blob: 969beb0e22311a4f7e7f4155524709d5dd8269bf [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>
Jens Axboeb9d128f2009-10-29 13:59:26 +010018#include <linux/backing-dev.h>
Badari Pulavarty027445c2006-09-30 23:28:46 -070019#include <linux/uio.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070020
21#define DEBUG 0
22
23#include <linux/sched.h>
24#include <linux/fs.h>
25#include <linux/file.h>
26#include <linux/mm.h>
27#include <linux/mman.h>
Michael S. Tsirkin3d2d8272009-09-21 17:03:51 -070028#include <linux/mmu_context.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029#include <linux/slab.h>
30#include <linux/timer.h>
31#include <linux/aio.h>
32#include <linux/highmem.h>
33#include <linux/workqueue.h>
34#include <linux/security.h>
Davide Libenzi9c3060b2007-05-10 22:23:21 -070035#include <linux/eventfd.h>
Jeff Moyercfb1e332009-10-02 18:57:36 -040036#include <linux/blkdev.h>
Jeff Moyer9d85cba2010-05-26 14:44:26 -070037#include <linux/compat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038
39#include <asm/kmap_types.h>
40#include <asm/uaccess.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070041
42#if DEBUG > 1
43#define dprintk printk
44#else
45#define dprintk(x...) do { ; } while (0)
46#endif
47
Linus Torvalds1da177e2005-04-16 15:20:36 -070048/*------ sysctl variables----*/
Zach Brownd55b5fd2005-11-07 00:59:31 -080049static DEFINE_SPINLOCK(aio_nr_lock);
50unsigned long aio_nr; /* current system wide number of aio requests */
51unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
Linus Torvalds1da177e2005-04-16 15:20:36 -070052/*----end sysctl variables---*/
53
Christoph Lametere18b8902006-12-06 20:33:20 -080054static struct kmem_cache *kiocb_cachep;
55static struct kmem_cache *kioctx_cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -070056
57static struct workqueue_struct *aio_wq;
58
59/* Used for rare fput completion. */
David Howells65f27f32006-11-22 14:55:48 +000060static void aio_fput_routine(struct work_struct *);
61static DECLARE_WORK(fput_work, aio_fput_routine);
Linus Torvalds1da177e2005-04-16 15:20:36 -070062
63static DEFINE_SPINLOCK(fput_lock);
Adrian Bunk25ee7e32005-04-25 08:18:14 -070064static LIST_HEAD(fput_head);
Linus Torvalds1da177e2005-04-16 15:20:36 -070065
David Howells65f27f32006-11-22 14:55:48 +000066static void aio_kick_handler(struct work_struct *);
Sébastien Duguc016e222005-06-28 20:44:59 -070067static void aio_queue_work(struct kioctx *);
Linus Torvalds1da177e2005-04-16 15:20:36 -070068
69/* aio_setup
70 * Creates the slab caches used by the aio routines, panic on
71 * failure as this is done early during the boot sequence.
72 */
73static int __init aio_setup(void)
74{
Christoph Lameter0a31bd52007-05-06 14:49:57 -070075 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
76 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -070077
Tejun Heod37adaa2011-01-26 17:42:27 +010078 aio_wq = alloc_workqueue("aio", 0, 1); /* used to limit concurrency */
Jens Axboecf159002011-03-02 20:12:18 -050079 BUG_ON(!aio_wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -070080
81 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
82
83 return 0;
84}
H Hartley Sweeten385773e2009-09-22 16:43:53 -070085__initcall(aio_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -070086
87static void aio_free_ring(struct kioctx *ctx)
88{
89 struct aio_ring_info *info = &ctx->ring_info;
90 long i;
91
92 for (i=0; i<info->nr_pages; i++)
93 put_page(info->ring_pages[i]);
94
95 if (info->mmap_size) {
96 down_write(&ctx->mm->mmap_sem);
97 do_munmap(ctx->mm, info->mmap_base, info->mmap_size);
98 up_write(&ctx->mm->mmap_sem);
99 }
100
101 if (info->ring_pages && info->ring_pages != info->internal_pages)
102 kfree(info->ring_pages);
103 info->ring_pages = NULL;
104 info->nr = 0;
105}
106
107static int aio_setup_ring(struct kioctx *ctx)
108{
109 struct aio_ring *ring;
110 struct aio_ring_info *info = &ctx->ring_info;
111 unsigned nr_events = ctx->max_reqs;
112 unsigned long size;
113 int nr_pages;
114
115 /* Compensate for the ring buffer's head/tail overlap entry */
116 nr_events += 2; /* 1 is required, 2 for good luck */
117
118 size = sizeof(struct aio_ring);
119 size += sizeof(struct io_event) * nr_events;
120 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
121
122 if (nr_pages < 0)
123 return -EINVAL;
124
125 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
126
127 info->nr = 0;
128 info->ring_pages = info->internal_pages;
129 if (nr_pages > AIO_RING_PAGES) {
Oliver Neukum11b0b5a2006-03-25 03:08:13 -0800130 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131 if (!info->ring_pages)
132 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133 }
134
135 info->mmap_size = nr_pages * PAGE_SIZE;
136 dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
137 down_write(&ctx->mm->mmap_sem);
138 info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
Robert P. J. Daye10a4432007-02-10 01:42:56 -0800139 PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700140 0);
141 if (IS_ERR((void *)info->mmap_base)) {
142 up_write(&ctx->mm->mmap_sem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143 info->mmap_size = 0;
144 aio_free_ring(ctx);
145 return -EAGAIN;
146 }
147
148 dprintk("mmap address: 0x%08lx\n", info->mmap_base);
149 info->nr_pages = get_user_pages(current, ctx->mm,
150 info->mmap_base, nr_pages,
151 1, 0, info->ring_pages, NULL);
152 up_write(&ctx->mm->mmap_sem);
153
154 if (unlikely(info->nr_pages != nr_pages)) {
155 aio_free_ring(ctx);
156 return -EAGAIN;
157 }
158
159 ctx->user_id = info->mmap_base;
160
161 info->nr = nr_events; /* trusted copy */
162
163 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
164 ring->nr = nr_events; /* user copy */
165 ring->id = ctx->user_id;
166 ring->head = ring->tail = 0;
167 ring->magic = AIO_RING_MAGIC;
168 ring->compat_features = AIO_RING_COMPAT_FEATURES;
169 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
170 ring->header_length = sizeof(struct aio_ring);
171 kunmap_atomic(ring, KM_USER0);
172
173 return 0;
174}
175
176
177/* aio_ring_event: returns a pointer to the event at the given index from
178 * kmap_atomic(, km). Release the pointer with put_aio_ring_event();
179 */
180#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
181#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
182#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
183
184#define aio_ring_event(info, nr, km) ({ \
185 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
186 struct io_event *__event; \
187 __event = kmap_atomic( \
188 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
189 __event += pos % AIO_EVENTS_PER_PAGE; \
190 __event; \
191})
192
193#define put_aio_ring_event(event, km) do { \
194 struct io_event *__event = (event); \
195 (void)__event; \
196 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
197} while(0)
198
Jens Axboeabf137d2008-12-09 08:11:22 +0100199static void ctx_rcu_free(struct rcu_head *head)
Adrian Bunkd5470b52008-04-29 00:58:57 -0700200{
Jens Axboeabf137d2008-12-09 08:11:22 +0100201 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
Adrian Bunkd5470b52008-04-29 00:58:57 -0700202 unsigned nr_events = ctx->max_reqs;
203
Adrian Bunkd5470b52008-04-29 00:58:57 -0700204 kmem_cache_free(kioctx_cachep, ctx);
205
206 if (nr_events) {
207 spin_lock(&aio_nr_lock);
208 BUG_ON(aio_nr - nr_events > aio_nr);
209 aio_nr -= nr_events;
210 spin_unlock(&aio_nr_lock);
211 }
212}
213
Jens Axboeabf137d2008-12-09 08:11:22 +0100214/* __put_ioctx
215 * Called when the last user of an aio context has gone away,
216 * and the struct needs to be freed.
