active) {
308 schedule();
309 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
310 }
311 __set_task_state(tsk, TASK_RUNNING);
312 remove_wait_queue(&ctx->wait, &wait);
313}
314
315/* wait_on_sync_kiocb:
316 * Waits on the given sync kiocb to complete.
317 */
318ssize_t fastcall wait_on_sync_kiocb(struct kiocb *iocb)
319{
320 while (iocb->ki_users) {
321 set_current_state(TASK_UNINTERRUPTIBLE);
322 if (!iocb->ki_users)
323 break;
324 schedule();
325 }
326 __set_current_state(TASK_RUNNING);
327 return iocb->ki_user_data;
328}
329
330/* exit_aio: called when the last user of mm goes away. At this point,
331 * there is no way for any new requests to be submited or any of the
332 * io_* syscalls to be called on the context. However, there may be
333 * outstanding requests which hold references to the context; as they
334 * go away, they will call put_ioctx and release any pinned memory
335 * associated with the request (held via struct page * references).
336 */
337void fastcall exit_aio(struct mm_struct *mm)
338{
339 struct kioctx *ctx = mm->ioctx_list;
340 mm->ioctx_list = NULL;
341 while (ctx) {
342 struct kioctx *next = ctx->next;
343 ctx->next = NULL;
344 aio_cancel_all(ctx);
345
346 wait_for_all_aios(ctx);
347 /*
348 * this is an overkill, but ensures we don't leave
349 * the ctx on the aio_wq
350 */
351 flush_workqueue(aio_wq);
352
353 if (1 != atomic_read(&ctx->users))
354 printk(KERN_DEBUG
355 "exit_aio:ioctx still alive: %d %d %d\n",
356 atomic_read(&ctx->users), ctx->dead,
357 ctx->reqs_active);
358 put_ioctx(ctx);
359 ctx = next;
360 }
361}
362
363/* __put_ioctx
364 * Called when the last user of an aio context has gone away,
365 * and the struct needs to be freed.
366 */
367void fastcall __put_ioctx(struct kioctx *ctx)
368{
369 unsigned nr_events = ctx->max_reqs;
370
371 if (unlikely(ctx->reqs_active))
372 BUG();
373
374 cancel_delayed_work(&ctx->wq);
375 flush_workqueue(aio_wq);
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
Zach Brownd55b5fd2005-11-07 00:59:31 -0800382 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 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388}
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 */
400static struct kiocb *FASTCALL(__aio_get_req(struct kioctx *ctx));
401static struct kiocb fastcall *__aio_get_req(struct kioctx *ctx)
402{
403 struct kiocb *req = NULL;
404 struct aio_ring *ring;
405 int okay = 0;
406
407 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
408 if (unlikely(!req))
409 return NULL;
410
Zach Brown4faa5282005-10-17 16:43:33 -0700411 req->ki_flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412 req->ki_users = 2;
413 req->ki_key = 0;
414 req->ki_ctx = ctx;
415 req->ki_cancel = NULL;
416 req->ki_retry = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417 req->ki_dtor = NULL;
418 req->private = NULL;
419 INIT_LIST_HEAD(&req->ki_run_list);
420
421 /* Check if the completion queue has enough free space to
422 * accept an event from this io.
423 */
424 spin_lock_irq(&ctx->ctx_lock);
425 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
426 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
427 list_add(&req->ki_list, &ctx->active_reqs);
428 get_ioctx(ctx);
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
443static 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
458static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
459{
Zach Brownd00689a2005-11-13 16:07:34 -0800460 assert_spin_locked(&ctx->ctx_lock);
461
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 if (req->ki_dtor)
463 req->ki_dtor(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464 kmem_cache_free(kiocb_cachep, req);
465 ctx->reqs_active--;
466
467 if (unlikely(!ctx->reqs_active && ctx->dead))
468 wake_up(&ctx->wait);
469}
470
471static void aio_fput_routine(void *data)
472{
473 spin_lock_irq(&fput_lock);
474 while (likely(!list_empty(&fput_head))) {
475 struct kiocb *req = list_kiocb(fput_head.next);
476 struct kioctx *ctx = req->ki_ctx;
477
478 list_del(&req->ki_list);
479 spin_unlock_irq(&fput_lock);
480
481 /* Complete the fput */
482 __fput(req->ki_filp);
483
484 /* Link the iocb into the context's free list */
485 spin_lock_irq(&ctx->ctx_lock);
486 really_put_req(ctx, req);
487 spin_unlock_irq(&ctx->ctx_lock);
488
489 put_ioctx(ctx);
490 spin_lock_irq(&fput_lock);
491 }
492 spin_unlock_irq(&fput_lock);
493}
494
495/* __aio_put_req
496 * Returns true if this put was the last user of the request.
497 */
498static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
499{
500 dprintk(KERN_DEBUG "aio_put(%p): f_count=%d\n",
501 req, atomic_read(&req->ki_filp->f_count));
502
Zach Brownd00689a2005-11-13 16:07:34 -0800503 assert_spin_locked(&ctx->ctx_lock);
504
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505 req->ki_users --;
506 if (unlikely(req->ki_users < 0))
507 BUG();
508 if (likely(req->ki_users))
509 return 0;
510 list_del(&req->ki_list); /* remove from active_reqs */
511 req->ki_cancel = NULL;
512 req->ki_retry = NULL;
513
514 /* Must be done under the lock to serialise against cancellation.
515 * Call this aio_fput as it duplicates fput via the fput_work.
