blob: 94766599db00d3ca530eb37da388e1e63b900d35 [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
370 if (unlikely(ctx->reqs_active))
371 BUG();
372
373 cancel_delayed_work(&ctx->wq);
374 flush_workqueue(aio_wq);
375 aio_free_ring(ctx);
376 mmdrop(ctx->mm);
377 ctx->mm = NULL;
378 pr_debug("__put_ioctx: freeing %p\n", ctx);
379 kmem_cache_free(kioctx_cachep, ctx);
380
Zach Brownd55b5fd2005-11-07 00:59:31 -0800381 if (nr_events) {
382 spin_lock(&aio_nr_lock);
383 BUG_ON(aio_nr - nr_events > aio_nr);
384 aio_nr -= nr_events;
385 spin_unlock(&aio_nr_lock);
386 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387}
388
389/* aio_get_req
390 * Allocate a slot for an aio request. Increments the users count
391 * of the kioctx so that the kioctx stays around until all requests are
392 * complete. Returns NULL if no requests are free.
393 *
394 * Returns with kiocb->users set to 2. The io submit code path holds
395 * an extra reference while submitting the i/o.
396 * This prevents races between the aio code path referencing the
397 * req (after submitting it) and aio_complete() freeing the req.
398 */
399static struct kiocb *FASTCALL(__aio_get_req(struct kioctx *ctx));
400static struct kiocb fastcall *__aio_get_req(struct kioctx *ctx)
401{
402 struct kiocb *req = NULL;
403 struct aio_ring *ring;
404 int okay = 0;
405
406 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
407 if (unlikely(!req))
408 return NULL;
409
Zach Brown4faa5282005-10-17 16:43:33 -0700410 req->ki_flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411 req->ki_users = 2;
412 req->ki_key = 0;
413 req->ki_ctx = ctx;
414 req->ki_cancel = NULL;
415 req->ki_retry = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416 req->ki_dtor = NULL;
417 req->private = NULL;
Badari Pulavartyeed4e512006-09-30 23:28:49 -0700418 req->ki_iovec = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419 INIT_LIST_HEAD(&req->ki_run_list);
420
421 /* Check if the completion queue has enough free space to
422 * accept an event from this io.
423 */
424 spin_lock_irq(&ctx->ctx_lock);
425 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
426 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
427 list_add(&req->ki_list, &ctx->active_reqs);
428 get_ioctx(ctx);
429 ctx->reqs_active++;
430 okay = 1;
431 }
432 kunmap_atomic(ring, KM_USER0);
433 spin_unlock_irq(&ctx->ctx_lock);
434
435 if (!okay) {
436 kmem_cache_free(kiocb_cachep, req);
437 req = NULL;
438 }
439
440 return req;
441}
442
443static inline struct kiocb *aio_get_req(struct kioctx *ctx)
444{
445 struct kiocb *req;
446 /* Handle a potential starvation case -- should be exceedingly rare as
447 * requests will be stuck on fput_head only if the aio_fput_routine is
448 * delayed and the requests were the last user of the struct file.
449 */
450 req = __aio_get_req(ctx);
451 if (unlikely(NULL == req)) {
452 aio_fput_routine(NULL);
453 req = __aio_get_req(ctx);
454 }
455 return req;
456}
457
458static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
459{
Zach Brownd00689a2005-11-13 16:07:34 -0800460 assert_spin_locked(&ctx->ctx_lock);
461
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 if (req->ki_dtor)
463 req->ki_dtor(req);
Badari Pulavartyeed4e512006-09-30 23:28:49 -0700464 if (req->ki_iovec != &req->ki_inline_vec)
465 kfree(req->ki_iovec);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466 kmem_cache_free(kiocb_cachep, req);
467 ctx->reqs_active--;
468
469 if (unlikely(!ctx->reqs_active && ctx->dead))
470 wake_up(&ctx->wait);
471}
472
473static void aio_fput_routine(void *data)
474{
475 spin_lock_irq(&fput_lock);
476 while (likely(!list_empty(&fput_head))) {
477 struct kiocb *req = list_kiocb(fput_head.next);
478 struct kioctx *ctx = req->ki_ctx;
479
480 list_del(&req->ki_list);
481 spin_unlock_irq(&fput_lock);
482
483 /* Complete the fput */
484 __fput(req->ki_filp);
485
486 /* Link the iocb into the context's free list */
487 spin_lock_irq(&ctx->ctx_lock);
488 really_put_req(ctx, req);
489 spin_unlock_irq(&ctx->ctx_lock);
490
491 put_ioctx(ctx);
492 spin_lock_irq(&fput_lock);
493 }
494 spin_unlock_irq(&fput_lock);
495}
496
497/* __aio_put_req
498 * Returns true if this put was the last user of the request.
499 */
500static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
501{
502 dprintk(KERN_DEBUG "aio_put(%p): f_count=%d\n",
503 req, atomic_read(&req->ki_filp->f_count));
504
Zach Brownd00689a2005-11-13 16:07:34 -0800505 assert_spin_locked(&ctx->ctx_lock);
506
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507 req->ki_users --;
508 if (unlikely(req->ki_users < 0))
509 BUG();
510 if (likely(req->ki_users))
511 return 0;
512 list_del(&req->ki_list); /* remove from active_reqs */
513 req->ki_cancel = NULL;
514 req->ki_retry = NULL;
515
516 /* Must be done under the lock to serialise against cancellation.
517 * Call this aio_fput as it duplicates fput via the fput_work.
518 */
Nick Piggin095975d2006-01-08 01:02:19 -0800519 if (unlikely(atomic_dec_and_test(&req->ki_filp->f_count))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520 get_ioctx(ctx);
521 spin_lock(&fput_lock);
522 list_add(&req->ki_list, &fput_head);
523 spin_unlock(&fput_lock);
524 queue_work(aio_wq, &fput_work);
525 } else
526 really_put_req(ctx, req);
527 return 1;
528}
529
530/* aio_put_req
531 * Returns true if this put was the last user of the kiocb,
532 * false if the request is still in use.