217 */
218static void __put_ioctx(struct kioctx *ctx)
219{
220 BUG_ON(ctx->reqs_active);
221
222 cancel_delayed_work(&ctx->wq);
223 cancel_work_sync(&ctx->wq.work);
224 aio_free_ring(ctx);
225 mmdrop(ctx->mm);
226 ctx->mm = NULL;
227 pr_debug("__put_ioctx: freeing %p\n", ctx);
228 call_rcu(&ctx->rcu_head, ctx_rcu_free);
229}
230
Nick Piggin3bd9a5d2011-02-25 14:44:26 -0800231static inline void get_ioctx(struct kioctx *kioctx)
232{
233 BUG_ON(atomic_read(&kioctx->users) <= 0);
234 atomic_inc(&kioctx->users);
235}
236
237static inline int try_get_ioctx(struct kioctx *kioctx)
238{
239 return atomic_inc_not_zero(&kioctx->users);
240}
241
242static inline void put_ioctx(struct kioctx *kioctx)
243{
244 BUG_ON(atomic_read(&kioctx->users) <= 0);
245 if (unlikely(atomic_dec_and_test(&kioctx->users)))
246 __put_ioctx(kioctx);
247}
Adrian Bunkd5470b52008-04-29 00:58:57 -0700248
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249/* ioctx_alloc
250 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
251 */
252static struct kioctx *ioctx_alloc(unsigned nr_events)
253{
254 struct mm_struct *mm;
255 struct kioctx *ctx;
Jens Axboeabf137d2008-12-09 08:11:22 +0100256 int did_sync = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257
258 /* Prevent overflows */
259 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
260 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
261 pr_debug("ENOMEM: nr_events too high\n");
262 return ERR_PTR(-EINVAL);
263 }
264
Zach Brownd55b5fd2005-11-07 00:59:31 -0800265 if ((unsigned long)nr_events > aio_max_nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266 return ERR_PTR(-EAGAIN);
267
Robert P. J. Dayc3762222007-02-10 01:45:03 -0800268 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269 if (!ctx)
270 return ERR_PTR(-ENOMEM);
271
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272 ctx->max_reqs = nr_events;
273 mm = ctx->mm = current->mm;
274 atomic_inc(&mm->mm_count);
275
276 atomic_set(&ctx->users, 1);
277 spin_lock_init(&ctx->ctx_lock);
278 spin_lock_init(&ctx->ring_info.ring_lock);
279 init_waitqueue_head(&ctx->wait);
280
281 INIT_LIST_HEAD(&ctx->active_reqs);
282 INIT_LIST_HEAD(&ctx->run_list);
David Howells65f27f32006-11-22 14:55:48 +0000283 INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284
285 if (aio_setup_ring(ctx) < 0)
286 goto out_freectx;
287
288 /* limit the number of system wide aios */
Jens Axboeabf137d2008-12-09 08:11:22 +0100289 do {
290 spin_lock_bh(&aio_nr_lock);
291 if (aio_nr + nr_events > aio_max_nr ||
292 aio_nr + nr_events < aio_nr)
293 ctx->max_reqs = 0;
294 else
295 aio_nr += ctx->max_reqs;
296 spin_unlock_bh(&aio_nr_lock);
297 if (ctx->max_reqs || did_sync)
298 break;
299
300 /* wait for rcu callbacks to have completed before giving up */
301 synchronize_rcu();
302 did_sync = 1;
303 ctx->max_reqs = nr_events;
304 } while (1);
305
Zach Brownd55b5fd2005-11-07 00:59:31 -0800306 if (ctx->max_reqs == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307 goto out_cleanup;
308
Jeff Moyer39fa0032008-04-29 01:03:48 -0700309 /* now link into global list. */
Jens Axboeabf137d2008-12-09 08:11:22 +0100310 spin_lock(&mm->ioctx_lock);
311 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
312 spin_unlock(&mm->ioctx_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313
314 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
315 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
316 return ctx;
317
318out_cleanup:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700319 __put_ioctx(ctx);
320 return ERR_PTR(-EAGAIN);
321
322out_freectx:
323 mmdrop(mm);
324 kmem_cache_free(kioctx_cachep, ctx);
325 ctx = ERR_PTR(-ENOMEM);
326
327 dprintk("aio: error allocating ioctx %p\n", ctx);
328 return ctx;
329}
330
331/* aio_cancel_all
332 * Cancels all outstanding aio requests on an aio context. Used
333 * when the processes owning a context have all exited to encourage
334 * the rapid destruction of the kioctx.
335 */
336static void aio_cancel_all(struct kioctx *ctx)
337{
338 int (*cancel)(struct kiocb *, struct io_event *);
339 struct io_event res;
340 spin_lock_irq(&ctx->ctx_lock);
341 ctx->dead = 1;
342 while (!list_empty(&ctx->active_reqs)) {
343 struct list_head *pos = ctx->active_reqs.next;
344 struct kiocb *iocb = list_kiocb(pos);
345 list_del_init(&iocb->ki_list);
346 cancel = iocb->ki_cancel;
347 kiocbSetCancelled(iocb);
348 if (cancel) {
349 iocb->ki_users++;
350 spin_unlock_irq(&ctx->ctx_lock);
351 cancel(iocb, &res);
352 spin_lock_irq(&ctx->ctx_lock);
353 }
354 }
355 spin_unlock_irq(&ctx->ctx_lock);
356}
357
Adrian Bunk25ee7e32005-04-25 08:18:14 -0700358static void wait_for_all_aios(struct kioctx *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359{
360 struct task_struct *tsk = current;
361 DECLARE_WAITQUEUE(wait, tsk);
362
Ken Chendee11c22007-02-03 01:13:45 -0800363 spin_lock_irq(&ctx->ctx_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364 if (!ctx->reqs_active)
Ken Chendee11c22007-02-03 01:13:45 -0800365 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366
367 add_wait_queue(&ctx->wait, &wait);
368 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
369 while (ctx->reqs_active) {
Ken Chendee11c22007-02-03 01:13:45 -0800370 spin_unlock_irq(&ctx->ctx_lock);
Jeff Moyer41d10da2007-10-16 23:27:20 -0700371 io_schedule();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
Ken Chendee11c22007-02-03 01:13:45 -0800373 spin_lock_irq(&ctx->ctx_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374 }
375 __set_task_state(tsk, TASK_RUNNING);
376 remove_wait_queue(&ctx->wait, &wait);
Ken Chendee11c22007-02-03 01:13:45 -0800377
378out:
379 spin_unlock_irq(&ctx->ctx_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380}
381
382/* wait_on_sync_kiocb:
383 * Waits on the given sync kiocb to complete.
384 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800385ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386{
387 while (iocb->ki_users) {
388 set_current_state(TASK_UNINTERRUPTIBLE);
389 if (!iocb->ki_users)
390 break;
Jeff Moyer41d10da2007-10-16 23:27:20 -0700391 io_schedule();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392 }
393 __set_current_state(TASK_RUNNING);
394 return iocb->ki_user_data;
395}
H Hartley Sweeten385773e2009-09-22 16:43:53 -0700396EXPORT_SYMBOL(wait_on_sync_kiocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397
398/* exit_aio: called when the last user of mm goes away. At this point,
399 * there is no way for any new requests to be submited or any of the
400 * io_* syscalls to be called on the context. However, there may be
401 * outstanding requests which hold references to the context; as they
402 * go away, they will call put_ioctx and release any pinned memory
403 * associated with the request (held via struct page * references).
404 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800405void exit_aio(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406{
Jens Axboeabf137d2008-12-09 08:11:22 +0100407 struct kioctx *ctx;
408
409 while (!hlist_empty(&mm->ioctx_list)) {
410 ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
411 hlist_del_rcu(&ctx->list);
412
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413 aio_cancel_all(ctx);
414
415 wait_for_all_aios(ctx);
416 /*
Andrew Mortona9df62c2007-05-09 02:33:54 -0700417 * Ensure we don't leave the ctx on the aio_wq
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418 */
Oleg Nesterov28e53bd2007-05-09 02:34:22 -0700419 cancel_work_sync(&ctx->wq.work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420
421 if (1 != atomic_read(&ctx->users))
422 printk(KERN_DEBUG
423 "exit_aio:ioctx still alive: %d %d %d\n",
424 atomic_read(&ctx->users), ctx->dead,
425 ctx->reqs_active);
426 put_ioctx(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427 }
428}
429
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430/* aio_get_req
431 * Allocate a slot for an aio request. Increments the users count
432 * of the kioctx so that the kioctx stays around until all requests are
433 * complete. Returns NULL if no requests are free.
434 *
435 * Returns with kiocb->users set to 2. The io submit code path holds
436 * an extra reference while submitting the i/o.
437 * This prevents races between the aio code path referencing the
438 * req (after submitting it) and aio_complete() freeing the req.
439 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800440static struct kiocb *__aio_get_req(struct kioctx *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441{
442 struct kiocb *req = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443
444 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
445 if (unlikely(!req))
446 return NULL;
447
Zach Brown4faa5282005-10-17 16:43:33 -0700448 req->ki_flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449 req->ki_users = 2;
450 req->ki_key = 0;
451 req->ki_ctx = ctx;
452 req->ki_cancel = NULL;
453 req->ki_retry = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454 req->ki_dtor = NULL;
455 req->private = NULL;
Badari Pulavartyeed4e512006-09-30 23:28:49 -0700456 req->ki_iovec = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 INIT_LIST_HEAD(&req->ki_run_list);
Davide Libenzi87c3a862009-03-18 17:04:19 -0700458 req->ki_eventfd = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460 return req;
461}
462
Jeff Moyer080d6762011-11-02 13:40:10 -0700463/*
464 * struct kiocb's are allocated in batches to reduce the number of
465 * times the ctx lock is acquired and released.
