blob: 7afa222f68028a338eae68907ae231a446cc398c [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 *);
61
62/* aio_setup
63 * Creates the slab caches used by the aio routines, panic on
64 * failure as this is done early during the boot sequence.
65 */
66static int __init aio_setup(void)
67{
68 kiocb_cachep = kmem_cache_create("kiocb", sizeof(struct kiocb),
69 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
70 kioctx_cachep = kmem_cache_create("kioctx", sizeof(struct kioctx),
71 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
72
73 aio_wq = create_workqueue("aio");
74
75 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
76
77 return 0;
78}
79
80static void aio_free_ring(struct kioctx *ctx)
81{
82 struct aio_ring_info *info = &ctx->ring_info;
83 long i;
84
85 for (i=0; i<info->nr_pages; i++)
86 put_page(info->ring_pages[i]);
87
88 if (info->mmap_size) {
89 down_write(&ctx->mm->mmap_sem);
90 do_munmap(ctx->mm, info->mmap_base, info->mmap_size);
91 up_write(&ctx->mm->mmap_sem);
92 }
93
94 if (info->ring_pages && info->ring_pages != info->internal_pages)
95 kfree(info->ring_pages);
96 info->ring_pages = NULL;
97 info->nr = 0;
98}
99
100static int aio_setup_ring(struct kioctx *ctx)
101{
102 struct aio_ring *ring;
103 struct aio_ring_info *info = &ctx->ring_info;
104 unsigned nr_events = ctx->max_reqs;
105 unsigned long size;
106 int nr_pages;
107
108 /* Compensate for the ring buffer's head/tail overlap entry */
109 nr_events += 2; /* 1 is required, 2 for good luck */
110
111 size = sizeof(struct aio_ring);
112 size += sizeof(struct io_event) * nr_events;
113 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
114
115 if (nr_pages < 0)
116 return -EINVAL;
117
118 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
119
120 info->nr = 0;
121 info->ring_pages = info->internal_pages;
122 if (nr_pages > AIO_RING_PAGES) {
123 info->ring_pages = kmalloc(sizeof(struct page *) * nr_pages, GFP_KERNEL);
124 if (!info->ring_pages)
125 return -ENOMEM;
126 memset(info->ring_pages, 0, sizeof(struct page *) * nr_pages);
127 }
128
129 info->mmap_size = nr_pages * PAGE_SIZE;
130 dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
131 down_write(&ctx->mm->mmap_sem);
132 info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
133 PROT_READ|PROT_WRITE, MAP_ANON|MAP_PRIVATE,
134 0);
135 if (IS_ERR((void *)info->mmap_base)) {
136 up_write(&ctx->mm->mmap_sem);
137 printk("mmap err: %ld\n", -info->mmap_base);
138 info->mmap_size = 0;
139 aio_free_ring(ctx);
140 return -EAGAIN;
141 }
142
143 dprintk("mmap address: 0x%08lx\n", info->mmap_base);
144 info->nr_pages = get_user_pages(current, ctx->mm,
145 info->mmap_base, nr_pages,
146 1, 0, info->ring_pages, NULL);
147 up_write(&ctx->mm->mmap_sem);
148
149 if (unlikely(info->nr_pages != nr_pages)) {
150 aio_free_ring(ctx);
151 return -EAGAIN;
152 }
153
154 ctx->user_id = info->mmap_base;
155
156 info->nr = nr_events; /* trusted copy */
157
158 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
159 ring->nr = nr_events; /* user copy */
160 ring->id = ctx->user_id;
161 ring->head = ring->tail = 0;
162 ring->magic = AIO_RING_MAGIC;
163 ring->compat_features = AIO_RING_COMPAT_FEATURES;
164 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
165 ring->header_length = sizeof(struct aio_ring);
166 kunmap_atomic(ring, KM_USER0);
167
168 return 0;
169}
170
171
172/* aio_ring_event: returns a pointer to the event at the given index from
173 * kmap_atomic(, km). Release the pointer with put_aio_ring_event();
174 */
175#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
176#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
177#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
178
179#define aio_ring_event(info, nr, km) ({ \
180 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
181 struct io_event *__event; \
182 __event = kmap_atomic( \
183 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
184 __event += pos % AIO_EVENTS_PER_PAGE; \
185 __event; \
186})
187
188#define put_aio_ring_event(event, km) do { \
189 struct io_event *__event = (event); \
190 (void)__event; \
191 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
192} while(0)
193
194/* ioctx_alloc
195 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
196 */
197static struct kioctx *ioctx_alloc(unsigned nr_events)
198{
199 struct mm_struct *mm;
200 struct kioctx *ctx;
201
202 /* Prevent overflows */
203 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
204 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
205 pr_debug("ENOMEM: nr_events too high\n");
206 return ERR_PTR(-EINVAL);
207 }
208
209 if (nr_events > aio_max_nr)
210 return ERR_PTR(-EAGAIN);
211
212 ctx = kmem_cache_alloc(kioctx_cachep, GFP_KERNEL);
213 if (!ctx)
214 return ERR_PTR(-ENOMEM);
215
216 memset(ctx, 0, sizeof(*ctx));
217 ctx->max_reqs = nr_events;
218 mm = ctx->mm = current->mm;
219 atomic_inc(&mm->mm_count);
220
221 atomic_set(&ctx->users, 1);
222 spin_lock_init(&ctx->ctx_lock);
223 spin_lock_init(&ctx->ring_info.ring_lock);
224 init_waitqueue_head(&ctx->wait);
225
226 INIT_LIST_HEAD(&ctx->active_reqs);
227 INIT_LIST_HEAD(&ctx->run_list);
228 INIT_WORK(&ctx->wq, aio_kick_handler, ctx);
229
230 if (aio_setup_ring(ctx) < 0)
231 goto out_freectx;
232
233 /* limit the number of system wide aios */
234 atomic_add(ctx->max_reqs, &aio_nr); /* undone by __put_ioctx */
235 if (unlikely(atomic_read(&aio_nr) > aio_max_nr))
236 goto out_cleanup;
237
238 /* now link into global list. kludge. FIXME */
239 write_lock(&mm->ioctx_list_lock);
240 ctx->next = mm->ioctx_list;
241 mm->ioctx_list = ctx;
242 write_unlock(&mm->ioctx_list_lock);
243
244 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
245 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
246 return ctx;
247
248out_cleanup:
249 atomic_sub(ctx->max_reqs, &aio_nr);
250 ctx->max_reqs = 0; /* prevent __put_ioctx from sub'ing aio_nr */
251 __put_ioctx(ctx);
252 return ERR_PTR(-EAGAIN);
253
254out_freectx:
255 mmdrop(mm);
256 kmem_cache_free(kioctx_cachep, ctx);
257 ctx = ERR_PTR(-ENOMEM);
258
259 dprintk("aio: error allocating ioctx %p\n", ctx);
260 return ctx;
261}
262
263/* aio_cancel_all
264 * Cancels all outstanding aio requests on an aio context. Used
265 * when the processes owning a context have all exited to encourage
266 * the rapid destruction of the kioctx.
