blob: 769791df36b4f357bdcbbd2fa891cb23d6f477cd [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <bcrl@kvack.org>
4 *
5 * Implements an efficient asynchronous io interface.
6 *
7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved.
8 *
9 * See ../COPYING for licensing terms.
10 */
11#include <linux/kernel.h>
12#include <linux/init.h>
13#include <linux/errno.h>
14#include <linux/time.h>
15#include <linux/aio_abi.h>
16#include <linux/module.h>
17#include <linux/syscalls.h>
18
19#define DEBUG 0
20
21#include <linux/sched.h>
22#include <linux/fs.h>
23#include <linux/file.h>
24#include <linux/mm.h>
25#include <linux/mman.h>
26#include <linux/slab.h>
27#include <linux/timer.h>
28#include <linux/aio.h>
29#include <linux/highmem.h>
30#include <linux/workqueue.h>
31#include <linux/security.h>
32
33#include <asm/kmap_types.h>
34#include <asm/uaccess.h>
35#include <asm/mmu_context.h>
36
37#if DEBUG > 1
38#define dprintk printk
39#else
40#define dprintk(x...) do { ; } while (0)
41#endif
42
Linus Torvalds1da177e2005-04-16 15:20:36 -070043/*------ sysctl variables----*/
44atomic_t aio_nr = ATOMIC_INIT(0); /* current system wide number of aio requests */
45unsigned aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
46/*----end sysctl variables---*/
47
48static kmem_cache_t *kiocb_cachep;
49static kmem_cache_t *kioctx_cachep;
50
51static struct workqueue_struct *aio_wq;
52
53/* Used for rare fput completion. */
54static void aio_fput_routine(void *);
55static DECLARE_WORK(fput_work, aio_fput_routine, NULL);
56
57static DEFINE_SPINLOCK(fput_lock);
Adrian Bunk25ee7e32005-04-25 08:18:14 -070058static LIST_HEAD(fput_head);
Linus Torvalds1da177e2005-04-16 15:20:36 -070059
60static void aio_kick_handler(void *);
Sébastien Duguc016e222005-06-28 20:44:59 -070061static void aio_queue_work(struct kioctx *);
Linus Torvalds1da177e2005-04-16 15:20:36 -070062
63/* aio_setup
64 * Creates the slab caches used by the aio routines, panic on
65 * failure as this is done early during the boot sequence.
66 */
67static int __init aio_setup(void)
68{
69 kiocb_cachep = kmem_cache_create("kiocb", sizeof(struct kiocb),
70 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
71 kioctx_cachep = kmem_cache_create("kioctx", sizeof(struct kioctx),
72 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
73
74 aio_wq = create_workqueue("aio");
75
76 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
77
78 return 0;
79}
80
81static void aio_free_ring(struct kioctx *ctx)
82{
83 struct aio_ring_info *info = &ctx->ring_info;
84 long i;
85
86 for (i=0; i<info->nr_pages; i++)
87 put_page(info->ring_pages[i]);
88
89 if (info->mmap_size) {
90 down_write(&ctx->mm->mmap_sem);
91 do_munmap(ctx->mm, info->mmap_base, info->mmap_size);
92 up_write(&ctx->mm->mmap_sem);
93 }
94
95 if (info->ring_pages && info->ring_pages != info->internal_pages)
96 kfree(info->ring_pages);
97 info->ring_pages = NULL;
98 info->nr = 0;
99}
100
101static int aio_setup_ring(struct kioctx *ctx)
102{
103 struct aio_ring *ring;
104 struct aio_ring_info *info = &ctx->ring_info;
105 unsigned nr_events = ctx->max_reqs;
106 unsigned long size;
107 int nr_pages;
108
109 /* Compensate for the ring buffer's head/tail overlap entry */
110 nr_events += 2; /* 1 is required, 2 for good luck */
111
112 size = sizeof(struct aio_ring);
113 size += sizeof(struct io_event) * nr_events;
114 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
115
116 if (nr_pages < 0)
117 return -EINVAL;
118
119 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
120
121 info->nr = 0;
122 info->ring_pages = info->internal_pages;
123 if (nr_pages > AIO_RING_PAGES) {
124 info->ring_pages = kmalloc(sizeof(struct page *) * nr_pages, GFP_KERNEL);
125 if (!info->ring_pages)
126 return -ENOMEM;
127 memset(info->ring_pages, 0, sizeof(struct page *) * nr_pages);
128 }
129
130 info->mmap_size = nr_pages * PAGE_SIZE;
131 dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
132 down_write(&ctx->mm->mmap_sem);
133 info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
134 PROT_READ|PROT_WRITE, MAP_ANON|MAP_PRIVATE,
135 0);
136 if (IS_ERR((void *)info->mmap_base)) {
137 up_write(&ctx->mm->mmap_sem);
138 printk("mmap err: %ld\n", -info->mmap_base);
139 info->mmap_size = 0;
140 aio_free_ring(ctx);
141 return -EAGAIN;
142 }
143
144 dprintk("mmap address: 0x%08lx\n", info->mmap_base);
145 info->nr_pages = get_user_pages(current, ctx->mm,
146 info->mmap_base, nr_pages,
147 1, 0, info->ring_pages, NULL);
148 up_write(&ctx->mm->mmap_sem);
149
150 if (unlikely(info->nr_pages != nr_pages)) {
151 aio_free_ring(ctx);
152 return -EAGAIN;
153 }
154
155 ctx->user_id = info->mmap_base;
156
157 info->nr = nr_events; /* trusted copy */
158
159 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
160 ring->nr = nr_events; /* user copy */
161 ring->id = ctx->user_id;
162 ring->head = ring->tail = 0;
163 ring->magic = AIO_RING_MAGIC;
164 ring->compat_features = AIO_RING_COMPAT_FEATURES;
165 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
166 ring->header_length = sizeof(struct aio_ring);
167 kunmap_atomic(ring, KM_USER0);
168
169 return 0;
170}
171
172
173/* aio_ring_event: returns a pointer to the event at the given index from
174 * kmap_atomic(, km). Release the pointer with put_aio_ring_event();
175 */
176#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
177#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
178#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
179
180#define aio_ring_event(info, nr, km) ({ \
181 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
182 struct io_event *__event; \
183 __event = kmap_atomic( \
184 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
185 __event += pos % AIO_EVENTS_PER_PAGE; \
186 __event; \
187})
188
189#define put_aio_ring_event(event, km) do { \
190 struct io_event *__event = (event); \
191 (void)__event; \
192 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
193} while(0)
194
195/* ioctx_alloc
196 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
197 */
198static struct kioctx *ioctx_alloc(unsigned nr_events)
199{
200 struct mm_struct *mm;
201 struct kioctx *ctx;
202
203 /* Prevent overflows */
204 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
205 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
206 pr_debug("ENOMEM: nr_events too high\n");
207 return ERR_PTR(-EINVAL);
208 }
209
210 if (nr_events > aio_max_nr)
211 return ERR_PTR(-EAGAIN);
212
213 ctx = kmem_cache_alloc(kioctx_cachep, GFP_KERNEL);
214 if (!ctx)
215 return ERR_PTR(-ENOMEM);
216
217 memset(ctx, 0, sizeof(*ctx));
218 ctx->max_reqs = nr_events;
219 mm = ctx->mm = current->mm;
220 atomic_inc(&mm->mm_count);
221
222 atomic_set(&ctx->users, 1);
223 spin_lock_init(&ctx->ctx_lock);
224 spin_lock_init(&ctx->ring_info.ring_lock);
225 init_waitqueue_head(&ctx->wait);
226
227 INIT_LIST_HEAD(&ctx->active_reqs);
228 INIT_LIST_HEAD(&ctx->run_list);
229 INIT_WORK(&ctx->wq, aio_kick_handler, ctx);
230
231 if (aio_setup_ring(ctx) < 0)
232 goto out_freectx;
233
234 /* limit the number of system wide aios */
235 atomic_add(ctx->max_reqs, &aio_nr); /* undone by __put_ioctx */
236 if (unlikely(atomic_read(&aio_nr) > aio_max_nr))
237 goto out_cleanup;
238
239 /* now link into global list. kludge. FIXME */
240 write_lock(&mm->ioctx_list_lock);
241 ctx->next = mm->ioctx_list;
242 mm->ioctx_list = ctx;
243 write_unlock(&mm->ioctx_list_lock);
244
245 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
246 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
247 return ctx;
248
249out_cleanup:
250 atomic_sub(ctx->max_reqs, &aio_nr);
251 ctx->max_reqs = 0; /* prevent __put_ioctx from sub'ing aio_nr */
252 __put_ioctx(ctx);
253 return ERR_PTR(-EAGAIN);
254
255out_freectx:
256 mmdrop(mm);
257 kmem_cache_free(kioctx_cachep, ctx);
258 ctx = ERR_PTR(-ENOMEM);
259
260 dprintk("aio: error allocating ioctx %p\n", ctx);
261 return ctx;
262}
263
264/* aio_cancel_all
265 * Cancels all outstanding aio requests on an aio context. Used
266 * when the processes owning a context have all exited to encourage
267 * the rapid destruction of the kioctx.
