blob: 277a5f2d18ad7b8ca66a3aef65551379054fc724 [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>
Badari Pulavarty027445c2006-09-30 23:28:46 -070018#include <linux/uio.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070019
20#define DEBUG 0
21
22#include <linux/sched.h>
23#include <linux/fs.h>
24#include <linux/file.h>
25#include <linux/mm.h>
26#include <linux/mman.h>
27#include <linux/slab.h>
28#include <linux/timer.h>
29#include <linux/aio.h>
30#include <linux/highmem.h>
31#include <linux/workqueue.h>
32#include <linux/security.h>
33
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) {
Oliver Neukum11b0b5a2006-03-25 03:08:13 -0800126 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127 if (!info->ring_pages)
128 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129 }
130
131 info->mmap_size = nr_pages * PAGE_SIZE;
132 dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
133 down_write(&ctx->mm->mmap_sem);
134 info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
135 PROT_READ|PROT_WRITE, MAP_ANON|MAP_PRIVATE,
136 0);
137 if (IS_ERR((void *)info->mmap_base)) {
138 up_write(&ctx->mm->mmap_sem);
139 printk("mmap err: %ld\n", -info->mmap_base);
140 info->mmap_size = 0;
141 aio_free_ring(ctx);
142 return -EAGAIN;
143 }
144
145 dprintk("mmap address: 0x%08lx\n", info->mmap_base);
146 info->nr_pages = get_user_pages(current, ctx->mm,
147 info->mmap_base, nr_pages,
148 1, 0, info->ring_pages, NULL);
149 up_write(&ctx->mm->mmap_sem);
150
151 if (unlikely(info->nr_pages != nr_pages)) {
152 aio_free_ring(ctx);
153 return -EAGAIN;
154 }
155
156 ctx->user_id = info->mmap_base;
157
158 info->nr = nr_events; /* trusted copy */
159
160 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
161 ring->nr = nr_events; /* user copy */
162 ring->id = ctx->user_id;
163 ring->head = ring->tail = 0;
164 ring->magic = AIO_RING_MAGIC;
165 ring->compat_features = AIO_RING_COMPAT_FEATURES;
166 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
167 ring->header_length = sizeof(struct aio_ring);
168 kunmap_atomic(ring, KM_USER0);
169
170 return 0;
171}
172
173
174/* aio_ring_event: returns a pointer to the event at the given index from
175 * kmap_atomic(, km). Release the pointer with put_aio_ring_event();
176 */
177#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
178#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
179#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
180
181#define aio_ring_event(info, nr, km) ({ \
182 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
183 struct io_event *__event; \
184 __event = kmap_atomic( \
185 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
186 __event += pos % AIO_EVENTS_PER_PAGE; \
187 __event; \
188})
189
190#define put_aio_ring_event(event, km) do { \
191 struct io_event *__event = (event); \
192 (void)__event; \
193 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
194} while(0)
195
196/* ioctx_alloc
197 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
198 */
199static struct kioctx *ioctx_alloc(unsigned nr_events)
200{
201 struct mm_struct *mm;
202 struct kioctx *ctx;
203
204 /* Prevent overflows */
205 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
206 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
207 pr_debug("ENOMEM: nr_events too high\n");
208 return ERR_PTR(-EINVAL);
209 }
210
Zach Brownd55b5fd2005-11-07 00:59:31 -0800211 if ((unsigned long)nr_events > aio_max_nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212 return ERR_PTR(-EAGAIN);
213
214 ctx = kmem_cache_alloc(kioctx_cachep, GFP_KERNEL);
215 if (!ctx)
216 return ERR_PTR(-ENOMEM);
217
218 memset(ctx, 0, sizeof(*ctx));
