blob: 26869cde3953588ff2c0b0aa524966d5c02519d2 [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>
37#include <linux/mempool.h>
38#include <linux/hash.h>
Jeff Moyer9d85cba2010-05-26 14:44:26 -070039#include <linux/compat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070040
41#include <asm/kmap_types.h>
42#include <asm/uaccess.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070043
44#if DEBUG > 1
45#define dprintk printk
46#else
47#define dprintk(x...) do { ; } while (0)
48#endif
49
Linus Torvalds1da177e2005-04-16 15:20:36 -070050/*------ sysctl variables----*/
Zach Brownd55b5fd2005-11-07 00:59:31 -080051static DEFINE_SPINLOCK(aio_nr_lock);
52unsigned long aio_nr; /* current system wide number of aio requests */
53unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
Linus Torvalds1da177e2005-04-16 15:20:36 -070054/*----end sysctl variables---*/
55
Christoph Lametere18b8902006-12-06 20:33:20 -080056static struct kmem_cache *kiocb_cachep;
57static struct kmem_cache *kioctx_cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -070058
59static struct workqueue_struct *aio_wq;
60
61/* Used for rare fput completion. */
David Howells65f27f32006-11-22 14:55:48 +000062static void aio_fput_routine(struct work_struct *);
63static DECLARE_WORK(fput_work, aio_fput_routine);
Linus Torvalds1da177e2005-04-16 15:20:36 -070064
65static DEFINE_SPINLOCK(fput_lock);
Adrian Bunk25ee7e32005-04-25 08:18:14 -070066static LIST_HEAD(fput_head);
Linus Torvalds1da177e2005-04-16 15:20:36 -070067
Jeff Moyercfb1e332009-10-02 18:57:36 -040068#define AIO_BATCH_HASH_BITS 3 /* allocated on-stack, so don't go crazy */
69#define AIO_BATCH_HASH_SIZE (1 << AIO_BATCH_HASH_BITS)
70struct aio_batch_entry {
71 struct hlist_node list;
72 struct address_space *mapping;
73};
74mempool_t *abe_pool;
75
David Howells65f27f32006-11-22 14:55:48 +000076static void aio_kick_handler(struct work_struct *);
Sébastien Duguc016e222005-06-28 20:44:59 -070077static void aio_queue_work(struct kioctx *);
Linus Torvalds1da177e2005-04-16 15:20:36 -070078
79/* aio_setup
80 * Creates the slab caches used by the aio routines, panic on
81 * failure as this is done early during the boot sequence.
82 */
83static int __init aio_setup(void)
84{
Christoph Lameter0a31bd52007-05-06 14:49:57 -070085 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
86 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -070087
88 aio_wq = create_workqueue("aio");
Jeff Moyercfb1e332009-10-02 18:57:36 -040089 abe_pool = mempool_create_kmalloc_pool(1, sizeof(struct aio_batch_entry));
Namhyung Kim27eaa1c2010-12-14 00:06:25 +090090 BUG_ON(!aio_wq || !abe_pool);
Linus Torvalds1da177e2005-04-16 15:20:36 -070091
92 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
93
94 return 0;
95}
H Hartley Sweeten385773e2009-09-22 16:43:53 -070096__initcall(aio_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -070097
98static void aio_free_ring(struct kioctx *ctx)
99{
100 struct aio_ring_info *info = &ctx->ring_info;
101 long i;
102
103 for (i=0; i<info->nr_pages; i++)
104 put_page(info->ring_pages[i]);
105
106 if (info->mmap_size) {
107 down_write(&ctx->mm->mmap_sem);
108 do_munmap(ctx->mm, info->mmap_base, info->mmap_size);
109 up_write(&ctx->mm->mmap_sem);
110 }
111
112 if (info->ring_pages && info->ring_pages != info->internal_pages)
113 kfree(info->ring_pages);
114 info->ring_pages = NULL;
115 info->nr = 0;
116}
117
118static int aio_setup_ring(struct kioctx *ctx)
119{
120 struct aio_ring *ring;
121 struct aio_ring_info *info = &ctx->ring_info;
122 unsigned nr_events = ctx->max_reqs;
123 unsigned long size;
124 int nr_pages;
125
126 /* Compensate for the ring buffer's head/tail overlap entry */
127 nr_events += 2; /* 1 is required, 2 for good luck */
128
129 size = sizeof(struct aio_ring);
130 size += sizeof(struct io_event) * nr_events;
131 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
132
133 if (nr_pages < 0)
134 return -EINVAL;
135
136 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
137
138 info->nr = 0;
139 info->ring_pages = info->internal_pages;
140 if (nr_pages > AIO_RING_PAGES) {
Oliver Neukum11b0b5a2006-03-25 03:08:13 -0800141 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142 if (!info->ring_pages)
143 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144 }
145
146 info->mmap_size = nr_pages * PAGE_SIZE;
147 dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
148 down_write(&ctx->mm->mmap_sem);
149 info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
Robert P. J. Daye10a4432007-02-10 01:42:56 -0800150 PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151 0);
152 if (IS_ERR((void *)info->mmap_base)) {
153 up_write(&ctx->mm->mmap_sem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154 info->mmap_size = 0;
155 aio_free_ring(ctx);
156 return -EAGAIN;
157 }
158
159 dprintk("mmap address: 0x%08lx\n", info->mmap_base);
160 info->nr_pages = get_user_pages(current, ctx->mm,
161 info->mmap_base, nr_pages,
162 1, 0, info->ring_pages, NULL);
163 up_write(&ctx->mm->mmap_sem);
164
165 if (unlikely(info->nr_pages != nr_pages)) {
166 aio_free_ring(ctx);
167 return -EAGAIN;
168 }
169
170 ctx->user_id = info->mmap_base;
171
172 info->nr = nr_events; /* trusted copy */
173
174 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
175 ring->nr = nr_events; /* user copy */
176 ring->id = ctx->user_id;
177 ring->head = ring->tail = 0;
178 ring->magic = AIO_RING_MAGIC;
179 ring->compat_features = AIO_RING_COMPAT_FEATURES;
180 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
181 ring->header_length = sizeof(struct aio_ring);
182 kunmap_atomic(ring, KM_USER0);
183
184 return 0;
185}
186
187
188/* aio_ring_event: returns a pointer to the event at the given index from
189 * kmap_atomic(, km). Release the pointer with put_aio_ring_event();
190 */
191#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
192#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
193#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
194
195#define aio_ring_event(info, nr, km) ({ \
196 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
197 struct io_event *__event; \
198 __event = kmap_atomic( \
199 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
200 __event += pos % AIO_EVENTS_PER_PAGE; \
201 __event; \
202})
203
204#define put_aio_ring_event(event, km) do { \
205 struct io_event *__event = (event); \
206 (void)__event; \
207 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
208} while(0)
209
Jens Axboeabf137d2008-12-09 08:11:22 +0100210static void ctx_rcu_free(struct rcu_head *head)
Adrian Bunkd5470b52008-04-29 00:58:57 -0700211{
Jens Axboeabf137d2008-12-09 08:11:22 +0100212 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
Adrian Bunkd5470b52008-04-29 00:58:57 -0700213 unsigned nr_events = ctx->max_reqs;
214
Adrian Bunkd5470b52008-04-29 00:58:57 -0700215 kmem_cache_free(kioctx_cachep, ctx);
216
217 if (nr_events) {
218 spin_lock(&aio_nr_lock);
219 BUG_ON(aio_nr - nr_events > aio_nr);
220 aio_nr -= nr_events;
221 spin_unlock(&aio_nr_lock);
222 }
223}
224
Jens Axboeabf137d2008-12-09 08:11:22 +0100225/* __put_ioctx
226 * Called when the last user of an aio context has gone away,
227 * and the struct needs to be freed.