268 */
269static void aio_cancel_all(struct kioctx *ctx)
270{
271 int (*cancel)(struct kiocb *, struct io_event *);
272 struct io_event res;
273 spin_lock_irq(&ctx->ctx_lock);
274 ctx->dead = 1;
275 while (!list_empty(&ctx->active_reqs)) {
276 struct list_head *pos = ctx->active_reqs.next;
277 struct kiocb *iocb = list_kiocb(pos);
278 list_del_init(&iocb->ki_list);
279 cancel = iocb->ki_cancel;
280 kiocbSetCancelled(iocb);
281 if (cancel) {
282 iocb->ki_users++;
283 spin_unlock_irq(&ctx->ctx_lock);
284 cancel(iocb, &res);
285 spin_lock_irq(&ctx->ctx_lock);
286 }
287 }
288 spin_unlock_irq(&ctx->ctx_lock);
289}
290
Adrian Bunk25ee7e32005-04-25 08:18:14 -0700291static void wait_for_all_aios(struct kioctx *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292{
293 struct task_struct *tsk = current;
294 DECLARE_WAITQUEUE(wait, tsk);
295
296 if (!ctx->reqs_active)
297 return;
298
299 add_wait_queue(&ctx->wait, &wait);
300 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
301 while (ctx->reqs_active) {
302 schedule();
303 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
304 }
305 __set_task_state(tsk, TASK_RUNNING);
306 remove_wait_queue(&ctx->wait, &wait);
307}
308
309/* wait_on_sync_kiocb:
310 * Waits on the given sync kiocb to complete.
311 */
312ssize_t fastcall wait_on_sync_kiocb(struct kiocb *iocb)
313{
314 while (iocb->ki_users) {
315 set_current_state(TASK_UNINTERRUPTIBLE);
316 if (!iocb->ki_users)
317 break;
318 schedule();
319 }
320 __set_current_state(TASK_RUNNING);
321 return iocb->ki_user_data;
322}
323
324/* exit_aio: called when the last user of mm goes away. At this point,
325 * there is no way for any new requests to be submited or any of the
326 * io_* syscalls to be called on the context. However, there may be
327 * outstanding requests which hold references to the context; as they
328 * go away, they will call put_ioctx and release any pinned memory
329 * associated with the request (held via struct page * references).
330 */
331void fastcall exit_aio(struct mm_struct *mm)
332{
333 struct kioctx *ctx = mm->ioctx_list;
334 mm->ioctx_list = NULL;
335 while (ctx) {
336 struct kioctx *next = ctx->next;
337 ctx->next = NULL;
338 aio_cancel_all(ctx);
339
340 wait_for_all_aios(ctx);
341 /*
342 * this is an overkill, but ensures we don't leave
343 * the ctx on the aio_wq
344 */
345 flush_workqueue(aio_wq);
346
347 if (1 != atomic_read(&ctx->users))
348 printk(KERN_DEBUG
349 "exit_aio:ioctx still alive: %d %d %d\n",
350 atomic_read(&ctx->users), ctx->dead,
351 ctx->reqs_active);
352 put_ioctx(ctx);
353 ctx = next;
354 }
355}
356
357/* __put_ioctx
358 * Called when the last user of an aio context has gone away,
359 * and the struct needs to be freed.
360 */
361void fastcall __put_ioctx(struct kioctx *ctx)
362{
363 unsigned nr_events = ctx->max_reqs;
364
365 if (unlikely(ctx->reqs_active))
366 BUG();
367
368 cancel_delayed_work(&ctx->wq);
369 flush_workqueue(aio_wq);
370 aio_free_ring(ctx);
371 mmdrop(ctx->mm);
372 ctx->mm = NULL;
373 pr_debug("__put_ioctx: freeing %p\n", ctx);
374 kmem_cache_free(kioctx_cachep, ctx);
375
376 atomic_sub(nr_events, &aio_nr);
377}
378
379/* aio_get_req
380 * Allocate a slot for an aio request. Increments the users count
381 * of the kioctx so that the kioctx stays around until all requests are
382 * complete. Returns NULL if no requests are free.
383 *
384 * Returns with kiocb->users set to 2. The io submit code path holds
385 * an extra reference while submitting the i/o.
386 * This prevents races between the aio code path referencing the
387 * req (after submitting it) and aio_complete() freeing the req.
388 */
389static struct kiocb *FASTCALL(__aio_get_req(struct kioctx *ctx));
390static struct kiocb fastcall *__aio_get_req(struct kioctx *ctx)
391{
392 struct kiocb *req = NULL;
393 struct aio_ring *ring;
394 int okay = 0;
395
396 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
397 if (unlikely(!req))
398 return NULL;
399
400 req->ki_flags = 1 << KIF_LOCKED;
401 req->ki_users = 2;
402 req->ki_key = 0;
403 req->ki_ctx = ctx;
404 req->ki_cancel = NULL;
405 req->ki_retry = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406 req->ki_dtor = NULL;
407 req->private = NULL;
408 INIT_LIST_HEAD(&req->ki_run_list);
409
410 /* Check if the completion queue has enough free space to
411 * accept an event from this io.