516 */
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700517 if (unlikely(rcuref_dec_and_test(&req->ki_filp->f_count))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518 get_ioctx(ctx);
519 spin_lock(&fput_lock);
520 list_add(&req->ki_list, &fput_head);
521 spin_unlock(&fput_lock);
522 queue_work(aio_wq, &fput_work);
523 } else
524 really_put_req(ctx, req);
525 return 1;
526}
527
528/* aio_put_req
529 * Returns true if this put was the last user of the kiocb,
530 * false if the request is still in use.
531 */
532int fastcall aio_put_req(struct kiocb *req)
533{
534 struct kioctx *ctx = req->ki_ctx;
535 int ret;
536 spin_lock_irq(&ctx->ctx_lock);
537 ret = __aio_put_req(ctx, req);
538 spin_unlock_irq(&ctx->ctx_lock);
539 if (ret)
540 put_ioctx(ctx);
541 return ret;
542}
543
544/* Lookup an ioctx id. ioctx_list is lockless for reads.
545 * FIXME: this is O(n) and is only suitable for development.
546 */
547struct kioctx *lookup_ioctx(unsigned long ctx_id)
548{
549 struct kioctx *ioctx;
550 struct mm_struct *mm;
551
552 mm = current->mm;
553 read_lock(&mm->ioctx_list_lock);
554 for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next)
555 if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) {
556 get_ioctx(ioctx);
557 break;
558 }
559 read_unlock(&mm->ioctx_list_lock);
560
561 return ioctx;
562}
563
564/*
565 * use_mm
566 * Makes the calling kernel thread take on the specified
567 * mm context.
568 * Called by the retry thread execute retries within the
569 * iocb issuer's mm context, so that copy_from/to_user
570 * operations work seamlessly for aio.
571 * (Note: this routine is intended to be called only
572 * from a kernel thread context)
573 */
574static void use_mm(struct mm_struct *mm)
575{
576 struct mm_struct *active_mm;
577 struct task_struct *tsk = current;
578
579 task_lock(tsk);
580 tsk->flags |= PF_BORROWED_MM;
581 active_mm = tsk->active_mm;
582 atomic_inc(&mm->mm_count);
583 tsk->mm = mm;
584 tsk->active_mm = mm;
Paolo 'Blaisorblade' Giarrusso1e40cd32005-09-03 15:57:25 -0700585 /*
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 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589 activate_mm(active_mm, mm);
590 task_unlock(tsk);
591
592 mmdrop(active_mm);
593}
594
595/*
596 * unuse_mm
597 * Reverses the effect of use_mm, i.e. releases the
598 * specified mm context which was earlier taken on
599 * by the calling kernel thread
600 * (Note: this routine is intended to be called only
601 * from a kernel thread context)
602 *
603 * Comments: Called with ctx->ctx_lock held. This nests
604 * task_lock instead ctx_lock.
605 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -0700606static void unuse_mm(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607{
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.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626 */
627static inline int __queue_kicked_iocb(struct kiocb *iocb)
628{
629 struct kioctx *ctx = iocb->ki_ctx;
630
Zach Brownd00689a2005-11-13 16:07:34 -0800631 assert_spin_locked(&ctx->ctx_lock);
632
Linus Torvalds1da177e2005-04-16 15:20:36 -0700633 if (list_empty(&iocb->ki_run_list)) {
634 list_add_tail(&iocb->ki_run_list,
635 &ctx->run_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636 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 reaquired
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 */
663static 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 (iocb->ki_retried++ > 1024*1024) {
670 printk("Maximal retry count. Bytes done %Zd\n",
671 iocb->ki_nbytes - iocb->ki_left);
672 return -EAGAIN;
673 }
674
675 if (!(iocb->ki_retried & 0xff)) {
Ken Chen644d3a02005-05-01 08:59:15 -0700676 pr_debug("%ld retry: %d of %d\n", iocb->ki_retried,
677 iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678 }
679
680 if (!(retry = iocb->ki_retry)) {
681 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
682 return 0;
683 }
684
685 /*
686 * We don't want the next retry iteration for this
687 * operation to start until this one has returned and
688 * updated the iocb state. However, wait_queue functions
689 * can trigger a kick_iocb from interrupt context in the
690 * meantime, indicating that data is available for the next
691 * iteration. We want to remember that and enable the
692 * next retry iteration _after_ we are through with
693 * this one.
694 *
695 * So, in order to be able to register a "kick", but
696 * prevent it from being queued now, we clear the kick
697 * flag, but make the kick code *think* that the iocb is
698 * still on the run list until we are actually done.
699 * When we are done with this iteration, we check if
700 * the iocb was kicked in the meantime and if so, queue
701 * it up afresh.
702 */
703
704 kiocbClearKicked(iocb);
705
706 /*
707 * This is so that aio_complete knows it doesn't need to
708 * pull the iocb off the run list (We can't just call
709 * INIT_LIST_HEAD because we don't want a kick_iocb to
710 * queue this on the run list yet)
711 */
712 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
713 spin_unlock_irq(&ctx->ctx_lock);
714
715 /* Quit retrying if the i/o has been cancelled */
716 if (kiocbIsCancelled(iocb)) {
717 ret = -EINTR;
718 aio_complete(iocb, ret, 0);
719 /* must not access the iocb after this */
720 goto out;
721 }
722
723 /*
724 * Now we are all set to call the retry method in async
725 * context. By setting this thread's io_wait context
726 * to point to the wait queue entry inside the currently
727 * running iocb for the duration of the retry, we ensure
728 * that async notification wakeups are queued by the
729 * operation instead of blocking waits, and when notified,
730 * cause the iocb to be kicked for continuation (through
731 * the aio_wake_function callback).