466 */
467#define KIOCB_BATCH_SIZE 32L
468struct kiocb_batch {
469 struct list_head head;
470 long count; /* number of requests left to allocate */
471};
472
473static void kiocb_batch_init(struct kiocb_batch *batch, long total)
474{
475 INIT_LIST_HEAD(&batch->head);
476 batch->count = total;
477}
478
Gleb Natapov69e47472012-01-08 17:07:28 +0200479static void kiocb_batch_free(struct kioctx *ctx, struct kiocb_batch *batch)
Jeff Moyer080d6762011-11-02 13:40:10 -0700480{
481 struct kiocb *req, *n;
482
Gleb Natapov69e47472012-01-08 17:07:28 +0200483 if (list_empty(&batch->head))
484 return;
485
486 spin_lock_irq(&ctx->ctx_lock);
Jeff Moyer080d6762011-11-02 13:40:10 -0700487 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
488 list_del(&req->ki_batch);
Gleb Natapov69e47472012-01-08 17:07:28 +0200489 list_del(&req->ki_list);
Jeff Moyer080d6762011-11-02 13:40:10 -0700490 kmem_cache_free(kiocb_cachep, req);
Gleb Natapov69e47472012-01-08 17:07:28 +0200491 ctx->reqs_active--;
Jeff Moyer080d6762011-11-02 13:40:10 -0700492 }
Gleb Natapov69e47472012-01-08 17:07:28 +0200493 spin_unlock_irq(&ctx->ctx_lock);
Jeff Moyer080d6762011-11-02 13:40:10 -0700494}
495
496/*
497 * Allocate a batch of kiocbs. This avoids taking and dropping the
498 * context lock a lot during setup.
499 */
500static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch)
501{
502 unsigned short allocated, to_alloc;
503 long avail;
504 bool called_fput = false;
505 struct kiocb *req, *n;
506 struct aio_ring *ring;
507
508 to_alloc = min(batch->count, KIOCB_BATCH_SIZE);
509 for (allocated = 0; allocated < to_alloc; allocated++) {
510 req = __aio_get_req(ctx);
511 if (!req)
512 /* allocation failed, go with what we've got */
513 break;
514 list_add(&req->ki_batch, &batch->head);
515 }
516
517 if (allocated == 0)
518 goto out;
519
520retry:
521 spin_lock_irq(&ctx->ctx_lock);
522 ring = kmap_atomic(ctx->ring_info.ring_pages[0]);
523
524 avail = aio_ring_avail(&ctx->ring_info, ring) - ctx->reqs_active;
525 BUG_ON(avail < 0);
526 if (avail == 0 && !called_fput) {
527 /*
528 * Handle a potential starvation case. It is possible that
529 * we hold the last reference on a struct file, causing us
530 * to delay the final fput to non-irq context. In this case,
531 * ctx->reqs_active is artificially high. Calling the fput
532 * routine here may free up a slot in the event completion
533 * ring, allowing this allocation to succeed.
534 */
535 kunmap_atomic(ring);
536 spin_unlock_irq(&ctx->ctx_lock);
537 aio_fput_routine(NULL);
538 called_fput = true;
539 goto retry;
540 }
541
542 if (avail < allocated) {
543 /* Trim back the number of requests. */
544 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
545 list_del(&req->ki_batch);
546 kmem_cache_free(kiocb_cachep, req);
547 if (--allocated <= avail)
548 break;
549 }
550 }
551
552 batch->count -= allocated;
553 list_for_each_entry(req, &batch->head, ki_batch) {
554 list_add(&req->ki_list, &ctx->active_reqs);
555 ctx->reqs_active++;
556 }
557
558 kunmap_atomic(ring);
559 spin_unlock_irq(&ctx->ctx_lock);
560
561out:
562 return allocated;
563}
564
565static inline struct kiocb *aio_get_req(struct kioctx *ctx,
566 struct kiocb_batch *batch)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567{
568 struct kiocb *req;
Jeff Moyer080d6762011-11-02 13:40:10 -0700569
570 if (list_empty(&batch->head))
571 if (kiocb_batch_refill(ctx, batch) == 0)
572 return NULL;
573 req = list_first_entry(&batch->head, struct kiocb, ki_batch);
574 list_del(&req->ki_batch);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575 return req;
576}
577
578static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
579{
Zach Brownd00689a2005-11-13 16:07:34 -0800580 assert_spin_locked(&ctx->ctx_lock);
581
Davide Libenzi13389012009-06-30 11:41:11 -0700582 if (req->ki_eventfd != NULL)
583 eventfd_ctx_put(req->ki_eventfd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584 if (req->ki_dtor)
585 req->ki_dtor(req);
Badari Pulavartyeed4e512006-09-30 23:28:49 -0700586 if (req->ki_iovec != &req->ki_inline_vec)
587 kfree(req->ki_iovec);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588 kmem_cache_free(kiocb_cachep, req);
589 ctx->reqs_active--;
590
591 if (unlikely(!ctx->reqs_active && ctx->dead))
Roland Dreiere91f90b2011-03-22 16:35:10 -0700592 wake_up_all(&ctx->wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593}
594
David Howells65f27f32006-11-22 14:55:48 +0000595static void aio_fput_routine(struct work_struct *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700596{
597 spin_lock_irq(&fput_lock);
598 while (likely(!list_empty(&fput_head))) {
599 struct kiocb *req = list_kiocb(fput_head.next);
600 struct kioctx *ctx = req->ki_ctx;
601
602 list_del(&req->ki_list);
603 spin_unlock_irq(&fput_lock);
604
Davide Libenzi87c3a862009-03-18 17:04:19 -0700605 /* Complete the fput(s) */
606 if (req->ki_filp != NULL)
Al Virod7065da2010-05-26 15:13:55 -0400607 fput(req->ki_filp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608
609 /* Link the iocb into the context's free list */
610 spin_lock_irq(&ctx->ctx_lock);
611 really_put_req(ctx, req);
612 spin_unlock_irq(&ctx->ctx_lock);
613
614 put_ioctx(ctx);
615 spin_lock_irq(&fput_lock);
616 }
617 spin_unlock_irq(&fput_lock);
618}
619
620/* __aio_put_req
621 * Returns true if this put was the last user of the request.
622 */
623static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
624{
Al Viro516e0cc2008-07-26 00:39:17 -0400625 dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
626 req, atomic_long_read(&req->ki_filp->f_count));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627
Zach Brownd00689a2005-11-13 16:07:34 -0800628 assert_spin_locked(&ctx->ctx_lock);
629
Davide Libenzi87c3a862009-03-18 17:04:19 -0700630 req->ki_users--;
Eric Sesterhenn93e06b42006-11-30 05:29:23 +0100631 BUG_ON(req->ki_users < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632 if (likely(req->ki_users))
633 return 0;
634 list_del(&req->ki_list); /* remove from active_reqs */
635 req->ki_cancel = NULL;
636 req->ki_retry = NULL;
637
Davide Libenzi87c3a862009-03-18 17:04:19 -0700638 /*
639 * Try to optimize the aio and eventfd file* puts, by avoiding to
Al Virod7065da2010-05-26 15:13:55 -0400640 * schedule work in case it is not final fput() time. In normal cases,
Davide Libenzi87c3a862009-03-18 17:04:19 -0700641 * we would not be holding the last reference to the file*, so
642 * this function will be executed w/out any aio kthread wakeup.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643 */
Al Virod7065da2010-05-26 15:13:55 -0400644 if (unlikely(!fput_atomic(req->ki_filp))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645 get_ioctx(ctx);
646 spin_lock(&fput_lock);
647 list_add(&req->ki_list, &fput_head);
648 spin_unlock(&fput_lock);
Tejun Heod37adaa2011-01-26 17:42:27 +0100649 schedule_work(&fput_work);
Davide Libenzi13389012009-06-30 11:41:11 -0700650 } else {
651 req->ki_filp = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652 really_put_req(ctx, req);
Davide Libenzi13389012009-06-30 11:41:11 -0700653 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654 return 1;
655}
656
657/* aio_put_req
658 * Returns true if this put was the last user of the kiocb,
659 * false if the request is still in use.