412 */
413 spin_lock_irq(&ctx->ctx_lock);
414 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
415 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
416 list_add(&req->ki_list, &ctx->active_reqs);
417 get_ioctx(ctx);
418 ctx->reqs_active++;
419 okay = 1;
420 }
421 kunmap_atomic(ring, KM_USER0);
422 spin_unlock_irq(&ctx->ctx_lock);
423
424 if (!okay) {
425 kmem_cache_free(kiocb_cachep, req);
426 req = NULL;
427 }
428
429 return req;
430}
431
432static inline struct kiocb *aio_get_req(struct kioctx *ctx)
433{
434 struct kiocb *req;
435 /* Handle a potential starvation case -- should be exceedingly rare as
436 * requests will be stuck on fput_head only if the aio_fput_routine is
437 * delayed and the requests were the last user of the struct file.
438 */
439 req = __aio_get_req(ctx);
440 if (unlikely(NULL == req)) {
441 aio_fput_routine(NULL);
442 req = __aio_get_req(ctx);
443 }
444 return req;
445}
446
447static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
448{
449 if (req->ki_dtor)
450 req->ki_dtor(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451 kmem_cache_free(kiocb_cachep, req);
452 ctx->reqs_active--;
453
454 if (unlikely(!ctx->reqs_active && ctx->dead))
455 wake_up(&ctx->wait);
456}
457
458static void aio_fput_routine(void *data)
459{
460 spin_lock_irq(&fput_lock);
461 while (likely(!list_empty(&fput_head))) {
462 struct kiocb *req = list_kiocb(fput_head.next);
463 struct kioctx *ctx = req->ki_ctx;
464
465 list_del(&req->ki_list);
466 spin_unlock_irq(&fput_lock);
467
468 /* Complete the fput */
469 __fput(req->ki_filp);
470
471 /* Link the iocb into the context's free list */
472 spin_lock_irq(&ctx->ctx_lock);
473 really_put_req(ctx, req);
474 spin_unlock_irq(&ctx->ctx_lock);
475
476 put_ioctx(ctx);
477 spin_lock_irq(&fput_lock);
478 }
479 spin_unlock_irq(&fput_lock);
480}
481
482/* __aio_put_req
483 * Returns true if this put was the last user of the request.
484 */
485static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
486{
487 dprintk(KERN_DEBUG "aio_put(%p): f_count=%d\n",
488 req, atomic_read(&req->ki_filp->f_count));
489
490 req->ki_users --;
491 if (unlikely(req->ki_users < 0))
492 BUG();
493 if (likely(req->ki_users))
494 return 0;
495 list_del(&req->ki_list); /* remove from active_reqs */
496 req->ki_cancel = NULL;
497 req->ki_retry = NULL;
498
499 /* Must be done under the lock to serialise against cancellation.
500 * Call this aio_fput as it duplicates fput via the fput_work.
501 */
502 if (unlikely(atomic_dec_and_test(&req->ki_filp->f_count))) {
503 get_ioctx(ctx);
504 spin_lock(&fput_lock);
505 list_add(&req->ki_list, &fput_head);
506 spin_unlock(&fput_lock);
507 queue_work(aio_wq, &fput_work);
508 } else
509 really_put_req(ctx, req);
510 return 1;
511}
512
513/* aio_put_req
514 * Returns true if this put was the last user of the kiocb,
515 * false if the request is still in use.
516 */
517int fastcall aio_put_req(struct kiocb *req)
518{
519 struct kioctx *ctx = req->ki_ctx;
520 int ret;
521 spin_lock_irq(&ctx->ctx_lock);
522 ret = __aio_put_req(ctx, req);
523 spin_unlock_irq(&ctx->ctx_lock);
524 if (ret)
525 put_ioctx(ctx);
526 return ret;
527}
528
529/* Lookup an ioctx id. ioctx_list is lockless for reads.
530 * FIXME: this is O(n) and is only suitable for development.
531 */
532struct kioctx *lookup_ioctx(unsigned long ctx_id)
533{
534 struct kioctx *ioctx;
535 struct mm_struct *mm;
536
537 mm = current->mm;
538 read_lock(&mm->ioctx_list_lock);
539 for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next)
540 if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) {
541 get_ioctx(ioctx);
542 break;
543 }
544 read_unlock(&mm->ioctx_list_lock);
545
546 return ioctx;
547}
548
549/*
550 * use_mm
551 * Makes the calling kernel thread take on the specified
552 * mm context.
553 * Called by the retry thread execute retries within the
554 * iocb issuer's mm context, so that copy_from/to_user
555 * operations work seamlessly for aio.
556 * (Note: this routine is intended to be called only
557 * from a kernel thread context)
558 */
559static void use_mm(struct mm_struct *mm)
560{
561 struct mm_struct *active_mm;
562 struct task_struct *tsk = current;
563
564 task_lock(tsk);
565 tsk->flags |= PF_BORROWED_MM;
566 active_mm = tsk->active_mm;
567 atomic_inc(&mm->mm_count);
568 tsk->mm = mm;
569 tsk->active_mm = mm;
Paolo 'Blaisorblade' Giarrusso1e40cd32005-09-03 15:57:25 -0700570 /*
571 * Note that on UML this *requires* PF_BORROWED_MM to be set, otherwise
572 * it won't work. Update it accordingly if you change it here
573 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574 activate_mm(active_mm, mm);
575 task_unlock(tsk);
576
577 mmdrop(active_mm);
578}
579
580/*
581 * unuse_mm
582 * Reverses the effect of use_mm, i.e. releases the
583 * specified mm context which was earlier taken on
584 * by the calling kernel thread
585 * (Note: this routine is intended to be called only
586 * from a kernel thread context)
587 *
588 * Comments: Called with ctx->ctx_lock held. This nests
589 * task_lock instead ctx_lock.
590 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -0700591static void unuse_mm(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592{
593 struct task_struct *tsk = current;
594
595 task_lock(tsk);
596 tsk->flags &= ~PF_BORROWED_MM;
597 tsk->mm = NULL;
598 /* active_mm is still 'mm' */
599 enter_lazy_tlb(mm, tsk);
600 task_unlock(tsk);
601}
602
603/*
604 * Queue up a kiocb to be retried. Assumes that the kiocb
605 * has already been marked as kicked, and places it on
606 * the retry run list for the corresponding ioctx, if it
607 * isn't already queued. Returns 1 if it actually queued
608 * the kiocb (to tell the caller to activate the work
609 * queue to process it), or 0, if it found that it was
610 * already queued.
611 *
612 * Should be called with the spin lock iocb->ki_ctx->ctx_lock
613 * held
614 */
615static inline int __queue_kicked_iocb(struct kiocb *iocb)
616{
617 struct kioctx *ctx = iocb->ki_ctx;
618
619 if (list_empty(&iocb->ki_run_list)) {
620 list_add_tail(&iocb->ki_run_list,
621 &ctx->run_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622 return 1;
623 }
624 return 0;
625}
626
627/* aio_run_iocb
628 * This is the core aio execution routine. It is
629 * invoked both for initial i/o submission and
630 * subsequent retries via the aio_kick_handler.
