blob: e41e932ba489ff040fe9ed43385fc164a8dd6515 [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>
32
33#include <asm/kmap_types.h>
34#include <asm/uaccess.h>
35#include <asm/mmu_context.h>
36
37#if DEBUG > 1
38#define dprintk printk
39#else
40#define dprintk(x...) do { ; } while (0)
41#endif
42
Linus Torvalds1da177e2005-04-16 15:20:36 -070043/*------ sysctl variables----*/
Zach Brownd55b5fd2005-11-07 00:59:31 -080044static DEFINE_SPINLOCK(aio_nr_lock);
45unsigned long aio_nr; /* current system wide number of aio requests */
46unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
Linus Torvalds1da177e2005-04-16 15:20:36 -070047/*----end sysctl variables---*/
48
49static kmem_cache_t *kiocb_cachep;
50static kmem_cache_t *kioctx_cachep;
51
52static struct workqueue_struct *aio_wq;
53
54/* Used for rare fput completion. */
55static void aio_fput_routine(void *);
56static DECLARE_WORK(fput_work, aio_fput_routine, NULL);
57
58static DEFINE_SPINLOCK(fput_lock);
Adrian Bunk25ee7e32005-04-25 08:18:14 -070059static LIST_HEAD(fput_head);
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
61static void aio_kick_handler(void *);
Sébastien Duguc016e222005-06-28 20:44:59 -070062static void aio_queue_work(struct kioctx *);
Linus Torvalds1da177e2005-04-16 15:20:36 -070063
64/* aio_setup
65 * Creates the slab caches used by the aio routines, panic on
66 * failure as this is done early during the boot sequence.
67 */
68static int __init aio_setup(void)
69{
70 kiocb_cachep = kmem_cache_create("kiocb", sizeof(struct kiocb),
71 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
72 kioctx_cachep = kmem_cache_create("kioctx", sizeof(struct kioctx),
73 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
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
82static 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
102static 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) {
Oliver Neukum11b0b5a2006-03-25 03:08:13 -0800125 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126 if (!info->ring_pages)
127 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700128 }
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_ANON|MAP_PRIVATE,
135 0);
136 if (IS_ERR((void *)info->mmap_base)) {
137 up_write(&ctx->mm->mmap_sem);
138 printk("mmap err: %ld\n", -info->mmap_base);
139 info->mmap_size = 0;
140 aio_free_ring(ctx);
141 return -EAGAIN;
142 }
143
144 dprintk("mmap address: 0x%08lx\n", info->mmap_base);
145 info->nr_pages = get_user_pages(current, ctx->mm,
146 info->mmap_base, nr_pages,
147 1, 0, info->ring_pages, NULL);
148 up_write(&ctx->mm->mmap_sem);
149
150 if (unlikely(info->nr_pages != nr_pages)) {
151 aio_free_ring(ctx);
152 return -EAGAIN;
153 }
154
155 ctx->user_id = info->mmap_base;
156
157 info->nr = nr_events; /* trusted copy */
158
159 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
160 ring->nr = nr_events; /* user copy */
161 ring->id = ctx->user_id;
162 ring->head = ring->tail = 0;
163 ring->magic = AIO_RING_MAGIC;
164 ring->compat_features = AIO_RING_COMPAT_FEATURES;
165 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
166 ring->header_length = sizeof(struct aio_ring);
167 kunmap_atomic(ring, KM_USER0);
168
169 return 0;
170}
171
172
173/* aio_ring_event: returns a pointer to the event at the given index from
174 * kmap_atomic(, km). Release the pointer with put_aio_ring_event();
175 */
176#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
177#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
178#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
179
180#define aio_ring_event(info, nr, km) ({ \
181 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
182 struct io_event *__event; \
183 __event = kmap_atomic( \
184 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
185 __event += pos % AIO_EVENTS_PER_PAGE; \
186 __event; \
187})
188
189#define put_aio_ring_event(event, km) do { \
190 struct io_event *__event = (event); \
191 (void)__event; \
192 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
193} while(0)
194
195/* ioctx_alloc
196 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
197 */
198static struct kioctx *ioctx_alloc(unsigned nr_events)
199{
200 struct mm_struct *mm;
201 struct kioctx *ctx;
202
203 /* Prevent overflows */
204 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
205 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
206 pr_debug("ENOMEM: nr_events too high\n");
207 return ERR_PTR(-EINVAL);
208 }
209
Zach Brownd55b5fd2005-11-07 00:59:31 -0800210 if ((unsigned long)nr_events > aio_max_nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211 return ERR_PTR(-EAGAIN);
212
213 ctx = kmem_cache_alloc(kioctx_cachep, GFP_KERNEL);
214 if (!ctx)
215 return ERR_PTR(-ENOMEM);
216
217 memset(ctx, 0, sizeof(*ctx));
218 ctx->max_reqs = nr_events;
219 mm = ctx->mm = current->mm;
220 atomic_inc(&mm->mm_count);
221
222 atomic_set(&ctx->users, 1);
223 spin_lock_init(&ctx->ctx_lock);
224 spin_lock_init(&ctx->ring_info.ring_lock);
225 init_waitqueue_head(&ctx->wait);
226
227 INIT_LIST_HEAD(&ctx->active_reqs);
228 INIT_LIST_HEAD(&ctx->run_list);
229 INIT_WORK(&ctx->wq, aio_kick_handler, ctx);
230
231 if (aio_setup_ring(ctx) < 0)
232 goto out_freectx;
233
234 /* limit the number of system wide aios */
Zach Brownd55b5fd2005-11-07 00:59:31 -0800235 spin_lock(&aio_nr_lock);
236 if (aio_nr + ctx->max_reqs > aio_max_nr ||
237 aio_nr + ctx->max_reqs < aio_nr)
238 ctx->max_reqs = 0;
239 else
240 aio_nr += ctx->max_reqs;
241 spin_unlock(&aio_nr_lock);
242 if (ctx->max_reqs == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700243 goto out_cleanup;
244
245 /* now link into global list. kludge. FIXME */
246 write_lock(&mm->ioctx_list_lock);
247 ctx->next = mm->ioctx_list;
248 mm->ioctx_list = ctx;
249 write_unlock(&mm->ioctx_list_lock);
250
251 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
252 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
253 return ctx;
254
255out_cleanup:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256 __put_ioctx(ctx);
257 return ERR_PTR(-EAGAIN);
258
259out_freectx:
260 mmdrop(mm);
261 kmem_cache_free(kioctx_cachep, ctx);
262 ctx = ERR_PTR(-ENOMEM);
263
264 dprintk("aio: error allocating ioctx %p\n", ctx);
265 return ctx;
266}
267
268/* aio_cancel_all
269 * Cancels all outstanding aio requests on an aio context. Used
270 * when the processes owning a context have all exited to encourage
271 * the rapid destruction of the kioctx.