228 */
229static void __put_ioctx(struct kioctx *ctx)
230{
231 BUG_ON(ctx->reqs_active);
232
233 cancel_delayed_work(&ctx->wq);
234 cancel_work_sync(&ctx->wq.work);
235 aio_free_ring(ctx);
236 mmdrop(ctx->mm);
237 ctx->mm = NULL;
238 pr_debug("__put_ioctx: freeing %p\n", ctx);
239 call_rcu(&ctx->rcu_head, ctx_rcu_free);
240}
241
Nick Piggin3bd9a5d2011-02-25 14:44:26 -0800242static inline void get_ioctx(struct kioctx *kioctx)
243{
244 BUG_ON(atomic_read(&kioctx->users) <= 0);
245 atomic_inc(&kioctx->users);
246}
247
248static inline int try_get_ioctx(struct kioctx *kioctx)
249{
250 return atomic_inc_not_zero(&kioctx->users);
251}
252
253static inline void put_ioctx(struct kioctx *kioctx)
254{
255 BUG_ON(atomic_read(&kioctx->users) <= 0);
256 if (unlikely(atomic_dec_and_test(&kioctx->users)))
257 __put_ioctx(kioctx);
258}
Adrian Bunkd5470b52008-04-29 00:58:57 -0700259
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260/* ioctx_alloc
261 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
262 */
263static struct kioctx *ioctx_alloc(unsigned nr_events)
264{
265 struct mm_struct *mm;
266 struct kioctx *ctx;
Jens Axboeabf137d2008-12-09 08:11:22 +0100267 int did_sync = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268
269 /* Prevent overflows */
270 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
271 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
272 pr_debug("ENOMEM: nr_events too high\n");
273 return ERR_PTR(-EINVAL);
274 }
275
Zach Brownd55b5fd2005-11-07 00:59:31 -0800276 if ((unsigned long)nr_events > aio_max_nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277 return ERR_PTR(-EAGAIN);
278
Robert P. J. Dayc3762222007-02-10 01:45:03 -0800279 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280 if (!ctx)
281 return ERR_PTR(-ENOMEM);
282
Linus Torvalds1da177e2005-04-16 15:20:36 -0700283 ctx->max_reqs = nr_events;
284 mm = ctx->mm = current->mm;
285 atomic_inc(&mm->mm_count);
286
287 atomic_set(&ctx->users, 1);
288 spin_lock_init(&ctx->ctx_lock);
289 spin_lock_init(&ctx->ring_info.ring_lock);
290 init_waitqueue_head(&ctx->wait);
291
292 INIT_LIST_HEAD(&ctx->active_reqs);
293 INIT_LIST_HEAD(&ctx->run_list);
David Howells65f27f32006-11-22 14:55:48 +0000294 INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700295
296 if (aio_setup_ring(ctx) < 0)
297 goto out_freectx;
298
299 /* limit the number of system wide aios */
Jens Axboeabf137d2008-12-09 08:11:22 +0100300 do {
301 spin_lock_bh(&aio_nr_lock);
302 if (aio_nr + nr_events > aio_max_nr ||
303 aio_nr + nr_events < aio_nr)
304 ctx->max_reqs = 0;
305 else
306 aio_nr += ctx->max_reqs;
307 spin_unlock_bh(&aio_nr_lock);
308 if (ctx->max_reqs || did_sync)
309 break;
310
311 /* wait for rcu callbacks to have completed before giving up */
312 synchronize_rcu();
313 did_sync = 1;
314 ctx->max_reqs = nr_events;
315 } while (1);
316
Zach Brownd55b5fd2005-11-07 00:59:31 -0800317 if (ctx->max_reqs == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318 goto out_cleanup;
319
Jeff Moyer39fa0032008-04-29 01:03:48 -0700320 /* now link into global list. */
Jens Axboeabf137d2008-12-09 08:11:22 +0100321 spin_lock(&mm->ioctx_lock);
322 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
323 spin_unlock(&mm->ioctx_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700324
325 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
326 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
327 return ctx;
328
329out_cleanup:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330 __put_ioctx(ctx);
331 return ERR_PTR(-EAGAIN);
332
333out_freectx:
334 mmdrop(mm);
335 kmem_cache_free(kioctx_cachep, ctx);
336 ctx = ERR_PTR(-ENOMEM);
337
338 dprintk("aio: error allocating ioctx %p\n", ctx);
339 return ctx;
340}
341
342/* aio_cancel_all
343 * Cancels all outstanding aio requests on an aio context. Used
344 * when the processes owning a context have all exited to encourage
345 * the rapid destruction of the kioctx.
346 */
347static void aio_cancel_all(struct kioctx *ctx)
348{
349 int (*cancel)(struct kiocb *, struct io_event *);
350 struct io_event res;
351 spin_lock_irq(&ctx->ctx_lock);
352 ctx->dead = 1;
353 while (!list_empty(&ctx->active_reqs)) {
354 struct list_head *pos = ctx->active_reqs.next;
355 struct kiocb *iocb = list_kiocb(pos);
356 list_del_init(&iocb->ki_list);
357 cancel = iocb->ki_cancel;
358 kiocbSetCancelled(iocb);
359 if (cancel) {
360 iocb->ki_users++;
361 spin_unlock_irq(&ctx->ctx_lock);
362 cancel(iocb, &res);
363 spin_lock_irq(&ctx->ctx_lock);
364 }
365 }
366 spin_unlock_irq(&ctx->ctx_lock);
367}
368
Adrian Bunk25ee7e32005-04-25 08:18:14 -0700369static void wait_for_all_aios(struct kioctx *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370{
371 struct task_struct *tsk = current;
372 DECLARE_WAITQUEUE(wait, tsk);
373
Ken Chendee11c22007-02-03 01:13:45 -0800374 spin_lock_irq(&ctx->ctx_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375 if (!ctx->reqs_active)
Ken Chendee11c22007-02-03 01:13:45 -0800376 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377
378 add_wait_queue(&ctx->wait, &wait);
379 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
380 while (ctx->reqs_active) {
Ken Chendee11c22007-02-03 01:13:45 -0800381 spin_unlock_irq(&ctx->ctx_lock);
Jeff Moyer41d10da2007-10-16 23:27:20 -0700382 io_schedule();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
Ken Chendee11c22007-02-03 01:13:45 -0800384 spin_lock_irq(&ctx->ctx_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385 }
386 __set_task_state(tsk, TASK_RUNNING);
387 remove_wait_queue(&ctx->wait, &wait);
Ken Chendee11c22007-02-03 01:13:45 -0800388
389out:
390 spin_unlock_irq(&ctx->ctx_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391}
392
393/* wait_on_sync_kiocb:
394 * Waits on the given sync kiocb to complete.
395 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800396ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397{
398 while (iocb->ki_users) {
399 set_current_state(TASK_UNINTERRUPTIBLE);
400 if (!iocb->ki_users)
401 break;
Jeff Moyer41d10da2007-10-16 23:27:20 -0700402 io_schedule();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403 }
404 __set_current_state(TASK_RUNNING);
405 return iocb->ki_user_data;
406}
H Hartley Sweeten385773e2009-09-22 16:43:53 -0700407EXPORT_SYMBOL(wait_on_sync_kiocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408
409/* exit_aio: called when the last user of mm goes away. At this point,
410 * there is no way for any new requests to be submited or any of the
411 * io_* syscalls to be called on the context. However, there may be
412 * outstanding requests which hold references to the context; as they
413 * go away, they will call put_ioctx and release any pinned memory
414 * associated with the request (held via struct page * references).
415 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800416void exit_aio(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417{
Jens Axboeabf137d2008-12-09 08:11:22 +0100418 struct kioctx *ctx;
419
420 while (!hlist_empty(&mm->ioctx_list)) {
421 ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
422 hlist_del_rcu(&ctx->list);
423
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424 aio_cancel_all(ctx);
425
426 wait_for_all_aios(ctx);
427 /*
Andrew Mortona9df62c2007-05-09 02:33:54 -0700428 * Ensure we don't leave the ctx on the aio_wq
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429 */
Oleg Nesterov28e53bd2007-05-09 02:34:22 -0700430 cancel_work_sync(&ctx->wq.work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431
432 if (1 != atomic_read(&ctx->users))
433 printk(KERN_DEBUG
434 "exit_aio:ioctx still alive: %d %d %d\n",
435 atomic_read(&ctx->users), ctx->dead,
436 ctx->reqs_active);
437 put_ioctx(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438 }
439}
440
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441/* aio_get_req
442 * Allocate a slot for an aio request. Increments the users count
443 * of the kioctx so that the kioctx stays around until all requests are
444 * complete. Returns NULL if no requests are free.