267 */
268static void aio_cancel_all(struct kioctx *ctx)
269{
270 int (*cancel)(struct kiocb *, struct io_event *);
271 struct io_event res;
272 spin_lock_irq(&ctx->ctx_lock);
273 ctx->dead = 1;
274 while (!list_empty(&ctx->active_reqs)) {
275 struct list_head *pos = ctx->active_reqs.next;
276 struct kiocb *iocb = list_kiocb(pos);
277 list_del_init(&iocb->ki_list);
278 cancel = iocb->ki_cancel;
279 kiocbSetCancelled(iocb);
280 if (cancel) {
281 iocb->ki_users++;
282 spin_unlock_irq(&ctx->ctx_lock);
283 cancel(iocb, &res);
284 spin_lock_irq(&ctx->ctx_lock);
285 }
286 }
287 spin_unlock_irq(&ctx->ctx_lock);
288}
289
Adrian Bunk25ee7e32005-04-25 08:18:14 -0700290static void wait_for_all_aios(struct kioctx *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291{
292 struct task_struct *tsk = current;
293 DECLARE_WAITQUEUE(wait, tsk);
294
295 if (!ctx->reqs_active)
296 return;
297
298 add_wait_queue(&ctx->wait, &wait);
299 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
300 while (ctx->reqs_active) {
301 schedule();
302 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
303 }
304 __set_task_state(tsk, TASK_RUNNING);
305 remove_wait_queue(&ctx->wait, &wait);
306}
307
308/* wait_on_sync_kiocb:
309 * Waits on the given sync kiocb to complete.
310 */
311ssize_t fastcall wait_on_sync_kiocb(struct kiocb *iocb)
312{
313 while (iocb->ki_users) {
314 set_current_state(TASK_UNINTERRUPTIBLE);
315 if (!iocb->ki_users)
316 break;
317 schedule();
318 }
319 __set_current_state(TASK_RUNNING);
320 return iocb->ki_user_data;
321}
322
323/* exit_aio: called when the last user of mm goes away. At this point,
324 * there is no way for any new requests to be submited or any of the
325 * io_* syscalls to be called on the context. However, there may be
326 * outstanding requests which hold references to the context; as they
327 * go away, they will call put_ioctx and release any pinned memory
328 * associated with the request (held via struct page * references).
329 */
330void fastcall exit_aio(struct mm_struct *mm)
331{
332 struct kioctx *ctx = mm->ioctx_list;
333 mm->ioctx_list = NULL;
334 while (ctx) {
335 struct kioctx *next = ctx->next;
336 ctx->next = NULL;
337 aio_cancel_all(ctx);
338
339 wait_for_all_aios(ctx);
340 /*
341 * this is an overkill, but ensures we don't leave
342 * the ctx on the aio_wq
343 */
344 flush_workqueue(aio_wq);
345
346 if (1 != atomic_read(&ctx->users))
347 printk(KERN_DEBUG
348 "exit_aio:ioctx still alive: %d %d %d\n",
349 atomic_read(&ctx->users), ctx->dead,
350 ctx->reqs_active);
351 put_ioctx(ctx);
352 ctx = next;
353 }
354}
355
356/* __put_ioctx
357 * Called when the last user of an aio context has gone away,
358 * and the struct needs to be freed.
359 */
360void fastcall __put_ioctx(struct kioctx *ctx)
361{
362 unsigned nr_events = ctx->max_reqs;
363
364 if (unlikely(ctx->reqs_active))
365 BUG();
366
367 cancel_delayed_work(&ctx->wq);
368 flush_workqueue(aio_wq);
369 aio_free_ring(ctx);
370 mmdrop(ctx->mm);
371 ctx->mm = NULL;
372 pr_debug("__put_ioctx: freeing %p\n", ctx);
373 kmem_cache_free(kioctx_cachep, ctx);
374
375 atomic_sub(nr_events, &aio_nr);
376}
377
378/* aio_get_req
379 * Allocate a slot for an aio request. Increments the users count
380 * of the kioctx so that the kioctx stays around until all requests are
381 * complete. Returns NULL if no requests are free.
382 *
383 * Returns with kiocb->users set to 2. The io submit code path holds
384 * an extra reference while submitting the i/o.
385 * This prevents races between the aio code path referencing the
386 * req (after submitting it) and aio_complete() freeing the req.
387 */
388static struct kiocb *FASTCALL(__aio_get_req(struct kioctx *ctx));
389static struct kiocb fastcall *__aio_get_req(struct kioctx *ctx)
390{
391 struct kiocb *req = NULL;
392 struct aio_ring *ring;
393 int okay = 0;
394
395 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
396 if (unlikely(!req))
397 return NULL;
398
399 req->ki_flags = 1 << KIF_LOCKED;
400 req->ki_users = 2;
401 req->ki_key = 0;
402 req->ki_ctx = ctx;
403 req->ki_cancel = NULL;
404 req->ki_retry = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405 req->ki_dtor = NULL;
406 req->private = NULL;
407 INIT_LIST_HEAD(&req->ki_run_list);
408
409 /* Check if the completion queue has enough free space to
410 * accept an event from this io.
411 */
412 spin_lock_irq(&ctx->ctx_lock);
413 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
414 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
415 list_add(&req->ki_list, &ctx->active_reqs);
416 get_ioctx(ctx);
417 ctx->reqs_active++;
418 okay = 1;
419 }
420 kunmap_atomic(ring, KM_USER0);
421 spin_unlock_irq(&ctx->ctx_lock);
422
423 if (!okay) {
424 kmem_cache_free(kiocb_cachep, req);
425 req = NULL;
426 }
427
428 return req;
429}
430
431static inline struct kiocb *aio_get_req(struct kioctx *ctx)
432{
433 struct kiocb *req;
434 /* Handle a potential starvation case -- should be exceedingly rare as
435 * requests will be stuck on fput_head only if the aio_fput_routine is
436 * delayed and the requests were the last user of the struct file.