272 */
273static void aio_cancel_all(struct kioctx *ctx)
274{
275 int (*cancel)(struct kiocb *, struct io_event *);
276 struct io_event res;
277 spin_lock_irq(&ctx->ctx_lock);
278 ctx->dead = 1;
279 while (!list_empty(&ctx->active_reqs)) {
280 struct list_head *pos = ctx->active_reqs.next;
281 struct kiocb *iocb = list_kiocb(pos);
282 list_del_init(&iocb->ki_list);
283 cancel = iocb->ki_cancel;
284 kiocbSetCancelled(iocb);
285 if (cancel) {
286 iocb->ki_users++;
287 spin_unlock_irq(&ctx->ctx_lock);
288 cancel(iocb, &res);
289 spin_lock_irq(&ctx->ctx_lock);
290 }
291 }
292 spin_unlock_irq(&ctx->ctx_lock);
293}
294
Adrian Bunk25ee7e32005-04-25 08:18:14 -0700295static void wait_for_all_aios(struct kioctx *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296{
297 struct task_struct *tsk = current;
298 DECLARE_WAITQUEUE(wait, tsk);
299
300 if (!ctx->reqs_active)
301 return;
302
303 add_wait_queue(&ctx->wait, &wait);
304 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
305 while (ctx->reqs_active) {
306 schedule();
307 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
308 }
309 __set_task_state(tsk, TASK_RUNNING);
310 remove_wait_queue(&ctx->wait, &wait);
311}
312
313/* wait_on_sync_kiocb:
314 * Waits on the given sync kiocb to complete.
315 */
316ssize_t fastcall wait_on_sync_kiocb(struct kiocb *iocb)
317{
318 while (iocb->ki_users) {
319 set_current_state(TASK_UNINTERRUPTIBLE);
320 if (!iocb->ki_users)
321 break;
322 schedule();
323 }
324 __set_current_state(TASK_RUNNING);
325 return iocb->ki_user_data;
326}
327
328/* exit_aio: called when the last user of mm goes away. At this point,
329 * there is no way for any new requests to be submited or any of the
330 * io_* syscalls to be called on the context. However, there may be
331 * outstanding requests which hold references to the context; as they
332 * go away, they will call put_ioctx and release any pinned memory
333 * associated with the request (held via struct page * references).
334 */
335void fastcall exit_aio(struct mm_struct *mm)
336{
337 struct kioctx *ctx = mm->ioctx_list;
338 mm->ioctx_list = NULL;
339 while (ctx) {
340 struct kioctx *next = ctx->next;
341 ctx->next = NULL;
342 aio_cancel_all(ctx);
343
344 wait_for_all_aios(ctx);
345 /*
346 * this is an overkill, but ensures we don't leave
347 * the ctx on the aio_wq
348 */
349 flush_workqueue(aio_wq);
350
351 if (1 != atomic_read(&ctx->users))
352 printk(KERN_DEBUG
353 "exit_aio:ioctx still alive: %d %d %d\n",
354 atomic_read(&ctx->users), ctx->dead,
355 ctx->reqs_active);
356 put_ioctx(ctx);
357 ctx = next;
358 }
359}
360
361/* __put_ioctx
362 * Called when the last user of an aio context has gone away,
363 * and the struct needs to be freed.
364 */
365void fastcall __put_ioctx(struct kioctx *ctx)
366{
367 unsigned nr_events = ctx->max_reqs;
368
369 if (unlikely(ctx->reqs_active))
370 BUG();
371
372 cancel_delayed_work(&ctx->wq);
373 flush_workqueue(aio_wq);
374 aio_free_ring(ctx);
375 mmdrop(ctx->mm);
376 ctx->mm = NULL;
377 pr_debug("__put_ioctx: freeing %p\n", ctx);
378 kmem_cache_free(kioctx_cachep, ctx);
379
Zach Brownd55b5fd2005-11-07 00:59:31 -0800380 if (nr_events) {
381 spin_lock(&aio_nr_lock);
382 BUG_ON(aio_nr - nr_events > aio_nr);
383 aio_nr -= nr_events;
384 spin_unlock(&aio_nr_lock);
385 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386}
387
388/* aio_get_req
389 * Allocate a slot for an aio request. Increments the users count
390 * of the kioctx so that the kioctx stays around until all requests are
391 * complete. Returns NULL if no requests are free.
392 *
393 * Returns with kiocb->users set to 2. The io submit code path holds
394 * an extra reference while submitting the i/o.
395 * This prevents races between the aio code path referencing the
396 * req (after submitting it) and aio_complete() freeing the req.
397 */
398static struct kiocb *FASTCALL(__aio_get_req(struct kioctx *ctx));
399static struct kiocb fastcall *__aio_get_req(struct kioctx *ctx)
400{
401 struct kiocb *req = NULL;
402 struct aio_ring *ring;
403 int okay = 0;
404
405 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
406 if (unlikely(!req))
407 return NULL;
408
Zach Brown4faa5282005-10-17 16:43:33 -0700409 req->ki_flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410 req->ki_users = 2;
411 req->ki_key = 0;
412 req->ki_ctx = ctx;
413 req->ki_cancel = NULL;
414 req->ki_retry = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415 req->ki_dtor = NULL;
416 req->private = NULL;
417 INIT_LIST_HEAD(&req->ki_run_list);
418
419 /* Check if the completion queue has enough free space to
420 * accept an event from this io.
421 */
422 spin_lock_irq(&ctx->ctx_lock);
423 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
424 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
425 list_add(&req->ki_list, &ctx->active_reqs);
426 get_ioctx(ctx);
427 ctx->reqs_active++;
428 okay = 1;
429 }
430 kunmap_atomic(ring, KM_USER0);
431 spin_unlock_irq(&ctx->ctx_lock);
432
433 if (!okay) {
434 kmem_cache_free(kiocb_cachep, req);
435 req = NULL;
436 }
437
438 return req;
439}
440
441static inline struct kiocb *aio_get_req(struct kioctx *ctx)
442{
443 struct kiocb *req;
444 /* Handle a potential starvation case -- should be exceedingly rare as
445 * requests will be stuck on fput_head only if the aio_fput_routine is
446 * delayed and the requests were the last user of the struct file.
447 */
448 req = __aio_get_req(ctx);
449 if (unlikely(NULL == req)) {
450 aio_fput_routine(NULL);
451 req = __aio_get_req(ctx);
452 }
453 return req;
454}
455
456static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
457{
Zach Brownd00689a2005-11-13 16:07:34 -0800458 assert_spin_locked(&ctx->ctx_lock);
459
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460 if (req->ki_dtor)
461 req->ki_dtor(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 kmem_cache_free(kiocb_cachep, req);
463 ctx->reqs_active--;
464
465 if (unlikely(!ctx->reqs_active && ctx->dead))
466 wake_up(&ctx->wait);
467}
468
469static void aio_fput_routine(void *data)
470{
471 spin_lock_irq(&fput_lock);
472 while (likely(!list_empty(&fput_head))) {
473 struct kiocb *req = list_kiocb(fput_head.next);
474 struct kioctx *ctx = req->ki_ctx;
475
476 list_del(&req->ki_list);
477 spin_unlock_irq(&fput_lock);
478
479 /* Complete the fput */
480 __fput(req->ki_filp);
481
482 /* Link the iocb into the context's free list */
483 spin_lock_irq(&ctx->ctx_lock);
484 really_put_req(ctx, req);
485 spin_unlock_irq(&ctx->ctx_lock);
486
487 put_ioctx(ctx);
488 spin_lock_irq(&fput_lock);
489 }
490 spin_unlock_irq(&fput_lock);
491}
492
493/* __aio_put_req
494 * Returns true if this put was the last user of the request.