445 *
446 * Returns with kiocb->users set to 2. The io submit code path holds
447 * an extra reference while submitting the i/o.
448 * This prevents races between the aio code path referencing the
449 * req (after submitting it) and aio_complete() freeing the req.
450 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800451static struct kiocb *__aio_get_req(struct kioctx *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452{
453 struct kiocb *req = NULL;
454 struct aio_ring *ring;
455 int okay = 0;
456
457 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
458 if (unlikely(!req))
459 return NULL;
460
Zach Brown4faa5282005-10-17 16:43:33 -0700461 req->ki_flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 req->ki_users = 2;
463 req->ki_key = 0;
464 req->ki_ctx = ctx;
465 req->ki_cancel = NULL;
466 req->ki_retry = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467 req->ki_dtor = NULL;
468 req->private = NULL;
Badari Pulavartyeed4e512006-09-30 23:28:49 -0700469 req->ki_iovec = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470 INIT_LIST_HEAD(&req->ki_run_list);
Davide Libenzi87c3a862009-03-18 17:04:19 -0700471 req->ki_eventfd = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472
473 /* Check if the completion queue has enough free space to
474 * accept an event from this io.
475 */
476 spin_lock_irq(&ctx->ctx_lock);
477 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
478 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
479 list_add(&req->ki_list, &ctx->active_reqs);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480 ctx->reqs_active++;
481 okay = 1;
482 }
483 kunmap_atomic(ring, KM_USER0);
484 spin_unlock_irq(&ctx->ctx_lock);
485
486 if (!okay) {
487 kmem_cache_free(kiocb_cachep, req);
488 req = NULL;
489 }
490
491 return req;
492}
493
494static inline struct kiocb *aio_get_req(struct kioctx *ctx)
495{
496 struct kiocb *req;
497 /* Handle a potential starvation case -- should be exceedingly rare as
498 * requests will be stuck on fput_head only if the aio_fput_routine is
499 * delayed and the requests were the last user of the struct file.
500 */
501 req = __aio_get_req(ctx);
502 if (unlikely(NULL == req)) {
503 aio_fput_routine(NULL);
504 req = __aio_get_req(ctx);
505 }
506 return req;
507}
508
509static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
510{
Zach Brownd00689a2005-11-13 16:07:34 -0800511 assert_spin_locked(&ctx->ctx_lock);
512
Davide Libenzi13389012009-06-30 11:41:11 -0700513 if (req->ki_eventfd != NULL)
514 eventfd_ctx_put(req->ki_eventfd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515 if (req->ki_dtor)
516 req->ki_dtor(req);
Badari Pulavartyeed4e512006-09-30 23:28:49 -0700517 if (req->ki_iovec != &req->ki_inline_vec)
518 kfree(req->ki_iovec);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 kmem_cache_free(kiocb_cachep, req);
520 ctx->reqs_active--;
521
522 if (unlikely(!ctx->reqs_active && ctx->dead))
523 wake_up(&ctx->wait);
524}
525
David Howells65f27f32006-11-22 14:55:48 +0000526static void aio_fput_routine(struct work_struct *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700527{
528 spin_lock_irq(&fput_lock);
529 while (likely(!list_empty(&fput_head))) {
530 struct kiocb *req = list_kiocb(fput_head.next);
531 struct kioctx *ctx = req->ki_ctx;
532
533 list_del(&req->ki_list);
534 spin_unlock_irq(&fput_lock);
535
Davide Libenzi87c3a862009-03-18 17:04:19 -0700536 /* Complete the fput(s) */
537 if (req->ki_filp != NULL)
Al Virod7065da2010-05-26 15:13:55 -0400538 fput(req->ki_filp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539
540 /* Link the iocb into the context's free list */
541 spin_lock_irq(&ctx->ctx_lock);
542 really_put_req(ctx, req);
543 spin_unlock_irq(&ctx->ctx_lock);
544
545 put_ioctx(ctx);
546 spin_lock_irq(&fput_lock);
547 }
548 spin_unlock_irq(&fput_lock);
549}
550
551/* __aio_put_req
552 * Returns true if this put was the last user of the request.
553 */
554static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
555{
Al Viro516e0cc2008-07-26 00:39:17 -0400556 dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
557 req, atomic_long_read(&req->ki_filp->f_count));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700558
Zach Brownd00689a2005-11-13 16:07:34 -0800559 assert_spin_locked(&ctx->ctx_lock);
560
Davide Libenzi87c3a862009-03-18 17:04:19 -0700561 req->ki_users--;
Eric Sesterhenn93e06b42006-11-30 05:29:23 +0100562 BUG_ON(req->ki_users < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563 if (likely(req->ki_users))
564 return 0;
565 list_del(&req->ki_list); /* remove from active_reqs */
566 req->ki_cancel = NULL;
567 req->ki_retry = NULL;
568
Davide Libenzi87c3a862009-03-18 17:04:19 -0700569 /*
570 * Try to optimize the aio and eventfd file* puts, by avoiding to
Al Virod7065da2010-05-26 15:13:55 -0400571 * schedule work in case it is not final fput() time. In normal cases,
Davide Libenzi87c3a862009-03-18 17:04:19 -0700572 * we would not be holding the last reference to the file*, so
573 * this function will be executed w/out any aio kthread wakeup.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574 */
Al Virod7065da2010-05-26 15:13:55 -0400575 if (unlikely(!fput_atomic(req->ki_filp))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576 get_ioctx(ctx);
577 spin_lock(&fput_lock);
578 list_add(&req->ki_list, &fput_head);
579 spin_unlock(&fput_lock);
580 queue_work(aio_wq, &fput_work);
Davide Libenzi13389012009-06-30 11:41:11 -0700581 } else {
582 req->ki_filp = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583 really_put_req(ctx, req);
Davide Libenzi13389012009-06-30 11:41:11 -0700584 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 return 1;
586}
587
588/* aio_put_req
589 * Returns true if this put was the last user of the kiocb,
590 * false if the request is still in use.
591 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800592int aio_put_req(struct kiocb *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593{
594 struct kioctx *ctx = req->ki_ctx;
595 int ret;
596 spin_lock_irq(&ctx->ctx_lock);
597 ret = __aio_put_req(ctx, req);
598 spin_unlock_irq(&ctx->ctx_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599 return ret;
600}
H Hartley Sweeten385773e2009-09-22 16:43:53 -0700601EXPORT_SYMBOL(aio_put_req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602
Adrian Bunkd5470b52008-04-29 00:58:57 -0700603static struct kioctx *lookup_ioctx(unsigned long ctx_id)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604{
Jens Axboeabf137d2008-12-09 08:11:22 +0100605 struct mm_struct *mm = current->mm;
Jeff Moyer65c24492009-03-18 17:04:21 -0700606 struct kioctx *ctx, *ret = NULL;
Jens Axboeabf137d2008-12-09 08:11:22 +0100607 struct hlist_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608
Jens Axboeabf137d2008-12-09 08:11:22 +0100609 rcu_read_lock();
610
611 hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) {
Nick Piggin3bd9a5d2011-02-25 14:44:26 -0800612 /*
613 * RCU protects us against accessing freed memory but
614 * we have to be careful not to get a reference when the
615 * reference count already dropped to 0 (ctx->dead test
616 * is unreliable because of races).