437 */
438 req = __aio_get_req(ctx);
439 if (unlikely(NULL == req)) {
440 aio_fput_routine(NULL);
441 req = __aio_get_req(ctx);
442 }
443 return req;
444}
445
446static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
447{
448 if (req->ki_dtor)
449 req->ki_dtor(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450 kmem_cache_free(kiocb_cachep, req);
451 ctx->reqs_active--;
452
453 if (unlikely(!ctx->reqs_active && ctx->dead))
454 wake_up(&ctx->wait);
455}
456
457static void aio_fput_routine(void *data)
458{
459 spin_lock_irq(&fput_lock);
460 while (likely(!list_empty(&fput_head))) {
461 struct kiocb *req = list_kiocb(fput_head.next);
462 struct kioctx *ctx = req->ki_ctx;
463
464 list_del(&req->ki_list);
465 spin_unlock_irq(&fput_lock);
466
467 /* Complete the fput */
468 __fput(req->ki_filp);
469
470 /* Link the iocb into the context's free list */
471 spin_lock_irq(&ctx->ctx_lock);
472 really_put_req(ctx, req);
473 spin_unlock_irq(&ctx->ctx_lock);
474
475 put_ioctx(ctx);
476 spin_lock_irq(&fput_lock);
477 }
478 spin_unlock_irq(&fput_lock);
479}
480
481/* __aio_put_req
482 * Returns true if this put was the last user of the request.
483 */
484static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
485{
486 dprintk(KERN_DEBUG "aio_put(%p): f_count=%d\n",
487 req, atomic_read(&req->ki_filp->f_count));
488
489 req->ki_users --;
490 if (unlikely(req->ki_users < 0))
491 BUG();
492 if (likely(req->ki_users))
493 return 0;
494 list_del(&req->ki_list); /* remove from active_reqs */
495 req->ki_cancel = NULL;
496 req->ki_retry = NULL;
497
498 /* Must be done under the lock to serialise against cancellation.
499 * Call this aio_fput as it duplicates fput via the fput_work.
500 */
501 if (unlikely(atomic_dec_and_test(&req->ki_filp->f_count))) {
502 get_ioctx(ctx);
503 spin_lock(&fput_lock);
504 list_add(&req->ki_list, &fput_head);
505 spin_unlock(&fput_lock);
506 queue_work(aio_wq, &fput_work);
507 } else
508 really_put_req(ctx, req);
509 return 1;
510}
511
512/* aio_put_req
513 * Returns true if this put was the last user of the kiocb,
514 * false if the request is still in use.
515 */
516int fastcall aio_put_req(struct kiocb *req)
517{
518 struct kioctx *ctx = req->ki_ctx;
519 int ret;
520 spin_lock_irq(&ctx->ctx_lock);
521 ret = __aio_put_req(ctx, req);
522 spin_unlock_irq(&ctx->ctx_lock);
523 if (ret)
524 put_ioctx(ctx);
525 return ret;
526}
527
528/* Lookup an ioctx id. ioctx_list is lockless for reads.
529 * FIXME: this is O(n) and is only suitable for development.
530 */
531struct kioctx *lookup_ioctx(unsigned long ctx_id)
532{
533 struct kioctx *ioctx;
534 struct mm_struct *mm;
535
536 mm = current->mm;
537 read_lock(&mm->ioctx_list_lock);
538 for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next)
539 if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) {
540 get_ioctx(ioctx);
541 break;
542 }
543 read_unlock(&mm->ioctx_list_lock);
544
545 return ioctx;
546}
547
548/*
549 * use_mm
550 * Makes the calling kernel thread take on the specified
551 * mm context.
552 * Called by the retry thread execute retries within the
553 * iocb issuer's mm context, so that copy_from/to_user
554 * operations work seamlessly for aio.
555 * (Note: this routine is intended to be called only
556 * from a kernel thread context)
557 */
558static void use_mm(struct mm_struct *mm)
559{
560 struct mm_struct *active_mm;
561 struct task_struct *tsk = current;
562
563 task_lock(tsk);
564 tsk->flags |= PF_BORROWED_MM;
565 active_mm = tsk->active_mm;
566 atomic_inc(&mm->mm_count);
567 tsk->mm = mm;
568 tsk->active_mm = mm;
569 activate_mm(active_mm, mm);
570 task_unlock(tsk);
571
572 mmdrop(active_mm);
573}
574
575/*
576 * unuse_mm
577 * Reverses the effect of use_mm, i.e. releases the
578 * specified mm context which was earlier taken on
579 * by the calling kernel thread
580 * (Note: this routine is intended to be called only
581 * from a kernel thread context)
582 *
583 * Comments: Called with ctx->ctx_lock held. This nests
584 * task_lock instead ctx_lock.
585 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -0700586static void unuse_mm(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587{
588 struct task_struct *tsk = current;
589
590 task_lock(tsk);
591 tsk->flags &= ~PF_BORROWED_MM;
592 tsk->mm = NULL;
593 /* active_mm is still 'mm' */
594 enter_lazy_tlb(mm, tsk);
595 task_unlock(tsk);
596}
597
598/*
599 * Queue up a kiocb to be retried. Assumes that the kiocb
600 * has already been marked as kicked, and places it on
601 * the retry run list for the corresponding ioctx, if it
602 * isn't already queued. Returns 1 if it actually queued
603 * the kiocb (to tell the caller to activate the work
604 * queue to process it), or 0, if it found that it was
605 * already queued.
606 *
607 * Should be called with the spin lock iocb->ki_ctx->ctx_lock
608 * held
609 */
610static inline int __queue_kicked_iocb(struct kiocb *iocb)
611{
612 struct kioctx *ctx = iocb->ki_ctx;
613
614 if (list_empty(&iocb->ki_run_list)) {
615 list_add_tail(&iocb->ki_run_list,
616 &ctx->run_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617 return 1;
618 }
619 return 0;
620}
621
622/* aio_run_iocb
623 * This is the core aio execution routine. It is
624 * invoked both for initial i/o submission and
625 * subsequent retries via the aio_kick_handler.
626 * Expects to be invoked with iocb->ki_ctx->lock
627 * already held. The lock is released and reaquired
628 * as needed during processing.
629 *
630 * Calls the iocb retry method (already setup for the
631 * iocb on initial submission) for operation specific
632 * handling, but takes care of most of common retry
633 * execution details for a given iocb. The retry method
634 * needs to be non-blocking as far as possible, to avoid
635 * holding up other iocbs waiting to be serviced by the
636 * retry kernel thread.
637 *
638 * The trickier parts in this code have to do with
639 * ensuring that only one retry instance is in progress
640 * for a given iocb at any time. Providing that guarantee
641 * simplifies the coding of individual aio operations as
642 * it avoids various potential races.
643 */
644static ssize_t aio_run_iocb(struct kiocb *iocb)
645{
646 struct kioctx *ctx = iocb->ki_ctx;
647 ssize_t (*retry)(struct kiocb *);
648 ssize_t ret;
649
650 if (iocb->ki_retried++ > 1024*1024) {
651 printk("Maximal retry count. Bytes done %Zd\n",
652 iocb->ki_nbytes - iocb->ki_left);
653 return -EAGAIN;
654 }
655
656 if (!(iocb->ki_retried & 0xff)) {
Ken Chen644d3a02005-05-01 08:59:15 -0700657 pr_debug("%ld retry: %d of %d\n", iocb->ki_retried,
658 iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659 }
660
661 if (!(retry = iocb->ki_retry)) {
662 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
663 return 0;
664 }
665
666 /*
667 * We don't want the next retry iteration for this
668 * operation to start until this one has returned and
669 * updated the iocb state. However, wait_queue functions
670 * can trigger a kick_iocb from interrupt context in the
671 * meantime, indicating that data is available for the next
672 * iteration. We want to remember that and enable the
673 * next retry iteration _after_ we are through with
674 * this one.