495 */
496static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
497{
498 dprintk(KERN_DEBUG "aio_put(%p): f_count=%d\n",
499 req, atomic_read(&req->ki_filp->f_count));
500
Zach Brownd00689a2005-11-13 16:07:34 -0800501 assert_spin_locked(&ctx->ctx_lock);
502
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503 req->ki_users --;
504 if (unlikely(req->ki_users < 0))
505 BUG();
506 if (likely(req->ki_users))
507 return 0;
508 list_del(&req->ki_list); /* remove from active_reqs */
509 req->ki_cancel = NULL;
510 req->ki_retry = NULL;
511
512 /* Must be done under the lock to serialise against cancellation.
513 * Call this aio_fput as it duplicates fput via the fput_work.
514 */
Nick Piggin095975d2006-01-08 01:02:19 -0800515 if (unlikely(atomic_dec_and_test(&req->ki_filp->f_count))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516 get_ioctx(ctx);
517 spin_lock(&fput_lock);
518 list_add(&req->ki_list, &fput_head);
519 spin_unlock(&fput_lock);
520 queue_work(aio_wq, &fput_work);
521 } else
522 really_put_req(ctx, req);
523 return 1;
524}
525
526/* aio_put_req
527 * Returns true if this put was the last user of the kiocb,
528 * false if the request is still in use.
529 */
530int fastcall aio_put_req(struct kiocb *req)
531{
532 struct kioctx *ctx = req->ki_ctx;
533 int ret;
534 spin_lock_irq(&ctx->ctx_lock);
535 ret = __aio_put_req(ctx, req);
536 spin_unlock_irq(&ctx->ctx_lock);
537 if (ret)
538 put_ioctx(ctx);
539 return ret;
540}
541
542/* Lookup an ioctx id. ioctx_list is lockless for reads.
543 * FIXME: this is O(n) and is only suitable for development.
544 */
545struct kioctx *lookup_ioctx(unsigned long ctx_id)
546{
547 struct kioctx *ioctx;
548 struct mm_struct *mm;
549
550 mm = current->mm;
551 read_lock(&mm->ioctx_list_lock);
552 for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next)
553 if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) {
554 get_ioctx(ioctx);
555 break;
556 }
557 read_unlock(&mm->ioctx_list_lock);
558
559 return ioctx;
560}
561
562/*
563 * use_mm
564 * Makes the calling kernel thread take on the specified
565 * mm context.
566 * Called by the retry thread execute retries within the
567 * iocb issuer's mm context, so that copy_from/to_user
568 * operations work seamlessly for aio.
569 * (Note: this routine is intended to be called only
570 * from a kernel thread context)
571 */
572static void use_mm(struct mm_struct *mm)
573{
574 struct mm_struct *active_mm;
575 struct task_struct *tsk = current;
576
577 task_lock(tsk);
578 tsk->flags |= PF_BORROWED_MM;
579 active_mm = tsk->active_mm;
580 atomic_inc(&mm->mm_count);
581 tsk->mm = mm;
582 tsk->active_mm = mm;
Paolo 'Blaisorblade' Giarrusso1e40cd32005-09-03 15:57:25 -0700583 /*
584 * Note that on UML this *requires* PF_BORROWED_MM to be set, otherwise
585 * it won't work. Update it accordingly if you change it here
586 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587 activate_mm(active_mm, mm);
588 task_unlock(tsk);
589
590 mmdrop(active_mm);
591}
592
593/*
594 * unuse_mm
595 * Reverses the effect of use_mm, i.e. releases the
596 * specified mm context which was earlier taken on
597 * by the calling kernel thread
598 * (Note: this routine is intended to be called only
599 * from a kernel thread context)
600 *
601 * Comments: Called with ctx->ctx_lock held. This nests
602 * task_lock instead ctx_lock.
603 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -0700604static void unuse_mm(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605{
606 struct task_struct *tsk = current;
607
608 task_lock(tsk);
609 tsk->flags &= ~PF_BORROWED_MM;
610 tsk->mm = NULL;
611 /* active_mm is still 'mm' */
612 enter_lazy_tlb(mm, tsk);
613 task_unlock(tsk);
614}
615
616/*
617 * Queue up a kiocb to be retried. Assumes that the kiocb
618 * has already been marked as kicked, and places it on
619 * the retry run list for the corresponding ioctx, if it
620 * isn't already queued. Returns 1 if it actually queued
621 * the kiocb (to tell the caller to activate the work
622 * queue to process it), or 0, if it found that it was
623 * already queued.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624 */
625static inline int __queue_kicked_iocb(struct kiocb *iocb)
626{
627 struct kioctx *ctx = iocb->ki_ctx;
628
Zach Brownd00689a2005-11-13 16:07:34 -0800629 assert_spin_locked(&ctx->ctx_lock);
630
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631 if (list_empty(&iocb->ki_run_list)) {
632 list_add_tail(&iocb->ki_run_list,
633 &ctx->run_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634 return 1;
635 }
636 return 0;
637}
638
639/* aio_run_iocb
640 * This is the core aio execution routine. It is
641 * invoked both for initial i/o submission and
642 * subsequent retries via the aio_kick_handler.
643 * Expects to be invoked with iocb->ki_ctx->lock
644 * already held. The lock is released and reaquired
645 * as needed during processing.
646 *
647 * Calls the iocb retry method (already setup for the
648 * iocb on initial submission) for operation specific
649 * handling, but takes care of most of common retry
650 * execution details for a given iocb. The retry method
651 * needs to be non-blocking as far as possible, to avoid
652 * holding up other iocbs waiting to be serviced by the
653 * retry kernel thread.
654 *
655 * The trickier parts in this code have to do with
656 * ensuring that only one retry instance is in progress
657 * for a given iocb at any time. Providing that guarantee
658 * simplifies the coding of individual aio operations as
659 * it avoids various potential races.
660 */
661static ssize_t aio_run_iocb(struct kiocb *iocb)
662{
663 struct kioctx *ctx = iocb->ki_ctx;
664 ssize_t (*retry)(struct kiocb *);
665 ssize_t ret;
666
667 if (iocb->ki_retried++ > 1024*1024) {
668 printk("Maximal retry count. Bytes done %Zd\n",
669 iocb->ki_nbytes - iocb->ki_left);
670 return -EAGAIN;
671 }
672
673 if (!(iocb->ki_retried & 0xff)) {
Ken Chen644d3a02005-05-01 08:59:15 -0700674 pr_debug("%ld retry: %d of %d\n", iocb->ki_retried,
675 iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676 }
677
678 if (!(retry = iocb->ki_retry)) {
679 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
680 return 0;
681 }
682
683 /*
684 * We don't want the next retry iteration for this
685 * operation to start until this one has returned and
686 * updated the iocb state. However, wait_queue functions
687 * can trigger a kick_iocb from interrupt context in the
688 * meantime, indicating that data is available for the next
689 * iteration. We want to remember that and enable the
690 * next retry iteration _after_ we are through with
691 * this one.