631 * Expects to be invoked with iocb->ki_ctx->lock
632 * already held. The lock is released and reaquired
633 * as needed during processing.
634 *
635 * Calls the iocb retry method (already setup for the
636 * iocb on initial submission) for operation specific
637 * handling, but takes care of most of common retry
638 * execution details for a given iocb. The retry method
639 * needs to be non-blocking as far as possible, to avoid
640 * holding up other iocbs waiting to be serviced by the
641 * retry kernel thread.
642 *
643 * The trickier parts in this code have to do with
644 * ensuring that only one retry instance is in progress
645 * for a given iocb at any time. Providing that guarantee
646 * simplifies the coding of individual aio operations as
647 * it avoids various potential races.
648 */
649static ssize_t aio_run_iocb(struct kiocb *iocb)
650{
651 struct kioctx *ctx = iocb->ki_ctx;
652 ssize_t (*retry)(struct kiocb *);
653 ssize_t ret;
654
655 if (iocb->ki_retried++ > 1024*1024) {
656 printk("Maximal retry count. Bytes done %Zd\n",
657 iocb->ki_nbytes - iocb->ki_left);
658 return -EAGAIN;
659 }
660
661 if (!(iocb->ki_retried & 0xff)) {
Ken Chen644d3a02005-05-01 08:59:15 -0700662 pr_debug("%ld retry: %d of %d\n", iocb->ki_retried,
663 iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664 }
665
666 if (!(retry = iocb->ki_retry)) {
667 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
668 return 0;
669 }
670
671 /*
672 * We don't want the next retry iteration for this
673 * operation to start until this one has returned and
674 * updated the iocb state. However, wait_queue functions
675 * can trigger a kick_iocb from interrupt context in the
676 * meantime, indicating that data is available for the next
677 * iteration. We want to remember that and enable the
678 * next retry iteration _after_ we are through with
679 * this one.
680 *
681 * So, in order to be able to register a "kick", but
682 * prevent it from being queued now, we clear the kick
683 * flag, but make the kick code *think* that the iocb is
684 * still on the run list until we are actually done.
685 * When we are done with this iteration, we check if
686 * the iocb was kicked in the meantime and if so, queue
687 * it up afresh.
688 */
689
690 kiocbClearKicked(iocb);
691
692 /*
693 * This is so that aio_complete knows it doesn't need to
694 * pull the iocb off the run list (We can't just call
695 * INIT_LIST_HEAD because we don't want a kick_iocb to
696 * queue this on the run list yet)
697 */
698 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
699 spin_unlock_irq(&ctx->ctx_lock);
700
701 /* Quit retrying if the i/o has been cancelled */
702 if (kiocbIsCancelled(iocb)) {
703 ret = -EINTR;
704 aio_complete(iocb, ret, 0);
705 /* must not access the iocb after this */
706 goto out;
707 }
708
709 /*
710 * Now we are all set to call the retry method in async
711 * context. By setting this thread's io_wait context
712 * to point to the wait queue entry inside the currently
713 * running iocb for the duration of the retry, we ensure
714 * that async notification wakeups are queued by the
715 * operation instead of blocking waits, and when notified,
716 * cause the iocb to be kicked for continuation (through
717 * the aio_wake_function callback).
718 */
719 BUG_ON(current->io_wait != NULL);
720 current->io_wait = &iocb->ki_wait;
721 ret = retry(iocb);
722 current->io_wait = NULL;
723
724 if (-EIOCBRETRY != ret) {
725 if (-EIOCBQUEUED != ret) {
726 BUG_ON(!list_empty(&iocb->ki_wait.task_list));
727 aio_complete(iocb, ret, 0);
728 /* must not access the iocb after this */
729 }
730 } else {
731 /*
732 * Issue an additional retry to avoid waiting forever if
733 * no waits were queued (e.g. in case of a short read).
734 */
735 if (list_empty(&iocb->ki_wait.task_list))
736 kiocbSetKicked(iocb);
737 }
738out:
739 spin_lock_irq(&ctx->ctx_lock);
740
741 if (-EIOCBRETRY == ret) {
742 /*
743 * OK, now that we are done with this iteration
744 * and know that there is more left to go,
745 * this is where we let go so that a subsequent
746 * "kick" can start the next iteration
747 */
748
749 /* will make __queue_kicked_iocb succeed from here on */
750 INIT_LIST_HEAD(&iocb->ki_run_list);
751 /* we must queue the next iteration ourselves, if it
752 * has already been kicked */
753 if (kiocbIsKicked(iocb)) {
754 __queue_kicked_iocb(iocb);
Sébastien Duguc016e222005-06-28 20:44:59 -0700755
756 /*
757 * __queue_kicked_iocb will always return 1 here, because
758 * iocb->ki_run_list is empty at this point so it should
759 * be safe to unconditionally queue the context into the
760 * work queue.
761 */
762 aio_queue_work(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763 }
764 }
765 return ret;
766}
767
768/*
769 * __aio_run_iocbs:
770 * Process all pending retries queued on the ioctx
771 * run list.
772 * Assumes it is operating within the aio issuer's mm
773 * context. Expects to be called with ctx->ctx_lock held
774 */
775static int __aio_run_iocbs(struct kioctx *ctx)
776{
777 struct kiocb *iocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778 LIST_HEAD(run_list);
779
780 list_splice_init(&ctx->run_list, &run_list);
781 while (!list_empty(&run_list)) {
782 iocb = list_entry(run_list.next, struct kiocb,
783 ki_run_list);
784 list_del(&iocb->ki_run_list);
785 /*
786 * Hold an extra reference while retrying i/o.
787 */
788 iocb->ki_users++; /* grab extra reference */
789 aio_run_iocb(iocb);
790 if (__aio_put_req(ctx, iocb)) /* drop extra ref */
791 put_ioctx(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793 if (!list_empty(&ctx->run_list))
794 return 1;
795 return 0;
796}
797
798static void aio_queue_work(struct kioctx * ctx)
799{
800 unsigned long timeout;
801 /*
802 * if someone is waiting, get the work started right
803 * away, otherwise, use a longer delay
804 */
805 smp_mb();
806 if (waitqueue_active(&ctx->wait))
807 timeout = 1;
808 else
809 timeout = HZ/10;
810 queue_delayed_work(aio_wq, &ctx->wq, timeout);
811}
812
813
814/*
815 * aio_run_iocbs:
816 * Process all pending retries queued on the ioctx
817 * run list.
818 * Assumes it is operating within the aio issuer's mm
819 * context.