675 *
676 * So, in order to be able to register a "kick", but
677 * prevent it from being queued now, we clear the kick
678 * flag, but make the kick code *think* that the iocb is
679 * still on the run list until we are actually done.
680 * When we are done with this iteration, we check if
681 * the iocb was kicked in the meantime and if so, queue
682 * it up afresh.
683 */
684
685 kiocbClearKicked(iocb);
686
687 /*
688 * This is so that aio_complete knows it doesn't need to
689 * pull the iocb off the run list (We can't just call
690 * INIT_LIST_HEAD because we don't want a kick_iocb to
691 * queue this on the run list yet)
692 */
693 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
694 spin_unlock_irq(&ctx->ctx_lock);
695
696 /* Quit retrying if the i/o has been cancelled */
697 if (kiocbIsCancelled(iocb)) {
698 ret = -EINTR;
699 aio_complete(iocb, ret, 0);
700 /* must not access the iocb after this */
701 goto out;
702 }
703
704 /*
705 * Now we are all set to call the retry method in async
706 * context. By setting this thread's io_wait context
707 * to point to the wait queue entry inside the currently
708 * running iocb for the duration of the retry, we ensure
709 * that async notification wakeups are queued by the
710 * operation instead of blocking waits, and when notified,
711 * cause the iocb to be kicked for continuation (through
712 * the aio_wake_function callback).
713 */
714 BUG_ON(current->io_wait != NULL);
715 current->io_wait = &iocb->ki_wait;
716 ret = retry(iocb);
717 current->io_wait = NULL;
718
719 if (-EIOCBRETRY != ret) {
720 if (-EIOCBQUEUED != ret) {
721 BUG_ON(!list_empty(&iocb->ki_wait.task_list));
722 aio_complete(iocb, ret, 0);
723 /* must not access the iocb after this */
724 }
725 } else {
726 /*
727 * Issue an additional retry to avoid waiting forever if
728 * no waits were queued (e.g. in case of a short read).
729 */
730 if (list_empty(&iocb->ki_wait.task_list))
731 kiocbSetKicked(iocb);
732 }
733out:
734 spin_lock_irq(&ctx->ctx_lock);
735
736 if (-EIOCBRETRY == ret) {
737 /*
738 * OK, now that we are done with this iteration
739 * and know that there is more left to go,
740 * this is where we let go so that a subsequent
741 * "kick" can start the next iteration
742 */
743
744 /* will make __queue_kicked_iocb succeed from here on */
745 INIT_LIST_HEAD(&iocb->ki_run_list);
746 /* we must queue the next iteration ourselves, if it
747 * has already been kicked */
748 if (kiocbIsKicked(iocb)) {
749 __queue_kicked_iocb(iocb);
750 }
751 }
752 return ret;
753}
754
755/*
756 * __aio_run_iocbs:
757 * Process all pending retries queued on the ioctx
758 * run list.
759 * Assumes it is operating within the aio issuer's mm
760 * context. Expects to be called with ctx->ctx_lock held
761 */
762static int __aio_run_iocbs(struct kioctx *ctx)
763{
764 struct kiocb *iocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765 LIST_HEAD(run_list);
766
767 list_splice_init(&ctx->run_list, &run_list);
768 while (!list_empty(&run_list)) {
769 iocb = list_entry(run_list.next, struct kiocb,
770 ki_run_list);
771 list_del(&iocb->ki_run_list);
772 /*
773 * Hold an extra reference while retrying i/o.
774 */
775 iocb->ki_users++; /* grab extra reference */
776 aio_run_iocb(iocb);
777 if (__aio_put_req(ctx, iocb)) /* drop extra ref */
778 put_ioctx(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780 if (!list_empty(&ctx->run_list))
781 return 1;
782 return 0;
783}
784
785static void aio_queue_work(struct kioctx * ctx)
786{
787 unsigned long timeout;
788 /*
789 * if someone is waiting, get the work started right
790 * away, otherwise, use a longer delay
791 */
792 smp_mb();
793 if (waitqueue_active(&ctx->wait))
794 timeout = 1;
795 else
796 timeout = HZ/10;
797 queue_delayed_work(aio_wq, &ctx->wq, timeout);
798}
799
800
801/*
802 * aio_run_iocbs:
803 * Process all pending retries queued on the ioctx
804 * run list.
805 * Assumes it is operating within the aio issuer's mm
806 * context.
807 */
808static inline void aio_run_iocbs(struct kioctx *ctx)
809{
810 int requeue;
811
812 spin_lock_irq(&ctx->ctx_lock);
813
814 requeue = __aio_run_iocbs(ctx);
815 spin_unlock_irq(&ctx->ctx_lock);
816 if (requeue)
817 aio_queue_work(ctx);
818}
819
820/*
821 * just like aio_run_iocbs, but keeps running them until
822 * the list stays empty
823 */
824static inline void aio_run_all_iocbs(struct kioctx *ctx)
825{
826 spin_lock_irq(&ctx->ctx_lock);
827 while (__aio_run_iocbs(ctx))
828 ;
829 spin_unlock_irq(&ctx->ctx_lock);
830}
831
832/*
833 * aio_kick_handler:
834 * Work queue handler triggered to process pending
835 * retries on an ioctx. Takes on the aio issuer's
836 * mm context before running the iocbs, so that
837 * copy_xxx_user operates on the issuer's address
838 * space.
839 * Run on aiod's context.
840 */
841static void aio_kick_handler(void *data)
842{
843 struct kioctx *ctx = data;
844 mm_segment_t oldfs = get_fs();
845 int requeue;
846
847 set_fs(USER_DS);
848 use_mm(ctx->mm);
849 spin_lock_irq(&ctx->ctx_lock);
850 requeue =__aio_run_iocbs(ctx);
851 unuse_mm(ctx->mm);
852 spin_unlock_irq(&ctx->ctx_lock);
853 set_fs(oldfs);
854 /*
855 * we're in a worker thread already, don't use queue_delayed_work,
856 */
857 if (requeue)
858 queue_work(aio_wq, &ctx->wq);
859}
860
861
862/*
863 * Called by kick_iocb to queue the kiocb for retry
864 * and if required activate the aio work queue to process
865 * it
866 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -0700867static void queue_kicked_iocb(struct kiocb *iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700868{
869 struct kioctx *ctx = iocb->ki_ctx;
870 unsigned long flags;
871 int run = 0;
872
873 WARN_ON((!list_empty(&iocb->ki_wait.task_list)));
874
875 spin_lock_irqsave(&ctx->ctx_lock, flags);
876 run = __queue_kicked_iocb(iocb);
877 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
Ken Chen644d3a02005-05-01 08:59:15 -0700878 if (run)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879 aio_queue_work(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880}
881
882/*
883 * kick_iocb:
884 * Called typically from a wait queue callback context
885 * (aio_wake_function) to trigger a retry of the iocb.