219 ctx->max_reqs = nr_events;
220 mm = ctx->mm = current->mm;
221 atomic_inc(&mm->mm_count);
222
223 atomic_set(&ctx->users, 1);
224 spin_lock_init(&ctx->ctx_lock);
225 spin_lock_init(&ctx->ring_info.ring_lock);
226 init_waitqueue_head(&ctx->wait);
227
228 INIT_LIST_HEAD(&ctx->active_reqs);
229 INIT_LIST_HEAD(&ctx->run_list);
230 INIT_WORK(&ctx->wq, aio_kick_handler, ctx);
231
232 if (aio_setup_ring(ctx) < 0)
233 goto out_freectx;
234
235 /* limit the number of system wide aios */
Zach Brownd55b5fd2005-11-07 00:59:31 -0800236 spin_lock(&aio_nr_lock);
237 if (aio_nr + ctx->max_reqs > aio_max_nr ||
238 aio_nr + ctx->max_reqs < aio_nr)
239 ctx->max_reqs = 0;
240 else
241 aio_nr += ctx->max_reqs;
242 spin_unlock(&aio_nr_lock);
243 if (ctx->max_reqs == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244 goto out_cleanup;
245
246 /* now link into global list. kludge. FIXME */
247 write_lock(&mm->ioctx_list_lock);
248 ctx->next = mm->ioctx_list;
249 mm->ioctx_list = ctx;
250 write_unlock(&mm->ioctx_list_lock);
251
252 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
253 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
254 return ctx;
255
256out_cleanup:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257 __put_ioctx(ctx);
258 return ERR_PTR(-EAGAIN);
259
260out_freectx:
261 mmdrop(mm);
262 kmem_cache_free(kioctx_cachep, ctx);
263 ctx = ERR_PTR(-ENOMEM);
264
265 dprintk("aio: error allocating ioctx %p\n", ctx);
266 return ctx;
267}
268
269/* aio_cancel_all
270 * Cancels all outstanding aio requests on an aio context. Used
271 * when the processes owning a context have all exited to encourage
272 * the rapid destruction of the kioctx.
273 */
274static void aio_cancel_all(struct kioctx *ctx)
275{
276 int (*cancel)(struct kiocb *, struct io_event *);
277 struct io_event res;
278 spin_lock_irq(&ctx->ctx_lock);
279 ctx->dead = 1;
280 while (!list_empty(&ctx->active_reqs)) {
281 struct list_head *pos = ctx->active_reqs.next;
282 struct kiocb *iocb = list_kiocb(pos);
283 list_del_init(&iocb->ki_list);
284 cancel = iocb->ki_cancel;
285 kiocbSetCancelled(iocb);
286 if (cancel) {
287 iocb->ki_users++;
288 spin_unlock_irq(&ctx->ctx_lock);
289 cancel(iocb, &res);
290 spin_lock_irq(&ctx->ctx_lock);
291 }
292 }
293 spin_unlock_irq(&ctx->ctx_lock);
294}
295
Adrian Bunk25ee7e32005-04-25 08:18:14 -0700296static void wait_for_all_aios(struct kioctx *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297{
298 struct task_struct *tsk = current;
299 DECLARE_WAITQUEUE(wait, tsk);
300
301 if (!ctx->reqs_active)
302 return;
303
304 add_wait_queue(&ctx->wait, &wait);
305 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
306 while (ctx->reqs_active) {
307 schedule();
308 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
309 }
310 __set_task_state(tsk, TASK_RUNNING);
311 remove_wait_queue(&ctx->wait, &wait);
312}
313
314/* wait_on_sync_kiocb:
315 * Waits on the given sync kiocb to complete.
316 */
317ssize_t fastcall wait_on_sync_kiocb(struct kiocb *iocb)
318{
319 while (iocb->ki_users) {
320 set_current_state(TASK_UNINTERRUPTIBLE);
321 if (!iocb->ki_users)
322 break;
323 schedule();
324 }
325 __set_current_state(TASK_RUNNING);
326 return iocb->ki_user_data;
327}
328
329/* exit_aio: called when the last user of mm goes away. At this point,
330 * there is no way for any new requests to be submited or any of the
331 * io_* syscalls to be called on the context. However, there may be
332 * outstanding requests which hold references to the context; as they
333 * go away, they will call put_ioctx and release any pinned memory
334 * associated with the request (held via struct page * references).