617 */
618 if (ctx->user_id == ctx_id && !ctx->dead && try_get_ioctx(ctx)){
Jeff Moyer65c24492009-03-18 17:04:21 -0700619 ret = ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620 break;
621 }
Jens Axboeabf137d2008-12-09 08:11:22 +0100622 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700623
Jens Axboeabf137d2008-12-09 08:11:22 +0100624 rcu_read_unlock();
Jeff Moyer65c24492009-03-18 17:04:21 -0700625 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626}
627
628/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629 * Queue up a kiocb to be retried. Assumes that the kiocb
630 * has already been marked as kicked, and places it on
631 * the retry run list for the corresponding ioctx, if it
632 * isn't already queued. Returns 1 if it actually queued
633 * the kiocb (to tell the caller to activate the work
634 * queue to process it), or 0, if it found that it was
635 * already queued.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636 */
637static inline int __queue_kicked_iocb(struct kiocb *iocb)
638{
639 struct kioctx *ctx = iocb->ki_ctx;
640
Zach Brownd00689a2005-11-13 16:07:34 -0800641 assert_spin_locked(&ctx->ctx_lock);
642
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643 if (list_empty(&iocb->ki_run_list)) {
644 list_add_tail(&iocb->ki_run_list,
645 &ctx->run_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646 return 1;
647 }
648 return 0;
649}
650
651/* aio_run_iocb
652 * This is the core aio execution routine. It is
653 * invoked both for initial i/o submission and
654 * subsequent retries via the aio_kick_handler.
655 * Expects to be invoked with iocb->ki_ctx->lock
Andreas Mohrd6e05ed2006-06-26 18:35:02 +0200656 * already held. The lock is released and reacquired
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657 * as needed during processing.
658 *
659 * Calls the iocb retry method (already setup for the
660 * iocb on initial submission) for operation specific
661 * handling, but takes care of most of common retry
662 * execution details for a given iocb. The retry method
663 * needs to be non-blocking as far as possible, to avoid
664 * holding up other iocbs waiting to be serviced by the
665 * retry kernel thread.
666 *
667 * The trickier parts in this code have to do with
668 * ensuring that only one retry instance is in progress
669 * for a given iocb at any time. Providing that guarantee
670 * simplifies the coding of individual aio operations as
671 * it avoids various potential races.
672 */
673static ssize_t aio_run_iocb(struct kiocb *iocb)
674{
675 struct kioctx *ctx = iocb->ki_ctx;
676 ssize_t (*retry)(struct kiocb *);
677 ssize_t ret;
678
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679 if (!(retry = iocb->ki_retry)) {
680 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
681 return 0;
682 }
683
684 /*
685 * We don't want the next retry iteration for this
686 * operation to start until this one has returned and
687 * updated the iocb state. However, wait_queue functions
688 * can trigger a kick_iocb from interrupt context in the
689 * meantime, indicating that data is available for the next
690 * iteration. We want to remember that and enable the
691 * next retry iteration _after_ we are through with
692 * this one.
693 *
694 * So, in order to be able to register a "kick", but
695 * prevent it from being queued now, we clear the kick
696 * flag, but make the kick code *think* that the iocb is
697 * still on the run list until we are actually done.
698 * When we are done with this iteration, we check if
699 * the iocb was kicked in the meantime and if so, queue
700 * it up afresh.
701 */
702
703 kiocbClearKicked(iocb);
704
705 /*
706 * This is so that aio_complete knows it doesn't need to
707 * pull the iocb off the run list (We can't just call
708 * INIT_LIST_HEAD because we don't want a kick_iocb to
709 * queue this on the run list yet)
710 */
711 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
712 spin_unlock_irq(&ctx->ctx_lock);
713
714 /* Quit retrying if the i/o has been cancelled */
715 if (kiocbIsCancelled(iocb)) {
716 ret = -EINTR;
717 aio_complete(iocb, ret, 0);
718 /* must not access the iocb after this */
719 goto out;
720 }
721
722 /*
723 * Now we are all set to call the retry method in async
Alexey Dobriyan6212e3a2007-10-18 03:04:56 -0700724 * context.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700726 ret = retry(iocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727
Jan Karaa0c42ba2010-09-22 13:05:03 -0700728 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
729 /*
730 * There's no easy way to restart the syscall since other AIO's
731 * may be already running. Just fail this IO with EINTR.
732 */
733 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
734 ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
735 ret = -EINTR;
Zach Brown897f15f2005-09-30 11:58:55 -0700736 aio_complete(iocb, ret, 0);
Jan Karaa0c42ba2010-09-22 13:05:03 -0700737 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738out:
739 spin_lock_irq(&ctx->ctx_lock);
740
741 if (-EIOCBRETRY == ret) {
742 /*
743 * OK, now that we are done with this iteration
744 * and know that there is more left to go,
745 * this is where we let go so that a subsequent
746 * "kick" can start the next iteration
747 */
748
749 /* will make __queue_kicked_iocb succeed from here on */
750 INIT_LIST_HEAD(&iocb->ki_run_list);
751 /* we must queue the next iteration ourselves, if it
752 * has already been kicked */
753 if (kiocbIsKicked(iocb)) {
754 __queue_kicked_iocb(iocb);
Sébastien Duguc016e222005-06-28 20:44:59 -0700755
756 /*
757 * __queue_kicked_iocb will always return 1 here, because
758 * iocb->ki_run_list is empty at this point so it should
759 * be safe to unconditionally queue the context into the
760 * work queue.
761 */
762 aio_queue_work(ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763 }
764 }
765 return ret;
766}
767
768/*
769 * __aio_run_iocbs:
770 * Process all pending retries queued on the ioctx
771 * run list.
772 * Assumes it is operating within the aio issuer's mm
Zach Brownd00689a2005-11-13 16:07:34 -0800773 * context.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774 */
775static int __aio_run_iocbs(struct kioctx *ctx)
776{
777 struct kiocb *iocb;
Oleg Nesterov626ab0e2006-06-23 02:05:55 -0700778 struct list_head run_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779
Zach Brownd00689a2005-11-13 16:07:34 -0800780 assert_spin_locked(&ctx->ctx_lock);
781
Oleg Nesterov626ab0e2006-06-23 02:05:55 -0700782 list_replace_init(&ctx->run_list, &run_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700783 while (!list_empty(&run_list)) {
784 iocb = list_entry(run_list.next, struct kiocb,
785 ki_run_list);
786 list_del(&iocb->ki_run_list);
787 /*
788 * Hold an extra reference while retrying i/o.
789 */
790 iocb->ki_users++; /* grab extra reference */
791 aio_run_iocb(iocb);
Ken Chendee11c22007-02-03 01:13:45 -0800792 __aio_put_req(ctx, iocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794 if (!list_empty(&ctx->run_list))
795 return 1;
796 return 0;
797}
798
799static void aio_queue_work(struct kioctx * ctx)
800{
801 unsigned long timeout;
802 /*
803 * if someone is waiting, get the work started right
804 * away, otherwise, use a longer delay
805 */
806 smp_mb();
807 if (waitqueue_active(&ctx->wait))
808 timeout = 1;
809 else
810 timeout = HZ/10;
811 queue_delayed_work(aio_wq, &ctx->wq, timeout);
812}
813
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814/*
Jeff Moyerd3486f82011-01-12 17:01:08 -0800815 * aio_run_all_iocbs:
816 * Process all pending retries queued on the ioctx
817 * run list, and keep running them until the list
818 * stays empty.
819 * Assumes it is operating within the aio issuer's mm context.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820 */
821static inline void aio_run_all_iocbs(struct kioctx *ctx)
822{
823 spin_lock_irq(&ctx->ctx_lock);
824 while (__aio_run_iocbs(ctx))
825 ;
826 spin_unlock_irq(&ctx->ctx_lock);
827}
828
829/*
830 * aio_kick_handler:
831 * Work queue handler triggered to process pending
832 * retries on an ioctx. Takes on the aio issuer's
833 * mm context before running the iocbs, so that
834 * copy_xxx_user operates on the issuer's address
835 * space.
836 * Run on aiod's context.