660 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800661int aio_put_req(struct kiocb *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662{
663 struct kioctx *ctx = req->ki_ctx;
664 int ret;
665 spin_lock_irq(&ctx->ctx_lock);
666 ret = __aio_put_req(ctx, req);
667 spin_unlock_irq(&ctx->ctx_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668 return ret;
669}
H Hartley Sweeten385773e2009-09-22 16:43:53 -0700670EXPORT_SYMBOL(aio_put_req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671
Adrian Bunkd5470b52008-04-29 00:58:57 -0700672static struct kioctx *lookup_ioctx(unsigned long ctx_id)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673{
Jens Axboeabf137d2008-12-09 08:11:22 +0100674 struct mm_struct *mm = current->mm;
Jeff Moyer65c24492009-03-18 17:04:21 -0700675 struct kioctx *ctx, *ret = NULL;
Jens Axboeabf137d2008-12-09 08:11:22 +0100676 struct hlist_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677
Jens Axboeabf137d2008-12-09 08:11:22 +0100678 rcu_read_lock();
679
680 hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) {
Nick Piggin3bd9a5d2011-02-25 14:44:26 -0800681 /*
682 * RCU protects us against accessing freed memory but
683 * we have to be careful not to get a reference when the
684 * reference count already dropped to 0 (ctx->dead test
685 * is unreliable because of races).
686 */
687 if (ctx->user_id == ctx_id && !ctx->dead && try_get_ioctx(ctx)){
Jeff Moyer65c24492009-03-18 17:04:21 -0700688 ret = ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689 break;
690 }
Jens Axboeabf137d2008-12-09 08:11:22 +0100691 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692
Jens Axboeabf137d2008-12-09 08:11:22 +0100693 rcu_read_unlock();
Jeff Moyer65c24492009-03-18 17:04:21 -0700694 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695}
696
697/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698 * Queue up a kiocb to be retried. Assumes that the kiocb
699 * has already been marked as kicked, and places it on
700 * the retry run list for the corresponding ioctx, if it
701 * isn't already queued. Returns 1 if it actually queued
702 * the kiocb (to tell the caller to activate the work
703 * queue to process it), or 0, if it found that it was
704 * already queued.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705 */
706static inline int __queue_kicked_iocb(struct kiocb *iocb)
707{
708 struct kioctx *ctx = iocb->ki_ctx;
709
Zach Brownd00689a2005-11-13 16:07:34 -0800710 assert_spin_locked(&ctx->ctx_lock);
711
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712 if (list_empty(&iocb->ki_run_list)) {
713 list_add_tail(&iocb->ki_run_list,
714 &ctx->run_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715 return 1;
716 }
717 return 0;
718}
719
720/* aio_run_iocb
721 * This is the core aio execution routine. It is
722 * invoked both for initial i/o submission and
723 * subsequent retries via the aio_kick_handler.
724 * Expects to be invoked with iocb->ki_ctx->lock
Andreas Mohrd6e05ed2006-06-26 18:35:02 +0200725 * already held. The lock is released and reacquired
Linus Torvalds1da177e2005-04-16 15:20:36 -0700726 * as needed during processing.
727 *
728 * Calls the iocb retry method (already setup for the
729 * iocb on initial submission) for operation specific
730 * handling, but takes care of most of common retry
731 * execution details for a given iocb. The retry method
732 * needs to be non-blocking as far as possible, to avoid
733 * holding up other iocbs waiting to be serviced by the
734 * retry kernel thread.
735 *
736 * The trickier parts in this code have to do with
737 * ensuring that only one retry instance is in progress
738 * for a given iocb at any time. Providing that guarantee
739 * simplifies the coding of individual aio operations as
740 * it avoids various potential races.
741 */
742static ssize_t aio_run_iocb(struct kiocb *iocb)
743{
744 struct kioctx *ctx = iocb->ki_ctx;
745 ssize_t (*retry)(struct kiocb *);
746 ssize_t ret;
747
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748 if (!(retry = iocb->ki_retry)) {
749 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
750 return 0;
751 }
752
753 /*
754 * We don't want the next retry iteration for this
755 * operation to start until this one has returned and
756 * updated the iocb state. However, wait_queue functions
757 * can trigger a kick_iocb from interrupt context in the
758 * meantime, indicating that data is available for the next
759 * iteration. We want to remember that and enable the
760 * next retry iteration _after_ we are through with
761 * this one.
762 *
763 * So, in order to be able to register a "kick", but
764 * prevent it from being queued now, we clear the kick
765 * flag, but make the kick code *think* that the iocb is
766 * still on the run list until we are actually done.
767 * When we are done with this iteration, we check if
768 * the iocb was kicked in the meantime and if so, queue
769 * it up afresh.
770 */
771
772 kiocbClearKicked(iocb);
773
774 /*
775 * This is so that aio_complete knows it doesn't need to
776 * pull the iocb off the run list (We can't just call
777 * INIT_LIST_HEAD because we don't want a kick_iocb to
778 * queue this on the run list yet)
779 */
780 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
781 spin_unlock_irq(&ctx->ctx_lock);
782
783 /* Quit retrying if the i/o has been cancelled */
784 if (kiocbIsCancelled(iocb)) {
785 ret = -EINTR;
786 aio_complete(iocb, ret, 0);
787 /* must not access the iocb after this */
788 goto out;
789 }
790
791 /*
792 * Now we are all set to call the retry method in async
Alexey Dobriyan6212e3a2007-10-18 03:04:56 -0700793 * context.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795 ret = retry(iocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796
Jan Karaa0c42ba2010-09-22 13:05:03 -0700797 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
798 /*
799 * There's no easy way to restart the syscall since other AIO's
800 * may be already running. Just fail this IO with EINTR.
801 */
802 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
803 ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
804 ret = -EINTR;
Zach Brown897f15f2005-09-30 11:58:55 -0700805 aio_complete(iocb, ret, 0);
Jan Karaa0c42ba2010-09-22 13:05:03 -0700806 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807out:
808 spin_lock_irq(&ctx->ctx_lock);
809
810 if (-EIOCBRETRY == ret) {
811 /*
812 * OK, now that we are done with this iteration
813 * and know that there is more left to go,
814 * this is where we let go so that a subsequent
815 * "kick" can start the next iteration
816 */
817
818 /* will make __queue_kicked_iocb succeed from here on */
819 INIT_LIST_HEAD(&iocb->ki_run_list);
820 /* we must queue the next iteration ourselves, if it
821 * has already been kicked */
822 if (kiocbIsKicked(iocb)) {
823 __queue_kicked_iocb(iocb);
Sébastien Duguc016e222005-06-28 20:44:59 -0700824
825 /*
826 * __queue_kicked_iocb will always return 1 here, because
827 * iocb->ki_run_list is empty at this point so it should
828 * be safe to unconditionally queue the context into the
829 * work queue.
830 */
831 aio_queue_work(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832 }
833 }
834 return ret;
835}
836
837/*
838 * __aio_run_iocbs:
839 * Process all pending retries queued on the ioctx
840 * run list.
841 * Assumes it is operating within the aio issuer's mm
Zach Brownd00689a2005-11-13 16:07:34 -0800842 * context.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843 */
844static int __aio_run_iocbs(struct kioctx *ctx)
845{
846 struct kiocb *iocb;
Oleg Nesterov626ab0e2006-06-23 02:05:55 -0700847 struct list_head run_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848
Zach Brownd00689a2005-11-13 16:07:34 -0800849 assert_spin_locked(&ctx->ctx_lock);
850
Oleg Nesterov626ab0e2006-06-23 02:05:55 -0700851 list_replace_init(&ctx->run_list, &run_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852 while (!list_empty(&run_list)) {
853 iocb = list_entry(run_list.next, struct kiocb,
854 ki_run_list);
855 list_del(&iocb->ki_run_list);
856 /*
857 * Hold an extra reference while retrying i/o.
858 */
859 iocb->ki_users++; /* grab extra reference */
860 aio_run_iocb(iocb);
Ken Chendee11c22007-02-03 01:13:45 -0800861 __aio_put_req(ctx, iocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863 if (!list_empty(&ctx->run_list))
864 return 1;
865 return 0;
866}
867
868static void aio_queue_work(struct kioctx * ctx)
869{
870 unsigned long timeout;
871 /*
872 * if someone is waiting, get the work started right
873 * away, otherwise, use a longer delay
874 */
875 smp_mb();
876 if (waitqueue_active(&ctx->wait))
877 timeout = 1;
878 else
879 timeout = HZ/10;
880 queue_delayed_work(aio_wq, &ctx->wq, timeout);
881}
882
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883/*
Jeff Moyerd3486f82011-01-12 17:01:08 -0800884 * aio_run_all_iocbs:
885 * Process all pending retries queued on the ioctx
886 * run list, and keep running them until the list
887 * stays empty.
888 * Assumes it is operating within the aio issuer's mm context.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889 */
890static inline void aio_run_all_iocbs(struct kioctx *ctx)
891{
892 spin_lock_irq(&ctx->ctx_lock);
893 while (__aio_run_iocbs(ctx))
894 ;
895 spin_unlock_irq(&ctx->ctx_lock);
896}
897
898/*
899 * aio_kick_handler:
900 * Work queue handler triggered to process pending
901 * retries on an ioctx. Takes on the aio issuer's
902 * mm context before running the iocbs, so that
903 * copy_xxx_user operates on the issuer's address
904 * space.
905 * Run on aiod's context.