820 */
821static inline void aio_run_iocbs(struct kioctx *ctx)
822{
823 int requeue;
824
825 spin_lock_irq(&ctx->ctx_lock);
826
827 requeue = __aio_run_iocbs(ctx);
828 spin_unlock_irq(&ctx->ctx_lock);
829 if (requeue)
830 aio_queue_work(ctx);
831}
832
833/*
834 * just like aio_run_iocbs, but keeps running them until
835 * the list stays empty
836 */
837static inline void aio_run_all_iocbs(struct kioctx *ctx)
838{
839 spin_lock_irq(&ctx->ctx_lock);
840 while (__aio_run_iocbs(ctx))
841 ;
842 spin_unlock_irq(&ctx->ctx_lock);
843}
844
845/*
846 * aio_kick_handler:
847 * Work queue handler triggered to process pending
848 * retries on an ioctx. Takes on the aio issuer's
849 * mm context before running the iocbs, so that
850 * copy_xxx_user operates on the issuer's address
851 * space.
852 * Run on aiod's context.
853 */
854static void aio_kick_handler(void *data)
855{
856 struct kioctx *ctx = data;
857 mm_segment_t oldfs = get_fs();
858 int requeue;
859
860 set_fs(USER_DS);
861 use_mm(ctx->mm);
862 spin_lock_irq(&ctx->ctx_lock);
863 requeue =__aio_run_iocbs(ctx);
864 unuse_mm(ctx->mm);
865 spin_unlock_irq(&ctx->ctx_lock);
866 set_fs(oldfs);
867 /*
868 * we're in a worker thread already, don't use queue_delayed_work,
869 */
870 if (requeue)
871 queue_work(aio_wq, &ctx->wq);
872}
873
874
875/*
876 * Called by kick_iocb to queue the kiocb for retry
877 * and if required activate the aio work queue to process
878 * it
879 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -0700880static void queue_kicked_iocb(struct kiocb *iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881{
882 struct kioctx *ctx = iocb->ki_ctx;
883 unsigned long flags;
884 int run = 0;
885
886 WARN_ON((!list_empty(&iocb->ki_wait.task_list)));
887
888 spin_lock_irqsave(&ctx->ctx_lock, flags);
889 run = __queue_kicked_iocb(iocb);
890 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
Ken Chen644d3a02005-05-01 08:59:15 -0700891 if (run)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892 aio_queue_work(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893}
894
895/*
896 * kick_iocb:
897 * Called typically from a wait queue callback context
898 * (aio_wake_function) to trigger a retry of the iocb.
899 * The retry is usually executed by aio workqueue
900 * threads (See aio_kick_handler).
901 */
902void fastcall kick_iocb(struct kiocb *iocb)
903{
904 /* sync iocbs are easy: they can only ever be executing from a
905 * single context. */
906 if (is_sync_kiocb(iocb)) {
907 kiocbSetKicked(iocb);
908 wake_up_process(iocb->ki_obj.tsk);
909 return;
910 }
911
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912 /* If its already kicked we shouldn't queue it again */
913 if (!kiocbTryKick(iocb)) {
914 queue_kicked_iocb(iocb);
915 }
916}
917EXPORT_SYMBOL(kick_iocb);
918
919/* aio_complete
920 * Called when the io request on the given iocb is complete.
921 * Returns true if this is the last user of the request. The
922 * only other user of the request can be the cancellation code.
923 */
924int fastcall aio_complete(struct kiocb *iocb, long res, long res2)
925{
926 struct kioctx *ctx = iocb->ki_ctx;
927 struct aio_ring_info *info;
928 struct aio_ring *ring;
929 struct io_event *event;
930 unsigned long flags;
931 unsigned long tail;
932 int ret;
933
934 /* Special case handling for sync iocbs: events go directly
935 * into the iocb for fast handling. Note that this will not
936 * work if we allow sync kiocbs to be cancelled. in which
937 * case the usage count checks will have to move under ctx_lock
938 * for all cases.
939 */
940 if (is_sync_kiocb(iocb)) {
941 int ret;
942
943 iocb->ki_user_data = res;
944 if (iocb->ki_users == 1) {
945 iocb->ki_users = 0;
946 ret = 1;
947 } else {
948 spin_lock_irq(&ctx->ctx_lock);
949 iocb->ki_users--;
950 ret = (0 == iocb->ki_users);
951 spin_unlock_irq(&ctx->ctx_lock);
952 }
953 /* sync iocbs put the task here for us */
954 wake_up_process(iocb->ki_obj.tsk);
955 return ret;
956 }
957
958 info = &ctx->ring_info;
959
960 /* add a completion event to the ring buffer.
961 * must be done holding ctx->ctx_lock to prevent
962 * other code from messing with the tail
963 * pointer since we might be called from irq
964 * context.
965 */
966 spin_lock_irqsave(&ctx->ctx_lock, flags);
967
968 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
969 list_del_init(&iocb->ki_run_list);
970
971 /*
972 * cancelled requests don't get events, userland was given one
973 * when the event got cancelled.
974 */
975 if (kiocbIsCancelled(iocb))
976 goto put_rq;
977
978 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
979
980 tail = info->tail;
981 event = aio_ring_event(info, tail, KM_IRQ0);
Ken Chen4bf69b22005-05-01 08:59:15 -0700982 if (++tail >= info->nr)
983 tail = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984
985 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
986 event->data = iocb->ki_user_data;
987 event->res = res;
988 event->res2 = res2;
989
990 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
991 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
992 res, res2);
993
994 /* after flagging the request as done, we
995 * must never even look at it again
996 */
997 smp_wmb(); /* make event visible before updating tail */
998
999 info->tail = tail;
1000 ring->tail = tail;
1001
1002 put_aio_ring_event(event, KM_IRQ0);
1003 kunmap_atomic(ring, KM_IRQ1);
1004
1005 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
1006
Ken Chen644d3a02005-05-01 08:59:15 -07001007 pr_debug("%ld retries: %d of %d\n", iocb->ki_retried,
1008 iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001009put_rq:
1010 /* everything turned out well, dispose of the aiocb. */
1011 ret = __aio_put_req(ctx, iocb);
1012
1013 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1014
1015 if (waitqueue_active(&ctx->wait))
1016 wake_up(&ctx->wait);
1017
1018 if (ret)
1019 put_ioctx(ctx);
1020
1021 return ret;
1022}
1023
1024/* aio_read_evt
1025 * Pull an event off of the ioctx's event ring. Returns the number of
1026 * events fetched (0 or 1 ;-)
1027 * FIXME: make this use cmpxchg.