533 */
534int fastcall aio_put_req(struct kiocb *req)
535{
536 struct kioctx *ctx = req->ki_ctx;
537 int ret;
538 spin_lock_irq(&ctx->ctx_lock);
539 ret = __aio_put_req(ctx, req);
540 spin_unlock_irq(&ctx->ctx_lock);
541 if (ret)
542 put_ioctx(ctx);
543 return ret;
544}
545
546/* Lookup an ioctx id. ioctx_list is lockless for reads.
547 * FIXME: this is O(n) and is only suitable for development.
548 */
549struct kioctx *lookup_ioctx(unsigned long ctx_id)
550{
551 struct kioctx *ioctx;
552 struct mm_struct *mm;
553
554 mm = current->mm;
555 read_lock(&mm->ioctx_list_lock);
556 for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next)
557 if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) {
558 get_ioctx(ioctx);
559 break;
560 }
561 read_unlock(&mm->ioctx_list_lock);
562
563 return ioctx;
564}
565
566/*
567 * use_mm
568 * Makes the calling kernel thread take on the specified
569 * mm context.
570 * Called by the retry thread execute retries within the
571 * iocb issuer's mm context, so that copy_from/to_user
572 * operations work seamlessly for aio.
573 * (Note: this routine is intended to be called only
574 * from a kernel thread context)
575 */
576static void use_mm(struct mm_struct *mm)
577{
578 struct mm_struct *active_mm;
579 struct task_struct *tsk = current;
580
581 task_lock(tsk);
582 tsk->flags |= PF_BORROWED_MM;
583 active_mm = tsk->active_mm;
584 atomic_inc(&mm->mm_count);
585 tsk->mm = mm;
586 tsk->active_mm = mm;
Paolo 'Blaisorblade' Giarrusso1e40cd32005-09-03 15:57:25 -0700587 /*
588 * Note that on UML this *requires* PF_BORROWED_MM to be set, otherwise
589 * it won't work. Update it accordingly if you change it here
590 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591 activate_mm(active_mm, mm);
592 task_unlock(tsk);
593
594 mmdrop(active_mm);
595}
596
597/*
598 * unuse_mm
599 * Reverses the effect of use_mm, i.e. releases the
600 * specified mm context which was earlier taken on
601 * by the calling kernel thread
602 * (Note: this routine is intended to be called only
603 * from a kernel thread context)
604 *
605 * Comments: Called with ctx->ctx_lock held. This nests
606 * task_lock instead ctx_lock.
607 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -0700608static void unuse_mm(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609{
610 struct task_struct *tsk = current;
611
612 task_lock(tsk);
613 tsk->flags &= ~PF_BORROWED_MM;
614 tsk->mm = NULL;
615 /* active_mm is still 'mm' */
616 enter_lazy_tlb(mm, tsk);
617 task_unlock(tsk);
618}
619
620/*
621 * Queue up a kiocb to be retried. Assumes that the kiocb
622 * has already been marked as kicked, and places it on
623 * the retry run list for the corresponding ioctx, if it
624 * isn't already queued. Returns 1 if it actually queued
625 * the kiocb (to tell the caller to activate the work
626 * queue to process it), or 0, if it found that it was
627 * already queued.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628 */
629static inline int __queue_kicked_iocb(struct kiocb *iocb)
630{
631 struct kioctx *ctx = iocb->ki_ctx;
632
Zach Brownd00689a2005-11-13 16:07:34 -0800633 assert_spin_locked(&ctx->ctx_lock);
634
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635 if (list_empty(&iocb->ki_run_list)) {
636 list_add_tail(&iocb->ki_run_list,
637 &ctx->run_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638 return 1;
639 }
640 return 0;
641}
642
643/* aio_run_iocb
644 * This is the core aio execution routine. It is
645 * invoked both for initial i/o submission and
646 * subsequent retries via the aio_kick_handler.
647 * Expects to be invoked with iocb->ki_ctx->lock
Andreas Mohrd6e05ed2006-06-26 18:35:02 +0200648 * already held. The lock is released and reacquired
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649 * as needed during processing.
650 *
651 * Calls the iocb retry method (already setup for the
652 * iocb on initial submission) for operation specific
653 * handling, but takes care of most of common retry
654 * execution details for a given iocb. The retry method
655 * needs to be non-blocking as far as possible, to avoid
656 * holding up other iocbs waiting to be serviced by the
657 * retry kernel thread.
658 *
659 * The trickier parts in this code have to do with
660 * ensuring that only one retry instance is in progress
661 * for a given iocb at any time. Providing that guarantee
662 * simplifies the coding of individual aio operations as
663 * it avoids various potential races.
664 */
665static ssize_t aio_run_iocb(struct kiocb *iocb)
666{
667 struct kioctx *ctx = iocb->ki_ctx;
668 ssize_t (*retry)(struct kiocb *);
669 ssize_t ret;
670
671 if (iocb->ki_retried++ > 1024*1024) {
672 printk("Maximal retry count. Bytes done %Zd\n",
673 iocb->ki_nbytes - iocb->ki_left);
674 return -EAGAIN;
675 }
676
677 if (!(iocb->ki_retried & 0xff)) {
Zach Brown691578c2006-10-03 01:16:04 -0700678 pr_debug("%ld retry: %zd of %zd\n", iocb->ki_retried,
Ken Chen644d3a02005-05-01 08:59:15 -0700679 iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680 }
681
682 if (!(retry = iocb->ki_retry)) {
683 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
684 return 0;
685 }
686
687 /*
688 * We don't want the next retry iteration for this
689 * operation to start until this one has returned and
690 * updated the iocb state. However, wait_queue functions
691 * can trigger a kick_iocb from interrupt context in the
692 * meantime, indicating that data is available for the next
693 * iteration. We want to remember that and enable the
694 * next retry iteration _after_ we are through with
695 * this one.
696 *
697 * So, in order to be able to register a "kick", but
698 * prevent it from being queued now, we clear the kick
699 * flag, but make the kick code *think* that the iocb is
700 * still on the run list until we are actually done.
701 * When we are done with this iteration, we check if
702 * the iocb was kicked in the meantime and if so, queue
703 * it up afresh.