906 */
David Howells65f27f32006-11-22 14:55:48 +0000907static void aio_kick_handler(struct work_struct *work)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700908{
David Howells65f27f32006-11-22 14:55:48 +0000909 struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700910 mm_segment_t oldfs = get_fs();
Zach Brown1ebb1102006-12-29 16:47:02 -0800911 struct mm_struct *mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912 int requeue;
913
914 set_fs(USER_DS);
915 use_mm(ctx->mm);
916 spin_lock_irq(&ctx->ctx_lock);
917 requeue =__aio_run_iocbs(ctx);
Zach Brown1ebb1102006-12-29 16:47:02 -0800918 mm = ctx->mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700919 spin_unlock_irq(&ctx->ctx_lock);
Zach Brown1ebb1102006-12-29 16:47:02 -0800920 unuse_mm(mm);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921 set_fs(oldfs);
922 /*
923 * we're in a worker thread already, don't use queue_delayed_work,
924 */
925 if (requeue)
David Howells52bad642006-11-22 14:54:01 +0000926 queue_delayed_work(aio_wq, &ctx->wq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927}
928
929
930/*
931 * Called by kick_iocb to queue the kiocb for retry
932 * and if required activate the aio work queue to process
933 * it
934 */
Zach Brown998765e2005-09-30 11:58:54 -0700935static void try_queue_kicked_iocb(struct kiocb *iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936{
937 struct kioctx *ctx = iocb->ki_ctx;
938 unsigned long flags;
939 int run = 0;
940
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941 spin_lock_irqsave(&ctx->ctx_lock, flags);
Zach Brown998765e2005-09-30 11:58:54 -0700942 /* set this inside the lock so that we can't race with aio_run_iocb()
943 * testing it and putting the iocb on the run list under the lock */
944 if (!kiocbTryKick(iocb))
945 run = __queue_kicked_iocb(iocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
Ken Chen644d3a02005-05-01 08:59:15 -0700947 if (run)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700948 aio_queue_work(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949}
950
951/*
952 * kick_iocb:
953 * Called typically from a wait queue callback context
Shaohua Lifac046a2009-12-15 16:47:47 -0800954 * to trigger a retry of the iocb.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700955 * The retry is usually executed by aio workqueue
956 * threads (See aio_kick_handler).
957 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800958void kick_iocb(struct kiocb *iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959{
960 /* sync iocbs are easy: they can only ever be executing from a
961 * single context. */
962 if (is_sync_kiocb(iocb)) {
963 kiocbSetKicked(iocb);
964 wake_up_process(iocb->ki_obj.tsk);
965 return;
966 }
967
Zach Brown998765e2005-09-30 11:58:54 -0700968 try_queue_kicked_iocb(iocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969}
970EXPORT_SYMBOL(kick_iocb);
971
972/* aio_complete
973 * Called when the io request on the given iocb is complete.
974 * Returns true if this is the last user of the request. The
975 * only other user of the request can be the cancellation code.
976 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800977int aio_complete(struct kiocb *iocb, long res, long res2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700978{
979 struct kioctx *ctx = iocb->ki_ctx;
980 struct aio_ring_info *info;
981 struct aio_ring *ring;
982 struct io_event *event;
983 unsigned long flags;
984 unsigned long tail;
985 int ret;
986
Zach Brown20dcae32005-11-13 16:07:33 -0800987 /*
988 * Special case handling for sync iocbs:
989 * - events go directly into the iocb for fast handling
990 * - the sync task with the iocb in its stack holds the single iocb
991 * ref, no other paths have a way to get another ref
992 * - the sync task helpfully left a reference to itself in the iocb
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993 */
994 if (is_sync_kiocb(iocb)) {
Zach Brown20dcae32005-11-13 16:07:33 -0800995 BUG_ON(iocb->ki_users != 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996 iocb->ki_user_data = res;
Zach Brown20dcae32005-11-13 16:07:33 -0800997 iocb->ki_users = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700998 wake_up_process(iocb->ki_obj.tsk);
Zach Brown20dcae32005-11-13 16:07:33 -0800999 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000 }
1001
1002 info = &ctx->ring_info;
1003
1004 /* add a completion event to the ring buffer.
1005 * must be done holding ctx->ctx_lock to prevent
1006 * other code from messing with the tail
1007 * pointer since we might be called from irq
1008 * context.
1009 */
1010 spin_lock_irqsave(&ctx->ctx_lock, flags);
1011
1012 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
1013 list_del_init(&iocb->ki_run_list);
1014
1015 /*
1016 * cancelled requests don't get events, userland was given one
1017 * when the event got cancelled.
1018 */
1019 if (kiocbIsCancelled(iocb))
1020 goto put_rq;
1021
1022 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
1023
1024 tail = info->tail;
1025 event = aio_ring_event(info, tail, KM_IRQ0);
Ken Chen4bf69b22005-05-01 08:59:15 -07001026 if (++tail >= info->nr)
1027 tail = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001028
1029 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
1030 event->data = iocb->ki_user_data;
1031 event->res = res;
1032 event->res2 = res2;
1033
1034 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
1035 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
1036 res, res2);
1037
1038 /* after flagging the request as done, we
1039 * must never even look at it again
1040 */
1041 smp_wmb(); /* make event visible before updating tail */
1042
1043 info->tail = tail;
1044 ring->tail = tail;
1045
1046 put_aio_ring_event(event, KM_IRQ0);
1047 kunmap_atomic(ring, KM_IRQ1);
1048
1049 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
Davide Libenzi8d1c98b2008-04-10 21:29:19 -07001050
1051 /*
1052 * Check if the user asked us to deliver the result through an
1053 * eventfd. The eventfd_signal() function is safe to be called
1054 * from IRQ context.
1055 */
Davide Libenzi87c3a862009-03-18 17:04:19 -07001056 if (iocb->ki_eventfd != NULL)
Davide Libenzi8d1c98b2008-04-10 21:29:19 -07001057 eventfd_signal(iocb->ki_eventfd, 1);
1058
Linus Torvalds1da177e2005-04-16 15:20:36 -07001059put_rq:
1060 /* everything turned out well, dispose of the aiocb. */
1061 ret = __aio_put_req(ctx, iocb);
1062
Quentin Barnes6cb2a212008-03-19 17:00:39 -07001063 /*
1064 * We have to order our ring_info tail store above and test
1065 * of the wait list below outside the wait lock. This is
1066 * like in wake_up_bit() where clearing a bit has to be
1067 * ordered with the unlocked test.
1068 */
1069 smp_mb();
1070
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071 if (waitqueue_active(&ctx->wait))
1072 wake_up(&ctx->wait);
1073
Ken Chendee11c22007-02-03 01:13:45 -08001074 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001075 return ret;
1076}
H Hartley Sweeten385773e2009-09-22 16:43:53 -07001077EXPORT_SYMBOL(aio_complete);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001078
1079/* aio_read_evt
1080 * Pull an event off of the ioctx's event ring. Returns the number of
1081 * events fetched (0 or 1 ;-)
1082 * FIXME: make this use cmpxchg.
1083 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1084 */
1085static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1086{
1087 struct aio_ring_info *info = &ioctx->ring_info;
1088 struct aio_ring *ring;
1089 unsigned long head;
1090 int ret = 0;
1091
1092 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1093 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1094 (unsigned long)ring->head, (unsigned long)ring->tail,
1095 (unsigned long)ring->nr);
1096
1097 if (ring->head == ring->tail)
1098 goto out;
1099
1100 spin_lock(&info->ring_lock);
1101
1102 head = ring->head % info->nr;
1103 if (head != ring->tail) {
1104 struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1105 *ent = *evp;
1106 head = (head + 1) % info->nr;
1107 smp_mb(); /* finish reading the event before updatng the head */
1108 ring->head = head;
1109 ret = 1;
1110 put_aio_ring_event(evp, KM_USER1);
1111 }
1112 spin_unlock(&info->ring_lock);
1113
1114out:
1115 kunmap_atomic(ring, KM_USER0);
1116 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1117 (unsigned long)ring->head, (unsigned long)ring->tail);
1118 return ret;
1119}
1120
1121struct aio_timeout {
1122 struct timer_list timer;
1123 int timed_out;
1124 struct task_struct *p;
1125};
1126
1127static void timeout_func(unsigned long data)
1128{
1129 struct aio_timeout *to = (struct aio_timeout *)data;
1130
1131 to->timed_out = 1;
1132 wake_up_process(to->p);
1133}
1134
1135static inline void init_timeout(struct aio_timeout *to)
1136{
Thomas Gleixnerc6f3a972008-04-30 00:55:03 -07001137 setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001138 to->timed_out = 0;
1139 to->p = current;
1140}
1141
1142static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1143 const struct timespec *ts)
1144{
1145 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1146 if (time_after(to->timer.expires, jiffies))
1147 add_timer(&to->timer);
1148 else
1149 to->timed_out = 1;
1150}
1151
1152static inline void clear_timeout(struct aio_timeout *to)
1153{
1154 del_singleshot_timer_sync(&to->timer);
1155}
1156
1157static int read_events(struct kioctx *ctx,
1158 long min_nr, long nr,
1159 struct io_event __user *event,
1160 struct timespec __user *timeout)
1161{
1162 long start_jiffies = jiffies;
1163 struct task_struct *tsk = current;
1164 DECLARE_WAITQUEUE(wait, tsk);
1165 int ret;
1166 int i = 0;
1167 struct io_event ent;
1168 struct aio_timeout to;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169 int retry = 0;
1170
1171 /* needed to zero any padding within an entry (there shouldn't be
1172 * any, but C is fun!