692 *
693 * So, in order to be able to register a "kick", but
694 * prevent it from being queued now, we clear the kick
695 * flag, but make the kick code *think* that the iocb is
696 * still on the run list until we are actually done.
697 * When we are done with this iteration, we check if
698 * the iocb was kicked in the meantime and if so, queue
699 * it up afresh.
700 */
701
702 kiocbClearKicked(iocb);
703
704 /*
705 * This is so that aio_complete knows it doesn't need to
706 * pull the iocb off the run list (We can't just call
707 * INIT_LIST_HEAD because we don't want a kick_iocb to
708 * queue this on the run list yet)
709 */
710 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
711 spin_unlock_irq(&ctx->ctx_lock);
712
713 /* Quit retrying if the i/o has been cancelled */
714 if (kiocbIsCancelled(iocb)) {
715 ret = -EINTR;
716 aio_complete(iocb, ret, 0);
717 /* must not access the iocb after this */
718 goto out;
719 }
720
721 /*
722 * Now we are all set to call the retry method in async
723 * context. By setting this thread's io_wait context
724 * to point to the wait queue entry inside the currently
725 * running iocb for the duration of the retry, we ensure
726 * that async notification wakeups are queued by the
727 * operation instead of blocking waits, and when notified,
728 * cause the iocb to be kicked for continuation (through
729 * the aio_wake_function callback).
730 */
731 BUG_ON(current->io_wait != NULL);
732 current->io_wait = &iocb->ki_wait;
733 ret = retry(iocb);
734 current->io_wait = NULL;
735
Zach Brown897f15f2005-09-30 11:58:55 -0700736 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
737 BUG_ON(!list_empty(&iocb->ki_wait.task_list));
738 aio_complete(iocb, ret, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739 }
740out:
741 spin_lock_irq(&ctx->ctx_lock);
742
743 if (-EIOCBRETRY == ret) {
744 /*
745 * OK, now that we are done with this iteration
746 * and know that there is more left to go,
747 * this is where we let go so that a subsequent
748 * "kick" can start the next iteration
749 */
750
751 /* will make __queue_kicked_iocb succeed from here on */
752 INIT_LIST_HEAD(&iocb->ki_run_list);
753 /* we must queue the next iteration ourselves, if it
754 * has already been kicked */
755 if (kiocbIsKicked(iocb)) {
756 __queue_kicked_iocb(iocb);
Sébastien Duguc016e222005-06-28 20:44:59 -0700757
758 /*
759 * __queue_kicked_iocb will always return 1 here, because
760 * iocb->ki_run_list is empty at this point so it should
761 * be safe to unconditionally queue the context into the
762 * work queue.
763 */
764 aio_queue_work(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765 }
766 }
767 return ret;
768}
769
770/*
771 * __aio_run_iocbs:
772 * Process all pending retries queued on the ioctx
773 * run list.
774 * Assumes it is operating within the aio issuer's mm
Zach Brownd00689a2005-11-13 16:07:34 -0800775 * context.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776 */
777static int __aio_run_iocbs(struct kioctx *ctx)
778{
779 struct kiocb *iocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780 LIST_HEAD(run_list);
781
Zach Brownd00689a2005-11-13 16:07:34 -0800782 assert_spin_locked(&ctx->ctx_lock);
783
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784 list_splice_init(&ctx->run_list, &run_list);
785 while (!list_empty(&run_list)) {
786 iocb = list_entry(run_list.next, struct kiocb,
787 ki_run_list);
788 list_del(&iocb->ki_run_list);
789 /*
790 * Hold an extra reference while retrying i/o.
791 */
792 iocb->ki_users++; /* grab extra reference */
793 aio_run_iocb(iocb);
794 if (__aio_put_req(ctx, iocb)) /* drop extra ref */
795 put_ioctx(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797 if (!list_empty(&ctx->run_list))
798 return 1;
799 return 0;
800}
801
802static void aio_queue_work(struct kioctx * ctx)
803{
804 unsigned long timeout;
805 /*
806 * if someone is waiting, get the work started right
807 * away, otherwise, use a longer delay
808 */
809 smp_mb();
810 if (waitqueue_active(&ctx->wait))
811 timeout = 1;
812 else
813 timeout = HZ/10;
814 queue_delayed_work(aio_wq, &ctx->wq, timeout);
815}
816
817
818/*
819 * aio_run_iocbs:
820 * Process all pending retries queued on the ioctx
821 * run list.
822 * Assumes it is operating within the aio issuer's mm
823 * context.
824 */
825static inline void aio_run_iocbs(struct kioctx *ctx)
826{
827 int requeue;
828
829 spin_lock_irq(&ctx->ctx_lock);
830
831 requeue = __aio_run_iocbs(ctx);
832 spin_unlock_irq(&ctx->ctx_lock);
833 if (requeue)
834 aio_queue_work(ctx);
835}
836
837/*
838 * just like aio_run_iocbs, but keeps running them until
839 * the list stays empty
840 */
841static inline void aio_run_all_iocbs(struct kioctx *ctx)
842{
843 spin_lock_irq(&ctx->ctx_lock);
844 while (__aio_run_iocbs(ctx))
845 ;
846 spin_unlock_irq(&ctx->ctx_lock);
847}
848
849/*
850 * aio_kick_handler:
851 * Work queue handler triggered to process pending
852 * retries on an ioctx. Takes on the aio issuer's
853 * mm context before running the iocbs, so that
854 * copy_xxx_user operates on the issuer's address
855 * space.
856 * Run on aiod's context.
857 */
858static void aio_kick_handler(void *data)
859{
860 struct kioctx *ctx = data;
861 mm_segment_t oldfs = get_fs();
862 int requeue;
863
864 set_fs(USER_DS);
865 use_mm(ctx->mm);
866 spin_lock_irq(&ctx->ctx_lock);
867 requeue =__aio_run_iocbs(ctx);
868 unuse_mm(ctx->mm);
869 spin_unlock_irq(&ctx->ctx_lock);
870 set_fs(oldfs);
871 /*
872 * we're in a worker thread already, don't use queue_delayed_work,
873 */
874 if (requeue)
875 queue_work(aio_wq, &ctx->wq);
876}
877
878
879/*
880 * Called by kick_iocb to queue the kiocb for retry
881 * and if required activate the aio work queue to process
882 * it
883 */
Zach Brown998765e2005-09-30 11:58:54 -0700884static void try_queue_kicked_iocb(struct kiocb *iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885{
886 struct kioctx *ctx = iocb->ki_ctx;
887 unsigned long flags;
888 int run = 0;
889
Zach Brown998765e2005-09-30 11:58:54 -0700890 /* We're supposed to be the only path putting the iocb back on the run
891 * list. If we find that the iocb is *back* on a wait queue already
892 * than retry has happened before we could queue the iocb. This also
893 * means that the retry could have completed and freed our iocb, no
894 * good. */
895 BUG_ON((!list_empty(&iocb->ki_wait.task_list)));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896
897 spin_lock_irqsave(&ctx->ctx_lock, flags);
Zach Brown998765e2005-09-30 11:58:54 -0700898 /* set this inside the lock so that we can't race with aio_run_iocb()
899 * testing it and putting the iocb on the run list under the lock */
900 if (!kiocbTryKick(iocb))
901 run = __queue_kicked_iocb(iocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
Ken Chen644d3a02005-05-01 08:59:15 -0700903 if (run)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904 aio_queue_work(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700905}
906
907/*
908 * kick_iocb:
909 * Called typically from a wait queue callback context
910 * (aio_wake_function) to trigger a retry of the iocb.