837 */
David Howells65f27f32006-11-22 14:55:48 +0000838static void aio_kick_handler(struct work_struct *work)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839{
David Howells65f27f32006-11-22 14:55:48 +0000840 struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700841 mm_segment_t oldfs = get_fs();
Zach Brown1ebb1102006-12-29 16:47:02 -0800842 struct mm_struct *mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843 int requeue;
844
845 set_fs(USER_DS);
846 use_mm(ctx->mm);
847 spin_lock_irq(&ctx->ctx_lock);
848 requeue =__aio_run_iocbs(ctx);
Zach Brown1ebb1102006-12-29 16:47:02 -0800849 mm = ctx->mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850 spin_unlock_irq(&ctx->ctx_lock);
Zach Brown1ebb1102006-12-29 16:47:02 -0800851 unuse_mm(mm);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852 set_fs(oldfs);
853 /*
854 * we're in a worker thread already, don't use queue_delayed_work,
855 */
856 if (requeue)
David Howells52bad642006-11-22 14:54:01 +0000857 queue_delayed_work(aio_wq, &ctx->wq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700858}
859
860
861/*
862 * Called by kick_iocb to queue the kiocb for retry
863 * and if required activate the aio work queue to process
864 * it
865 */
Zach Brown998765e2005-09-30 11:58:54 -0700866static void try_queue_kicked_iocb(struct kiocb *iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700867{
868 struct kioctx *ctx = iocb->ki_ctx;
869 unsigned long flags;
870 int run = 0;
871
Linus Torvalds1da177e2005-04-16 15:20:36 -0700872 spin_lock_irqsave(&ctx->ctx_lock, flags);
Zach Brown998765e2005-09-30 11:58:54 -0700873 /* set this inside the lock so that we can't race with aio_run_iocb()
874 * testing it and putting the iocb on the run list under the lock */
875 if (!kiocbTryKick(iocb))
876 run = __queue_kicked_iocb(iocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877 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
Shaohua Lifac046a2009-12-15 16:47:47 -0800885 * to trigger a retry of the iocb.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886 * The retry is usually executed by aio workqueue
887 * threads (See aio_kick_handler).
888 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800889void kick_iocb(struct kiocb *iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890{
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
Zach Brown998765e2005-09-30 11:58:54 -0700899 try_queue_kicked_iocb(iocb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900}
901EXPORT_SYMBOL(kick_iocb);
902
903/* aio_complete
904 * Called when the io request on the given iocb is complete.
905 * Returns true if this is the last user of the request. The
906 * only other user of the request can be the cancellation code.
907 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -0800908int aio_complete(struct kiocb *iocb, long res, long res2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909{
910 struct kioctx *ctx = iocb->ki_ctx;
911 struct aio_ring_info *info;
912 struct aio_ring *ring;
913 struct io_event *event;
914 unsigned long flags;
915 unsigned long tail;
916 int ret;
917
Zach Brown20dcae32005-11-13 16:07:33 -0800918 /*
919 * Special case handling for sync iocbs:
920 * - events go directly into the iocb for fast handling
921 * - the sync task with the iocb in its stack holds the single iocb
922 * ref, no other paths have a way to get another ref
923 * - the sync task helpfully left a reference to itself in the iocb
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924 */
925 if (is_sync_kiocb(iocb)) {
Zach Brown20dcae32005-11-13 16:07:33 -0800926 BUG_ON(iocb->ki_users != 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927 iocb->ki_user_data = res;
Zach Brown20dcae32005-11-13 16:07:33 -0800928 iocb->ki_users = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929 wake_up_process(iocb->ki_obj.tsk);
Zach Brown20dcae32005-11-13 16:07:33 -0800930 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 }
932
933 info = &ctx->ring_info;
934
935 /* add a completion event to the ring buffer.
936 * must be done holding ctx->ctx_lock to prevent
937 * other code from messing with the tail
938 * pointer since we might be called from irq
939 * context.
940 */
941 spin_lock_irqsave(&ctx->ctx_lock, flags);
942
943 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
944 list_del_init(&iocb->ki_run_list);
945
946 /*
947 * cancelled requests don't get events, userland was given one
948 * when the event got cancelled.
949 */
950 if (kiocbIsCancelled(iocb))
951 goto put_rq;
952
953 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
954
955 tail = info->tail;
956 event = aio_ring_event(info, tail, KM_IRQ0);
Ken Chen4bf69b22005-05-01 08:59:15 -0700957 if (++tail >= info->nr)
958 tail = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959
960 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
961 event->data = iocb->ki_user_data;
962 event->res = res;
963 event->res2 = res2;
964
965 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
966 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
967 res, res2);
968
969 /* after flagging the request as done, we
970 * must never even look at it again
971 */
972 smp_wmb(); /* make event visible before updating tail */
973
974 info->tail = tail;
975 ring->tail = tail;
976
977 put_aio_ring_event(event, KM_IRQ0);
978 kunmap_atomic(ring, KM_IRQ1);
979
980 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
Davide Libenzi8d1c98b2008-04-10 21:29:19 -0700981
982 /*
983 * Check if the user asked us to deliver the result through an
984 * eventfd. The eventfd_signal() function is safe to be called
985 * from IRQ context.
986 */
Davide Libenzi87c3a862009-03-18 17:04:19 -0700987 if (iocb->ki_eventfd != NULL)
Davide Libenzi8d1c98b2008-04-10 21:29:19 -0700988 eventfd_signal(iocb->ki_eventfd, 1);
989
Linus Torvalds1da177e2005-04-16 15:20:36 -0700990put_rq:
991 /* everything turned out well, dispose of the aiocb. */
992 ret = __aio_put_req(ctx, iocb);
993
Quentin Barnes6cb2a212008-03-19 17:00:39 -0700994 /*
995 * We have to order our ring_info tail store above and test
996 * of the wait list below outside the wait lock. This is
997 * like in wake_up_bit() where clearing a bit has to be
998 * ordered with the unlocked test.
999 */
1000 smp_mb();
1001
Linus Torvalds1da177e2005-04-16 15:20:36 -07001002 if (waitqueue_active(&ctx->wait))
1003 wake_up(&ctx->wait);
1004
Ken Chendee11c22007-02-03 01:13:45 -08001005 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001006 return ret;
1007}
H Hartley Sweeten385773e2009-09-22 16:43:53 -07001008EXPORT_SYMBOL(aio_complete);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001009
1010/* aio_read_evt
1011 * Pull an event off of the ioctx's event ring. Returns the number of
1012 * events fetched (0 or 1 ;-)
1013 * FIXME: make this use cmpxchg.
1014 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1015 */
1016static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1017{
1018 struct aio_ring_info *info = &ioctx->ring_info;
1019 struct aio_ring *ring;
1020 unsigned long head;
1021 int ret = 0;
1022
1023 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1024 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1025 (unsigned long)ring->head, (unsigned long)ring->tail,
1026 (unsigned long)ring->nr);
1027
1028 if (ring->head == ring->tail)
1029 goto out;
1030
1031 spin_lock(&info->ring_lock);
1032
1033 head = ring->head % info->nr;
1034 if (head != ring->tail) {
1035 struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1036 *ent = *evp;
1037 head = (head + 1) % info->nr;
1038 smp_mb(); /* finish reading the event before updatng the head */
1039 ring->head = head;
1040 ret = 1;
1041 put_aio_ring_event(evp, KM_USER1);
1042 }
1043 spin_unlock(&info->ring_lock);
1044
1045out:
1046 kunmap_atomic(ring, KM_USER0);
1047 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1048 (unsigned long)ring->head, (unsigned long)ring->tail);
1049 return ret;
1050}
1051
1052struct aio_timeout {
1053 struct timer_list timer;
1054 int timed_out;
1055 struct task_struct *p;
1056};
1057
1058static void timeout_func(unsigned long data)
1059{
1060 struct aio_timeout *to = (struct aio_timeout *)data;
1061
1062 to->timed_out = 1;
1063 wake_up_process(to->p);
1064}
1065
1066static inline void init_timeout(struct aio_timeout *to)
1067{
Thomas Gleixnerc6f3a972008-04-30 00:55:03 -07001068 setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001069 to->timed_out = 0;
1070 to->p = current;
1071}
1072
1073static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1074 const struct timespec *ts)
1075{
1076 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1077 if (time_after(to->timer.expires, jiffies))
1078 add_timer(&to->timer);
1079 else
1080 to->timed_out = 1;
1081}
1082
1083static inline void clear_timeout(struct aio_timeout *to)
1084{
1085 del_singleshot_timer_sync(&to->timer);
1086}
1087
1088static int read_events(struct kioctx *ctx,
1089 long min_nr, long nr,
1090 struct io_event __user *event,
1091 struct timespec __user *timeout)
1092{
1093 long start_jiffies = jiffies;
1094 struct task_struct *tsk = current;
1095 DECLARE_WAITQUEUE(wait, tsk);
1096 int ret;
1097 int i = 0;
1098 struct io_event ent;
1099 struct aio_timeout to;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001100 int retry = 0;
1101
1102 /* needed to zero any padding within an entry (there shouldn't be
1103 * any, but C is fun!