335 */
336void fastcall exit_aio(struct mm_struct *mm)
337{
338 struct kioctx *ctx = mm->ioctx_list;
339 mm->ioctx_list = NULL;
340 while (ctx) {
341 struct kioctx *next = ctx->next;
342 ctx->next = NULL;
343 aio_cancel_all(ctx);
344
345 wait_for_all_aios(ctx);
346 /*
347 * this is an overkill, but ensures we don't leave
348 * the ctx on the aio_wq
349 */
350 flush_workqueue(aio_wq);
351
352 if (1 != atomic_read(&ctx->users))
353 printk(KERN_DEBUG
354 "exit_aio:ioctx still alive: %d %d %d\n",
355 atomic_read(&ctx->users), ctx->dead,
356 ctx->reqs_active);
357 put_ioctx(ctx);
358 ctx = next;
359 }
360}
361
362/* __put_ioctx
363 * Called when the last user of an aio context has gone away,
364 * and the struct needs to be freed.
365 */
366void fastcall __put_ioctx(struct kioctx *ctx)
367{
368 unsigned nr_events = ctx->max_reqs;
369
Eric Sesterhenn93e06b42006-11-30 05:29:23 +0100370 BUG_ON(ctx->reqs_active);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700371
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;
Badari Pulavartyeed4e512006-09-30 23:28:49 -0700417 req->ki_iovec = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418 INIT_LIST_HEAD(&req->ki_run_list);
419
420 /* Check if the completion queue has enough free space to
421 * accept an event from this io.
422 */
423 spin_lock_irq(&ctx->ctx_lock);
424 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
425 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
426 list_add(&req->ki_list, &ctx->active_reqs);
427 get_ioctx(ctx);
428 ctx->reqs_active++;
429 okay = 1;
430 }
431 kunmap_atomic(ring, KM_USER0);
432 spin_unlock_irq(&ctx->ctx_lock);
433
434 if (!okay) {
435 kmem_cache_free(kiocb_cachep, req);
436 req = NULL;
437 }
438
439 return req;
440}
441
442static inline struct kiocb *aio_get_req(struct kioctx *ctx)
443{
444 struct kiocb *req;
445 /* Handle a potential starvation case -- should be exceedingly rare as
446 * requests will be stuck on fput_head only if the aio_fput_routine is
447 * delayed and the requests were the last user of the struct file.
448 */
449 req = __aio_get_req(ctx);
450 if (unlikely(NULL == req)) {
451 aio_fput_routine(NULL);
452 req = __aio_get_req(ctx);
453 }
454 return req;
455}
456
457static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
458{
Zach Brownd00689a2005-11-13 16:07:34 -0800459 assert_spin_locked(&ctx->ctx_lock);
460
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461 if (req->ki_dtor)
462 req->ki_dtor(req);
Badari Pulavartyeed4e512006-09-30 23:28:49 -0700463 if (req->ki_iovec != &req->ki_inline_vec)
464 kfree(req->ki_iovec);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465 kmem_cache_free(kiocb_cachep, req);
466 ctx->reqs_active--;
467
468 if (unlikely(!ctx->reqs_active && ctx->dead))
469 wake_up(&ctx->wait);
470}
471
472static void aio_fput_routine(void *data)
473{
474 spin_lock_irq(&fput_lock);
475 while (likely(!list_empty(&fput_head))) {
476 struct kiocb *req = list_kiocb(fput_head.next);
477 struct kioctx *ctx = req->ki_ctx;
478
479 list_del(&req->ki_list);
480 spin_unlock_irq(&fput_lock);
481
482 /* Complete the fput */
483 __fput(req->ki_filp);
484
485 /* Link the iocb into the context's free list */
486 spin_lock_irq(&ctx->ctx_lock);
487 really_put_req(ctx, req);
488 spin_unlock_irq(&ctx->ctx_lock);
489
490 put_ioctx(ctx);
491 spin_lock_irq(&fput_lock);
492 }
493 spin_unlock_irq(&fput_lock);
494}
495
496/* __aio_put_req
497 * Returns true if this put was the last user of the request.