704 */
705
706 kiocbClearKicked(iocb);
707
708 /*
709 * This is so that aio_complete knows it doesn't need to
710 * pull the iocb off the run list (We can't just call
711 * INIT_LIST_HEAD because we don't want a kick_iocb to
712 * queue this on the run list yet)
713 */
714 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
715 spin_unlock_irq(&ctx->ctx_lock);
716
717 /* Quit retrying if the i/o has been cancelled */
718 if (kiocbIsCancelled(iocb)) {
719 ret = -EINTR;
720 aio_complete(iocb, ret, 0);
721 /* must not access the iocb after this */
722 goto out;
723 }
724
725 /*
726 * Now we are all set to call the retry method in async
727 * context. By setting this thread's io_wait context
728 * to point to the wait queue entry inside the currently
729 * running iocb for the duration of the retry, we ensure
730 * that async notification wakeups are queued by the
731 * operation instead of blocking waits, and when notified,
732 * cause the iocb to be kicked for continuation (through
733 * the aio_wake_function callback).
734 */
735 BUG_ON(current->io_wait != NULL);
736 current->io_wait = &iocb->ki_wait;
737 ret = retry(iocb);
738 current->io_wait = NULL;
739
Zach Brown897f15f2005-09-30 11:58:55 -0700740 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
741 BUG_ON(!list_empty(&iocb->ki_wait.task_list));
742 aio_complete(iocb, ret, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743 }
744out:
745 spin_lock_irq(&ctx->ctx_lock);
746
747 if (-EIOCBRETRY == ret) {
748 /*
749 * OK, now that we are done with this iteration
750 * and know that there is more left to go,
751 * this is where we let go so that a subsequent
752 * "kick" can start the next iteration
753 */
754
755 /* will make __queue_kicked_iocb succeed from here on */
756 INIT_LIST_HEAD(&iocb->ki_run_list);
757 /* we must queue the next iteration ourselves, if it
758 * has already been kicked */
759 if (kiocbIsKicked(iocb)) {
760 __queue_kicked_iocb(iocb);
Sébastien Duguc016e222005-06-28 20:44:59 -0700761
762 /*
763 * __queue_kicked_iocb will always return 1 here, because
764 * iocb->ki_run_list is empty at this point so it should
765 * be safe to unconditionally queue the context into the
766 * work queue.
767 */
768 aio_queue_work(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700769 }
770 }
771 return ret;
772}
773
774/*
775 * __aio_run_iocbs:
776 * Process all pending retries queued on the ioctx
777 * run list.
778 * Assumes it is operating within the aio issuer's mm
Zach Brownd00689a2005-11-13 16:07:34 -0800779 * context.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780 */
781static int __aio_run_iocbs(struct kioctx *ctx)
782{
783 struct kiocb *iocb;
Oleg Nesterov626ab0e2006-06-23 02:05:55 -0700784 struct list_head run_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785
Zach Brownd00689a2005-11-13 16:07:34 -0800786 assert_spin_locked(&ctx->ctx_lock);
787
Oleg Nesterov626ab0e2006-06-23 02:05:55 -0700788 list_replace_init(&ctx->run_list, &run_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789 while (!list_empty(&run_list)) {
790 iocb = list_entry(run_list.next, struct kiocb,
791 ki_run_list);
792 list_del(&iocb->ki_run_list);
793 /*
794 * Hold an extra reference while retrying i/o.
795 */
796 iocb->ki_users++; /* grab extra reference */
797 aio_run_iocb(iocb);
798 if (__aio_put_req(ctx, iocb)) /* drop extra ref */
799 put_ioctx(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801 if (!list_empty(&ctx->run_list))
802 return 1;
803 return 0;
804}
805
806static void aio_queue_work(struct kioctx * ctx)
807{
808 unsigned long timeout;
809 /*
810 * if someone is waiting, get the work started right
811 * away, otherwise, use a longer delay
812 */
813 smp_mb();
814 if (waitqueue_active(&ctx->wait))
815 timeout = 1;
816 else
817 timeout = HZ/10;
818 queue_delayed_work(aio_wq, &ctx->wq, timeout);
819}
820
821
822/*
823 * aio_run_iocbs:
824 * Process all pending retries queued on the ioctx
825 * run list.
826 * Assumes it is operating within the aio issuer's mm
827 * context.
828 */
829static inline void aio_run_iocbs(struct kioctx *ctx)
830{
831 int requeue;
832
833 spin_lock_irq(&ctx->ctx_lock);
834
835 requeue = __aio_run_iocbs(ctx);
836 spin_unlock_irq(&ctx->ctx_lock);
837 if (requeue)
838 aio_queue_work(ctx);
839}
840
841/*
842 * just like aio_run_iocbs, but keeps running them until
843 * the list stays empty
844 */
845static inline void aio_run_all_iocbs(struct kioctx *ctx)
846{
847 spin_lock_irq(&ctx->ctx_lock);
848 while (__aio_run_iocbs(ctx))
849 ;
850 spin_unlock_irq(&ctx->ctx_lock);
851}
852
853/*
854 * aio_kick_handler:
855 * Work queue handler triggered to process pending
856 * retries on an ioctx. Takes on the aio issuer's
857 * mm context before running the iocbs, so that
858 * copy_xxx_user operates on the issuer's address
859 * space.
860 * Run on aiod's context.