732 */
733 BUG_ON(current->io_wait != NULL);
734 current->io_wait = &iocb->ki_wait;
735 ret = retry(iocb);
736 current->io_wait = NULL;
737
Zach Brown897f15f2005-09-30 11:58:55 -0700738 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
739 BUG_ON(!list_empty(&iocb->ki_wait.task_list));
740 aio_complete(iocb, ret, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741 }
742out:
743 spin_lock_irq(&ctx->ctx_lock);
744
745 if (-EIOCBRETRY == ret) {
746 /*
747 * OK, now that we are done with this iteration
748 * and know that there is more left to go,
749 * this is where we let go so that a subsequent
750 * "kick" can start the next iteration
751 */
752
753 /* will make __queue_kicked_iocb succeed from here on */
754 INIT_LIST_HEAD(&iocb->ki_run_list);
755 /* we must queue the next iteration ourselves, if it
756 * has already been kicked */
757 if (kiocbIsKicked(iocb)) {
758 __queue_kicked_iocb(iocb);
Sébastien Duguc016e222005-06-28 20:44:59 -0700759
760 /*
761 * __queue_kicked_iocb will always return 1 here, because
762 * iocb->ki_run_list is empty at this point so it should
763 * be safe to unconditionally queue the context into the
764 * work queue.
765 */
766 aio_queue_work(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767 }
768 }
769 return ret;
770}
771
772/*
773 * __aio_run_iocbs:
774 * Process all pending retries queued on the ioctx
775 * run list.
776 * Assumes it is operating within the aio issuer's mm
Zach Brownd00689a2005-11-13 16:07:34 -0800777 * context.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778 */
779static int __aio_run_iocbs(struct kioctx *ctx)
780{
781 struct kiocb *iocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782 LIST_HEAD(run_list);
783
Zach Brownd00689a2005-11-13 16:07:34 -0800784 assert_spin_locked(&ctx->ctx_lock);
785
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786 list_splice_init(&ctx->run_list, &run_list);
787 while (!list_empty(&run_list)) {
788 iocb = list_entry(run_list.next, struct kiocb,
789 ki_run_list);
790 list_del(&iocb->ki_run_list);
791 /*
792 * Hold an extra reference while retrying i/o.
793 */
794 iocb->ki_users++; /* grab extra reference */
795 aio_run_iocb(iocb);
796 if (__aio_put_req(ctx, iocb)) /* drop extra ref */
797 put_ioctx(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799 if (!list_empty(&ctx->run_list))
800 return 1;
801 return 0;
802}
803
804static void aio_queue_work(struct kioctx * ctx)
805{
806 unsigned long timeout;
807 /*
808 * if someone is waiting, get the work started right
809 * away, otherwise, use a longer delay
810 */
811 smp_mb();
812 if (waitqueue_active(&ctx->wait))
813 timeout = 1;
814 else
815 timeout = HZ/10;
816 queue_delayed_work(aio_wq, &ctx->wq, timeout);
817}
818
819
820/*
821 * aio_run_iocbs:
822 * Process all pending retries queued on the ioctx
823 * run list.
824 * Assumes it is operating within the aio issuer's mm
825 * context.
826 */
827static inline void aio_run_iocbs(struct kioctx *ctx)
828{
829 int requeue;
830
831 spin_lock_irq(&ctx->ctx_lock);
832
833 requeue = __aio_run_iocbs(ctx);
834 spin_unlock_irq(&ctx->ctx_lock);
835 if (requeue)
836 aio_queue_work(ctx);
837}
838
839/*
840 * just like aio_run_iocbs, but keeps running them until
841 * the list stays empty
842 */
843static inline void aio_run_all_iocbs(struct kioctx *ctx)
844{
845 spin_lock_irq(&ctx->ctx_lock);
846 while (__aio_run_iocbs(ctx))
847 ;
848 spin_unlock_irq(&ctx->ctx_lock);
849}
850
851/*
852 * aio_kick_handler:
853 * Work queue handler triggered to process pending
854 * retries on an ioctx. Takes on the aio issuer's
855 * mm context before running the iocbs, so that
856 * copy_xxx_user operates on the issuer's address
857 * space.
858 * Run on aiod's context.
859 */
860static void aio_kick_handler(void *data)
861{
862 struct kioctx *ctx = data;
863 mm_segment_t oldfs = get_fs();
864 int requeue;
865
866 set_fs(USER_DS);
867 use_mm(ctx->mm);
868 spin_lock_irq(&ctx->ctx_lock);
869 requeue =__aio_run_iocbs(ctx);
870 unuse_mm(ctx->mm);
871 spin_unlock_irq(&ctx->ctx_lock);
872 set_fs(oldfs);
873 /*
874 * we're in a worker thread already, don't use queue_delayed_work,
875 */
876 if (requeue)
877 queue_work(aio_wq, &ctx->wq);
878}
879
880
881/*
882 * Called by kick_iocb to queue the kiocb for retry
883 * and if required activate the aio work queue to process
884 * it
885 */
Zach Brown998765e2005-09-30 11:58:54 -0700886static void try_queue_kicked_iocb(struct kiocb *iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700887{
888 struct kioctx *ctx = iocb->ki_ctx;
889 unsigned long flags;
890 int run = 0;
891
Zach Brown998765e2005-09-30 11:58:54 -0700892 /* We're supposed to be the only path putting the iocb back on the run
893 * list. If we find that the iocb is *back* on a wait queue already
894 * than retry has happened before we could queue the iocb. This also
895 * means that the retry could have completed and freed our iocb, no
896 * good. */
897 BUG_ON((!list_empty(&iocb->ki_wait.task_list)));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898
899 spin_lock_irqsave(&ctx->ctx_lock, flags);
Zach Brown998765e2005-09-30 11:58:54 -0700900 /* set this inside the lock so that we can't race with aio_run_iocb()
901 * testing it and putting the iocb on the run list under the lock */
902 if (!kiocbTryKick(iocb))
903 run = __queue_kicked_iocb(iocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
Ken Chen644d3a02005-05-01 08:59:15 -0700905 if (run)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906 aio_queue_work(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700907}
908
909/*
910 * kick_iocb:
911 * Called typically from a wait queue callback context
912 * (aio_wake_function) to trigger a retry of the iocb.