886 * The retry is usually executed by aio workqueue
887 * threads (See aio_kick_handler).
888 */
889void fastcall kick_iocb(struct kiocb *iocb)
890{
891 /* sync iocbs are easy: they can only ever be executing from a
892 * single context. */
893 if (is_sync_kiocb(iocb)) {
894 kiocbSetKicked(iocb);
895 wake_up_process(iocb->ki_obj.tsk);
896 return;
897 }
898
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899 /* If its already kicked we shouldn't queue it again */
900 if (!kiocbTryKick(iocb)) {
901 queue_kicked_iocb(iocb);
902 }
903}
904EXPORT_SYMBOL(kick_iocb);
905
906/* aio_complete
907 * Called when the io request on the given iocb is complete.
908 * Returns true if this is the last user of the request. The
909 * only other user of the request can be the cancellation code.
910 */
911int fastcall aio_complete(struct kiocb *iocb, long res, long res2)
912{
913 struct kioctx *ctx = iocb->ki_ctx;
914 struct aio_ring_info *info;
915 struct aio_ring *ring;
916 struct io_event *event;
917 unsigned long flags;
918 unsigned long tail;
919 int ret;
920
921 /* Special case handling for sync iocbs: events go directly
922 * into the iocb for fast handling. Note that this will not
923 * work if we allow sync kiocbs to be cancelled. in which
924 * case the usage count checks will have to move under ctx_lock
925 * for all cases.
926 */
927 if (is_sync_kiocb(iocb)) {
928 int ret;
929
930 iocb->ki_user_data = res;
931 if (iocb->ki_users == 1) {
932 iocb->ki_users = 0;
933 ret = 1;
934 } else {
935 spin_lock_irq(&ctx->ctx_lock);
936 iocb->ki_users--;
937 ret = (0 == iocb->ki_users);
938 spin_unlock_irq(&ctx->ctx_lock);
939 }
940 /* sync iocbs put the task here for us */
941 wake_up_process(iocb->ki_obj.tsk);
942 return ret;
943 }
944
945 info = &ctx->ring_info;
946
947 /* add a completion event to the ring buffer.
948 * must be done holding ctx->ctx_lock to prevent
949 * other code from messing with the tail
950 * pointer since we might be called from irq
951 * context.
952 */
953 spin_lock_irqsave(&ctx->ctx_lock, flags);
954
955 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
956 list_del_init(&iocb->ki_run_list);
957
958 /*
959 * cancelled requests don't get events, userland was given one
960 * when the event got cancelled.
961 */
962 if (kiocbIsCancelled(iocb))
963 goto put_rq;
964
965 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
966
967 tail = info->tail;
968 event = aio_ring_event(info, tail, KM_IRQ0);
Ken Chen4bf69b22005-05-01 08:59:15 -0700969 if (++tail >= info->nr)
970 tail = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971
972 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
973 event->data = iocb->ki_user_data;
974 event->res = res;
975 event->res2 = res2;
976
977 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
978 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
979 res, res2);
980
981 /* after flagging the request as done, we
982 * must never even look at it again
983 */
984 smp_wmb(); /* make event visible before updating tail */
985
986 info->tail = tail;
987 ring->tail = tail;
988
989 put_aio_ring_event(event, KM_IRQ0);
990 kunmap_atomic(ring, KM_IRQ1);
991
992 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
993
Ken Chen644d3a02005-05-01 08:59:15 -0700994 pr_debug("%ld retries: %d of %d\n", iocb->ki_retried,
995 iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996put_rq:
997 /* everything turned out well, dispose of the aiocb. */
998 ret = __aio_put_req(ctx, iocb);
999
1000 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1001
1002 if (waitqueue_active(&ctx->wait))
1003 wake_up(&ctx->wait);
1004
1005 if (ret)
1006 put_ioctx(ctx);
1007
1008 return ret;
1009}
1010
1011/* aio_read_evt
1012 * Pull an event off of the ioctx's event ring. Returns the number of
1013 * events fetched (0 or 1 ;-)
1014 * FIXME: make this use cmpxchg.
1015 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1016 */
1017static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1018{
1019 struct aio_ring_info *info = &ioctx->ring_info;
1020 struct aio_ring *ring;
1021 unsigned long head;
1022 int ret = 0;
1023
1024 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1025 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1026 (unsigned long)ring->head, (unsigned long)ring->tail,
1027 (unsigned long)ring->nr);
1028
1029 if (ring->head == ring->tail)
1030 goto out;
1031
1032 spin_lock(&info->ring_lock);
1033
1034 head = ring->head % info->nr;
1035 if (head != ring->tail) {
1036 struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1037 *ent = *evp;
1038 head = (head + 1) % info->nr;
1039 smp_mb(); /* finish reading the event before updatng the head */
1040 ring->head = head;
1041 ret = 1;
1042 put_aio_ring_event(evp, KM_USER1);
1043 }
1044 spin_unlock(&info->ring_lock);
1045
1046out:
1047 kunmap_atomic(ring, KM_USER0);
1048 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1049 (unsigned long)ring->head, (unsigned long)ring->tail);
1050 return ret;
1051}
1052
1053struct aio_timeout {
1054 struct timer_list timer;
1055 int timed_out;
1056 struct task_struct *p;
1057};
1058
1059static void timeout_func(unsigned long data)
1060{
1061 struct aio_timeout *to = (struct aio_timeout *)data;
1062
1063 to->timed_out = 1;
1064 wake_up_process(to->p);
1065}
1066
1067static inline void init_timeout(struct aio_timeout *to)
1068{
1069 init_timer(&to->timer);
1070 to->timer.data = (unsigned long)to;
1071 to->timer.function = timeout_func;
1072 to->timed_out = 0;
1073 to->p = current;
1074}
1075
1076static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1077 const struct timespec *ts)
1078{
1079 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1080 if (time_after(to->timer.expires, jiffies))
1081 add_timer(&to->timer);
1082 else
1083 to->timed_out = 1;
1084}
1085
1086static inline void clear_timeout(struct aio_timeout *to)
1087{
1088 del_singleshot_timer_sync(&to->timer);
1089}
1090
1091static int read_events(struct kioctx *ctx,
1092 long min_nr, long nr,
1093 struct io_event __user *event,
1094 struct timespec __user *timeout)
1095{
1096 long start_jiffies = jiffies;
1097 struct task_struct *tsk = current;
1098 DECLARE_WAITQUEUE(wait, tsk);
1099 int ret;
1100 int i = 0;
1101 struct io_event ent;
1102 struct aio_timeout to;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001103 int retry = 0;
1104
1105 /* needed to zero any padding within an entry (there shouldn't be
1106 * any, but C is fun!