498 */
499static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
500{
501 dprintk(KERN_DEBUG "aio_put(%p): f_count=%d\n",
502 req, atomic_read(&req->ki_filp->f_count));
503
Zach Brownd00689a2005-11-13 16:07:34 -0800504 assert_spin_locked(&ctx->ctx_lock);
505
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506 req->ki_users --;
Eric Sesterhenn93e06b42006-11-30 05:29:23 +0100507 BUG_ON(req->ki_users < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508 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 */
Nick Piggin095975d2006-01-08 01:02:19 -0800517 if (unlikely(atomic_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
Andreas Mohrd6e05ed2006-06-26 18:35:02 +0200646 * already held. The lock is released and reacquired
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 * 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)) {
Zach Brown691578c2006-10-03 01:16:04 -0700676 pr_debug("%ld retry: %zd of %zd\n", iocb->ki_retried,
Ken Chen644d3a02005-05-01 08:59:15 -0700677 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;
Oleg Nesterov626ab0e2006-06-23 02:05:55 -0700782 struct list_head run_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700783
Zach Brownd00689a2005-11-13 16:07:34 -0800784 assert_spin_locked(&ctx->ctx_lock);
785
Oleg Nesterov626ab0e2006-06-23 02:05:55 -0700786 list_replace_init(&ctx->run_list, &run_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787 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
Zach Brown691578c2006-10-03 01:16:04 -07001009 pr_debug("%ld retries: %zd of %zd\n", iocb->ki_retried,
Ken Chen644d3a02005-05-01 08:59:15 -07001010 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
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001305static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1306{
1307 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1308
1309 BUG_ON(ret <= 0);
1310
1311 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1312 ssize_t this = min((ssize_t)iov->iov_len, ret);
1313 iov->iov_base += this;
1314 iov->iov_len -= this;
1315 iocb->ki_left -= this;
1316 ret -= this;
1317 if (iov->iov_len == 0) {
1318 iocb->ki_cur_seg++;
1319 iov++;
1320 }
1321 }
1322
1323 /* the caller should not have done more io than what fit in
1324 * the remaining iovecs */
1325 BUG_ON(ret > 0 && iocb->ki_left == 0);
1326}
1327
1328static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001329{
1330 struct file *file = iocb->ki_filp;
1331 struct address_space *mapping = file->f_mapping;
1332 struct inode *inode = mapping->host;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001333 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1334 unsigned long, loff_t);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001335 ssize_t ret = 0;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001336 unsigned short opcode;
1337
1338 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1339 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1340 rw_op = file->f_op->aio_read;
1341 opcode = IOCB_CMD_PREADV;
1342 } else {
1343 rw_op = file->f_op->aio_write;
1344 opcode = IOCB_CMD_PWRITEV;
1345 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346
Zach Brown897f15f2005-09-30 11:58:55 -07001347 do {
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001348 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1349 iocb->ki_nr_segs - iocb->ki_cur_seg,
1350 iocb->ki_pos);
1351 if (ret > 0)
1352 aio_advance_iovec(iocb, ret);
Badari Pulavarty027445c2006-09-30 23:28:46 -07001353
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001354 /* retry all partial writes. retry partial reads as long as its a
1355 * regular file. */
Zach Brown353fb072005-09-30 11:58:56 -07001356 } while (ret > 0 && iocb->ki_left > 0 &&
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001357 (opcode == IOCB_CMD_PWRITEV ||
1358 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359
1360 /* This means we must have transferred all that we could */
1361 /* No need to retry anymore */
1362 if ((ret == 0) || (iocb->ki_left == 0))
1363 ret = iocb->ki_nbytes - iocb->ki_left;
1364
1365 return ret;
1366}
1367
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368static 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
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001388static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb)
1389{
1390 ssize_t ret;
1391
1392 ret = rw_copy_check_uvector(type, (struct iovec __user *)kiocb->ki_buf,
1393 kiocb->ki_nbytes, 1,
1394 &kiocb->ki_inline_vec, &kiocb->ki_iovec);
1395 if (ret < 0)
1396 goto out;
1397
1398 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1399 kiocb->ki_cur_seg = 0;
1400 /* ki_nbytes/left now reflect bytes instead of segs */
1401 kiocb->ki_nbytes = ret;
1402 kiocb->ki_left = ret;
1403
1404 ret = 0;
1405out:
1406 return ret;
1407}
1408
1409static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
1410{
1411 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1412 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1413 kiocb->ki_iovec->iov_len = kiocb->ki_left;
1414 kiocb->ki_nr_segs = 1;
1415 kiocb->ki_cur_seg = 0;
1416 kiocb->ki_nbytes = kiocb->ki_left;
1417 return 0;
1418}
1419
Linus Torvalds1da177e2005-04-16 15:20:36 -07001420/*
1421 * aio_setup_iocb:
1422 * Performs the initial checks and aio retry method
1423 * setup for the kiocb at the time of io submission.