911 * The retry is usually executed by aio workqueue
912 * threads (See aio_kick_handler).
913 */
914void fastcall kick_iocb(struct kiocb *iocb)
915{
916 /* sync iocbs are easy: they can only ever be executing from a
917 * single context. */
918 if (is_sync_kiocb(iocb)) {
919 kiocbSetKicked(iocb);
920 wake_up_process(iocb->ki_obj.tsk);
921 return;
922 }
923
Zach Brown998765e2005-09-30 11:58:54 -0700924 try_queue_kicked_iocb(iocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925}
926EXPORT_SYMBOL(kick_iocb);
927
928/* aio_complete
929 * Called when the io request on the given iocb is complete.
930 * Returns true if this is the last user of the request. The
931 * only other user of the request can be the cancellation code.
932 */
933int fastcall aio_complete(struct kiocb *iocb, long res, long res2)
934{
935 struct kioctx *ctx = iocb->ki_ctx;
936 struct aio_ring_info *info;
937 struct aio_ring *ring;
938 struct io_event *event;
939 unsigned long flags;
940 unsigned long tail;
941 int ret;
942
Zach Brown20dcae32005-11-13 16:07:33 -0800943 /*
944 * Special case handling for sync iocbs:
945 * - events go directly into the iocb for fast handling
946 * - the sync task with the iocb in its stack holds the single iocb
947 * ref, no other paths have a way to get another ref
948 * - the sync task helpfully left a reference to itself in the iocb
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949 */
950 if (is_sync_kiocb(iocb)) {
Zach Brown20dcae32005-11-13 16:07:33 -0800951 BUG_ON(iocb->ki_users != 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952 iocb->ki_user_data = res;
Zach Brown20dcae32005-11-13 16:07:33 -0800953 iocb->ki_users = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954 wake_up_process(iocb->ki_obj.tsk);
Zach Brown20dcae32005-11-13 16:07:33 -0800955 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 }
957
958 info = &ctx->ring_info;
959
960 /* add a completion event to the ring buffer.
961 * must be done holding ctx->ctx_lock to prevent
962 * other code from messing with the tail
963 * pointer since we might be called from irq
964 * context.
965 */
966 spin_lock_irqsave(&ctx->ctx_lock, flags);
967
968 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
969 list_del_init(&iocb->ki_run_list);
970
971 /*
972 * cancelled requests don't get events, userland was given one
973 * when the event got cancelled.
974 */
975 if (kiocbIsCancelled(iocb))
976 goto put_rq;
977
978 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
979
980 tail = info->tail;
981 event = aio_ring_event(info, tail, KM_IRQ0);
Ken Chen4bf69b22005-05-01 08:59:15 -0700982 if (++tail >= info->nr)
983 tail = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984
985 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
986 event->data = iocb->ki_user_data;
987 event->res = res;
988 event->res2 = res2;
989
990 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
991 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
992 res, res2);
993
994 /* after flagging the request as done, we
995 * must never even look at it again
996 */
997 smp_wmb(); /* make event visible before updating tail */
998
999 info->tail = tail;
1000 ring->tail = tail;
1001
1002 put_aio_ring_event(event, KM_IRQ0);
1003 kunmap_atomic(ring, KM_IRQ1);
1004
1005 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
1006
Ken Chen644d3a02005-05-01 08:59:15 -07001007 pr_debug("%ld retries: %d of %d\n", iocb->ki_retried,
1008 iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001009put_rq:
1010 /* everything turned out well, dispose of the aiocb. */
1011 ret = __aio_put_req(ctx, iocb);
1012
1013 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1014
1015 if (waitqueue_active(&ctx->wait))
1016 wake_up(&ctx->wait);
1017
1018 if (ret)
1019 put_ioctx(ctx);
1020
1021 return ret;
1022}
1023
1024/* aio_read_evt
1025 * Pull an event off of the ioctx's event ring. Returns the number of
1026 * events fetched (0 or 1 ;-)
1027 * FIXME: make this use cmpxchg.
1028 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1029 */
1030static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1031{
1032 struct aio_ring_info *info = &ioctx->ring_info;
1033 struct aio_ring *ring;
1034 unsigned long head;
1035 int ret = 0;
1036
1037 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1038 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1039 (unsigned long)ring->head, (unsigned long)ring->tail,
1040 (unsigned long)ring->nr);
1041
1042 if (ring->head == ring->tail)
1043 goto out;
1044
1045 spin_lock(&info->ring_lock);
1046
1047 head = ring->head % info->nr;
1048 if (head != ring->tail) {
1049 struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1050 *ent = *evp;
1051 head = (head + 1) % info->nr;
1052 smp_mb(); /* finish reading the event before updatng the head */
1053 ring->head = head;
1054 ret = 1;
1055 put_aio_ring_event(evp, KM_USER1);
1056 }
1057 spin_unlock(&info->ring_lock);
1058
1059out:
1060 kunmap_atomic(ring, KM_USER0);
1061 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1062 (unsigned long)ring->head, (unsigned long)ring->tail);
1063 return ret;
1064}
1065
1066struct aio_timeout {
1067 struct timer_list timer;
1068 int timed_out;
1069 struct task_struct *p;
1070};
1071
1072static void timeout_func(unsigned long data)
1073{
1074 struct aio_timeout *to = (struct aio_timeout *)data;
1075
1076 to->timed_out = 1;
1077 wake_up_process(to->p);
1078}
1079
1080static inline void init_timeout(struct aio_timeout *to)
1081{
1082 init_timer(&to->timer);
1083 to->timer.data = (unsigned long)to;
1084 to->timer.function = timeout_func;
1085 to->timed_out = 0;
1086 to->p = current;
1087}
1088
1089static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1090 const struct timespec *ts)
1091{
1092 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1093 if (time_after(to->timer.expires, jiffies))
1094 add_timer(&to->timer);
1095 else
1096 to->timed_out = 1;
1097}
1098
1099static inline void clear_timeout(struct aio_timeout *to)
1100{
1101 del_singleshot_timer_sync(&to->timer);
1102}
1103
1104static int read_events(struct kioctx *ctx,
1105 long min_nr, long nr,
1106 struct io_event __user *event,
1107 struct timespec __user *timeout)
1108{
1109 long start_jiffies = jiffies;
1110 struct task_struct *tsk = current;
1111 DECLARE_WAITQUEUE(wait, tsk);
1112 int ret;
1113 int i = 0;
1114 struct io_event ent;
1115 struct aio_timeout to;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116 int retry = 0;
1117
1118 /* needed to zero any padding within an entry (there shouldn't be
1119 * any, but C is fun!