1173 */
1174 memset(&ent, 0, sizeof(ent));
1175retry:
1176 ret = 0;
1177 while (likely(i < nr)) {
1178 ret = aio_read_evt(ctx, &ent);
1179 if (unlikely(ret <= 0))
1180 break;
1181
1182 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1183 ent.data, ent.obj, ent.res, ent.res2);
1184
1185 /* Could we split the check in two? */
1186 ret = -EFAULT;
1187 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1188 dprintk("aio: lost an event due to EFAULT.\n");
1189 break;
1190 }
1191 ret = 0;
1192
1193 /* Good, event copied to userland, update counts. */
1194 event ++;
1195 i ++;
1196 }
1197
1198 if (min_nr <= i)
1199 return i;
1200 if (ret)
1201 return ret;
1202
1203 /* End fast path */
1204
1205 /* racey check, but it gets redone */
1206 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1207 retry = 1;
1208 aio_run_all_iocbs(ctx);
1209 goto retry;
1210 }
1211
1212 init_timeout(&to);
1213 if (timeout) {
1214 struct timespec ts;
1215 ret = -EFAULT;
1216 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1217 goto out;
1218
1219 set_timeout(start_jiffies, &to, &ts);
1220 }
1221
1222 while (likely(i < nr)) {
1223 add_wait_queue_exclusive(&ctx->wait, &wait);
1224 do {
1225 set_task_state(tsk, TASK_INTERRUPTIBLE);
1226 ret = aio_read_evt(ctx, &ent);
1227 if (ret)
1228 break;
1229 if (min_nr <= i)
1230 break;
Jeff Moyere92adcb2008-04-28 02:12:04 -07001231 if (unlikely(ctx->dead)) {
1232 ret = -EINVAL;
1233 break;
1234 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235 if (to.timed_out) /* Only check after read evt */
1236 break;
Jeff Moyere00ba3d2007-12-04 23:45:02 -08001237 /* Try to only show up in io wait if there are ops
1238 * in flight */
1239 if (ctx->reqs_active)
1240 io_schedule();
1241 else
1242 schedule();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001243 if (signal_pending(tsk)) {
1244 ret = -EINTR;
1245 break;
1246 }
1247 /*ret = aio_read_evt(ctx, &ent);*/
1248 } while (1) ;
1249
1250 set_task_state(tsk, TASK_RUNNING);
1251 remove_wait_queue(&ctx->wait, &wait);
1252
1253 if (unlikely(ret <= 0))
1254 break;
1255
1256 ret = -EFAULT;
1257 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1258 dprintk("aio: lost an event due to EFAULT.\n");
1259 break;
1260 }
1261
1262 /* Good, event copied to userland, update counts. */
1263 event ++;
1264 i ++;
1265 }
1266
1267 if (timeout)
1268 clear_timeout(&to);
1269out:
Thomas Gleixnerc6f3a972008-04-30 00:55:03 -07001270 destroy_timer_on_stack(&to.timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271 return i ? i : ret;
1272}
1273
1274/* Take an ioctx and remove it from the list of ioctx's. Protects
1275 * against races with itself via ->dead.
1276 */
1277static void io_destroy(struct kioctx *ioctx)
1278{
1279 struct mm_struct *mm = current->mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280 int was_dead;
1281
1282 /* delete the entry from the list is someone else hasn't already */
Jens Axboeabf137d2008-12-09 08:11:22 +01001283 spin_lock(&mm->ioctx_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284 was_dead = ioctx->dead;
1285 ioctx->dead = 1;
Jens Axboeabf137d2008-12-09 08:11:22 +01001286 hlist_del_rcu(&ioctx->list);
1287 spin_unlock(&mm->ioctx_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288
1289 dprintk("aio_release(%p)\n", ioctx);
1290 if (likely(!was_dead))
1291 put_ioctx(ioctx); /* twice for the list */
1292
1293 aio_cancel_all(ioctx);
1294 wait_for_all_aios(ioctx);
Jeff Moyere92adcb2008-04-28 02:12:04 -07001295
1296 /*
1297 * Wake up any waiters. The setting of ctx->dead must be seen
1298 * by other CPUs at this point. Right now, we rely on the
1299 * locking done by the above calls to ensure this consistency.
1300 */
Roland Dreiere91f90b2011-03-22 16:35:10 -07001301 wake_up_all(&ioctx->wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001302 put_ioctx(ioctx); /* once for the lookup */
1303}
1304
1305/* sys_io_setup:
1306 * Create an aio_context capable of receiving at least nr_events.
1307 * ctxp must not point to an aio_context that already exists, and
1308 * must be initialized to 0 prior to the call. On successful
1309 * creation of the aio_context, *ctxp is filled in with the resulting
1310 * handle. May fail with -EINVAL if *ctxp is not initialized,
1311 * if the specified nr_events exceeds internal limits. May fail
1312 * with -EAGAIN if the specified nr_events exceeds the user's limit
1313 * of available events. May fail with -ENOMEM if insufficient kernel
1314 * resources are available. May fail with -EFAULT if an invalid
1315 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1316 * implemented.
1317 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001318SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001319{
1320 struct kioctx *ioctx = NULL;
1321 unsigned long ctx;
1322 long ret;
1323
1324 ret = get_user(ctx, ctxp);
1325 if (unlikely(ret))
1326 goto out;
1327
1328 ret = -EINVAL;
Zach Brownd55b5fd2005-11-07 00:59:31 -08001329 if (unlikely(ctx || nr_events == 0)) {
1330 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1331 ctx, nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332 goto out;
1333 }
1334
1335 ioctx = ioctx_alloc(nr_events);
1336 ret = PTR_ERR(ioctx);
1337 if (!IS_ERR(ioctx)) {
1338 ret = put_user(ioctx->user_id, ctxp);
1339 if (!ret)
1340 return 0;
1341
1342 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1343 io_destroy(ioctx);
1344 }
1345
1346out:
1347 return ret;
1348}
1349
1350/* sys_io_destroy:
1351 * Destroy the aio_context specified. May cancel any outstanding
1352 * AIOs and block on completion. Will fail with -ENOSYS if not
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001353 * implemented. May fail with -EINVAL if the context pointed to
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354 * is invalid.