1104 */
1105 memset(&ent, 0, sizeof(ent));
1106retry:
1107 ret = 0;
1108 while (likely(i < nr)) {
1109 ret = aio_read_evt(ctx, &ent);
1110 if (unlikely(ret <= 0))
1111 break;
1112
1113 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1114 ent.data, ent.obj, ent.res, ent.res2);
1115
1116 /* Could we split the check in two? */
1117 ret = -EFAULT;
1118 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1119 dprintk("aio: lost an event due to EFAULT.\n");
1120 break;
1121 }
1122 ret = 0;
1123
1124 /* Good, event copied to userland, update counts. */
1125 event ++;
1126 i ++;
1127 }
1128
1129 if (min_nr <= i)
1130 return i;
1131 if (ret)
1132 return ret;
1133
1134 /* End fast path */
1135
1136 /* racey check, but it gets redone */
1137 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1138 retry = 1;
1139 aio_run_all_iocbs(ctx);
1140 goto retry;
1141 }
1142
1143 init_timeout(&to);
1144 if (timeout) {
1145 struct timespec ts;
1146 ret = -EFAULT;
1147 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1148 goto out;
1149
1150 set_timeout(start_jiffies, &to, &ts);
1151 }
1152
1153 while (likely(i < nr)) {
1154 add_wait_queue_exclusive(&ctx->wait, &wait);
1155 do {
1156 set_task_state(tsk, TASK_INTERRUPTIBLE);
1157 ret = aio_read_evt(ctx, &ent);
1158 if (ret)
1159 break;
1160 if (min_nr <= i)
1161 break;
Jeff Moyere92adcb2008-04-28 02:12:04 -07001162 if (unlikely(ctx->dead)) {
1163 ret = -EINVAL;
1164 break;
1165 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001166 if (to.timed_out) /* Only check after read evt */
1167 break;
Jeff Moyere00ba3d2007-12-04 23:45:02 -08001168 /* Try to only show up in io wait if there are ops
1169 * in flight */
1170 if (ctx->reqs_active)
1171 io_schedule();
1172 else
1173 schedule();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174 if (signal_pending(tsk)) {
1175 ret = -EINTR;
1176 break;
1177 }
1178 /*ret = aio_read_evt(ctx, &ent);*/
1179 } while (1) ;
1180
1181 set_task_state(tsk, TASK_RUNNING);
1182 remove_wait_queue(&ctx->wait, &wait);
1183
1184 if (unlikely(ret <= 0))
1185 break;
1186
1187 ret = -EFAULT;
1188 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1189 dprintk("aio: lost an event due to EFAULT.\n");
1190 break;
1191 }
1192
1193 /* Good, event copied to userland, update counts. */
1194 event ++;
1195 i ++;
1196 }
1197
1198 if (timeout)
1199 clear_timeout(&to);
1200out:
Thomas Gleixnerc6f3a972008-04-30 00:55:03 -07001201 destroy_timer_on_stack(&to.timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001202 return i ? i : ret;
1203}
1204
1205/* Take an ioctx and remove it from the list of ioctx's. Protects
1206 * against races with itself via ->dead.
1207 */
1208static void io_destroy(struct kioctx *ioctx)
1209{
1210 struct mm_struct *mm = current->mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211 int was_dead;
1212
1213 /* delete the entry from the list is someone else hasn't already */
Jens Axboeabf137d2008-12-09 08:11:22 +01001214 spin_lock(&mm->ioctx_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001215 was_dead = ioctx->dead;
1216 ioctx->dead = 1;
Jens Axboeabf137d2008-12-09 08:11:22 +01001217 hlist_del_rcu(&ioctx->list);
1218 spin_unlock(&mm->ioctx_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219
1220 dprintk("aio_release(%p)\n", ioctx);
1221 if (likely(!was_dead))
1222 put_ioctx(ioctx); /* twice for the list */
1223
1224 aio_cancel_all(ioctx);
1225 wait_for_all_aios(ioctx);
Jeff Moyere92adcb2008-04-28 02:12:04 -07001226
1227 /*
1228 * Wake up any waiters. The setting of ctx->dead must be seen
1229 * by other CPUs at this point. Right now, we rely on the
1230 * locking done by the above calls to ensure this consistency.
1231 */
1232 wake_up(&ioctx->wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233 put_ioctx(ioctx); /* once for the lookup */
1234}
1235
1236/* sys_io_setup:
1237 * Create an aio_context capable of receiving at least nr_events.
1238 * ctxp must not point to an aio_context that already exists, and
1239 * must be initialized to 0 prior to the call. On successful
1240 * creation of the aio_context, *ctxp is filled in with the resulting
1241 * handle. May fail with -EINVAL if *ctxp is not initialized,
1242 * if the specified nr_events exceeds internal limits. May fail
1243 * with -EAGAIN if the specified nr_events exceeds the user's limit
1244 * of available events. May fail with -ENOMEM if insufficient kernel
1245 * resources are available. May fail with -EFAULT if an invalid
1246 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1247 * implemented.
1248 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001249SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001250{
1251 struct kioctx *ioctx = NULL;
1252 unsigned long ctx;
1253 long ret;
1254
1255 ret = get_user(ctx, ctxp);
1256 if (unlikely(ret))
1257 goto out;
1258
1259 ret = -EINVAL;
Zach Brownd55b5fd2005-11-07 00:59:31 -08001260 if (unlikely(ctx || nr_events == 0)) {
1261 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1262 ctx, nr_events);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263 goto out;
1264 }
1265
1266 ioctx = ioctx_alloc(nr_events);
1267 ret = PTR_ERR(ioctx);
1268 if (!IS_ERR(ioctx)) {
1269 ret = put_user(ioctx->user_id, ctxp);
1270 if (!ret)
1271 return 0;
1272
1273 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1274 io_destroy(ioctx);
1275 }
1276
1277out:
1278 return ret;
1279}
1280
1281/* sys_io_destroy:
1282 * Destroy the aio_context specified. May cancel any outstanding
1283 * AIOs and block on completion. Will fail with -ENOSYS if not
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001284 * implemented. May fail with -EINVAL if the context pointed to
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 * is invalid.