1120 */
1121 memset(&ent, 0, sizeof(ent));
1122retry:
1123 ret = 0;
1124 while (likely(i < nr)) {
1125 ret = aio_read_evt(ctx, &ent);
1126 if (unlikely(ret <= 0))
1127 break;
1128
1129 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1130 ent.data, ent.obj, ent.res, ent.res2);
1131
1132 /* Could we split the check in two? */
1133 ret = -EFAULT;
1134 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1135 dprintk("aio: lost an event due to EFAULT.\n");
1136 break;
1137 }
1138 ret = 0;
1139
1140 /* Good, event copied to userland, update counts. */
1141 event ++;
1142 i ++;
1143 }
1144
1145 if (min_nr <= i)
1146 return i;
1147 if (ret)
1148 return ret;
1149
1150 /* End fast path */
1151
1152 /* racey check, but it gets redone */
1153 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1154 retry = 1;
1155 aio_run_all_iocbs(ctx);
1156 goto retry;
1157 }
1158
1159 init_timeout(&to);
1160 if (timeout) {
1161 struct timespec ts;
1162 ret = -EFAULT;
1163 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1164 goto out;
1165
1166 set_timeout(start_jiffies, &to, &ts);
1167 }
1168
1169 while (likely(i < nr)) {
1170 add_wait_queue_exclusive(&ctx->wait, &wait);
1171 do {
1172 set_task_state(tsk, TASK_INTERRUPTIBLE);
1173 ret = aio_read_evt(ctx, &ent);
1174 if (ret)
1175 break;
1176 if (min_nr <= i)
1177 break;
1178 ret = 0;
1179 if (to.timed_out) /* Only check after read evt */
1180 break;
1181 schedule();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182 if (signal_pending(tsk)) {
1183 ret = -EINTR;
1184 break;
1185 }
1186 /*ret = aio_read_evt(ctx, &ent);*/
1187 } while (1) ;
1188
1189 set_task_state(tsk, TASK_RUNNING);
1190 remove_wait_queue(&ctx->wait, &wait);
1191
1192 if (unlikely(ret <= 0))
1193 break;
1194
1195 ret = -EFAULT;
1196 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1197 dprintk("aio: lost an event due to EFAULT.\n");
1198 break;
1199 }
1200
1201 /* Good, event copied to userland, update counts. */
1202 event ++;
1203 i ++;
1204 }
1205
1206 if (timeout)
1207 clear_timeout(&to);
1208out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001209 return i ? i : ret;
1210}
1211
1212/* Take an ioctx and remove it from the list of ioctx's. Protects
1213 * against races with itself via ->dead.
1214 */
1215static void io_destroy(struct kioctx *ioctx)
1216{
1217 struct mm_struct *mm = current->mm;
1218 struct kioctx **tmp;
1219 int was_dead;
1220
1221 /* delete the entry from the list is someone else hasn't already */
1222 write_lock(&mm->ioctx_list_lock);
1223 was_dead = ioctx->dead;
1224 ioctx->dead = 1;
1225 for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx;
1226 tmp = &(*tmp)->next)
1227 ;
1228 if (*tmp)
1229 *tmp = ioctx->next;
1230 write_unlock(&mm->ioctx_list_lock);
1231
1232 dprintk("aio_release(%p)\n", ioctx);
1233 if (likely(!was_dead))
1234 put_ioctx(ioctx); /* twice for the list */
1235
1236 aio_cancel_all(ioctx);
1237 wait_for_all_aios(ioctx);
1238 put_ioctx(ioctx); /* once for the lookup */
1239}
1240
1241/* sys_io_setup:
1242 * Create an aio_context capable of receiving at least nr_events.
1243 * ctxp must not point to an aio_context that already exists, and
1244 * must be initialized to 0 prior to the call. On successful
1245 * creation of the aio_context, *ctxp is filled in with the resulting
1246 * handle. May fail with -EINVAL if *ctxp is not initialized,
1247 * if the specified nr_events exceeds internal limits. May fail
1248 * with -EAGAIN if the specified nr_events exceeds the user's limit
1249 * of available events. May fail with -ENOMEM if insufficient kernel
1250 * resources are available. May fail with -EFAULT if an invalid
1251 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1252 * implemented.
1253 */
1254asmlinkage long sys_io_setup(unsigned nr_events, aio_context_t __user *ctxp)
1255{
1256 struct kioctx *ioctx = NULL;
1257 unsigned long ctx;
1258 long ret;
1259
1260 ret = get_user(ctx, ctxp);
1261 if (unlikely(ret))
1262 goto out;
1263
1264 ret = -EINVAL;
Zach Brownd55b5fd2005-11-07 00:59:31 -08001265 if (unlikely(ctx || nr_events == 0)) {
1266 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1267 ctx, nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268 goto out;
1269 }
1270
1271 ioctx = ioctx_alloc(nr_events);
1272 ret = PTR_ERR(ioctx);
1273 if (!IS_ERR(ioctx)) {
1274 ret = put_user(ioctx->user_id, ctxp);
1275 if (!ret)
1276 return 0;
1277
1278 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1279 io_destroy(ioctx);
1280 }
1281
1282out:
1283 return ret;
1284}
1285
1286/* sys_io_destroy:
1287 * Destroy the aio_context specified. May cancel any outstanding
1288 * AIOs and block on completion. Will fail with -ENOSYS if not
1289 * implemented. May fail with -EFAULT if the context pointed to
1290 * is invalid.
1291 */
1292asmlinkage long sys_io_destroy(aio_context_t ctx)
1293{
1294 struct kioctx *ioctx = lookup_ioctx(ctx);
1295 if (likely(NULL != ioctx)) {
1296 io_destroy(ioctx);
1297 return 0;
1298 }
1299 pr_debug("EINVAL: io_destroy: invalid context id\n");
1300 return -EINVAL;
1301}
1302
1303/*
Zach Brown897f15f2005-09-30 11:58:55 -07001304 * aio_p{read,write} are the default ki_retry methods for
1305 * IO_CMD_P{READ,WRITE}. They maintains kiocb retry state around potentially
1306 * multiple calls to f_op->aio_read(). They loop around partial progress
1307 * instead of returning -EIOCBRETRY because they don't have the means to call
1308 * kick_iocb().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001309 */
1310static ssize_t aio_pread(struct kiocb *iocb)
1311{
1312 struct file *file = iocb->ki_filp;
1313 struct address_space *mapping = file->f_mapping;
1314 struct inode *inode = mapping->host;
1315 ssize_t ret = 0;
1316
Zach Brown897f15f2005-09-30 11:58:55 -07001317 do {
1318 ret = file->f_op->aio_read(iocb, iocb->ki_buf,
1319 iocb->ki_left, iocb->ki_pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320 /*
Zach Brown897f15f2005-09-30 11:58:55 -07001321 * Can't just depend on iocb->ki_left to determine
1322 * whether we are done. This may have been a short read.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001323 */
Zach Brown897f15f2005-09-30 11:58:55 -07001324 if (ret > 0) {
1325 iocb->ki_buf += ret;
1326 iocb->ki_left -= ret;
1327 }
1328
1329 /*
1330 * For pipes and sockets we return once we have some data; for
1331 * regular files we retry till we complete the entire read or
1332 * find that we can't read any more data (e.g short reads).