913 * The retry is usually executed by aio workqueue
914 * threads (See aio_kick_handler).
915 */
916void fastcall kick_iocb(struct kiocb *iocb)
917{
918 /* sync iocbs are easy: they can only ever be executing from a
919 * single context. */
920 if (is_sync_kiocb(iocb)) {
921 kiocbSetKicked(iocb);
922 wake_up_process(iocb->ki_obj.tsk);
923 return;
924 }
925
Zach Brown998765e2005-09-30 11:58:54 -0700926 try_queue_kicked_iocb(iocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927}
928EXPORT_SYMBOL(kick_iocb);
929
930/* aio_complete
931 * Called when the io request on the given iocb is complete.
932 * Returns true if this is the last user of the request. The
933 * only other user of the request can be the cancellation code.
934 */
935int fastcall aio_complete(struct kiocb *iocb, long res, long res2)
936{
937 struct kioctx *ctx = iocb->ki_ctx;
938 struct aio_ring_info *info;
939 struct aio_ring *ring;
940 struct io_event *event;
941 unsigned long flags;
942 unsigned long tail;
943 int ret;
944
Zach Brown20dcae32005-11-13 16:07:33 -0800945 /*
946 * Special case handling for sync iocbs:
947 * - events go directly into the iocb for fast handling
948 * - the sync task with the iocb in its stack holds the single iocb
949 * ref, no other paths have a way to get another ref
950 * - the sync task helpfully left a reference to itself in the iocb
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951 */
952 if (is_sync_kiocb(iocb)) {
Zach Brown20dcae32005-11-13 16:07:33 -0800953 BUG_ON(iocb->ki_users != 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954 iocb->ki_user_data = res;
Zach Brown20dcae32005-11-13 16:07:33 -0800955 iocb->ki_users = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 wake_up_process(iocb->ki_obj.tsk);
Zach Brown20dcae32005-11-13 16:07:33 -0800957 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 }
959
960 info = &ctx->ring_info;
961
962 /* add a completion event to the ring buffer.
963 * must be done holding ctx->ctx_lock to prevent
964 * other code from messing with the tail
965 * pointer since we might be called from irq
966 * context.
967 */
968 spin_lock_irqsave(&ctx->ctx_lock, flags);
969
970 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
971 list_del_init(&iocb->ki_run_list);
972
973 /*
974 * cancelled requests don't get events, userland was given one
975 * when the event got cancelled.
976 */
977 if (kiocbIsCancelled(iocb))
978 goto put_rq;
979
980 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
981
982 tail = info->tail;
983 event = aio_ring_event(info, tail, KM_IRQ0);
Ken Chen4bf69b22005-05-01 08:59:15 -0700984 if (++tail >= info->nr)
985 tail = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986
987 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
988 event->data = iocb->ki_user_data;
989 event->res = res;
990 event->res2 = res2;
991
992 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
993 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
994 res, res2);
995
996 /* after flagging the request as done, we
997 * must never even look at it again
998 */
999 smp_wmb(); /* make event visible before updating tail */
1000
1001 info->tail = tail;
1002 ring->tail = tail;
1003
1004 put_aio_ring_event(event, KM_IRQ0);
1005 kunmap_atomic(ring, KM_IRQ1);
1006
1007 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
1008
Ken Chen644d3a02005-05-01 08:59:15 -07001009 pr_debug("%ld retries: %d of %d\n", iocb->ki_retried,
1010 iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001011put_rq:
1012 /* everything turned out well, dispose of the aiocb. */
1013 ret = __aio_put_req(ctx, iocb);
1014
1015 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1016
1017 if (waitqueue_active(&ctx->wait))
1018 wake_up(&ctx->wait);
1019
1020 if (ret)
1021 put_ioctx(ctx);
1022
1023 return ret;
1024}
1025
1026/* aio_read_evt
1027 * Pull an event off of the ioctx's event ring. Returns the number of
1028 * events fetched (0 or 1 ;-)
1029 * FIXME: make this use cmpxchg.