1424 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001425static ssize_t aio_setup_iocb(struct kiocb *kiocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001426{
1427 struct file *file = kiocb->ki_filp;
1428 ssize_t ret = 0;
1429
1430 switch (kiocb->ki_opcode) {
1431 case IOCB_CMD_PREAD:
1432 ret = -EBADF;
1433 if (unlikely(!(file->f_mode & FMODE_READ)))
1434 break;
1435 ret = -EFAULT;
1436 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1437 kiocb->ki_left)))
1438 break;
Kostik Belousov8766ce42005-10-23 12:57:13 -07001439 ret = security_file_permission(file, MAY_READ);
1440 if (unlikely(ret))
1441 break;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001442 ret = aio_setup_single_vector(kiocb);
1443 if (ret)
1444 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001445 ret = -EINVAL;
1446 if (file->f_op->aio_read)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001447 kiocb->ki_retry = aio_rw_vect_retry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001448 break;
1449 case IOCB_CMD_PWRITE:
1450 ret = -EBADF;
1451 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1452 break;
1453 ret = -EFAULT;
1454 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1455 kiocb->ki_left)))
1456 break;
Kostik Belousov8766ce42005-10-23 12:57:13 -07001457 ret = security_file_permission(file, MAY_WRITE);
1458 if (unlikely(ret))
1459 break;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001460 ret = aio_setup_single_vector(kiocb);
1461 if (ret)
1462 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463 ret = -EINVAL;
1464 if (file->f_op->aio_write)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001465 kiocb->ki_retry = aio_rw_vect_retry;
1466 break;
1467 case IOCB_CMD_PREADV:
1468 ret = -EBADF;
1469 if (unlikely(!(file->f_mode & FMODE_READ)))
1470 break;
1471 ret = security_file_permission(file, MAY_READ);
1472 if (unlikely(ret))
1473 break;
1474 ret = aio_setup_vectored_rw(READ, kiocb);
1475 if (ret)
1476 break;
1477 ret = -EINVAL;
1478 if (file->f_op->aio_read)
1479 kiocb->ki_retry = aio_rw_vect_retry;
1480 break;
1481 case IOCB_CMD_PWRITEV:
1482 ret = -EBADF;
1483 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1484 break;
1485 ret = security_file_permission(file, MAY_WRITE);
1486 if (unlikely(ret))
1487 break;
1488 ret = aio_setup_vectored_rw(WRITE, kiocb);
1489 if (ret)
1490 break;
1491 ret = -EINVAL;
1492 if (file->f_op->aio_write)
1493 kiocb->ki_retry = aio_rw_vect_retry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494 break;
1495 case IOCB_CMD_FDSYNC:
1496 ret = -EINVAL;
1497 if (file->f_op->aio_fsync)
1498 kiocb->ki_retry = aio_fdsync;
1499 break;
1500 case IOCB_CMD_FSYNC:
1501 ret = -EINVAL;
1502 if (file->f_op->aio_fsync)
1503 kiocb->ki_retry = aio_fsync;
1504 break;
1505 default:
1506 dprintk("EINVAL: io_submit: no operation provided\n");
1507 ret = -EINVAL;
1508 }
1509
1510 if (!kiocb->ki_retry)
1511 return ret;
1512
1513 return 0;
1514}
1515
1516/*
1517 * aio_wake_function:
1518 * wait queue callback function for aio notification,
1519 * Simply triggers a retry of the operation via kick_iocb.
1520 *
1521 * This callback is specified in the wait queue entry in
1522 * a kiocb (current->io_wait points to this wait queue
1523 * entry when an aio operation executes; it is used
1524 * instead of a synchronous wait when an i/o blocking
1525 * condition is encountered during aio).
1526 *
1527 * Note:
1528 * This routine is executed with the wait queue lock held.