1028 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1029 */
1030static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1031{
1032 struct aio_ring_info *info = &ioctx->ring_info;
1033 struct aio_ring *ring;
1034 unsigned long head;
1035 int ret = 0;
1036
1037 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1038 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1039 (unsigned long)ring->head, (unsigned long)ring->tail,
1040 (unsigned long)ring->nr);
1041
1042 if (ring->head == ring->tail)
1043 goto out;
1044
1045 spin_lock(&info->ring_lock);
1046
1047 head = ring->head % info->nr;
1048 if (head != ring->tail) {
1049 struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1050 *ent = *evp;
1051 head = (head + 1) % info->nr;
1052 smp_mb(); /* finish reading the event before updatng the head */
1053 ring->head = head;
1054 ret = 1;
1055 put_aio_ring_event(evp, KM_USER1);
1056 }
1057 spin_unlock(&info->ring_lock);
1058
1059out:
1060 kunmap_atomic(ring, KM_USER0);
1061 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1062 (unsigned long)ring->head, (unsigned long)ring->tail);
1063 return ret;
1064}
1065
1066struct aio_timeout {
1067 struct timer_list timer;
1068 int timed_out;
1069 struct task_struct *p;
1070};
1071
1072static void timeout_func(unsigned long data)
1073{
1074 struct aio_timeout *to = (struct aio_timeout *)data;
1075
1076 to->timed_out = 1;
1077 wake_up_process(to->p);
1078}
1079
1080static inline void init_timeout(struct aio_timeout *to)
1081{
1082 init_timer(&to->timer);
1083 to->timer.data = (unsigned long)to;
1084 to->timer.function = timeout_func;
1085 to->timed_out = 0;
1086 to->p = current;
1087}
1088
1089static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1090 const struct timespec *ts)
1091{
1092 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1093 if (time_after(to->timer.expires, jiffies))
1094 add_timer(&to->timer);
1095 else
1096 to->timed_out = 1;
1097}
1098
1099static inline void clear_timeout(struct aio_timeout *to)
1100{
1101 del_singleshot_timer_sync(&to->timer);
1102}
1103
1104static int read_events(struct kioctx *ctx,
1105 long min_nr, long nr,
1106 struct io_event __user *event,
1107 struct timespec __user *timeout)
1108{
1109 long start_jiffies = jiffies;
1110 struct task_struct *tsk = current;
1111 DECLARE_WAITQUEUE(wait, tsk);
1112 int ret;
1113 int i = 0;
1114 struct io_event ent;
1115 struct aio_timeout to;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116 int retry = 0;
1117
1118 /* needed to zero any padding within an entry (there shouldn't be
1119 * any, but C is fun!
1120 */
1121 memset(&ent, 0, sizeof(ent));
1122retry:
1123 ret = 0;
1124 while (likely(i < nr)) {
1125 ret = aio_read_evt(ctx, &ent);
1126 if (unlikely(ret <= 0))
1127 break;
1128
1129 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1130 ent.data, ent.obj, ent.res, ent.res2);
1131
1132 /* Could we split the check in two? */
1133 ret = -EFAULT;
1134 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1135 dprintk("aio: lost an event due to EFAULT.\n");
1136 break;
1137 }
1138 ret = 0;
1139
1140 /* Good, event copied to userland, update counts. */
1141 event ++;
1142 i ++;
1143 }
1144
1145 if (min_nr <= i)
1146 return i;
1147 if (ret)
1148 return ret;
1149
1150 /* End fast path */
1151
1152 /* racey check, but it gets redone */
1153 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1154 retry = 1;
1155 aio_run_all_iocbs(ctx);
1156 goto retry;
1157 }
1158
1159 init_timeout(&to);
1160 if (timeout) {
1161 struct timespec ts;
1162 ret = -EFAULT;
1163 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1164 goto out;
1165
1166 set_timeout(start_jiffies, &to, &ts);
1167 }
1168
1169 while (likely(i < nr)) {
1170 add_wait_queue_exclusive(&ctx->wait, &wait);
1171 do {
1172 set_task_state(tsk, TASK_INTERRUPTIBLE);
1173 ret = aio_read_evt(ctx, &ent);
1174 if (ret)
1175 break;
1176 if (min_nr <= i)
1177 break;
1178 ret = 0;
1179 if (to.timed_out) /* Only check after read evt */
1180 break;
1181 schedule();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182 if (signal_pending(tsk)) {
1183 ret = -EINTR;
1184 break;
1185 }
1186 /*ret = aio_read_evt(ctx, &ent);*/
1187 } while (1) ;
1188
1189 set_task_state(tsk, TASK_RUNNING);
1190 remove_wait_queue(&ctx->wait, &wait);
1191
1192 if (unlikely(ret <= 0))
1193 break;
1194
1195 ret = -EFAULT;
1196 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1197 dprintk("aio: lost an event due to EFAULT.\n");
1198 break;
1199 }
1200
1201 /* Good, event copied to userland, update counts. */
1202 event ++;
1203 i ++;
1204 }
1205
1206 if (timeout)
1207 clear_timeout(&to);
1208out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001209 return i ? i : ret;
1210}
1211
1212/* Take an ioctx and remove it from the list of ioctx's. Protects
1213 * against races with itself via ->dead.
1214 */
1215static void io_destroy(struct kioctx *ioctx)
1216{
1217 struct mm_struct *mm = current->mm;
1218 struct kioctx **tmp;
1219 int was_dead;
1220
1221 /* delete the entry from the list is someone else hasn't already */
1222 write_lock(&mm->ioctx_list_lock);
1223 was_dead = ioctx->dead;
1224 ioctx->dead = 1;
1225 for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx;
1226 tmp = &(*tmp)->next)
1227 ;
1228 if (*tmp)
1229 *tmp = ioctx->next;
1230 write_unlock(&mm->ioctx_list_lock);
1231
1232 dprintk("aio_release(%p)\n", ioctx);
1233 if (likely(!was_dead))
1234 put_ioctx(ioctx); /* twice for the list */
1235
1236 aio_cancel_all(ioctx);
1237 wait_for_all_aios(ioctx);
1238 put_ioctx(ioctx); /* once for the lookup */
1239}
1240
1241/* sys_io_setup:
1242 * Create an aio_context capable of receiving at least nr_events.
1243 * ctxp must not point to an aio_context that already exists, and
1244 * must be initialized to 0 prior to the call. On successful
1245 * creation of the aio_context, *ctxp is filled in with the resulting
1246 * handle. May fail with -EINVAL if *ctxp is not initialized,
1247 * if the specified nr_events exceeds internal limits. May fail
1248 * with -EAGAIN if the specified nr_events exceeds the user's limit
1249 * of available events. May fail with -ENOMEM if insufficient kernel
1250 * resources are available. May fail with -EFAULT if an invalid
1251 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1252 * implemented.
1253 */
1254asmlinkage long sys_io_setup(unsigned nr_events, aio_context_t __user *ctxp)
1255{
1256 struct kioctx *ioctx = NULL;
1257 unsigned long ctx;
1258 long ret;
1259
1260 ret = get_user(ctx, ctxp);
1261 if (unlikely(ret))
1262 goto out;
1263
1264 ret = -EINVAL;
1265 if (unlikely(ctx || (int)nr_events <= 0)) {
1266 pr_debug("EINVAL: io_setup: ctx or nr_events > max\n");
1267 goto out;
1268 }
1269
1270 ioctx = ioctx_alloc(nr_events);
1271 ret = PTR_ERR(ioctx);
1272 if (!IS_ERR(ioctx)) {
1273 ret = put_user(ioctx->user_id, ctxp);
1274 if (!ret)
1275 return 0;
1276
1277 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1278 io_destroy(ioctx);
1279 }
1280
1281out:
1282 return ret;
1283}
1284
1285/* sys_io_destroy:
1286 * Destroy the aio_context specified. May cancel any outstanding
1287 * AIOs and block on completion. Will fail with -ENOSYS if not
1288 * implemented. May fail with -EFAULT if the context pointed to
1289 * is invalid.