1030 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1031 */
1032static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1033{
1034 struct aio_ring_info *info = &ioctx->ring_info;
1035 struct aio_ring *ring;
1036 unsigned long head;
1037 int ret = 0;
1038
1039 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1040 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1041 (unsigned long)ring->head, (unsigned long)ring->tail,
1042 (unsigned long)ring->nr);
1043
1044 if (ring->head == ring->tail)
1045 goto out;
1046
1047 spin_lock(&info->ring_lock);
1048
1049 head = ring->head % info->nr;
1050 if (head != ring->tail) {
1051 struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1052 *ent = *evp;
1053 head = (head + 1) % info->nr;
1054 smp_mb(); /* finish reading the event before updatng the head */
1055 ring->head = head;
1056 ret = 1;
1057 put_aio_ring_event(evp, KM_USER1);
1058 }
1059 spin_unlock(&info->ring_lock);
1060
1061out:
1062 kunmap_atomic(ring, KM_USER0);
1063 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1064 (unsigned long)ring->head, (unsigned long)ring->tail);
1065 return ret;
1066}
1067
1068struct aio_timeout {
1069 struct timer_list timer;
1070 int timed_out;
1071 struct task_struct *p;
1072};
1073
1074static void timeout_func(unsigned long data)
1075{
1076 struct aio_timeout *to = (struct aio_timeout *)data;
1077
1078 to->timed_out = 1;
1079 wake_up_process(to->p);
1080}
1081
1082static inline void init_timeout(struct aio_timeout *to)
1083{
1084 init_timer(&to->timer);
1085 to->timer.data = (unsigned long)to;
1086 to->timer.function = timeout_func;
1087 to->timed_out = 0;
1088 to->p = current;
1089}
1090
1091static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1092 const struct timespec *ts)
1093{
1094 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1095 if (time_after(to->timer.expires, jiffies))
1096 add_timer(&to->timer);
1097 else
1098 to->timed_out = 1;
1099}
1100
1101static inline void clear_timeout(struct aio_timeout *to)
1102{
1103 del_singleshot_timer_sync(&to->timer);
1104}
1105
1106static int read_events(struct kioctx *ctx,
1107 long min_nr, long nr,
1108 struct io_event __user *event,
1109 struct timespec __user *timeout)
1110{
1111 long start_jiffies = jiffies;
1112 struct task_struct *tsk = current;
1113 DECLARE_WAITQUEUE(wait, tsk);
1114 int ret;
1115 int i = 0;
1116 struct io_event ent;
1117 struct aio_timeout to;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118 int retry = 0;
1119
1120 /* needed to zero any padding within an entry (there shouldn't be
1121 * any, but C is fun!
1122 */
1123 memset(&ent, 0, sizeof(ent));
1124retry:
1125 ret = 0;
1126 while (likely(i < nr)) {
1127 ret = aio_read_evt(ctx, &ent);
1128 if (unlikely(ret <= 0))
1129 break;
1130
1131 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1132 ent.data, ent.obj, ent.res, ent.res2);
1133
1134 /* Could we split the check in two? */
1135 ret = -EFAULT;
1136 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1137 dprintk("aio: lost an event due to EFAULT.\n");
1138 break;
1139 }
1140 ret = 0;
1141
1142 /* Good, event copied to userland, update counts. */
1143 event ++;
1144 i ++;
1145 }
1146
1147 if (min_nr <= i)
1148 return i;
1149 if (ret)
1150 return ret;
1151
1152 /* End fast path */
1153
1154 /* racey check, but it gets redone */
1155 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1156 retry = 1;
1157 aio_run_all_iocbs(ctx);
1158 goto retry;
1159 }
1160
1161 init_timeout(&to);
1162 if (timeout) {
1163 struct timespec ts;
1164 ret = -EFAULT;
1165 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1166 goto out;
1167
1168 set_timeout(start_jiffies, &to, &ts);
1169 }
1170
1171 while (likely(i < nr)) {
1172 add_wait_queue_exclusive(&ctx->wait, &wait);
1173 do {
1174 set_task_state(tsk, TASK_INTERRUPTIBLE);
1175 ret = aio_read_evt(ctx, &ent);
1176 if (ret)
1177 break;
1178 if (min_nr <= i)
1179 break;
1180 ret = 0;
1181 if (to.timed_out) /* Only check after read evt */
1182 break;
1183 schedule();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184 if (signal_pending(tsk)) {
1185 ret = -EINTR;
1186 break;
1187 }
1188 /*ret = aio_read_evt(ctx, &ent);*/
1189 } while (1) ;
1190
1191 set_task_state(tsk, TASK_RUNNING);
1192 remove_wait_queue(&ctx->wait, &wait);
1193
1194 if (unlikely(ret <= 0))
1195 break;
1196
1197 ret = -EFAULT;
1198 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1199 dprintk("aio: lost an event due to EFAULT.\n");
1200 break;
1201 }
1202
1203 /* Good, event copied to userland, update counts. */
1204 event ++;
1205 i ++;
1206 }
1207
1208 if (timeout)
1209 clear_timeout(&to);
1210out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211 return i ? i : ret;
1212}
1213
1214/* Take an ioctx and remove it from the list of ioctx's. Protects
1215 * against races with itself via ->dead.
1216 */
1217static void io_destroy(struct kioctx *ioctx)
1218{
1219 struct mm_struct *mm = current->mm;
1220 struct kioctx **tmp;
1221 int was_dead;
1222
1223 /* delete the entry from the list is someone else hasn't already */
1224 write_lock(&mm->ioctx_list_lock);
1225 was_dead = ioctx->dead;
1226 ioctx->dead = 1;
1227 for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx;
1228 tmp = &(*tmp)->next)
1229 ;
1230 if (*tmp)
1231 *tmp = ioctx->next;
1232 write_unlock(&mm->ioctx_list_lock);
1233
1234 dprintk("aio_release(%p)\n", ioctx);
1235 if (likely(!was_dead))
1236 put_ioctx(ioctx); /* twice for the list */
1237
1238 aio_cancel_all(ioctx);
1239 wait_for_all_aios(ioctx);
1240 put_ioctx(ioctx); /* once for the lookup */
1241}
1242
1243/* sys_io_setup:
1244 * Create an aio_context capable of receiving at least nr_events.
1245 * ctxp must not point to an aio_context that already exists, and
1246 * must be initialized to 0 prior to the call. On successful
1247 * creation of the aio_context, *ctxp is filled in with the resulting
1248 * handle. May fail with -EINVAL if *ctxp is not initialized,
1249 * if the specified nr_events exceeds internal limits. May fail
1250 * with -EAGAIN if the specified nr_events exceeds the user's limit
1251 * of available events. May fail with -ENOMEM if insufficient kernel
1252 * resources are available. May fail with -EFAULT if an invalid
1253 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1254 * implemented.