1355 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001356SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357{
1358 struct kioctx *ioctx = lookup_ioctx(ctx);
1359 if (likely(NULL != ioctx)) {
1360 io_destroy(ioctx);
1361 return 0;
1362 }
1363 pr_debug("EINVAL: io_destroy: invalid context id\n");
1364 return -EINVAL;
1365}
1366
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001367static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1368{
1369 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1370
1371 BUG_ON(ret <= 0);
1372
1373 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1374 ssize_t this = min((ssize_t)iov->iov_len, ret);
1375 iov->iov_base += this;
1376 iov->iov_len -= this;
1377 iocb->ki_left -= this;
1378 ret -= this;
1379 if (iov->iov_len == 0) {
1380 iocb->ki_cur_seg++;
1381 iov++;
1382 }
1383 }
1384
1385 /* the caller should not have done more io than what fit in
1386 * the remaining iovecs */
1387 BUG_ON(ret > 0 && iocb->ki_left == 0);
1388}
1389
1390static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391{
1392 struct file *file = iocb->ki_filp;
1393 struct address_space *mapping = file->f_mapping;
1394 struct inode *inode = mapping->host;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001395 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1396 unsigned long, loff_t);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397 ssize_t ret = 0;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001398 unsigned short opcode;
1399
1400 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1401 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1402 rw_op = file->f_op->aio_read;
1403 opcode = IOCB_CMD_PREADV;
1404 } else {
1405 rw_op = file->f_op->aio_write;
1406 opcode = IOCB_CMD_PWRITEV;
1407 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001408
Rusty Russellc2ec6682008-02-08 04:20:15 -08001409 /* This matches the pread()/pwrite() logic */
1410 if (iocb->ki_pos < 0)
1411 return -EINVAL;
1412
Zach Brown897f15f2005-09-30 11:58:55 -07001413 do {
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001414 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1415 iocb->ki_nr_segs - iocb->ki_cur_seg,
1416 iocb->ki_pos);
1417 if (ret > 0)
1418 aio_advance_iovec(iocb, ret);
Badari Pulavarty027445c2006-09-30 23:28:46 -07001419
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001420 /* retry all partial writes. retry partial reads as long as its a
1421 * regular file. */
Zach Brown353fb072005-09-30 11:58:56 -07001422 } while (ret > 0 && iocb->ki_left > 0 &&
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001423 (opcode == IOCB_CMD_PWRITEV ||
1424 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001425
1426 /* This means we must have transferred all that we could */
1427 /* No need to retry anymore */
1428 if ((ret == 0) || (iocb->ki_left == 0))
1429 ret = iocb->ki_nbytes - iocb->ki_left;
1430
Rusty Russell7adfa2f2008-02-08 04:20:14 -08001431 /* If we managed to write some out we return that, rather than
1432 * the eventual error. */
1433 if (opcode == IOCB_CMD_PWRITEV
1434 && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
1435 && iocb->ki_nbytes - iocb->ki_left)
1436 ret = iocb->ki_nbytes - iocb->ki_left;
1437
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438 return ret;
1439}
1440
Linus Torvalds1da177e2005-04-16 15:20:36 -07001441static ssize_t aio_fdsync(struct kiocb *iocb)
1442{
1443 struct file *file = iocb->ki_filp;
1444 ssize_t ret = -EINVAL;
1445
1446 if (file->f_op->aio_fsync)
1447 ret = file->f_op->aio_fsync(iocb, 1);
1448 return ret;
1449}
1450
1451static ssize_t aio_fsync(struct kiocb *iocb)
1452{
1453 struct file *file = iocb->ki_filp;
1454 ssize_t ret = -EINVAL;
1455
1456 if (file->f_op->aio_fsync)
1457 ret = file->f_op->aio_fsync(iocb, 0);
1458 return ret;
1459}
1460
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001461static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001462{
1463 ssize_t ret;
1464
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001465#ifdef CONFIG_COMPAT
1466 if (compat)
1467 ret = compat_rw_copy_check_uvector(type,
1468 (struct compat_iovec __user *)kiocb->ki_buf,
1469 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
Christopher Yeohfcf63402011-10-31 17:06:39 -07001470 &kiocb->ki_iovec, 1);
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001471 else
1472#endif
1473 ret = rw_copy_check_uvector(type,
1474 (struct iovec __user *)kiocb->ki_buf,
1475 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
Christopher Yeohfcf63402011-10-31 17:06:39 -07001476 &kiocb->ki_iovec, 1);
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001477 if (ret < 0)
1478 goto out;
1479
1480 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1481 kiocb->ki_cur_seg = 0;
1482 /* ki_nbytes/left now reflect bytes instead of segs */
1483 kiocb->ki_nbytes = ret;
1484 kiocb->ki_left = ret;
1485
1486 ret = 0;
1487out:
1488 return ret;
1489}
1490
1491static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
1492{
1493 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1494 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1495 kiocb->ki_iovec->iov_len = kiocb->ki_left;
1496 kiocb->ki_nr_segs = 1;
1497 kiocb->ki_cur_seg = 0;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001498 return 0;
1499}
1500
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501/*
1502 * aio_setup_iocb:
1503 * Performs the initial checks and aio retry method
1504 * setup for the kiocb at the time of io submission.
1505 */
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001506static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001507{
1508 struct file *file = kiocb->ki_filp;
1509 ssize_t ret = 0;
1510
1511 switch (kiocb->ki_opcode) {
1512 case IOCB_CMD_PREAD:
1513 ret = -EBADF;
1514 if (unlikely(!(file->f_mode & FMODE_READ)))
1515 break;
1516 ret = -EFAULT;
1517 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1518 kiocb->ki_left)))
1519 break;
Kostik Belousov8766ce42005-10-23 12:57:13 -07001520 ret = security_file_permission(file, MAY_READ);
1521 if (unlikely(ret))
1522 break;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001523 ret = aio_setup_single_vector(kiocb);
1524 if (ret)
1525 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526 ret = -EINVAL;
1527 if (file->f_op->aio_read)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001528 kiocb->ki_retry = aio_rw_vect_retry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001529 break;
1530 case IOCB_CMD_PWRITE:
1531 ret = -EBADF;
1532 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1533 break;
1534 ret = -EFAULT;
1535 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1536 kiocb->ki_left)))
1537 break;
Kostik Belousov8766ce42005-10-23 12:57:13 -07001538 ret = security_file_permission(file, MAY_WRITE);
1539 if (unlikely(ret))
1540 break;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001541 ret = aio_setup_single_vector(kiocb);
1542 if (ret)
1543 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001544 ret = -EINVAL;
1545 if (file->f_op->aio_write)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001546 kiocb->ki_retry = aio_rw_vect_retry;
1547 break;
1548 case IOCB_CMD_PREADV:
1549 ret = -EBADF;
1550 if (unlikely(!(file->f_mode & FMODE_READ)))
1551 break;
1552 ret = security_file_permission(file, MAY_READ);
1553 if (unlikely(ret))
1554 break;
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001555 ret = aio_setup_vectored_rw(READ, kiocb, compat);
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001556 if (ret)
1557 break;
1558 ret = -EINVAL;
1559 if (file->f_op->aio_read)
1560 kiocb->ki_retry = aio_rw_vect_retry;
1561 break;
1562 case IOCB_CMD_PWRITEV:
1563 ret = -EBADF;
1564 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1565 break;
1566 ret = security_file_permission(file, MAY_WRITE);
1567 if (unlikely(ret))
1568 break;
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001569 ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001570 if (ret)
1571 break;
1572 ret = -EINVAL;
1573 if (file->f_op->aio_write)
1574 kiocb->ki_retry = aio_rw_vect_retry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001575 break;
1576 case IOCB_CMD_FDSYNC:
1577 ret = -EINVAL;
1578 if (file->f_op->aio_fsync)
1579 kiocb->ki_retry = aio_fdsync;
1580 break;
1581 case IOCB_CMD_FSYNC:
1582 ret = -EINVAL;
1583 if (file->f_op->aio_fsync)
1584 kiocb->ki_retry = aio_fsync;
1585 break;
1586 default:
1587 dprintk("EINVAL: io_submit: no operation provided\n");
1588 ret = -EINVAL;
1589 }
1590
1591 if (!kiocb->ki_retry)
1592 return ret;
1593
1594 return 0;
1595}
1596
Adrian Bunkd5470b52008-04-29 00:58:57 -07001597static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
Jeff Moyer080d6762011-11-02 13:40:10 -07001598 struct iocb *iocb, struct kiocb_batch *batch,
1599 bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001600{
1601 struct kiocb *req;
1602 struct file *file;
1603 ssize_t ret;
1604
1605 /* enforce forwards compatibility on users */
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001606 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001607 pr_debug("EINVAL: io_submit: reserve field set\n");
1608 return -EINVAL;
1609 }
1610
1611 /* prevent overflows */
1612 if (unlikely(
1613 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1614 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1615 ((ssize_t)iocb->aio_nbytes < 0)
1616 )) {
1617 pr_debug("EINVAL: io_submit: overflow check\n");
1618 return -EINVAL;
1619 }
1620
1621 file = fget(iocb->aio_fildes);
1622 if (unlikely(!file))
1623 return -EBADF;
1624
Jeff Moyer080d6762011-11-02 13:40:10 -07001625 req = aio_get_req(ctx, batch); /* returns with 2 references to req */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626 if (unlikely(!req)) {
1627 fput(file);
1628 return -EAGAIN;
1629 }
Yan Zheng87e28312007-10-08 12:16:20 -07001630 req->ki_filp = file;
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001631 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1632 /*
1633 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1634 * instance of the file* now. The file descriptor must be
1635 * an eventfd() fd, and will be signaled for each completed
1636 * event using the eventfd_signal() function.
1637 */
Davide Libenzi13389012009-06-30 11:41:11 -07001638 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
Hirofumi Nakagawa801678c2008-04-29 01:03:09 -07001639 if (IS_ERR(req->ki_eventfd)) {
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001640 ret = PTR_ERR(req->ki_eventfd);
Davide Libenzi87c3a862009-03-18 17:04:19 -07001641 req->ki_eventfd = NULL;
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001642 goto out_put_req;
1643 }
1644 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001645
Ken Chen212079c2005-05-01 08:59:15 -07001646 ret = put_user(req->ki_key, &user_iocb->aio_key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647 if (unlikely(ret)) {
1648 dprintk("EFAULT: aio_key\n");
1649 goto out_put_req;
1650 }
1651
1652 req->ki_obj.user = user_iocb;
1653 req->ki_user_data = iocb->aio_data;
1654 req->ki_pos = iocb->aio_offset;
1655
1656 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1657 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1658 req->ki_opcode = iocb->aio_lio_opcode;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001660 ret = aio_setup_iocb(req, compat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661
1662 if (ret)
1663 goto out_put_req;
1664
1665 spin_lock_irq(&ctx->ctx_lock);
Jan Kara7137c6bd2011-02-25 14:44:27 -08001666 /*
1667 * We could have raced with io_destroy() and are currently holding a
1668 * reference to ctx which should be destroyed. We cannot submit IO
1669 * since ctx gets freed as soon as io_submit() puts its reference. The
1670 * check here is reliable: io_destroy() sets ctx->dead before waiting
1671 * for outstanding IO and the barrier between these two is realized by
1672 * unlock of mm->ioctx_lock and lock of ctx->ctx_lock. Analogously we
1673 * increment ctx->reqs_active before checking for ctx->dead and the
1674 * barrier is realized by unlock and lock of ctx->ctx_lock. Thus if we
1675 * don't see ctx->dead set here, io_destroy() waits for our IO to
1676 * finish.