1290 */
1291asmlinkage long sys_io_destroy(aio_context_t ctx)
1292{
1293 struct kioctx *ioctx = lookup_ioctx(ctx);
1294 if (likely(NULL != ioctx)) {
1295 io_destroy(ioctx);
1296 return 0;
1297 }
1298 pr_debug("EINVAL: io_destroy: invalid context id\n");
1299 return -EINVAL;
1300}
1301
1302/*
1303 * Default retry method for aio_read (also used for first time submit)
1304 * Responsible for updating iocb state as retries progress
1305 */
1306static ssize_t aio_pread(struct kiocb *iocb)
1307{
1308 struct file *file = iocb->ki_filp;
1309 struct address_space *mapping = file->f_mapping;
1310 struct inode *inode = mapping->host;
1311 ssize_t ret = 0;
1312
1313 ret = file->f_op->aio_read(iocb, iocb->ki_buf,
1314 iocb->ki_left, iocb->ki_pos);
1315
1316 /*
1317 * Can't just depend on iocb->ki_left to determine
1318 * whether we are done. This may have been a short read.
1319 */
1320 if (ret > 0) {
1321 iocb->ki_buf += ret;
1322 iocb->ki_left -= ret;
1323 /*
1324 * For pipes and sockets we return once we have
1325 * some data; for regular files we retry till we
1326 * complete the entire read or find that we can't
1327 * read any more data (e.g short reads).
1328 */
1329 if (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))
1330 ret = -EIOCBRETRY;
1331 }
1332
1333 /* This means we must have transferred all that we could */
1334 /* No need to retry anymore */
1335 if ((ret == 0) || (iocb->ki_left == 0))
1336 ret = iocb->ki_nbytes - iocb->ki_left;
1337
1338 return ret;
1339}
1340
1341/*
1342 * Default retry method for aio_write (also used for first time submit)
1343 * Responsible for updating iocb state as retries progress
1344 */
1345static ssize_t aio_pwrite(struct kiocb *iocb)
1346{
1347 struct file *file = iocb->ki_filp;
1348 ssize_t ret = 0;
1349
1350 ret = file->f_op->aio_write(iocb, iocb->ki_buf,
1351 iocb->ki_left, iocb->ki_pos);
1352
1353 if (ret > 0) {
1354 iocb->ki_buf += ret;
1355 iocb->ki_left -= ret;
1356
1357 ret = -EIOCBRETRY;
1358 }
1359
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
1368static ssize_t aio_fdsync(struct kiocb *iocb)
1369{
1370 struct file *file = iocb->ki_filp;
1371 ssize_t ret = -EINVAL;
1372
1373 if (file->f_op->aio_fsync)
1374 ret = file->f_op->aio_fsync(iocb, 1);
1375 return ret;
1376}
1377
1378static ssize_t aio_fsync(struct kiocb *iocb)
1379{
1380 struct file *file = iocb->ki_filp;
1381 ssize_t ret = -EINVAL;
1382
1383 if (file->f_op->aio_fsync)
1384 ret = file->f_op->aio_fsync(iocb, 0);
1385 return ret;
1386}
1387
1388/*
1389 * aio_setup_iocb:
1390 * Performs the initial checks and aio retry method
1391 * setup for the kiocb at the time of io submission.
1392 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001393static ssize_t aio_setup_iocb(struct kiocb *kiocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394{
1395 struct file *file = kiocb->ki_filp;
1396 ssize_t ret = 0;
1397
1398 switch (kiocb->ki_opcode) {
1399 case IOCB_CMD_PREAD:
1400 ret = -EBADF;
1401 if (unlikely(!(file->f_mode & FMODE_READ)))
1402 break;
1403 ret = -EFAULT;
1404 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1405 kiocb->ki_left)))
1406 break;
1407 ret = -EINVAL;
1408 if (file->f_op->aio_read)
1409 kiocb->ki_retry = aio_pread;
1410 break;
1411 case IOCB_CMD_PWRITE:
1412 ret = -EBADF;
1413 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1414 break;
1415 ret = -EFAULT;
1416 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1417 kiocb->ki_left)))
1418 break;
1419 ret = -EINVAL;
1420 if (file->f_op->aio_write)
1421 kiocb->ki_retry = aio_pwrite;
1422 break;
1423 case IOCB_CMD_FDSYNC:
1424 ret = -EINVAL;
1425 if (file->f_op->aio_fsync)
1426 kiocb->ki_retry = aio_fdsync;
1427 break;
1428 case IOCB_CMD_FSYNC:
1429 ret = -EINVAL;
1430 if (file->f_op->aio_fsync)
1431 kiocb->ki_retry = aio_fsync;
1432 break;
1433 default:
1434 dprintk("EINVAL: io_submit: no operation provided\n");
1435 ret = -EINVAL;
1436 }
1437
1438 if (!kiocb->ki_retry)
1439 return ret;
1440
1441 return 0;
1442}
1443
1444/*
1445 * aio_wake_function:
1446 * wait queue callback function for aio notification,
1447 * Simply triggers a retry of the operation via kick_iocb.
1448 *
1449 * This callback is specified in the wait queue entry in
1450 * a kiocb (current->io_wait points to this wait queue
1451 * entry when an aio operation executes; it is used
1452 * instead of a synchronous wait when an i/o blocking
1453 * condition is encountered during aio).
1454 *
1455 * Note:
1456 * This routine is executed with the wait queue lock held.