861 */
862static void aio_kick_handler(void *data)
863{
864 struct kioctx *ctx = data;
865 mm_segment_t oldfs = get_fs();
866 int requeue;
867
868 set_fs(USER_DS);
869 use_mm(ctx->mm);
870 spin_lock_irq(&ctx->ctx_lock);
871 requeue =__aio_run_iocbs(ctx);
872 unuse_mm(ctx->mm);
873 spin_unlock_irq(&ctx->ctx_lock);
874 set_fs(oldfs);
875 /*
876 * we're in a worker thread already, don't use queue_delayed_work,
877 */
878 if (requeue)
879 queue_work(aio_wq, &ctx->wq);
880}
881
882
883/*
884 * Called by kick_iocb to queue the kiocb for retry
885 * and if required activate the aio work queue to process
886 * it
887 */
Zach Brown998765e2005-09-30 11:58:54 -0700888static void try_queue_kicked_iocb(struct kiocb *iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889{
890 struct kioctx *ctx = iocb->ki_ctx;
891 unsigned long flags;
892 int run = 0;
893
Zach Brown998765e2005-09-30 11:58:54 -0700894 /* We're supposed to be the only path putting the iocb back on the run
895 * list. If we find that the iocb is *back* on a wait queue already
896 * than retry has happened before we could queue the iocb. This also
897 * means that the retry could have completed and freed our iocb, no
898 * good. */
899 BUG_ON((!list_empty(&iocb->ki_wait.task_list)));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900
901 spin_lock_irqsave(&ctx->ctx_lock, flags);
Zach Brown998765e2005-09-30 11:58:54 -0700902 /* set this inside the lock so that we can't race with aio_run_iocb()
903 * testing it and putting the iocb on the run list under the lock */
904 if (!kiocbTryKick(iocb))
905 run = __queue_kicked_iocb(iocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
Ken Chen644d3a02005-05-01 08:59:15 -0700907 if (run)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700908 aio_queue_work(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909}
910
911/*
912 * kick_iocb:
913 * Called typically from a wait queue callback context
914 * (aio_wake_function) to trigger a retry of the iocb.
915 * The retry is usually executed by aio workqueue
916 * threads (See aio_kick_handler).
917 */
918void fastcall kick_iocb(struct kiocb *iocb)
919{
920 /* sync iocbs are easy: they can only ever be executing from a
921 * single context. */
922 if (is_sync_kiocb(iocb)) {
923 kiocbSetKicked(iocb);
924 wake_up_process(iocb->ki_obj.tsk);
925 return;
926 }
927
Zach Brown998765e2005-09-30 11:58:54 -0700928 try_queue_kicked_iocb(iocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929}
930EXPORT_SYMBOL(kick_iocb);
931
932/* aio_complete
933 * Called when the io request on the given iocb is complete.
934 * Returns true if this is the last user of the request. The
935 * only other user of the request can be the cancellation code.
936 */
937int fastcall aio_complete(struct kiocb *iocb, long res, long res2)
938{
939 struct kioctx *ctx = iocb->ki_ctx;
940 struct aio_ring_info *info;
941 struct aio_ring *ring;
942 struct io_event *event;
943 unsigned long flags;
944 unsigned long tail;
945 int ret;
946
Zach Brown20dcae32005-11-13 16:07:33 -0800947 /*
948 * Special case handling for sync iocbs:
949 * - events go directly into the iocb for fast handling
950 * - the sync task with the iocb in its stack holds the single iocb
951 * ref, no other paths have a way to get another ref
952 * - the sync task helpfully left a reference to itself in the iocb
Linus Torvalds1da177e2005-04-16 15:20:36 -0700953 */
954 if (is_sync_kiocb(iocb)) {
Zach Brown20dcae32005-11-13 16:07:33 -0800955 BUG_ON(iocb->ki_users != 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 iocb->ki_user_data = res;
Zach Brown20dcae32005-11-13 16:07:33 -0800957 iocb->ki_users = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 wake_up_process(iocb->ki_obj.tsk);
Zach Brown20dcae32005-11-13 16:07:33 -0800959 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960 }
961
962 info = &ctx->ring_info;
963
964 /* add a completion event to the ring buffer.
965 * must be done holding ctx->ctx_lock to prevent
966 * other code from messing with the tail
967 * pointer since we might be called from irq
968 * context.
969 */
970 spin_lock_irqsave(&ctx->ctx_lock, flags);
971
972 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
973 list_del_init(&iocb->ki_run_list);
974
975 /*
976 * cancelled requests don't get events, userland was given one
977 * when the event got cancelled.
978 */
979 if (kiocbIsCancelled(iocb))
980 goto put_rq;
981
982 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
983
984 tail = info->tail;
985 event = aio_ring_event(info, tail, KM_IRQ0);
Ken Chen4bf69b22005-05-01 08:59:15 -0700986 if (++tail >= info->nr)
987 tail = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700988
989 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
990 event->data = iocb->ki_user_data;
991 event->res = res;
992 event->res2 = res2;
993
994 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
995 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
996 res, res2);
997
998 /* after flagging the request as done, we
999 * must never even look at it again
1000 */
1001 smp_wmb(); /* make event visible before updating tail */
1002
1003 info->tail = tail;
1004 ring->tail = tail;
1005
1006 put_aio_ring_event(event, KM_IRQ0);
1007 kunmap_atomic(ring, KM_IRQ1);
1008
1009 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
1010
Zach Brown691578c2006-10-03 01:16:04 -07001011 pr_debug("%ld retries: %zd of %zd\n", iocb->ki_retried,
Ken Chen644d3a02005-05-01 08:59:15 -07001012 iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001013put_rq:
1014 /* everything turned out well, dispose of the aiocb. */
1015 ret = __aio_put_req(ctx, iocb);
1016
1017 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1018
1019 if (waitqueue_active(&ctx->wait))
1020 wake_up(&ctx->wait);
1021
1022 if (ret)
1023 put_ioctx(ctx);
1024
1025 return ret;
1026}
1027
1028/* aio_read_evt
1029 * Pull an event off of the ioctx's event ring. Returns the number of
1030 * events fetched (0 or 1 ;-)
1031 * FIXME: make this use cmpxchg.