1107 */
1108 memset(&ent, 0, sizeof(ent));
1109retry:
1110 ret = 0;
1111 while (likely(i < nr)) {
1112 ret = aio_read_evt(ctx, &ent);
1113 if (unlikely(ret <= 0))
1114 break;
1115
1116 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1117 ent.data, ent.obj, ent.res, ent.res2);
1118
1119 /* Could we split the check in two? */
1120 ret = -EFAULT;
1121 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1122 dprintk("aio: lost an event due to EFAULT.\n");
1123 break;
1124 }
1125 ret = 0;
1126
1127 /* Good, event copied to userland, update counts. */
1128 event ++;
1129 i ++;
1130 }
1131
1132 if (min_nr <= i)
1133 return i;
1134 if (ret)
1135 return ret;
1136
1137 /* End fast path */
1138
1139 /* racey check, but it gets redone */
1140 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1141 retry = 1;
1142 aio_run_all_iocbs(ctx);
1143 goto retry;
1144 }
1145
1146 init_timeout(&to);
1147 if (timeout) {
1148 struct timespec ts;
1149 ret = -EFAULT;
1150 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1151 goto out;
1152
1153 set_timeout(start_jiffies, &to, &ts);
1154 }
1155
1156 while (likely(i < nr)) {
1157 add_wait_queue_exclusive(&ctx->wait, &wait);
1158 do {
1159 set_task_state(tsk, TASK_INTERRUPTIBLE);
1160 ret = aio_read_evt(ctx, &ent);
1161 if (ret)
1162 break;
1163 if (min_nr <= i)
1164 break;
1165 ret = 0;
1166 if (to.timed_out) /* Only check after read evt */
1167 break;
1168 schedule();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169 if (signal_pending(tsk)) {
1170 ret = -EINTR;
1171 break;
1172 }
1173 /*ret = aio_read_evt(ctx, &ent);*/
1174 } while (1) ;
1175
1176 set_task_state(tsk, TASK_RUNNING);
1177 remove_wait_queue(&ctx->wait, &wait);
1178
1179 if (unlikely(ret <= 0))
1180 break;
1181
1182 ret = -EFAULT;
1183 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1184 dprintk("aio: lost an event due to EFAULT.\n");
1185 break;
1186 }
1187
1188 /* Good, event copied to userland, update counts. */
1189 event ++;
1190 i ++;
1191 }
1192
1193 if (timeout)
1194 clear_timeout(&to);
1195out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196 return i ? i : ret;
1197}
1198
1199/* Take an ioctx and remove it from the list of ioctx's. Protects
1200 * against races with itself via ->dead.
1201 */
1202static void io_destroy(struct kioctx *ioctx)
1203{
1204 struct mm_struct *mm = current->mm;
1205 struct kioctx **tmp;
1206 int was_dead;
1207
1208 /* delete the entry from the list is someone else hasn't already */
1209 write_lock(&mm->ioctx_list_lock);
1210 was_dead = ioctx->dead;
1211 ioctx->dead = 1;
1212 for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx;
1213 tmp = &(*tmp)->next)
1214 ;
1215 if (*tmp)
1216 *tmp = ioctx->next;
1217 write_unlock(&mm->ioctx_list_lock);
1218
1219 dprintk("aio_release(%p)\n", ioctx);
1220 if (likely(!was_dead))
1221 put_ioctx(ioctx); /* twice for the list */
1222
1223 aio_cancel_all(ioctx);
1224 wait_for_all_aios(ioctx);
1225 put_ioctx(ioctx); /* once for the lookup */
1226}
1227
1228/* sys_io_setup:
1229 * Create an aio_context capable of receiving at least nr_events.
1230 * ctxp must not point to an aio_context that already exists, and
1231 * must be initialized to 0 prior to the call. On successful
1232 * creation of the aio_context, *ctxp is filled in with the resulting
1233 * handle. May fail with -EINVAL if *ctxp is not initialized,
1234 * if the specified nr_events exceeds internal limits. May fail
1235 * with -EAGAIN if the specified nr_events exceeds the user's limit
1236 * of available events. May fail with -ENOMEM if insufficient kernel
1237 * resources are available. May fail with -EFAULT if an invalid
1238 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1239 * implemented.
1240 */
1241asmlinkage long sys_io_setup(unsigned nr_events, aio_context_t __user *ctxp)
1242{
1243 struct kioctx *ioctx = NULL;
1244 unsigned long ctx;
1245 long ret;
1246
1247 ret = get_user(ctx, ctxp);
1248 if (unlikely(ret))
1249 goto out;
1250
1251 ret = -EINVAL;
1252 if (unlikely(ctx || (int)nr_events <= 0)) {
1253 pr_debug("EINVAL: io_setup: ctx or nr_events > max\n");
1254 goto out;
1255 }
1256
1257 ioctx = ioctx_alloc(nr_events);
1258 ret = PTR_ERR(ioctx);
1259 if (!IS_ERR(ioctx)) {
1260 ret = put_user(ioctx->user_id, ctxp);
1261 if (!ret)
1262 return 0;
1263
1264 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1265 io_destroy(ioctx);
1266 }
1267
1268out:
1269 return ret;
1270}
1271
1272/* sys_io_destroy:
1273 * Destroy the aio_context specified. May cancel any outstanding
1274 * AIOs and block on completion. Will fail with -ENOSYS if not
1275 * implemented. May fail with -EFAULT if the context pointed to
1276 * is invalid.
1277 */
1278asmlinkage long sys_io_destroy(aio_context_t ctx)
1279{
1280 struct kioctx *ioctx = lookup_ioctx(ctx);
1281 if (likely(NULL != ioctx)) {
1282 io_destroy(ioctx);
1283 return 0;
1284 }
1285 pr_debug("EINVAL: io_destroy: invalid context id\n");
1286 return -EINVAL;
1287}
1288
1289/*
1290 * Default retry method for aio_read (also used for first time submit)
1291 * Responsible for updating iocb state as retries progress
1292 */
1293static ssize_t aio_pread(struct kiocb *iocb)
1294{
1295 struct file *file = iocb->ki_filp;
1296 struct address_space *mapping = file->f_mapping;
1297 struct inode *inode = mapping->host;
1298 ssize_t ret = 0;
1299
1300 ret = file->f_op->aio_read(iocb, iocb->ki_buf,
1301 iocb->ki_left, iocb->ki_pos);
1302
1303 /*
1304 * Can't just depend on iocb->ki_left to determine
1305 * whether we are done. This may have been a short read.