1677 */
1678 if (ctx->dead) {
1679 spin_unlock_irq(&ctx->ctx_lock);
1680 ret = -EINVAL;
1681 goto out_put_req;
1682 }
Benjamin LaHaiseac0b1bc2005-09-09 13:02:09 -07001683 aio_run_iocb(req);
Benjamin LaHaiseac0b1bc2005-09-09 13:02:09 -07001684 if (!list_empty(&ctx->run_list)) {
Ken Chen954d3e92005-05-01 08:59:16 -07001685 /* drain the run list */
1686 while (__aio_run_iocbs(ctx))
1687 ;
1688 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689 spin_unlock_irq(&ctx->ctx_lock);
Jeff Moyercfb1e332009-10-02 18:57:36 -04001690
Linus Torvalds1da177e2005-04-16 15:20:36 -07001691 aio_put_req(req); /* drop extra ref to req */
1692 return 0;
1693
1694out_put_req:
1695 aio_put_req(req); /* drop extra ref to req */
1696 aio_put_req(req); /* drop i/o ref to req */
1697 return ret;
1698}
1699
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001700long do_io_submit(aio_context_t ctx_id, long nr,
1701 struct iocb __user *__user *iocbpp, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702{
1703 struct kioctx *ctx;
1704 long ret = 0;
Jeff Moyer080d6762011-11-02 13:40:10 -07001705 int i = 0;
Shaohua Li9f5b9422010-07-01 07:55:01 +02001706 struct blk_plug plug;
Jeff Moyer080d6762011-11-02 13:40:10 -07001707 struct kiocb_batch batch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001708
1709 if (unlikely(nr < 0))
1710 return -EINVAL;
1711
Jeff Moyer75e1c702010-09-10 14:16:00 -07001712 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1713 nr = LONG_MAX/sizeof(*iocbpp);
1714
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1716 return -EFAULT;
1717
1718 ctx = lookup_ioctx(ctx_id);
1719 if (unlikely(!ctx)) {
1720 pr_debug("EINVAL: io_submit: invalid context id\n");
1721 return -EINVAL;
1722 }
1723
Jeff Moyer080d6762011-11-02 13:40:10 -07001724 kiocb_batch_init(&batch, nr);
1725
Shaohua Li9f5b9422010-07-01 07:55:01 +02001726 blk_start_plug(&plug);
1727
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728 /*
1729 * AKPM: should this return a partial result if some of the IOs were
1730 * successfully submitted?
1731 */
1732 for (i=0; i<nr; i++) {
1733 struct iocb __user *user_iocb;
1734 struct iocb tmp;
1735
1736 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1737 ret = -EFAULT;
1738 break;
1739 }
1740
1741 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1742 ret = -EFAULT;
1743 break;
1744 }
1745
Jeff Moyer080d6762011-11-02 13:40:10 -07001746 ret = io_submit_one(ctx, user_iocb, &tmp, &batch, compat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001747 if (ret)
1748 break;
1749 }
Shaohua Li9f5b9422010-07-01 07:55:01 +02001750 blk_finish_plug(&plug);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751
Gleb Natapov69e47472012-01-08 17:07:28 +02001752 kiocb_batch_free(ctx, &batch);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001753 put_ioctx(ctx);
1754 return i ? i : ret;
1755}
1756
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001757/* sys_io_submit:
1758 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1759 * the number of iocbs queued. May return -EINVAL if the aio_context
1760 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1761 * *iocbpp[0] is not properly initialized, if the operation specified
1762 * is invalid for the file descriptor in the iocb. May fail with
1763 * -EFAULT if any of the data structures point to invalid data. May
1764 * fail with -EBADF if the file descriptor specified in the first
1765 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1766 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1767 * fail with -ENOSYS if not implemented.
1768 */
1769SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1770 struct iocb __user * __user *, iocbpp)
1771{
1772 return do_io_submit(ctx_id, nr, iocbpp, 0);
1773}
1774
Linus Torvalds1da177e2005-04-16 15:20:36 -07001775/* lookup_kiocb
1776 * Finds a given iocb for cancellation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001778static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1779 u32 key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001780{
1781 struct list_head *pos;
Zach Brownd00689a2005-11-13 16:07:34 -08001782
1783 assert_spin_locked(&ctx->ctx_lock);
1784
Linus Torvalds1da177e2005-04-16 15:20:36 -07001785 /* TODO: use a hash or array, this sucks. */
1786 list_for_each(pos, &ctx->active_reqs) {
1787 struct kiocb *kiocb = list_kiocb(pos);
1788 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1789 return kiocb;
1790 }
1791 return NULL;
1792}
1793
1794/* sys_io_cancel:
1795 * Attempts to cancel an iocb previously passed to io_submit. If
1796 * the operation is successfully cancelled, the resulting event is
1797 * copied into the memory pointed to by result without being placed
1798 * into the completion queue and 0 is returned. May fail with
1799 * -EFAULT if any of the data structures pointed to are invalid.
1800 * May fail with -EINVAL if aio_context specified by ctx_id is
1801 * invalid. May fail with -EAGAIN if the iocb specified was not
1802 * cancelled. Will fail with -ENOSYS if not implemented.
1803 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001804SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1805 struct io_event __user *, result)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001806{
1807 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1808 struct kioctx *ctx;
1809 struct kiocb *kiocb;
1810 u32 key;
1811 int ret;
1812
1813 ret = get_user(key, &iocb->aio_key);
1814 if (unlikely(ret))
1815 return -EFAULT;
1816
1817 ctx = lookup_ioctx(ctx_id);
1818 if (unlikely(!ctx))
1819 return -EINVAL;
1820
1821 spin_lock_irq(&ctx->ctx_lock);
1822 ret = -EAGAIN;
1823 kiocb = lookup_kiocb(ctx, iocb, key);
1824 if (kiocb && kiocb->ki_cancel) {
1825 cancel = kiocb->ki_cancel;
1826 kiocb->ki_users ++;
1827 kiocbSetCancelled(kiocb);
1828 } else
1829 cancel = NULL;
1830 spin_unlock_irq(&ctx->ctx_lock);
1831
1832 if (NULL != cancel) {
1833 struct io_event tmp;
1834 pr_debug("calling cancel\n");
1835 memset(&tmp, 0, sizeof(tmp));
1836 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1837 tmp.data = kiocb->ki_user_data;
1838 ret = cancel(kiocb, &tmp);
1839 if (!ret) {
1840 /* Cancellation succeeded -- copy the result
1841 * into the user's buffer.
1842 */
1843 if (copy_to_user(result, &tmp, sizeof(tmp)))
1844 ret = -EFAULT;
1845 }
1846 } else
Wendy Cheng8f582022005-09-09 13:02:08 -07001847 ret = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001848
1849 put_ioctx(ctx);
1850
1851 return ret;
1852}
1853
1854/* io_getevents:
1855 * Attempts to read at least min_nr events and up to nr events from
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001856 * the completion queue for the aio_context specified by ctx_id. If
1857 * it succeeds, the number of read events is returned. May fail with
1858 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1859 * out of range, if timeout is out of range. May fail with -EFAULT
1860 * if any of the memory specified is invalid. May return 0 or
1861 * < min_nr if the timeout specified by timeout has elapsed
1862 * before sufficient events are available, where timeout == NULL
1863 * specifies an infinite timeout. Note that the timeout pointed to by
1864 * timeout is relative and will be updated if not NULL and the
1865 * operation blocks. Will fail with -ENOSYS if not implemented.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001867SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1868 long, min_nr,
1869 long, nr,
1870 struct io_event __user *, events,
1871 struct timespec __user *, timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001872{
1873 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1874 long ret = -EINVAL;
1875
1876 if (likely(ioctx)) {
Namhyung Kim2e410252011-01-12 17:01:08 -08001877 if (likely(min_nr <= nr && min_nr >= 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001878 ret = read_events(ioctx, min_nr, nr, events, timeout);
1879 put_ioctx(ioctx);
1880 }
1881
Roland McGrath598af052008-04-10 15:38:45 -07001882 asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883 return ret;
1884}