1032 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1033 */
1034static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1035{
1036 struct aio_ring_info *info = &ioctx->ring_info;
1037 struct aio_ring *ring;
1038 unsigned long head;
1039 int ret = 0;
1040
1041 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1042 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1043 (unsigned long)ring->head, (unsigned long)ring->tail,
1044 (unsigned long)ring->nr);
1045
1046 if (ring->head == ring->tail)
1047 goto out;
1048
1049 spin_lock(&info->ring_lock);
1050
1051 head = ring->head % info->nr;
1052 if (head != ring->tail) {
1053 struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1054 *ent = *evp;
1055 head = (head + 1) % info->nr;
1056 smp_mb(); /* finish reading the event before updatng the head */
1057 ring->head = head;
1058 ret = 1;
1059 put_aio_ring_event(evp, KM_USER1);
1060 }
1061 spin_unlock(&info->ring_lock);
1062
1063out:
1064 kunmap_atomic(ring, KM_USER0);
1065 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1066 (unsigned long)ring->head, (unsigned long)ring->tail);
1067 return ret;
1068}
1069
1070struct aio_timeout {
1071 struct timer_list timer;
1072 int timed_out;
1073 struct task_struct *p;
1074};
1075
1076static void timeout_func(unsigned long data)
1077{
1078 struct aio_timeout *to = (struct aio_timeout *)data;
1079
1080 to->timed_out = 1;
1081 wake_up_process(to->p);
1082}
1083
1084static inline void init_timeout(struct aio_timeout *to)
1085{
1086 init_timer(&to->timer);
1087 to->timer.data = (unsigned long)to;
1088 to->timer.function = timeout_func;
1089 to->timed_out = 0;
1090 to->p = current;
1091}
1092
1093static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1094 const struct timespec *ts)
1095{
1096 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1097 if (time_after(to->timer.expires, jiffies))
1098 add_timer(&to->timer);
1099 else
1100 to->timed_out = 1;
1101}
1102
1103static inline void clear_timeout(struct aio_timeout *to)
1104{
1105 del_singleshot_timer_sync(&to->timer);
1106}
1107
1108static int read_events(struct kioctx *ctx,
1109 long min_nr, long nr,
1110 struct io_event __user *event,
1111 struct timespec __user *timeout)
1112{
1113 long start_jiffies = jiffies;
1114 struct task_struct *tsk = current;
1115 DECLARE_WAITQUEUE(wait, tsk);
1116 int ret;
1117 int i = 0;
1118 struct io_event ent;
1119 struct aio_timeout to;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120 int retry = 0;
1121
1122 /* needed to zero any padding within an entry (there shouldn't be
1123 * any, but C is fun!
1124 */
1125 memset(&ent, 0, sizeof(ent));
1126retry:
1127 ret = 0;
1128 while (likely(i < nr)) {
1129 ret = aio_read_evt(ctx, &ent);
1130 if (unlikely(ret <= 0))
1131 break;
1132
1133 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1134 ent.data, ent.obj, ent.res, ent.res2);
1135
1136 /* Could we split the check in two? */
1137 ret = -EFAULT;
1138 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1139 dprintk("aio: lost an event due to EFAULT.\n");
1140 break;
1141 }
1142 ret = 0;
1143
1144 /* Good, event copied to userland, update counts. */
1145 event ++;
1146 i ++;
1147 }
1148
1149 if (min_nr <= i)
1150 return i;
1151 if (ret)
1152 return ret;
1153
1154 /* End fast path */
1155
1156 /* racey check, but it gets redone */
1157 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1158 retry = 1;
1159 aio_run_all_iocbs(ctx);
1160 goto retry;
1161 }
1162
1163 init_timeout(&to);
1164 if (timeout) {
1165 struct timespec ts;
1166 ret = -EFAULT;
1167 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1168 goto out;
1169
1170 set_timeout(start_jiffies, &to, &ts);
1171 }
1172
1173 while (likely(i < nr)) {
1174 add_wait_queue_exclusive(&ctx->wait, &wait);
1175 do {
1176 set_task_state(tsk, TASK_INTERRUPTIBLE);
1177 ret = aio_read_evt(ctx, &ent);
1178 if (ret)
1179 break;
1180 if (min_nr <= i)
1181 break;
1182 ret = 0;
1183 if (to.timed_out) /* Only check after read evt */
1184 break;
1185 schedule();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186 if (signal_pending(tsk)) {
1187 ret = -EINTR;
1188 break;
1189 }
1190 /*ret = aio_read_evt(ctx, &ent);*/
1191 } while (1) ;
1192
1193 set_task_state(tsk, TASK_RUNNING);
1194 remove_wait_queue(&ctx->wait, &wait);
1195
1196 if (unlikely(ret <= 0))
1197 break;
1198
1199 ret = -EFAULT;
1200 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1201 dprintk("aio: lost an event due to EFAULT.\n");
1202 break;
1203 }
1204
1205 /* Good, event copied to userland, update counts. */
1206 event ++;
1207 i ++;
1208 }
1209
1210 if (timeout)
1211 clear_timeout(&to);
1212out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001213 return i ? i : ret;
1214}
1215
1216/* Take an ioctx and remove it from the list of ioctx's. Protects
1217 * against races with itself via ->dead.
1218 */
1219static void io_destroy(struct kioctx *ioctx)
1220{
1221 struct mm_struct *mm = current->mm;
1222 struct kioctx **tmp;
1223 int was_dead;
1224
1225 /* delete the entry from the list is someone else hasn't already */
1226 write_lock(&mm->ioctx_list_lock);
1227 was_dead = ioctx->dead;
1228 ioctx->dead = 1;
1229 for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx;
1230 tmp = &(*tmp)->next)
1231 ;
1232 if (*tmp)
1233 *tmp = ioctx->next;
1234 write_unlock(&mm->ioctx_list_lock);
1235
1236 dprintk("aio_release(%p)\n", ioctx);
1237 if (likely(!was_dead))
1238 put_ioctx(ioctx); /* twice for the list */
1239
1240 aio_cancel_all(ioctx);
1241 wait_for_all_aios(ioctx);
1242 put_ioctx(ioctx); /* once for the lookup */
1243}
1244
1245/* sys_io_setup:
1246 * Create an aio_context capable of receiving at least nr_events.
1247 * ctxp must not point to an aio_context that already exists, and
1248 * must be initialized to 0 prior to the call. On successful
1249 * creation of the aio_context, *ctxp is filled in with the resulting
1250 * handle. May fail with -EINVAL if *ctxp is not initialized,
1251 * if the specified nr_events exceeds internal limits. May fail
1252 * with -EAGAIN if the specified nr_events exceeds the user's limit
1253 * of available events. May fail with -ENOMEM if insufficient kernel
1254 * resources are available. May fail with -EFAULT if an invalid
1255 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1256 * implemented.