1306 */
1307 if (ret > 0) {
1308 iocb->ki_buf += ret;
1309 iocb->ki_left -= ret;
1310 /*
1311 * For pipes and sockets we return once we have
1312 * some data; for regular files we retry till we
1313 * complete the entire read or find that we can't
1314 * read any more data (e.g short reads).
1315 */
1316 if (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))
1317 ret = -EIOCBRETRY;
1318 }
1319
1320 /* This means we must have transferred all that we could */
1321 /* No need to retry anymore */
1322 if ((ret == 0) || (iocb->ki_left == 0))
1323 ret = iocb->ki_nbytes - iocb->ki_left;
1324
1325 return ret;
1326}
1327
1328/*
1329 * Default retry method for aio_write (also used for first time submit)
1330 * Responsible for updating iocb state as retries progress
1331 */
1332static ssize_t aio_pwrite(struct kiocb *iocb)
1333{
1334 struct file *file = iocb->ki_filp;
1335 ssize_t ret = 0;
1336
1337 ret = file->f_op->aio_write(iocb, iocb->ki_buf,
1338 iocb->ki_left, iocb->ki_pos);
1339
1340 if (ret > 0) {
1341 iocb->ki_buf += ret;
1342 iocb->ki_left -= ret;
1343
1344 ret = -EIOCBRETRY;
1345 }
1346
1347 /* This means we must have transferred all that we could */
1348 /* No need to retry anymore */
1349 if ((ret == 0) || (iocb->ki_left == 0))
1350 ret = iocb->ki_nbytes - iocb->ki_left;
1351
1352 return ret;
1353}
1354
1355static ssize_t aio_fdsync(struct kiocb *iocb)
1356{
1357 struct file *file = iocb->ki_filp;
1358 ssize_t ret = -EINVAL;
1359
1360 if (file->f_op->aio_fsync)
1361 ret = file->f_op->aio_fsync(iocb, 1);
1362 return ret;
1363}
1364
1365static ssize_t aio_fsync(struct kiocb *iocb)
1366{
1367 struct file *file = iocb->ki_filp;
1368 ssize_t ret = -EINVAL;
1369
1370 if (file->f_op->aio_fsync)
1371 ret = file->f_op->aio_fsync(iocb, 0);
1372 return ret;
1373}
1374
1375/*
1376 * aio_setup_iocb:
1377 * Performs the initial checks and aio retry method
1378 * setup for the kiocb at the time of io submission.
1379 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001380static ssize_t aio_setup_iocb(struct kiocb *kiocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381{
1382 struct file *file = kiocb->ki_filp;
1383 ssize_t ret = 0;
1384
1385 switch (kiocb->ki_opcode) {
1386 case IOCB_CMD_PREAD:
1387 ret = -EBADF;
1388 if (unlikely(!(file->f_mode & FMODE_READ)))
1389 break;
1390 ret = -EFAULT;
1391 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1392 kiocb->ki_left)))
1393 break;
1394 ret = -EINVAL;
1395 if (file->f_op->aio_read)
1396 kiocb->ki_retry = aio_pread;
1397 break;
1398 case IOCB_CMD_PWRITE:
1399 ret = -EBADF;
1400 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1401 break;
1402 ret = -EFAULT;
1403 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1404 kiocb->ki_left)))
1405 break;
1406 ret = -EINVAL;
1407 if (file->f_op->aio_write)
1408 kiocb->ki_retry = aio_pwrite;
1409 break;
1410 case IOCB_CMD_FDSYNC:
1411 ret = -EINVAL;
1412 if (file->f_op->aio_fsync)
1413 kiocb->ki_retry = aio_fdsync;
1414 break;
1415 case IOCB_CMD_FSYNC:
1416 ret = -EINVAL;
1417 if (file->f_op->aio_fsync)
1418 kiocb->ki_retry = aio_fsync;
1419 break;
1420 default:
1421 dprintk("EINVAL: io_submit: no operation provided\n");
1422 ret = -EINVAL;
1423 }
1424
1425 if (!kiocb->ki_retry)
1426 return ret;
1427
1428 return 0;
1429}
1430
1431/*
1432 * aio_wake_function:
1433 * wait queue callback function for aio notification,
1434 * Simply triggers a retry of the operation via kick_iocb.
1435 *
1436 * This callback is specified in the wait queue entry in
1437 * a kiocb (current->io_wait points to this wait queue
1438 * entry when an aio operation executes; it is used
1439 * instead of a synchronous wait when an i/o blocking
1440 * condition is encountered during aio).
1441 *
1442 * Note:
1443 * This routine is executed with the wait queue lock held.
1444 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests
1445 * the ioctx lock inside the wait queue lock. This is safe
1446 * because this callback isn't used for wait queues which
1447 * are nested inside ioctx lock (i.e. ctx->wait)
1448 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001449static int aio_wake_function(wait_queue_t *wait, unsigned mode,
1450 int sync, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001451{
1452 struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait);
1453
1454 list_del_init(&wait->task_list);
1455 kick_iocb(iocb);
1456 return 1;
1457}
1458
1459int fastcall io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1460 struct iocb *iocb)
1461{
1462 struct kiocb *req;
1463 struct file *file;
1464 ssize_t ret;
1465
1466 /* enforce forwards compatibility on users */
1467 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2 ||
1468 iocb->aio_reserved3)) {
1469 pr_debug("EINVAL: io_submit: reserve field set\n");
1470 return -EINVAL;
1471 }
1472
1473 /* prevent overflows */
1474 if (unlikely(
1475 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1476 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1477 ((ssize_t)iocb->aio_nbytes < 0)
1478 )) {
1479 pr_debug("EINVAL: io_submit: overflow check\n");
1480 return -EINVAL;
1481 }
1482
1483 file = fget(iocb->aio_fildes);
1484 if (unlikely(!file))
1485 return -EBADF;
1486
1487 req = aio_get_req(ctx); /* returns with 2 references to req */
1488 if (unlikely(!req)) {
1489 fput(file);
1490 return -EAGAIN;
1491 }
1492
1493 req->ki_filp = file;
Ken Chen212079c2005-05-01 08:59:15 -07001494 ret = put_user(req->ki_key, &user_iocb->aio_key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495 if (unlikely(ret)) {
1496 dprintk("EFAULT: aio_key\n");
1497 goto out_put_req;
1498 }
1499
1500 req->ki_obj.user = user_iocb;
1501 req->ki_user_data = iocb->aio_data;
1502 req->ki_pos = iocb->aio_offset;
1503
1504 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1505 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1506 req->ki_opcode = iocb->aio_lio_opcode;
1507 init_waitqueue_func_entry(&req->ki_wait, aio_wake_function);
1508 INIT_LIST_HEAD(&req->ki_wait.task_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509 req->ki_retried = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510
1511 ret = aio_setup_iocb(req);
1512
1513 if (ret)
1514 goto out_put_req;
1515
1516 spin_lock_irq(&ctx->ctx_lock);
Ken Chen954d3e92005-05-01 08:59:16 -07001517 if (likely(list_empty(&ctx->run_list))) {
1518 aio_run_iocb(req);
1519 } else {
1520 list_add_tail(&req->ki_run_list, &ctx->run_list);
1521 /* drain the run list */
1522 while (__aio_run_iocbs(ctx))
1523 ;
1524 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525 spin_unlock_irq(&ctx->ctx_lock);
1526 aio_put_req(req); /* drop extra ref to req */
1527 return 0;
1528
1529out_put_req:
1530 aio_put_req(req); /* drop extra ref to req */
1531 aio_put_req(req); /* drop i/o ref to req */
1532 return ret;
1533}
1534
1535/* sys_io_submit:
1536 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1537 * the number of iocbs queued. May return -EINVAL if the aio_context
1538 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1539 * *iocbpp[0] is not properly initialized, if the operation specified
1540 * is invalid for the file descriptor in the iocb. May fail with
1541 * -EFAULT if any of the data structures point to invalid data. May
1542 * fail with -EBADF if the file descriptor specified in the first
1543 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1544 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1545 * fail with -ENOSYS if not implemented.