1255 */
1256asmlinkage long sys_io_setup(unsigned nr_events, aio_context_t __user *ctxp)
1257{
1258 struct kioctx *ioctx = NULL;
1259 unsigned long ctx;
1260 long ret;
1261
1262 ret = get_user(ctx, ctxp);
1263 if (unlikely(ret))
1264 goto out;
1265
1266 ret = -EINVAL;
Zach Brownd55b5fd2005-11-07 00:59:31 -08001267 if (unlikely(ctx || nr_events == 0)) {
1268 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1269 ctx, nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270 goto out;
1271 }
1272
1273 ioctx = ioctx_alloc(nr_events);
1274 ret = PTR_ERR(ioctx);
1275 if (!IS_ERR(ioctx)) {
1276 ret = put_user(ioctx->user_id, ctxp);
1277 if (!ret)
1278 return 0;
1279
1280 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1281 io_destroy(ioctx);
1282 }
1283
1284out:
1285 return ret;
1286}
1287
1288/* sys_io_destroy:
1289 * Destroy the aio_context specified. May cancel any outstanding
1290 * AIOs and block on completion. Will fail with -ENOSYS if not
1291 * implemented. May fail with -EFAULT if the context pointed to
1292 * is invalid.
1293 */
1294asmlinkage long sys_io_destroy(aio_context_t ctx)
1295{
1296 struct kioctx *ioctx = lookup_ioctx(ctx);
1297 if (likely(NULL != ioctx)) {
1298 io_destroy(ioctx);
1299 return 0;
1300 }
1301 pr_debug("EINVAL: io_destroy: invalid context id\n");
1302 return -EINVAL;
1303}
1304
1305/*
Zach Brown897f15f2005-09-30 11:58:55 -07001306 * aio_p{read,write} are the default ki_retry methods for
1307 * IO_CMD_P{READ,WRITE}. They maintains kiocb retry state around potentially
1308 * multiple calls to f_op->aio_read(). They loop around partial progress
1309 * instead of returning -EIOCBRETRY because they don't have the means to call
1310 * kick_iocb().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001311 */
1312static ssize_t aio_pread(struct kiocb *iocb)
1313{
1314 struct file *file = iocb->ki_filp;
1315 struct address_space *mapping = file->f_mapping;
1316 struct inode *inode = mapping->host;
1317 ssize_t ret = 0;
1318
Zach Brown897f15f2005-09-30 11:58:55 -07001319 do {
1320 ret = file->f_op->aio_read(iocb, iocb->ki_buf,
1321 iocb->ki_left, iocb->ki_pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322 /*
Zach Brown897f15f2005-09-30 11:58:55 -07001323 * Can't just depend on iocb->ki_left to determine
1324 * whether we are done. This may have been a short read.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001325 */
Zach Brown897f15f2005-09-30 11:58:55 -07001326 if (ret > 0) {
1327 iocb->ki_buf += ret;
1328 iocb->ki_left -= ret;
1329 }
1330
1331 /*
1332 * For pipes and sockets we return once we have some data; for
1333 * regular files we retry till we complete the entire read or
1334 * find that we can't read any more data (e.g short reads).
1335 */
Zach Brown353fb072005-09-30 11:58:56 -07001336 } while (ret > 0 && iocb->ki_left > 0 &&
Zach Brown897f15f2005-09-30 11:58:55 -07001337 !S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338
1339 /* This means we must have transferred all that we could */
1340 /* No need to retry anymore */
1341 if ((ret == 0) || (iocb->ki_left == 0))
1342 ret = iocb->ki_nbytes - iocb->ki_left;
1343
1344 return ret;
1345}
1346
Zach Brown897f15f2005-09-30 11:58:55 -07001347/* see aio_pread() */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348static ssize_t aio_pwrite(struct kiocb *iocb)
1349{
1350 struct file *file = iocb->ki_filp;
1351 ssize_t ret = 0;
1352
Zach Brown897f15f2005-09-30 11:58:55 -07001353 do {
1354 ret = file->f_op->aio_write(iocb, iocb->ki_buf,
1355 iocb->ki_left, iocb->ki_pos);
1356 if (ret > 0) {
1357 iocb->ki_buf += ret;
1358 iocb->ki_left -= ret;
1359 }
Zach Brown353fb072005-09-30 11:58:56 -07001360 } while (ret > 0 && iocb->ki_left > 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362 if ((ret == 0) || (iocb->ki_left == 0))
1363 ret = iocb->ki_nbytes - iocb->ki_left;
1364
1365 return ret;
1366}
1367
1368static ssize_t aio_fdsync(struct kiocb *iocb)
1369{
1370 struct file *file = iocb->ki_filp;
1371 ssize_t ret = -EINVAL;
1372
1373 if (file->f_op->aio_fsync)
1374 ret = file->f_op->aio_fsync(iocb, 1);
1375 return ret;
1376}
1377
1378static ssize_t aio_fsync(struct kiocb *iocb)
1379{
1380 struct file *file = iocb->ki_filp;
1381 ssize_t ret = -EINVAL;
1382
1383 if (file->f_op->aio_fsync)
1384 ret = file->f_op->aio_fsync(iocb, 0);
1385 return ret;
1386}
1387
1388/*
1389 * aio_setup_iocb:
1390 * Performs the initial checks and aio retry method
1391 * setup for the kiocb at the time of io submission.