1257 */
1258asmlinkage long sys_io_setup(unsigned nr_events, aio_context_t __user *ctxp)
1259{
1260 struct kioctx *ioctx = NULL;
1261 unsigned long ctx;
1262 long ret;
1263
1264 ret = get_user(ctx, ctxp);
1265 if (unlikely(ret))
1266 goto out;
1267
1268 ret = -EINVAL;
Zach Brownd55b5fd2005-11-07 00:59:31 -08001269 if (unlikely(ctx || nr_events == 0)) {
1270 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1271 ctx, nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001272 goto out;
1273 }
1274
1275 ioctx = ioctx_alloc(nr_events);
1276 ret = PTR_ERR(ioctx);
1277 if (!IS_ERR(ioctx)) {
1278 ret = put_user(ioctx->user_id, ctxp);
1279 if (!ret)
1280 return 0;
1281
1282 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1283 io_destroy(ioctx);
1284 }
1285
1286out:
1287 return ret;
1288}
1289
1290/* sys_io_destroy:
1291 * Destroy the aio_context specified. May cancel any outstanding
1292 * AIOs and block on completion. Will fail with -ENOSYS if not
1293 * implemented. May fail with -EFAULT if the context pointed to
1294 * is invalid.
1295 */
1296asmlinkage long sys_io_destroy(aio_context_t ctx)
1297{
1298 struct kioctx *ioctx = lookup_ioctx(ctx);
1299 if (likely(NULL != ioctx)) {
1300 io_destroy(ioctx);
1301 return 0;
1302 }
1303 pr_debug("EINVAL: io_destroy: invalid context id\n");
1304 return -EINVAL;
1305}
1306
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001307static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1308{
1309 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1310
1311 BUG_ON(ret <= 0);
1312
1313 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1314 ssize_t this = min((ssize_t)iov->iov_len, ret);
1315 iov->iov_base += this;
1316 iov->iov_len -= this;
1317 iocb->ki_left -= this;
1318 ret -= this;
1319 if (iov->iov_len == 0) {
1320 iocb->ki_cur_seg++;
1321 iov++;
1322 }
1323 }
1324
1325 /* the caller should not have done more io than what fit in
1326 * the remaining iovecs */
1327 BUG_ON(ret > 0 && iocb->ki_left == 0);
1328}
1329
1330static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001331{
1332 struct file *file = iocb->ki_filp;
1333 struct address_space *mapping = file->f_mapping;
1334 struct inode *inode = mapping->host;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001335 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1336 unsigned long, loff_t);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337 ssize_t ret = 0;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001338 unsigned short opcode;
1339
1340 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1341 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1342 rw_op = file->f_op->aio_read;
1343 opcode = IOCB_CMD_PREADV;
1344 } else {
1345 rw_op = file->f_op->aio_write;
1346 opcode = IOCB_CMD_PWRITEV;
1347 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348
Zach Brown897f15f2005-09-30 11:58:55 -07001349 do {
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001350 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1351 iocb->ki_nr_segs - iocb->ki_cur_seg,
1352 iocb->ki_pos);
1353 if (ret > 0)
1354 aio_advance_iovec(iocb, ret);
Badari Pulavarty027445c2006-09-30 23:28:46 -07001355
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001356 /* retry all partial writes. retry partial reads as long as its a
1357 * regular file. */
Zach Brown353fb072005-09-30 11:58:56 -07001358 } while (ret > 0 && iocb->ki_left > 0 &&
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001359 (opcode == IOCB_CMD_PWRITEV ||
1360 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361
1362 /* This means we must have transferred all that we could */
1363 /* No need to retry anymore */
1364 if ((ret == 0) || (iocb->ki_left == 0))
1365 ret = iocb->ki_nbytes - iocb->ki_left;
1366
1367 return ret;
1368}
1369
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370static ssize_t aio_fdsync(struct kiocb *iocb)
1371{
1372 struct file *file = iocb->ki_filp;
1373 ssize_t ret = -EINVAL;
1374
1375 if (file->f_op->aio_fsync)
1376 ret = file->f_op->aio_fsync(iocb, 1);
1377 return ret;
1378}
1379
1380static ssize_t aio_fsync(struct kiocb *iocb)
1381{
1382 struct file *file = iocb->ki_filp;
1383 ssize_t ret = -EINVAL;
1384
1385 if (file->f_op->aio_fsync)
1386 ret = file->f_op->aio_fsync(iocb, 0);
1387 return ret;
1388}
1389
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001390static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb)
1391{
1392 ssize_t ret;
1393
1394 ret = rw_copy_check_uvector(type, (struct iovec __user *)kiocb->ki_buf,
1395 kiocb->ki_nbytes, 1,
1396 &kiocb->ki_inline_vec, &kiocb->ki_iovec);
1397 if (ret < 0)
1398 goto out;
1399
1400 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1401 kiocb->ki_cur_seg = 0;
1402 /* ki_nbytes/left now reflect bytes instead of segs */
1403 kiocb->ki_nbytes = ret;
1404 kiocb->ki_left = ret;
1405
1406 ret = 0;
1407out:
1408 return ret;
1409}
1410
1411static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
1412{
1413 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1414 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1415 kiocb->ki_iovec->iov_len = kiocb->ki_left;
1416 kiocb->ki_nr_segs = 1;
1417 kiocb->ki_cur_seg = 0;
1418 kiocb->ki_nbytes = kiocb->ki_left;
1419 return 0;
1420}
1421
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422/*
1423 * aio_setup_iocb:
1424 * Performs the initial checks and aio retry method
1425 * setup for the kiocb at the time of io submission.