1457 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests
1458 * the ioctx lock inside the wait queue lock. This is safe
1459 * because this callback isn't used for wait queues which
1460 * are nested inside ioctx lock (i.e. ctx->wait)
1461 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001462static int aio_wake_function(wait_queue_t *wait, unsigned mode,
1463 int sync, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001464{
1465 struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait);
1466
1467 list_del_init(&wait->task_list);
1468 kick_iocb(iocb);
1469 return 1;
1470}
1471
1472int fastcall io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1473 struct iocb *iocb)
1474{
1475 struct kiocb *req;
1476 struct file *file;
1477 ssize_t ret;
1478
1479 /* enforce forwards compatibility on users */
1480 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2 ||
1481 iocb->aio_reserved3)) {
1482 pr_debug("EINVAL: io_submit: reserve field set\n");
1483 return -EINVAL;
1484 }
1485
1486 /* prevent overflows */
1487 if (unlikely(
1488 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1489 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1490 ((ssize_t)iocb->aio_nbytes < 0)
1491 )) {
1492 pr_debug("EINVAL: io_submit: overflow check\n");
1493 return -EINVAL;
1494 }
1495
1496 file = fget(iocb->aio_fildes);
1497 if (unlikely(!file))
1498 return -EBADF;
1499
1500 req = aio_get_req(ctx); /* returns with 2 references to req */
1501 if (unlikely(!req)) {
1502 fput(file);
1503 return -EAGAIN;
1504 }
1505
1506 req->ki_filp = file;
Ken Chen212079c2005-05-01 08:59:15 -07001507 ret = put_user(req->ki_key, &user_iocb->aio_key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508 if (unlikely(ret)) {
1509 dprintk("EFAULT: aio_key\n");
1510 goto out_put_req;
1511 }
1512
1513 req->ki_obj.user = user_iocb;
1514 req->ki_user_data = iocb->aio_data;
1515 req->ki_pos = iocb->aio_offset;
1516
1517 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1518 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1519 req->ki_opcode = iocb->aio_lio_opcode;
1520 init_waitqueue_func_entry(&req->ki_wait, aio_wake_function);
1521 INIT_LIST_HEAD(&req->ki_wait.task_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001522 req->ki_retried = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523
1524 ret = aio_setup_iocb(req);
1525
1526 if (ret)
1527 goto out_put_req;
1528
1529 spin_lock_irq(&ctx->ctx_lock);
Ken Chen954d3e92005-05-01 08:59:16 -07001530 if (likely(list_empty(&ctx->run_list))) {
1531 aio_run_iocb(req);
1532 } else {
1533 list_add_tail(&req->ki_run_list, &ctx->run_list);
1534 /* drain the run list */
1535 while (__aio_run_iocbs(ctx))
1536 ;
1537 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001538 spin_unlock_irq(&ctx->ctx_lock);
1539 aio_put_req(req); /* drop extra ref to req */
1540 return 0;
1541
1542out_put_req:
1543 aio_put_req(req); /* drop extra ref to req */
1544 aio_put_req(req); /* drop i/o ref to req */
1545 return ret;
1546}
1547
1548/* sys_io_submit:
1549 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1550 * the number of iocbs queued. May return -EINVAL if the aio_context
1551 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1552 * *iocbpp[0] is not properly initialized, if the operation specified
1553 * is invalid for the file descriptor in the iocb. May fail with
1554 * -EFAULT if any of the data structures point to invalid data. May
1555 * fail with -EBADF if the file descriptor specified in the first
1556 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1557 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1558 * fail with -ENOSYS if not implemented.
1559 */
1560asmlinkage long sys_io_submit(aio_context_t ctx_id, long nr,
1561 struct iocb __user * __user *iocbpp)
1562{
1563 struct kioctx *ctx;
1564 long ret = 0;
1565 int i;
1566
1567 if (unlikely(nr < 0))
1568 return -EINVAL;
1569
1570 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1571 return -EFAULT;
1572
1573 ctx = lookup_ioctx(ctx_id);
1574 if (unlikely(!ctx)) {
1575 pr_debug("EINVAL: io_submit: invalid context id\n");
1576 return -EINVAL;
1577 }
1578
1579 /*
1580 * AKPM: should this return a partial result if some of the IOs were
1581 * successfully submitted?
1582 */
1583 for (i=0; i<nr; i++) {
1584 struct iocb __user *user_iocb;
1585 struct iocb tmp;
1586
1587 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1588 ret = -EFAULT;
1589 break;
1590 }
1591
1592 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1593 ret = -EFAULT;
1594 break;
1595 }
1596
1597 ret = io_submit_one(ctx, user_iocb, &tmp);
1598 if (ret)
1599 break;
1600 }
1601
1602 put_ioctx(ctx);
1603 return i ? i : ret;
1604}
1605
1606/* lookup_kiocb
1607 * Finds a given iocb for cancellation.
1608 * MUST be called with ctx->ctx_lock held.
1609 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001610static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1611 u32 key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001612{
1613 struct list_head *pos;
1614 /* TODO: use a hash or array, this sucks. */
1615 list_for_each(pos, &ctx->active_reqs) {
1616 struct kiocb *kiocb = list_kiocb(pos);
1617 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1618 return kiocb;
1619 }
1620 return NULL;
1621}
1622
1623/* sys_io_cancel:
1624 * Attempts to cancel an iocb previously passed to io_submit. If
1625 * the operation is successfully cancelled, the resulting event is
1626 * copied into the memory pointed to by result without being placed
1627 * into the completion queue and 0 is returned. May fail with
1628 * -EFAULT if any of the data structures pointed to are invalid.
1629 * May fail with -EINVAL if aio_context specified by ctx_id is
1630 * invalid. May fail with -EAGAIN if the iocb specified was not
1631 * cancelled. Will fail with -ENOSYS if not implemented.
1632 */
1633asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb,
1634 struct io_event __user *result)
1635{
1636 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1637 struct kioctx *ctx;
1638 struct kiocb *kiocb;
1639 u32 key;
1640 int ret;
1641
1642 ret = get_user(key, &iocb->aio_key);
1643 if (unlikely(ret))
1644 return -EFAULT;
1645
1646 ctx = lookup_ioctx(ctx_id);
1647 if (unlikely(!ctx))
1648 return -EINVAL;
1649
1650 spin_lock_irq(&ctx->ctx_lock);
1651 ret = -EAGAIN;
1652 kiocb = lookup_kiocb(ctx, iocb, key);
1653 if (kiocb && kiocb->ki_cancel) {
1654 cancel = kiocb->ki_cancel;
1655 kiocb->ki_users ++;
1656 kiocbSetCancelled(kiocb);
1657 } else
1658 cancel = NULL;
1659 spin_unlock_irq(&ctx->ctx_lock);
1660
1661 if (NULL != cancel) {
1662 struct io_event tmp;
1663 pr_debug("calling cancel\n");
1664 memset(&tmp, 0, sizeof(tmp));
1665 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1666 tmp.data = kiocb->ki_user_data;
1667 ret = cancel(kiocb, &tmp);
1668 if (!ret) {
1669 /* Cancellation succeeded -- copy the result
1670 * into the user's buffer.
1671 */
1672 if (copy_to_user(result, &tmp, sizeof(tmp)))
1673 ret = -EFAULT;
1674 }
1675 } else
Wendy Cheng8f582022005-09-09 13:02:08 -07001676 ret = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001677
1678 put_ioctx(ctx);
1679
1680 return ret;
1681}
1682
1683/* io_getevents:
1684 * Attempts to read at least min_nr events and up to nr events from
1685 * the completion queue for the aio_context specified by ctx_id. May
1686 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range,
1687 * if nr is out of range, if when is out of range. May fail with
1688 * -EFAULT if any of the memory specified to is invalid. May return
1689 * 0 or < min_nr if no events are available and the timeout specified
1690 * by when has elapsed, where when == NULL specifies an infinite
1691 * timeout. Note that the timeout pointed to by when is relative and
1692 * will be updated if not NULL and the operation blocks. Will fail
1693 * with -ENOSYS if not implemented.
1694 */
1695asmlinkage long sys_io_getevents(aio_context_t ctx_id,
1696 long min_nr,
1697 long nr,
1698 struct io_event __user *events,
1699 struct timespec __user *timeout)
1700{
1701 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1702 long ret = -EINVAL;
1703
1704 if (likely(ioctx)) {
1705 if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0))
1706 ret = read_events(ioctx, min_nr, nr, events, timeout);
1707 put_ioctx(ioctx);
1708 }
1709
1710 return ret;
1711}
1712
1713__initcall(aio_setup);
1714
1715EXPORT_SYMBOL(aio_complete);
1716EXPORT_SYMBOL(aio_put_req);
1717EXPORT_SYMBOL(wait_on_sync_kiocb);