1333 */
Zach Brown353fb072005-09-30 11:58:56 -07001334 } while (ret > 0 && iocb->ki_left > 0 &&
Zach Brown897f15f2005-09-30 11:58:55 -07001335 !S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336
1337 /* This means we must have transferred all that we could */
1338 /* No need to retry anymore */
1339 if ((ret == 0) || (iocb->ki_left == 0))
1340 ret = iocb->ki_nbytes - iocb->ki_left;
1341
1342 return ret;
1343}
1344
Zach Brown897f15f2005-09-30 11:58:55 -07001345/* see aio_pread() */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346static ssize_t aio_pwrite(struct kiocb *iocb)
1347{
1348 struct file *file = iocb->ki_filp;
1349 ssize_t ret = 0;
1350
Zach Brown897f15f2005-09-30 11:58:55 -07001351 do {
1352 ret = file->f_op->aio_write(iocb, iocb->ki_buf,
1353 iocb->ki_left, iocb->ki_pos);
1354 if (ret > 0) {
1355 iocb->ki_buf += ret;
1356 iocb->ki_left -= ret;
1357 }
Zach Brown353fb072005-09-30 11:58:56 -07001358 } while (ret > 0 && iocb->ki_left > 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360 if ((ret == 0) || (iocb->ki_left == 0))
1361 ret = iocb->ki_nbytes - iocb->ki_left;
1362
1363 return ret;
1364}
1365
1366static ssize_t aio_fdsync(struct kiocb *iocb)
1367{
1368 struct file *file = iocb->ki_filp;
1369 ssize_t ret = -EINVAL;
1370
1371 if (file->f_op->aio_fsync)
1372 ret = file->f_op->aio_fsync(iocb, 1);
1373 return ret;
1374}
1375
1376static ssize_t aio_fsync(struct kiocb *iocb)
1377{
1378 struct file *file = iocb->ki_filp;
1379 ssize_t ret = -EINVAL;
1380
1381 if (file->f_op->aio_fsync)
1382 ret = file->f_op->aio_fsync(iocb, 0);
1383 return ret;
1384}
1385
1386/*
1387 * aio_setup_iocb:
1388 * Performs the initial checks and aio retry method
1389 * setup for the kiocb at the time of io submission.
1390 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001391static ssize_t aio_setup_iocb(struct kiocb *kiocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392{
1393 struct file *file = kiocb->ki_filp;
1394 ssize_t ret = 0;
1395
1396 switch (kiocb->ki_opcode) {
1397 case IOCB_CMD_PREAD:
1398 ret = -EBADF;
1399 if (unlikely(!(file->f_mode & FMODE_READ)))
1400 break;
1401 ret = -EFAULT;
1402 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1403 kiocb->ki_left)))
1404 break;
Kostik Belousov8766ce42005-10-23 12:57:13 -07001405 ret = security_file_permission(file, MAY_READ);
1406 if (unlikely(ret))
1407 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001408 ret = -EINVAL;
1409 if (file->f_op->aio_read)
1410 kiocb->ki_retry = aio_pread;
1411 break;
1412 case IOCB_CMD_PWRITE:
1413 ret = -EBADF;
1414 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1415 break;
1416 ret = -EFAULT;
1417 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1418 kiocb->ki_left)))
1419 break;
Kostik Belousov8766ce42005-10-23 12:57:13 -07001420 ret = security_file_permission(file, MAY_WRITE);
1421 if (unlikely(ret))
1422 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001423 ret = -EINVAL;
1424 if (file->f_op->aio_write)
1425 kiocb->ki_retry = aio_pwrite;
1426 break;
1427 case IOCB_CMD_FDSYNC:
1428 ret = -EINVAL;
1429 if (file->f_op->aio_fsync)
1430 kiocb->ki_retry = aio_fdsync;
1431 break;
1432 case IOCB_CMD_FSYNC:
1433 ret = -EINVAL;
1434 if (file->f_op->aio_fsync)
1435 kiocb->ki_retry = aio_fsync;
1436 break;
1437 default:
1438 dprintk("EINVAL: io_submit: no operation provided\n");
1439 ret = -EINVAL;
1440 }
1441
1442 if (!kiocb->ki_retry)
1443 return ret;
1444
1445 return 0;
1446}
1447
1448/*
1449 * aio_wake_function:
1450 * wait queue callback function for aio notification,
1451 * Simply triggers a retry of the operation via kick_iocb.
1452 *
1453 * This callback is specified in the wait queue entry in
1454 * a kiocb (current->io_wait points to this wait queue
1455 * entry when an aio operation executes; it is used
1456 * instead of a synchronous wait when an i/o blocking
1457 * condition is encountered during aio).
1458 *
1459 * Note:
1460 * This routine is executed with the wait queue lock held.
1461 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests
1462 * the ioctx lock inside the wait queue lock. This is safe
1463 * because this callback isn't used for wait queues which
1464 * are nested inside ioctx lock (i.e. ctx->wait)
1465 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001466static int aio_wake_function(wait_queue_t *wait, unsigned mode,
1467 int sync, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468{
1469 struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait);
1470
1471 list_del_init(&wait->task_list);
1472 kick_iocb(iocb);
1473 return 1;
1474}
1475
1476int fastcall io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1477 struct iocb *iocb)
1478{
1479 struct kiocb *req;
1480 struct file *file;
1481 ssize_t ret;
1482
1483 /* enforce forwards compatibility on users */
1484 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2 ||
1485 iocb->aio_reserved3)) {
1486 pr_debug("EINVAL: io_submit: reserve field set\n");
1487 return -EINVAL;
1488 }
1489
1490 /* prevent overflows */
1491 if (unlikely(
1492 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1493 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1494 ((ssize_t)iocb->aio_nbytes < 0)
1495 )) {
1496 pr_debug("EINVAL: io_submit: overflow check\n");
1497 return -EINVAL;
1498 }
1499
1500 file = fget(iocb->aio_fildes);
1501 if (unlikely(!file))
1502 return -EBADF;
1503
1504 req = aio_get_req(ctx); /* returns with 2 references to req */
1505 if (unlikely(!req)) {
1506 fput(file);
1507 return -EAGAIN;
1508 }
1509
1510 req->ki_filp = file;
Ken Chen212079c2005-05-01 08:59:15 -07001511 ret = put_user(req->ki_key, &user_iocb->aio_key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001512 if (unlikely(ret)) {
1513 dprintk("EFAULT: aio_key\n");
1514 goto out_put_req;
1515 }
1516
1517 req->ki_obj.user = user_iocb;
1518 req->ki_user_data = iocb->aio_data;
1519 req->ki_pos = iocb->aio_offset;
1520
1521 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1522 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1523 req->ki_opcode = iocb->aio_lio_opcode;
1524 init_waitqueue_func_entry(&req->ki_wait, aio_wake_function);
1525 INIT_LIST_HEAD(&req->ki_wait.task_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526 req->ki_retried = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001527
1528 ret = aio_setup_iocb(req);
1529
1530 if (ret)
1531 goto out_put_req;
1532
1533 spin_lock_irq(&ctx->ctx_lock);
Benjamin LaHaiseac0b1bc2005-09-09 13:02:09 -07001534 aio_run_iocb(req);
Benjamin LaHaiseac0b1bc2005-09-09 13:02:09 -07001535 if (!list_empty(&ctx->run_list)) {
Ken Chen954d3e92005-05-01 08:59:16 -07001536 /* drain the run list */
1537 while (__aio_run_iocbs(ctx))
1538 ;
1539 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001540 spin_unlock_irq(&ctx->ctx_lock);
1541 aio_put_req(req); /* drop extra ref to req */
1542 return 0;
1543
1544out_put_req:
1545 aio_put_req(req); /* drop extra ref to req */
1546 aio_put_req(req); /* drop i/o ref to req */
1547 return ret;
1548}
1549
1550/* sys_io_submit:
1551 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1552 * the number of iocbs queued. May return -EINVAL if the aio_context
1553 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1554 * *iocbpp[0] is not properly initialized, if the operation specified
1555 * is invalid for the file descriptor in the iocb. May fail with
1556 * -EFAULT if any of the data structures point to invalid data. May
1557 * fail with -EBADF if the file descriptor specified in the first
1558 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1559 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1560 * fail with -ENOSYS if not implemented.