1392 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001393static ssize_t aio_setup_iocb(struct kiocb *kiocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394{
1395 struct file *file = kiocb->ki_filp;
1396 ssize_t ret = 0;
1397
1398 switch (kiocb->ki_opcode) {
1399 case IOCB_CMD_PREAD:
1400 ret = -EBADF;
1401 if (unlikely(!(file->f_mode & FMODE_READ)))
1402 break;
1403 ret = -EFAULT;
1404 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1405 kiocb->ki_left)))
1406 break;
Kostik Belousov8766ce42005-10-23 12:57:13 -07001407 ret = security_file_permission(file, MAY_READ);
1408 if (unlikely(ret))
1409 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410 ret = -EINVAL;
1411 if (file->f_op->aio_read)
1412 kiocb->ki_retry = aio_pread;
1413 break;
1414 case IOCB_CMD_PWRITE:
1415 ret = -EBADF;
1416 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1417 break;
1418 ret = -EFAULT;
1419 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1420 kiocb->ki_left)))
1421 break;
Kostik Belousov8766ce42005-10-23 12:57:13 -07001422 ret = security_file_permission(file, MAY_WRITE);
1423 if (unlikely(ret))
1424 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001425 ret = -EINVAL;
1426 if (file->f_op->aio_write)
1427 kiocb->ki_retry = aio_pwrite;
1428 break;
1429 case IOCB_CMD_FDSYNC:
1430 ret = -EINVAL;
1431 if (file->f_op->aio_fsync)
1432 kiocb->ki_retry = aio_fdsync;
1433 break;
1434 case IOCB_CMD_FSYNC:
1435 ret = -EINVAL;
1436 if (file->f_op->aio_fsync)
1437 kiocb->ki_retry = aio_fsync;
1438 break;
1439 default:
1440 dprintk("EINVAL: io_submit: no operation provided\n");
1441 ret = -EINVAL;
1442 }
1443
1444 if (!kiocb->ki_retry)
1445 return ret;
1446
1447 return 0;
1448}
1449
1450/*
1451 * aio_wake_function:
1452 * wait queue callback function for aio notification,
1453 * Simply triggers a retry of the operation via kick_iocb.
1454 *
1455 * This callback is specified in the wait queue entry in
1456 * a kiocb (current->io_wait points to this wait queue
1457 * entry when an aio operation executes; it is used
1458 * instead of a synchronous wait when an i/o blocking
1459 * condition is encountered during aio).
1460 *
1461 * Note:
1462 * This routine is executed with the wait queue lock held.
1463 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests
1464 * the ioctx lock inside the wait queue lock. This is safe
1465 * because this callback isn't used for wait queues which
1466 * are nested inside ioctx lock (i.e. ctx->wait)
1467 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001468static int aio_wake_function(wait_queue_t *wait, unsigned mode,
1469 int sync, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470{
1471 struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait);
1472
1473 list_del_init(&wait->task_list);
1474 kick_iocb(iocb);
1475 return 1;
1476}
1477
1478int fastcall io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1479 struct iocb *iocb)
1480{
1481 struct kiocb *req;
1482 struct file *file;
1483 ssize_t ret;
1484
1485 /* enforce forwards compatibility on users */
1486 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2 ||
1487 iocb->aio_reserved3)) {
1488 pr_debug("EINVAL: io_submit: reserve field set\n");
1489 return -EINVAL;
1490 }
1491
1492 /* prevent overflows */
1493 if (unlikely(
1494 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1495 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1496 ((ssize_t)iocb->aio_nbytes < 0)
1497 )) {
1498 pr_debug("EINVAL: io_submit: overflow check\n");
1499 return -EINVAL;
1500 }
1501
1502 file = fget(iocb->aio_fildes);
1503 if (unlikely(!file))
1504 return -EBADF;
1505
1506 req = aio_get_req(ctx); /* returns with 2 references to req */
1507 if (unlikely(!req)) {
1508 fput(file);
1509 return -EAGAIN;
1510 }
1511
1512 req->ki_filp = file;
Ken Chen212079c2005-05-01 08:59:15 -07001513 ret = put_user(req->ki_key, &user_iocb->aio_key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001514 if (unlikely(ret)) {
1515 dprintk("EFAULT: aio_key\n");
1516 goto out_put_req;
1517 }
1518
1519 req->ki_obj.user = user_iocb;
1520 req->ki_user_data = iocb->aio_data;
1521 req->ki_pos = iocb->aio_offset;
1522
1523 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1524 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1525 req->ki_opcode = iocb->aio_lio_opcode;
1526 init_waitqueue_func_entry(&req->ki_wait, aio_wake_function);
1527 INIT_LIST_HEAD(&req->ki_wait.task_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528 req->ki_retried = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001529
1530 ret = aio_setup_iocb(req);
1531
1532 if (ret)
1533 goto out_put_req;
1534
1535 spin_lock_irq(&ctx->ctx_lock);
Benjamin LaHaiseac0b1bc2005-09-09 13:02:09 -07001536 aio_run_iocb(req);
Benjamin LaHaiseac0b1bc2005-09-09 13:02:09 -07001537 if (!list_empty(&ctx->run_list)) {
Ken Chen954d3e92005-05-01 08:59:16 -07001538 /* drain the run list */
1539 while (__aio_run_iocbs(ctx))
1540 ;
1541 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542 spin_unlock_irq(&ctx->ctx_lock);
1543 aio_put_req(req); /* drop extra ref to req */
1544 return 0;
1545
1546out_put_req:
1547 aio_put_req(req); /* drop extra ref to req */
1548 aio_put_req(req); /* drop i/o ref to req */
1549 return ret;
1550}
1551
1552/* sys_io_submit:
1553 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1554 * the number of iocbs queued. May return -EINVAL if the aio_context
1555 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1556 * *iocbpp[0] is not properly initialized, if the operation specified
1557 * is invalid for the file descriptor in the iocb. May fail with
1558 * -EFAULT if any of the data structures point to invalid data. May
1559 * fail with -EBADF if the file descriptor specified in the first
1560 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1561 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1562 * fail with -ENOSYS if not implemented.