1286 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001287SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288{
1289 struct kioctx *ioctx = lookup_ioctx(ctx);
1290 if (likely(NULL != ioctx)) {
1291 io_destroy(ioctx);
1292 return 0;
1293 }
1294 pr_debug("EINVAL: io_destroy: invalid context id\n");
1295 return -EINVAL;
1296}
1297
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001298static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1299{
1300 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1301
1302 BUG_ON(ret <= 0);
1303
1304 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1305 ssize_t this = min((ssize_t)iov->iov_len, ret);
1306 iov->iov_base += this;
1307 iov->iov_len -= this;
1308 iocb->ki_left -= this;
1309 ret -= this;
1310 if (iov->iov_len == 0) {
1311 iocb->ki_cur_seg++;
1312 iov++;
1313 }
1314 }
1315
1316 /* the caller should not have done more io than what fit in
1317 * the remaining iovecs */
1318 BUG_ON(ret > 0 && iocb->ki_left == 0);
1319}
1320
1321static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322{
1323 struct file *file = iocb->ki_filp;
1324 struct address_space *mapping = file->f_mapping;
1325 struct inode *inode = mapping->host;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001326 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1327 unsigned long, loff_t);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328 ssize_t ret = 0;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001329 unsigned short opcode;
1330
1331 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1332 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1333 rw_op = file->f_op->aio_read;
1334 opcode = IOCB_CMD_PREADV;
1335 } else {
1336 rw_op = file->f_op->aio_write;
1337 opcode = IOCB_CMD_PWRITEV;
1338 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001339
Rusty Russellc2ec6682008-02-08 04:20:15 -08001340 /* This matches the pread()/pwrite() logic */
1341 if (iocb->ki_pos < 0)
1342 return -EINVAL;
1343
Zach Brown897f15f2005-09-30 11:58:55 -07001344 do {
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001345 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1346 iocb->ki_nr_segs - iocb->ki_cur_seg,
1347 iocb->ki_pos);
1348 if (ret > 0)
1349 aio_advance_iovec(iocb, ret);
Badari Pulavarty027445c2006-09-30 23:28:46 -07001350
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001351 /* retry all partial writes. retry partial reads as long as its a
1352 * regular file. */
Zach Brown353fb072005-09-30 11:58:56 -07001353 } while (ret > 0 && iocb->ki_left > 0 &&
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001354 (opcode == IOCB_CMD_PWRITEV ||
1355 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356
1357 /* This means we must have transferred all that we could */
1358 /* No need to retry anymore */
1359 if ((ret == 0) || (iocb->ki_left == 0))
1360 ret = iocb->ki_nbytes - iocb->ki_left;
1361
Rusty Russell7adfa2f2008-02-08 04:20:14 -08001362 /* If we managed to write some out we return that, rather than
1363 * the eventual error. */
1364 if (opcode == IOCB_CMD_PWRITEV
1365 && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
1366 && iocb->ki_nbytes - iocb->ki_left)
1367 ret = iocb->ki_nbytes - iocb->ki_left;
1368
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369 return ret;
1370}
1371
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372static ssize_t aio_fdsync(struct kiocb *iocb)
1373{
1374 struct file *file = iocb->ki_filp;
1375 ssize_t ret = -EINVAL;
1376
1377 if (file->f_op->aio_fsync)
1378 ret = file->f_op->aio_fsync(iocb, 1);
1379 return ret;
1380}
1381
1382static ssize_t aio_fsync(struct kiocb *iocb)
1383{
1384 struct file *file = iocb->ki_filp;
1385 ssize_t ret = -EINVAL;
1386
1387 if (file->f_op->aio_fsync)
1388 ret = file->f_op->aio_fsync(iocb, 0);
1389 return ret;
1390}
1391
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001392static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001393{
1394 ssize_t ret;
1395
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001396#ifdef CONFIG_COMPAT
1397 if (compat)
1398 ret = compat_rw_copy_check_uvector(type,
1399 (struct compat_iovec __user *)kiocb->ki_buf,
1400 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1401 &kiocb->ki_iovec);
1402 else
1403#endif
1404 ret = rw_copy_check_uvector(type,
1405 (struct iovec __user *)kiocb->ki_buf,
1406 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1407 &kiocb->ki_iovec);
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001408 if (ret < 0)
1409 goto out;
1410
1411 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1412 kiocb->ki_cur_seg = 0;
1413 /* ki_nbytes/left now reflect bytes instead of segs */
1414 kiocb->ki_nbytes = ret;
1415 kiocb->ki_left = ret;
1416
1417 ret = 0;
1418out:
1419 return ret;
1420}
1421
1422static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
1423{
1424 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1425 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1426 kiocb->ki_iovec->iov_len = kiocb->ki_left;
1427 kiocb->ki_nr_segs = 1;
1428 kiocb->ki_cur_seg = 0;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001429 return 0;
1430}
1431
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432/*
1433 * aio_setup_iocb:
1434 * Performs the initial checks and aio retry method
1435 * setup for the kiocb at the time of io submission.
1436 */
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001437static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438{
1439 struct file *file = kiocb->ki_filp;
1440 ssize_t ret = 0;
1441
1442 switch (kiocb->ki_opcode) {
1443 case IOCB_CMD_PREAD:
1444 ret = -EBADF;
1445 if (unlikely(!(file->f_mode & FMODE_READ)))
1446 break;
1447 ret = -EFAULT;
1448 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1449 kiocb->ki_left)))
1450 break;
Kostik Belousov8766ce42005-10-23 12:57:13 -07001451 ret = security_file_permission(file, MAY_READ);
1452 if (unlikely(ret))
1453 break;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001454 ret = aio_setup_single_vector(kiocb);
1455 if (ret)
1456 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457 ret = -EINVAL;
1458 if (file->f_op->aio_read)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001459 kiocb->ki_retry = aio_rw_vect_retry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001460 break;
1461 case IOCB_CMD_PWRITE:
1462 ret = -EBADF;
1463 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1464 break;
1465 ret = -EFAULT;
1466 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1467 kiocb->ki_left)))
1468 break;
Kostik Belousov8766ce42005-10-23 12:57:13 -07001469 ret = security_file_permission(file, MAY_WRITE);
1470 if (unlikely(ret))
1471 break;
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001472 ret = aio_setup_single_vector(kiocb);
1473 if (ret)
1474 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475 ret = -EINVAL;
1476 if (file->f_op->aio_write)
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001477 kiocb->ki_retry = aio_rw_vect_retry;
1478 break;
1479 case IOCB_CMD_PREADV:
1480 ret = -EBADF;
1481 if (unlikely(!(file->f_mode & FMODE_READ)))
1482 break;
1483 ret = security_file_permission(file, MAY_READ);
1484 if (unlikely(ret))
1485 break;
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001486 ret = aio_setup_vectored_rw(READ, kiocb, compat);
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001487 if (ret)
1488 break;
1489 ret = -EINVAL;
1490 if (file->f_op->aio_read)
1491 kiocb->ki_retry = aio_rw_vect_retry;
1492 break;
1493 case IOCB_CMD_PWRITEV:
1494 ret = -EBADF;
1495 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1496 break;
1497 ret = security_file_permission(file, MAY_WRITE);
1498 if (unlikely(ret))
1499 break;
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001500 ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
Badari Pulavartyeed4e512006-09-30 23:28:49 -07001501 if (ret)
1502 break;
1503 ret = -EINVAL;
1504 if (file->f_op->aio_write)
1505 kiocb->ki_retry = aio_rw_vect_retry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506 break;
1507 case IOCB_CMD_FDSYNC:
1508 ret = -EINVAL;
1509 if (file->f_op->aio_fsync)
1510 kiocb->ki_retry = aio_fdsync;
1511 break;
1512 case IOCB_CMD_FSYNC:
1513 ret = -EINVAL;
1514 if (file->f_op->aio_fsync)
1515 kiocb->ki_retry = aio_fsync;
1516 break;
1517 default:
1518 dprintk("EINVAL: io_submit: no operation provided\n");
1519 ret = -EINVAL;
1520 }
1521
1522 if (!kiocb->ki_retry)
1523 return ret;
1524
1525 return 0;
1526}
1527
Jeff Moyercfb1e332009-10-02 18:57:36 -04001528static void aio_batch_add(struct address_space *mapping,
1529 struct hlist_head *batch_hash)
1530{
1531 struct aio_batch_entry *abe;
1532 struct hlist_node *pos;
1533 unsigned bucket;
1534
1535 bucket = hash_ptr(mapping, AIO_BATCH_HASH_BITS);
1536 hlist_for_each_entry(abe, pos, &batch_hash[bucket], list) {
1537 if (abe->mapping == mapping)
1538 return;
1539 }
1540
1541 abe = mempool_alloc(abe_pool, GFP_KERNEL);
Chris Mason306fb092010-08-23 10:47:55 -04001542
1543 /*
1544 * we should be using igrab here, but
1545 * we don't want to hammer on the global
1546 * inode spinlock just to take an extra
1547 * reference on a file that we must already
1548 * have a reference to.
1549 *
1550 * When we're called, we always have a reference
1551 * on the file, so we must always have a reference
Al Viro7de9c6e2010-10-23 11:11:40 -04001552 * on the inode, so ihold() is safe here.