1561 */
1562asmlinkage long sys_io_submit(aio_context_t ctx_id, long nr,
1563 struct iocb __user * __user *iocbpp)
1564{
1565 struct kioctx *ctx;
1566 long ret = 0;
1567 int i;
1568
1569 if (unlikely(nr < 0))
1570 return -EINVAL;
1571
1572 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1573 return -EFAULT;
1574
1575 ctx = lookup_ioctx(ctx_id);
1576 if (unlikely(!ctx)) {
1577 pr_debug("EINVAL: io_submit: invalid context id\n");
1578 return -EINVAL;
1579 }
1580
1581 /*
1582 * AKPM: should this return a partial result if some of the IOs were
1583 * successfully submitted?
1584 */
1585 for (i=0; i<nr; i++) {
1586 struct iocb __user *user_iocb;
1587 struct iocb tmp;
1588
1589 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1590 ret = -EFAULT;
1591 break;
1592 }
1593
1594 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1595 ret = -EFAULT;
1596 break;
1597 }
1598
1599 ret = io_submit_one(ctx, user_iocb, &tmp);
1600 if (ret)
1601 break;
1602 }
1603
1604 put_ioctx(ctx);
1605 return i ? i : ret;
1606}
1607
1608/* lookup_kiocb
1609 * Finds a given iocb for cancellation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001610 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001611static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1612 u32 key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613{
1614 struct list_head *pos;
Zach Brownd00689a2005-11-13 16:07:34 -08001615
1616 assert_spin_locked(&ctx->ctx_lock);
1617
Linus Torvalds1da177e2005-04-16 15:20:36 -07001618 /* TODO: use a hash or array, this sucks. */
1619 list_for_each(pos, &ctx->active_reqs) {
1620 struct kiocb *kiocb = list_kiocb(pos);
1621 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1622 return kiocb;
1623 }
1624 return NULL;
1625}
1626
1627/* sys_io_cancel:
1628 * Attempts to cancel an iocb previously passed to io_submit. If
1629 * the operation is successfully cancelled, the resulting event is
1630 * copied into the memory pointed to by result without being placed
1631 * into the completion queue and 0 is returned. May fail with
1632 * -EFAULT if any of the data structures pointed to are invalid.
1633 * May fail with -EINVAL if aio_context specified by ctx_id is
1634 * invalid. May fail with -EAGAIN if the iocb specified was not
1635 * cancelled. Will fail with -ENOSYS if not implemented.
1636 */
1637asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb,
1638 struct io_event __user *result)
1639{
1640 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1641 struct kioctx *ctx;
1642 struct kiocb *kiocb;
1643 u32 key;
1644 int ret;
1645
1646 ret = get_user(key, &iocb->aio_key);
1647 if (unlikely(ret))
1648 return -EFAULT;
1649
1650 ctx = lookup_ioctx(ctx_id);
1651 if (unlikely(!ctx))
1652 return -EINVAL;
1653
1654 spin_lock_irq(&ctx->ctx_lock);
1655 ret = -EAGAIN;
1656 kiocb = lookup_kiocb(ctx, iocb, key);
1657 if (kiocb && kiocb->ki_cancel) {
1658 cancel = kiocb->ki_cancel;
1659 kiocb->ki_users ++;
1660 kiocbSetCancelled(kiocb);
1661 } else
1662 cancel = NULL;
1663 spin_unlock_irq(&ctx->ctx_lock);
1664
1665 if (NULL != cancel) {
1666 struct io_event tmp;
1667 pr_debug("calling cancel\n");
1668 memset(&tmp, 0, sizeof(tmp));
1669 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1670 tmp.data = kiocb->ki_user_data;
1671 ret = cancel(kiocb, &tmp);
1672 if (!ret) {
1673 /* Cancellation succeeded -- copy the result
1674 * into the user's buffer.
1675 */
1676 if (copy_to_user(result, &tmp, sizeof(tmp)))
1677 ret = -EFAULT;
1678 }
1679 } else
Wendy Cheng8f582022005-09-09 13:02:08 -07001680 ret = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681
1682 put_ioctx(ctx);
1683
1684 return ret;
1685}
1686
1687/* io_getevents:
1688 * Attempts to read at least min_nr events and up to nr events from
1689 * the completion queue for the aio_context specified by ctx_id. May
1690 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range,
1691 * if nr is out of range, if when is out of range. May fail with
1692 * -EFAULT if any of the memory specified to is invalid. May return
1693 * 0 or < min_nr if no events are available and the timeout specified
1694 * by when has elapsed, where when == NULL specifies an infinite
1695 * timeout. Note that the timeout pointed to by when is relative and
1696 * will be updated if not NULL and the operation blocks. Will fail
1697 * with -ENOSYS if not implemented.
1698 */
1699asmlinkage long sys_io_getevents(aio_context_t ctx_id,
1700 long min_nr,
1701 long nr,
1702 struct io_event __user *events,
1703 struct timespec __user *timeout)
1704{
1705 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1706 long ret = -EINVAL;
1707
1708 if (likely(ioctx)) {
1709 if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0))
1710 ret = read_events(ioctx, min_nr, nr, events, timeout);
1711 put_ioctx(ioctx);
1712 }
1713
1714 return ret;
1715}
1716
1717__initcall(aio_setup);
1718
1719EXPORT_SYMBOL(aio_complete);
1720EXPORT_SYMBOL(aio_put_req);
1721EXPORT_SYMBOL(wait_on_sync_kiocb);