1529 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests
1530 * the ioctx lock inside the wait queue lock. This is safe
1531 * because this callback isn't used for wait queues which
1532 * are nested inside ioctx lock (i.e. ctx->wait)
1533 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001534static int aio_wake_function(wait_queue_t *wait, unsigned mode,
1535 int sync, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001536{
1537 struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait);
1538
1539 list_del_init(&wait->task_list);
1540 kick_iocb(iocb);
1541 return 1;
1542}
1543
1544int fastcall io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1545 struct iocb *iocb)
1546{
1547 struct kiocb *req;
1548 struct file *file;
1549 ssize_t ret;
1550
1551 /* enforce forwards compatibility on users */
1552 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2 ||
1553 iocb->aio_reserved3)) {
1554 pr_debug("EINVAL: io_submit: reserve field set\n");
1555 return -EINVAL;
1556 }
1557
1558 /* prevent overflows */
1559 if (unlikely(
1560 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1561 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1562 ((ssize_t)iocb->aio_nbytes < 0)
1563 )) {
1564 pr_debug("EINVAL: io_submit: overflow check\n");
1565 return -EINVAL;
1566 }
1567
1568 file = fget(iocb->aio_fildes);
1569 if (unlikely(!file))
1570 return -EBADF;
1571
1572 req = aio_get_req(ctx); /* returns with 2 references to req */
1573 if (unlikely(!req)) {
1574 fput(file);
1575 return -EAGAIN;
1576 }
1577
1578 req->ki_filp = file;
Ken Chen212079c2005-05-01 08:59:15 -07001579 ret = put_user(req->ki_key, &user_iocb->aio_key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580 if (unlikely(ret)) {
1581 dprintk("EFAULT: aio_key\n");
1582 goto out_put_req;
1583 }
1584
1585 req->ki_obj.user = user_iocb;
1586 req->ki_user_data = iocb->aio_data;
1587 req->ki_pos = iocb->aio_offset;
1588
1589 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1590 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1591 req->ki_opcode = iocb->aio_lio_opcode;
1592 init_waitqueue_func_entry(&req->ki_wait, aio_wake_function);
1593 INIT_LIST_HEAD(&req->ki_wait.task_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001594 req->ki_retried = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001595
1596 ret = aio_setup_iocb(req);
1597
1598 if (ret)
1599 goto out_put_req;
1600
1601 spin_lock_irq(&ctx->ctx_lock);
Benjamin LaHaiseac0b1bc2005-09-09 13:02:09 -07001602 aio_run_iocb(req);
Benjamin LaHaiseac0b1bc2005-09-09 13:02:09 -07001603 if (!list_empty(&ctx->run_list)) {
Ken Chen954d3e92005-05-01 08:59:16 -07001604 /* drain the run list */
1605 while (__aio_run_iocbs(ctx))
1606 ;
1607 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001608 spin_unlock_irq(&ctx->ctx_lock);
1609 aio_put_req(req); /* drop extra ref to req */
1610 return 0;
1611
1612out_put_req:
1613 aio_put_req(req); /* drop extra ref to req */
1614 aio_put_req(req); /* drop i/o ref to req */
1615 return ret;
1616}
1617
1618/* sys_io_submit:
1619 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1620 * the number of iocbs queued. May return -EINVAL if the aio_context
1621 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1622 * *iocbpp[0] is not properly initialized, if the operation specified
1623 * is invalid for the file descriptor in the iocb. May fail with
1624 * -EFAULT if any of the data structures point to invalid data. May
1625 * fail with -EBADF if the file descriptor specified in the first
1626 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1627 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1628 * fail with -ENOSYS if not implemented.
1629 */
1630asmlinkage long sys_io_submit(aio_context_t ctx_id, long nr,
1631 struct iocb __user * __user *iocbpp)
1632{
1633 struct kioctx *ctx;
1634 long ret = 0;
1635 int i;
1636
1637 if (unlikely(nr < 0))
1638 return -EINVAL;
1639
1640 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1641 return -EFAULT;
1642
1643 ctx = lookup_ioctx(ctx_id);
1644 if (unlikely(!ctx)) {
1645 pr_debug("EINVAL: io_submit: invalid context id\n");
1646 return -EINVAL;
1647 }
1648
1649 /*
1650 * AKPM: should this return a partial result if some of the IOs were
1651 * successfully submitted?