1426 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001427static ssize_t aio_setup_iocb(struct kiocb *kiocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428{
1429 struct file *file = kiocb->ki_filp;
1430 ssize_t ret = 0;
1431
1432 switch (kiocb->ki_opcode) {
1433 case IOCB_CMD_PREAD:
1434 ret = -EBADF;
1435 if (unlikely(!(file->f_mode & FMODE_READ)))
1436 break;
1437 ret = -EFAULT;
1438 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1439 kiocb->ki_left)))
1440 break;
Kostik Belousov8766ce42005-10-23 12:57:13 -07001441 ret = security_file_permission(file, MAY_READ);
1442 if (unlikely(ret))
1443 break;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001444 ret = aio_setup_single_vector(kiocb);
1445 if (ret)
1446 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001447 ret = -EINVAL;
1448 if (file->f_op->aio_read)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001449 kiocb->ki_retry = aio_rw_vect_retry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450 break;
1451 case IOCB_CMD_PWRITE:
1452 ret = -EBADF;
1453 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1454 break;
1455 ret = -EFAULT;
1456 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1457 kiocb->ki_left)))
1458 break;
Kostik Belousov8766ce42005-10-23 12:57:13 -07001459 ret = security_file_permission(file, MAY_WRITE);
1460 if (unlikely(ret))
1461 break;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001462 ret = aio_setup_single_vector(kiocb);
1463 if (ret)
1464 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465 ret = -EINVAL;
1466 if (file->f_op->aio_write)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001467 kiocb->ki_retry = aio_rw_vect_retry;
1468 break;
1469 case IOCB_CMD_PREADV:
1470 ret = -EBADF;
1471 if (unlikely(!(file->f_mode & FMODE_READ)))
1472 break;
1473 ret = security_file_permission(file, MAY_READ);
1474 if (unlikely(ret))
1475 break;
1476 ret = aio_setup_vectored_rw(READ, kiocb);
1477 if (ret)
1478 break;
1479 ret = -EINVAL;
1480 if (file->f_op->aio_read)
1481 kiocb->ki_retry = aio_rw_vect_retry;
1482 break;
1483 case IOCB_CMD_PWRITEV:
1484 ret = -EBADF;
1485 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1486 break;
1487 ret = security_file_permission(file, MAY_WRITE);
1488 if (unlikely(ret))
1489 break;
1490 ret = aio_setup_vectored_rw(WRITE, kiocb);
1491 if (ret)
1492 break;
1493 ret = -EINVAL;
1494 if (file->f_op->aio_write)
1495 kiocb->ki_retry = aio_rw_vect_retry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496 break;
1497 case IOCB_CMD_FDSYNC:
1498 ret = -EINVAL;
1499 if (file->f_op->aio_fsync)
1500 kiocb->ki_retry = aio_fdsync;
1501 break;
1502 case IOCB_CMD_FSYNC:
1503 ret = -EINVAL;
1504 if (file->f_op->aio_fsync)
1505 kiocb->ki_retry = aio_fsync;
1506 break;
1507 default:
1508 dprintk("EINVAL: io_submit: no operation provided\n");
1509 ret = -EINVAL;
1510 }
1511
1512 if (!kiocb->ki_retry)
1513 return ret;
1514
1515 return 0;
1516}
1517
1518/*
1519 * aio_wake_function:
1520 * wait queue callback function for aio notification,
1521 * Simply triggers a retry of the operation via kick_iocb.
1522 *
1523 * This callback is specified in the wait queue entry in
1524 * a kiocb (current->io_wait points to this wait queue
1525 * entry when an aio operation executes; it is used
1526 * instead of a synchronous wait when an i/o blocking
1527 * condition is encountered during aio).
1528 *
1529 * Note:
1530 * This routine is executed with the wait queue lock held.
1531 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests
1532 * the ioctx lock inside the wait queue lock. This is safe
1533 * because this callback isn't used for wait queues which
1534 * are nested inside ioctx lock (i.e. ctx->wait)
1535 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001536static int aio_wake_function(wait_queue_t *wait, unsigned mode,
1537 int sync, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001538{
1539 struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait);
1540
1541 list_del_init(&wait->task_list);
1542 kick_iocb(iocb);
1543 return 1;
1544}
1545
1546int fastcall io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1547 struct iocb *iocb)
1548{
1549 struct kiocb *req;
1550 struct file *file;
1551 ssize_t ret;
1552
1553 /* enforce forwards compatibility on users */
1554 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2 ||
1555 iocb->aio_reserved3)) {
1556 pr_debug("EINVAL: io_submit: reserve field set\n");
1557 return -EINVAL;
1558 }
1559
1560 /* prevent overflows */
1561 if (unlikely(
1562 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1563 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1564 ((ssize_t)iocb->aio_nbytes < 0)
1565 )) {
1566 pr_debug("EINVAL: io_submit: overflow check\n");
1567 return -EINVAL;
1568 }
1569
1570 file = fget(iocb->aio_fildes);
1571 if (unlikely(!file))
1572 return -EBADF;
1573
1574 req = aio_get_req(ctx); /* returns with 2 references to req */
1575 if (unlikely(!req)) {
1576 fput(file);
1577 return -EAGAIN;
1578 }
1579
1580 req->ki_filp = file;
Ken Chen212079c2005-05-01 08:59:15 -07001581 ret = put_user(req->ki_key, &user_iocb->aio_key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001582 if (unlikely(ret)) {
1583 dprintk("EFAULT: aio_key\n");
1584 goto out_put_req;
1585 }
1586
1587 req->ki_obj.user = user_iocb;
1588 req->ki_user_data = iocb->aio_data;
1589 req->ki_pos = iocb->aio_offset;
1590
1591 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1592 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1593 req->ki_opcode = iocb->aio_lio_opcode;
1594 init_waitqueue_func_entry(&req->ki_wait, aio_wake_function);
1595 INIT_LIST_HEAD(&req->ki_wait.task_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001596 req->ki_retried = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597
1598 ret = aio_setup_iocb(req);
1599
1600 if (ret)
1601 goto out_put_req;
1602
1603 spin_lock_irq(&ctx->ctx_lock);
Benjamin LaHaiseac0b1bc2005-09-09 13:02:09 -07001604 aio_run_iocb(req);
Benjamin LaHaiseac0b1bc2005-09-09 13:02:09 -07001605 if (!list_empty(&ctx->run_list)) {
Ken Chen954d3e92005-05-01 08:59:16 -07001606 /* drain the run list */
1607 while (__aio_run_iocbs(ctx))
1608 ;
1609 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001610 spin_unlock_irq(&ctx->ctx_lock);
1611 aio_put_req(req); /* drop extra ref to req */
1612 return 0;
1613
1614out_put_req:
1615 aio_put_req(req); /* drop extra ref to req */
1616 aio_put_req(req); /* drop i/o ref to req */
1617 return ret;
1618}
1619
1620/* sys_io_submit:
1621 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1622 * the number of iocbs queued. May return -EINVAL if the aio_context
1623 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1624 * *iocbpp[0] is not properly initialized, if the operation specified
1625 * is invalid for the file descriptor in the iocb. May fail with
1626 * -EFAULT if any of the data structures point to invalid data. May
1627 * fail with -EBADF if the file descriptor specified in the first
1628 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1629 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1630 * fail with -ENOSYS if not implemented.