1546 */
1547asmlinkage long sys_io_submit(aio_context_t ctx_id, long nr,
1548 struct iocb __user * __user *iocbpp)
1549{
1550 struct kioctx *ctx;
1551 long ret = 0;
1552 int i;
1553
1554 if (unlikely(nr < 0))
1555 return -EINVAL;
1556
1557 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1558 return -EFAULT;
1559
1560 ctx = lookup_ioctx(ctx_id);
1561 if (unlikely(!ctx)) {
1562 pr_debug("EINVAL: io_submit: invalid context id\n");
1563 return -EINVAL;
1564 }
1565
1566 /*
1567 * AKPM: should this return a partial result if some of the IOs were
1568 * successfully submitted?
1569 */
1570 for (i=0; i<nr; i++) {
1571 struct iocb __user *user_iocb;
1572 struct iocb tmp;
1573
1574 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1575 ret = -EFAULT;
1576 break;
1577 }
1578
1579 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1580 ret = -EFAULT;
1581 break;
1582 }
1583
1584 ret = io_submit_one(ctx, user_iocb, &tmp);
1585 if (ret)
1586 break;
1587 }
1588
1589 put_ioctx(ctx);
1590 return i ? i : ret;
1591}
1592
1593/* lookup_kiocb
1594 * Finds a given iocb for cancellation.
1595 * MUST be called with ctx->ctx_lock held.
1596 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001597static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1598 u32 key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001599{
1600 struct list_head *pos;
1601 /* TODO: use a hash or array, this sucks. */
1602 list_for_each(pos, &ctx->active_reqs) {
1603 struct kiocb *kiocb = list_kiocb(pos);
1604 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1605 return kiocb;
1606 }
1607 return NULL;
1608}
1609
1610/* sys_io_cancel:
1611 * Attempts to cancel an iocb previously passed to io_submit. If
1612 * the operation is successfully cancelled, the resulting event is
1613 * copied into the memory pointed to by result without being placed
1614 * into the completion queue and 0 is returned. May fail with
1615 * -EFAULT if any of the data structures pointed to are invalid.
1616 * May fail with -EINVAL if aio_context specified by ctx_id is
1617 * invalid. May fail with -EAGAIN if the iocb specified was not
1618 * cancelled. Will fail with -ENOSYS if not implemented.
1619 */
1620asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb,
1621 struct io_event __user *result)
1622{
1623 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1624 struct kioctx *ctx;
1625 struct kiocb *kiocb;
1626 u32 key;
1627 int ret;
1628
1629 ret = get_user(key, &iocb->aio_key);
1630 if (unlikely(ret))
1631 return -EFAULT;
1632
1633 ctx = lookup_ioctx(ctx_id);
1634 if (unlikely(!ctx))
1635 return -EINVAL;
1636
1637 spin_lock_irq(&ctx->ctx_lock);
1638 ret = -EAGAIN;
1639 kiocb = lookup_kiocb(ctx, iocb, key);
1640 if (kiocb && kiocb->ki_cancel) {
1641 cancel = kiocb->ki_cancel;
1642 kiocb->ki_users ++;
1643 kiocbSetCancelled(kiocb);
1644 } else
1645 cancel = NULL;
1646 spin_unlock_irq(&ctx->ctx_lock);
1647
1648 if (NULL != cancel) {
1649 struct io_event tmp;
1650 pr_debug("calling cancel\n");
1651 memset(&tmp, 0, sizeof(tmp));
1652 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1653 tmp.data = kiocb->ki_user_data;
1654 ret = cancel(kiocb, &tmp);
1655 if (!ret) {
1656 /* Cancellation succeeded -- copy the result
1657 * into the user's buffer.
1658 */
1659 if (copy_to_user(result, &tmp, sizeof(tmp)))
1660 ret = -EFAULT;
1661 }
1662 } else
1663 printk(KERN_DEBUG "iocb has no cancel operation\n");
1664
1665 put_ioctx(ctx);
1666
1667 return ret;
1668}
1669
1670/* io_getevents:
1671 * Attempts to read at least min_nr events and up to nr events from
1672 * the completion queue for the aio_context specified by ctx_id. May
1673 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range,
1674 * if nr is out of range, if when is out of range. May fail with
1675 * -EFAULT if any of the memory specified to is invalid. May return
1676 * 0 or < min_nr if no events are available and the timeout specified
1677 * by when has elapsed, where when == NULL specifies an infinite
1678 * timeout. Note that the timeout pointed to by when is relative and
1679 * will be updated if not NULL and the operation blocks. Will fail
1680 * with -ENOSYS if not implemented.
1681 */
1682asmlinkage long sys_io_getevents(aio_context_t ctx_id,
1683 long min_nr,
1684 long nr,
1685 struct io_event __user *events,
1686 struct timespec __user *timeout)
1687{
1688 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1689 long ret = -EINVAL;
1690
1691 if (likely(ioctx)) {
1692 if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0))
1693 ret = read_events(ioctx, min_nr, nr, events, timeout);
1694 put_ioctx(ioctx);
1695 }
1696
1697 return ret;
1698}
1699
1700__initcall(aio_setup);
1701
1702EXPORT_SYMBOL(aio_complete);
1703EXPORT_SYMBOL(aio_put_req);
1704EXPORT_SYMBOL(wait_on_sync_kiocb);