1652 */
1653 for (i=0; i<nr; i++) {
1654 struct iocb __user *user_iocb;
1655 struct iocb tmp;
1656
1657 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1658 ret = -EFAULT;
1659 break;
1660 }
1661
1662 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1663 ret = -EFAULT;
1664 break;
1665 }
1666
1667 ret = io_submit_one(ctx, user_iocb, &tmp);
1668 if (ret)
1669 break;
1670 }
1671
1672 put_ioctx(ctx);
1673 return i ? i : ret;
1674}
1675
1676/* lookup_kiocb
1677 * Finds a given iocb for cancellation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001679static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1680 u32 key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681{
1682 struct list_head *pos;
Zach Brownd00689a2005-11-13 16:07:34 -08001683
1684 assert_spin_locked(&ctx->ctx_lock);
1685
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686 /* TODO: use a hash or array, this sucks. */
1687 list_for_each(pos, &ctx->active_reqs) {
1688 struct kiocb *kiocb = list_kiocb(pos);
1689 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1690 return kiocb;
1691 }
1692 return NULL;
1693}
1694
1695/* sys_io_cancel:
1696 * Attempts to cancel an iocb previously passed to io_submit. If
1697 * the operation is successfully cancelled, the resulting event is
1698 * copied into the memory pointed to by result without being placed
1699 * into the completion queue and 0 is returned. May fail with
1700 * -EFAULT if any of the data structures pointed to are invalid.
1701 * May fail with -EINVAL if aio_context specified by ctx_id is
1702 * invalid. May fail with -EAGAIN if the iocb specified was not
1703 * cancelled. Will fail with -ENOSYS if not implemented.
1704 */
1705asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb,
1706 struct io_event __user *result)
1707{
1708 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1709 struct kioctx *ctx;
1710 struct kiocb *kiocb;
1711 u32 key;
1712 int ret;
1713
1714 ret = get_user(key, &iocb->aio_key);
1715 if (unlikely(ret))
1716 return -EFAULT;
1717
1718 ctx = lookup_ioctx(ctx_id);
1719 if (unlikely(!ctx))
1720 return -EINVAL;
1721
1722 spin_lock_irq(&ctx->ctx_lock);
1723 ret = -EAGAIN;
1724 kiocb = lookup_kiocb(ctx, iocb, key);
1725 if (kiocb && kiocb->ki_cancel) {
1726 cancel = kiocb->ki_cancel;
1727 kiocb->ki_users ++;
1728 kiocbSetCancelled(kiocb);
1729 } else
1730 cancel = NULL;
1731 spin_unlock_irq(&ctx->ctx_lock);
1732
1733 if (NULL != cancel) {
1734 struct io_event tmp;
1735 pr_debug("calling cancel\n");
1736 memset(&tmp, 0, sizeof(tmp));
1737 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1738 tmp.data = kiocb->ki_user_data;
1739 ret = cancel(kiocb, &tmp);
1740 if (!ret) {
1741 /* Cancellation succeeded -- copy the result
1742 * into the user's buffer.
1743 */
1744 if (copy_to_user(result, &tmp, sizeof(tmp)))
1745 ret = -EFAULT;
1746 }
1747 } else
Wendy Cheng8f582022005-09-09 13:02:08 -07001748 ret = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001749
1750 put_ioctx(ctx);
1751
1752 return ret;
1753}
1754
1755/* io_getevents:
1756 * Attempts to read at least min_nr events and up to nr events from
1757 * the completion queue for the aio_context specified by ctx_id. May
1758 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range,
1759 * if nr is out of range, if when is out of range. May fail with
1760 * -EFAULT if any of the memory specified to is invalid. May return
1761 * 0 or < min_nr if no events are available and the timeout specified
1762 * by when has elapsed, where when == NULL specifies an infinite
1763 * timeout. Note that the timeout pointed to by when is relative and
1764 * will be updated if not NULL and the operation blocks. Will fail
1765 * with -ENOSYS if not implemented.
1766 */
1767asmlinkage long sys_io_getevents(aio_context_t ctx_id,
1768 long min_nr,
1769 long nr,
1770 struct io_event __user *events,
1771 struct timespec __user *timeout)
1772{
1773 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1774 long ret = -EINVAL;
1775
1776 if (likely(ioctx)) {
1777 if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0))
1778 ret = read_events(ioctx, min_nr, nr, events, timeout);
1779 put_ioctx(ioctx);
1780 }
1781
1782 return ret;
1783}
1784
1785__initcall(aio_setup);
1786
1787EXPORT_SYMBOL(aio_complete);
1788EXPORT_SYMBOL(aio_put_req);
1789EXPORT_SYMBOL(wait_on_sync_kiocb);