1631 */
1632asmlinkage long sys_io_submit(aio_context_t ctx_id, long nr,
1633 struct iocb __user * __user *iocbpp)
1634{
1635 struct kioctx *ctx;
1636 long ret = 0;
1637 int i;
1638
1639 if (unlikely(nr < 0))
1640 return -EINVAL;
1641
1642 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1643 return -EFAULT;
1644
1645 ctx = lookup_ioctx(ctx_id);
1646 if (unlikely(!ctx)) {
1647 pr_debug("EINVAL: io_submit: invalid context id\n");
1648 return -EINVAL;
1649 }
1650
1651 /*
1652 * AKPM: should this return a partial result if some of the IOs were
1653 * successfully submitted?
1654 */
1655 for (i=0; i<nr; i++) {
1656 struct iocb __user *user_iocb;
1657 struct iocb tmp;
1658
1659 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1660 ret = -EFAULT;
1661 break;
1662 }
1663
1664 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1665 ret = -EFAULT;
1666 break;
1667 }
1668
1669 ret = io_submit_one(ctx, user_iocb, &tmp);
1670 if (ret)
1671 break;
1672 }
1673
1674 put_ioctx(ctx);
1675 return i ? i : ret;
1676}
1677
1678/* lookup_kiocb
1679 * Finds a given iocb for cancellation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001680 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001681static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1682 u32 key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683{
1684 struct list_head *pos;
Zach Brownd00689a2005-11-13 16:07:34 -08001685
1686 assert_spin_locked(&ctx->ctx_lock);
1687
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688 /* TODO: use a hash or array, this sucks. */
1689 list_for_each(pos, &ctx->active_reqs) {
1690 struct kiocb *kiocb = list_kiocb(pos);
1691 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1692 return kiocb;
1693 }
1694 return NULL;
1695}
1696
1697/* sys_io_cancel:
1698 * Attempts to cancel an iocb previously passed to io_submit. If
1699 * the operation is successfully cancelled, the resulting event is
1700 * copied into the memory pointed to by result without being placed
1701 * into the completion queue and 0 is returned. May fail with
1702 * -EFAULT if any of the data structures pointed to are invalid.
1703 * May fail with -EINVAL if aio_context specified by ctx_id is
1704 * invalid. May fail with -EAGAIN if the iocb specified was not
1705 * cancelled. Will fail with -ENOSYS if not implemented.
1706 */
1707asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb,
1708 struct io_event __user *result)
1709{
1710 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1711 struct kioctx *ctx;
1712 struct kiocb *kiocb;
1713 u32 key;
1714 int ret;
1715
1716 ret = get_user(key, &iocb->aio_key);
1717 if (unlikely(ret))
1718 return -EFAULT;
1719
1720 ctx = lookup_ioctx(ctx_id);
1721 if (unlikely(!ctx))
1722 return -EINVAL;
1723
1724 spin_lock_irq(&ctx->ctx_lock);
1725 ret = -EAGAIN;
1726 kiocb = lookup_kiocb(ctx, iocb, key);
1727 if (kiocb && kiocb->ki_cancel) {
1728 cancel = kiocb->ki_cancel;
1729 kiocb->ki_users ++;
1730 kiocbSetCancelled(kiocb);
1731 } else
1732 cancel = NULL;
1733 spin_unlock_irq(&ctx->ctx_lock);
1734
1735 if (NULL != cancel) {
1736 struct io_event tmp;
1737 pr_debug("calling cancel\n");
1738 memset(&tmp, 0, sizeof(tmp));
1739 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1740 tmp.data = kiocb->ki_user_data;
1741 ret = cancel(kiocb, &tmp);
1742 if (!ret) {
1743 /* Cancellation succeeded -- copy the result
1744 * into the user's buffer.
1745 */
1746 if (copy_to_user(result, &tmp, sizeof(tmp)))
1747 ret = -EFAULT;
1748 }
1749 } else
Wendy Cheng8f582022005-09-09 13:02:08 -07001750 ret = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751
1752 put_ioctx(ctx);
1753
1754 return ret;
1755}
1756
1757/* io_getevents:
1758 * Attempts to read at least min_nr events and up to nr events from
1759 * the completion queue for the aio_context specified by ctx_id. May
1760 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range,
1761 * if nr is out of range, if when is out of range. May fail with
1762 * -EFAULT if any of the memory specified to is invalid. May return
1763 * 0 or < min_nr if no events are available and the timeout specified
1764 * by when has elapsed, where when == NULL specifies an infinite
1765 * timeout. Note that the timeout pointed to by when is relative and
1766 * will be updated if not NULL and the operation blocks. Will fail
1767 * with -ENOSYS if not implemented.
1768 */
1769asmlinkage long sys_io_getevents(aio_context_t ctx_id,
1770 long min_nr,
1771 long nr,
1772 struct io_event __user *events,
1773 struct timespec __user *timeout)
1774{
1775 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1776 long ret = -EINVAL;
1777
1778 if (likely(ioctx)) {
1779 if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0))
1780 ret = read_events(ioctx, min_nr, nr, events, timeout);
1781 put_ioctx(ioctx);
1782 }
1783
1784 return ret;
1785}
1786
1787__initcall(aio_setup);
1788
1789EXPORT_SYMBOL(aio_complete);
1790EXPORT_SYMBOL(aio_put_req);
1791EXPORT_SYMBOL(wait_on_sync_kiocb);