1563 */
1564asmlinkage long sys_io_submit(aio_context_t ctx_id, long nr,
1565 struct iocb __user * __user *iocbpp)
1566{
1567 struct kioctx *ctx;
1568 long ret = 0;
1569 int i;
1570
1571 if (unlikely(nr < 0))
1572 return -EINVAL;
1573
1574 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1575 return -EFAULT;
1576
1577 ctx = lookup_ioctx(ctx_id);
1578 if (unlikely(!ctx)) {
1579 pr_debug("EINVAL: io_submit: invalid context id\n");
1580 return -EINVAL;
1581 }
1582
1583 /*
1584 * AKPM: should this return a partial result if some of the IOs were
1585 * successfully submitted?
1586 */
1587 for (i=0; i<nr; i++) {
1588 struct iocb __user *user_iocb;
1589 struct iocb tmp;
1590
1591 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1592 ret = -EFAULT;
1593 break;
1594 }
1595
1596 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1597 ret = -EFAULT;
1598 break;
1599 }
1600
1601 ret = io_submit_one(ctx, user_iocb, &tmp);
1602 if (ret)
1603 break;
1604 }
1605
1606 put_ioctx(ctx);
1607 return i ? i : ret;
1608}
1609
1610/* lookup_kiocb
1611 * Finds a given iocb for cancellation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001612 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001613static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1614 u32 key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001615{
1616 struct list_head *pos;
Zach Brownd00689a2005-11-13 16:07:34 -08001617
1618 assert_spin_locked(&ctx->ctx_lock);
1619
Linus Torvalds1da177e2005-04-16 15:20:36 -07001620 /* TODO: use a hash or array, this sucks. */
1621 list_for_each(pos, &ctx->active_reqs) {
1622 struct kiocb *kiocb = list_kiocb(pos);
1623 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1624 return kiocb;
1625 }
1626 return NULL;
1627}
1628
1629/* sys_io_cancel:
1630 * Attempts to cancel an iocb previously passed to io_submit. If
1631 * the operation is successfully cancelled, the resulting event is
1632 * copied into the memory pointed to by result without being placed
1633 * into the completion queue and 0 is returned. May fail with
1634 * -EFAULT if any of the data structures pointed to are invalid.
1635 * May fail with -EINVAL if aio_context specified by ctx_id is
1636 * invalid. May fail with -EAGAIN if the iocb specified was not
1637 * cancelled. Will fail with -ENOSYS if not implemented.
1638 */
1639asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb,
1640 struct io_event __user *result)
1641{
1642 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1643 struct kioctx *ctx;
1644 struct kiocb *kiocb;
1645 u32 key;
1646 int ret;
1647
1648 ret = get_user(key, &iocb->aio_key);
1649 if (unlikely(ret))
1650 return -EFAULT;
1651
1652 ctx = lookup_ioctx(ctx_id);
1653 if (unlikely(!ctx))
1654 return -EINVAL;
1655
1656 spin_lock_irq(&ctx->ctx_lock);
1657 ret = -EAGAIN;
1658 kiocb = lookup_kiocb(ctx, iocb, key);
1659 if (kiocb && kiocb->ki_cancel) {
1660 cancel = kiocb->ki_cancel;
1661 kiocb->ki_users ++;
1662 kiocbSetCancelled(kiocb);
1663 } else
1664 cancel = NULL;
1665 spin_unlock_irq(&ctx->ctx_lock);
1666
1667 if (NULL != cancel) {
1668 struct io_event tmp;
1669 pr_debug("calling cancel\n");
1670 memset(&tmp, 0, sizeof(tmp));
1671 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1672 tmp.data = kiocb->ki_user_data;
1673 ret = cancel(kiocb, &tmp);
1674 if (!ret) {
1675 /* Cancellation succeeded -- copy the result
1676 * into the user's buffer.
1677 */
1678 if (copy_to_user(result, &tmp, sizeof(tmp)))
1679 ret = -EFAULT;
1680 }
1681 } else
Wendy Cheng8f582022005-09-09 13:02:08 -07001682 ret = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683
1684 put_ioctx(ctx);
1685
1686 return ret;
1687}
1688
1689/* io_getevents:
1690 * Attempts to read at least min_nr events and up to nr events from
1691 * the completion queue for the aio_context specified by ctx_id. May
1692 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range,
1693 * if nr is out of range, if when is out of range. May fail with
1694 * -EFAULT if any of the memory specified to is invalid. May return
1695 * 0 or < min_nr if no events are available and the timeout specified
1696 * by when has elapsed, where when == NULL specifies an infinite
1697 * timeout. Note that the timeout pointed to by when is relative and
1698 * will be updated if not NULL and the operation blocks. Will fail
1699 * with -ENOSYS if not implemented.
1700 */
1701asmlinkage long sys_io_getevents(aio_context_t ctx_id,
1702 long min_nr,
1703 long nr,
1704 struct io_event __user *events,
1705 struct timespec __user *timeout)
1706{
1707 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1708 long ret = -EINVAL;
1709
1710 if (likely(ioctx)) {
1711 if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0))
1712 ret = read_events(ioctx, min_nr, nr, events, timeout);
1713 put_ioctx(ioctx);
1714 }
1715
1716 return ret;
1717}
1718
1719__initcall(aio_setup);
1720
1721EXPORT_SYMBOL(aio_complete);
1722EXPORT_SYMBOL(aio_put_req);
1723EXPORT_SYMBOL(wait_on_sync_kiocb);