Chris Mason306fb092010-08-23 10:47:55 -04001553 */
Al Viro7de9c6e2010-10-23 11:11:40 -04001554 ihold(mapping->host);
Jeff Moyercfb1e332009-10-02 18:57:36 -04001555 abe->mapping = mapping;
1556 hlist_add_head(&abe->list, &batch_hash[bucket]);
1557 return;
1558}
1559
1560static void aio_batch_free(struct hlist_head *batch_hash)
1561{
1562 struct aio_batch_entry *abe;
1563 struct hlist_node *pos, *n;
1564 int i;
1565
1566 for (i = 0; i < AIO_BATCH_HASH_SIZE; i++) {
1567 hlist_for_each_entry_safe(abe, pos, n, &batch_hash[i], list) {
1568 blk_run_address_space(abe->mapping);
1569 iput(abe->mapping->host);
1570 hlist_del(&abe->list);
1571 mempool_free(abe, abe_pool);
1572 }
1573 }
1574}
1575
Adrian Bunkd5470b52008-04-29 00:58:57 -07001576static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001577 struct iocb *iocb, struct hlist_head *batch_hash,
1578 bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001579{
1580 struct kiocb *req;
1581 struct file *file;
1582 ssize_t ret;
1583
1584 /* enforce forwards compatibility on users */
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001585 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001586 pr_debug("EINVAL: io_submit: reserve field set\n");
1587 return -EINVAL;
1588 }
1589
1590 /* prevent overflows */
1591 if (unlikely(
1592 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1593 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1594 ((ssize_t)iocb->aio_nbytes < 0)
1595 )) {
1596 pr_debug("EINVAL: io_submit: overflow check\n");
1597 return -EINVAL;
1598 }
1599
1600 file = fget(iocb->aio_fildes);
1601 if (unlikely(!file))
1602 return -EBADF;
1603
1604 req = aio_get_req(ctx); /* returns with 2 references to req */
1605 if (unlikely(!req)) {
1606 fput(file);
1607 return -EAGAIN;
1608 }
Yan Zheng87e28312007-10-08 12:16:20 -07001609 req->ki_filp = file;
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001610 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1611 /*
1612 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1613 * instance of the file* now. The file descriptor must be
1614 * an eventfd() fd, and will be signaled for each completed
1615 * event using the eventfd_signal() function.
1616 */
Davide Libenzi13389012009-06-30 11:41:11 -07001617 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
Hirofumi Nakagawa801678c2008-04-29 01:03:09 -07001618 if (IS_ERR(req->ki_eventfd)) {
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001619 ret = PTR_ERR(req->ki_eventfd);
Davide Libenzi87c3a862009-03-18 17:04:19 -07001620 req->ki_eventfd = NULL;
Davide Libenzi9c3060b2007-05-10 22:23:21 -07001621 goto out_put_req;
1622 }
1623 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624
Ken Chen212079c2005-05-01 08:59:15 -07001625 ret = put_user(req->ki_key, &user_iocb->aio_key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626 if (unlikely(ret)) {
1627 dprintk("EFAULT: aio_key\n");
1628 goto out_put_req;
1629 }
1630
1631 req->ki_obj.user = user_iocb;
1632 req->ki_user_data = iocb->aio_data;
1633 req->ki_pos = iocb->aio_offset;
1634
1635 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1636 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1637 req->ki_opcode = iocb->aio_lio_opcode;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001638
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001639 ret = aio_setup_iocb(req, compat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640
1641 if (ret)
1642 goto out_put_req;
1643
1644 spin_lock_irq(&ctx->ctx_lock);
Jan Kara7137c6bd2011-02-25 14:44:27 -08001645 /*
1646 * We could have raced with io_destroy() and are currently holding a
1647 * reference to ctx which should be destroyed. We cannot submit IO
1648 * since ctx gets freed as soon as io_submit() puts its reference. The
1649 * check here is reliable: io_destroy() sets ctx->dead before waiting
1650 * for outstanding IO and the barrier between these two is realized by
1651 * unlock of mm->ioctx_lock and lock of ctx->ctx_lock. Analogously we
1652 * increment ctx->reqs_active before checking for ctx->dead and the
1653 * barrier is realized by unlock and lock of ctx->ctx_lock. Thus if we
1654 * don't see ctx->dead set here, io_destroy() waits for our IO to
1655 * finish.
1656 */
1657 if (ctx->dead) {
1658 spin_unlock_irq(&ctx->ctx_lock);
1659 ret = -EINVAL;
1660 goto out_put_req;
1661 }
Benjamin LaHaiseac0b1bc2005-09-09 13:02:09 -07001662 aio_run_iocb(req);
Benjamin LaHaiseac0b1bc2005-09-09 13:02:09 -07001663 if (!list_empty(&ctx->run_list)) {
Ken Chen954d3e92005-05-01 08:59:16 -07001664 /* drain the run list */
1665 while (__aio_run_iocbs(ctx))
1666 ;
1667 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668 spin_unlock_irq(&ctx->ctx_lock);
Jeff Moyercfb1e332009-10-02 18:57:36 -04001669 if (req->ki_opcode == IOCB_CMD_PREAD ||
1670 req->ki_opcode == IOCB_CMD_PREADV ||
1671 req->ki_opcode == IOCB_CMD_PWRITE ||
1672 req->ki_opcode == IOCB_CMD_PWRITEV)
1673 aio_batch_add(file->f_mapping, batch_hash);
1674
Linus Torvalds1da177e2005-04-16 15:20:36 -07001675 aio_put_req(req); /* drop extra ref to req */
1676 return 0;
1677
1678out_put_req:
1679 aio_put_req(req); /* drop extra ref to req */
1680 aio_put_req(req); /* drop i/o ref to req */
1681 return ret;
1682}
1683
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001684long do_io_submit(aio_context_t ctx_id, long nr,
1685 struct iocb __user *__user *iocbpp, bool compat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686{
1687 struct kioctx *ctx;
1688 long ret = 0;
1689 int i;
Jeff Moyercfb1e332009-10-02 18:57:36 -04001690 struct hlist_head batch_hash[AIO_BATCH_HASH_SIZE] = { { 0, }, };
Linus Torvalds1da177e2005-04-16 15:20:36 -07001691
1692 if (unlikely(nr < 0))
1693 return -EINVAL;
1694
Jeff Moyer75e1c702010-09-10 14:16:00 -07001695 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1696 nr = LONG_MAX/sizeof(*iocbpp);
1697
Linus Torvalds1da177e2005-04-16 15:20:36 -07001698 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1699 return -EFAULT;
1700
1701 ctx = lookup_ioctx(ctx_id);
1702 if (unlikely(!ctx)) {
1703 pr_debug("EINVAL: io_submit: invalid context id\n");
1704 return -EINVAL;
1705 }
1706
1707 /*
1708 * AKPM: should this return a partial result if some of the IOs were
1709 * successfully submitted?
1710 */
1711 for (i=0; i<nr; i++) {
1712 struct iocb __user *user_iocb;
1713 struct iocb tmp;
1714
1715 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1716 ret = -EFAULT;
1717 break;
1718 }
1719
1720 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1721 ret = -EFAULT;
1722 break;
1723 }
1724
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001725 ret = io_submit_one(ctx, user_iocb, &tmp, batch_hash, compat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726 if (ret)
1727 break;
1728 }
Jeff Moyercfb1e332009-10-02 18:57:36 -04001729 aio_batch_free(batch_hash);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001730
1731 put_ioctx(ctx);
1732 return i ? i : ret;
1733}
1734
Jeff Moyer9d85cba2010-05-26 14:44:26 -07001735/* sys_io_submit:
1736 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1737 * the number of iocbs queued. May return -EINVAL if the aio_context
1738 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1739 * *iocbpp[0] is not properly initialized, if the operation specified
1740 * is invalid for the file descriptor in the iocb. May fail with
1741 * -EFAULT if any of the data structures point to invalid data. May
1742 * fail with -EBADF if the file descriptor specified in the first
1743 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1744 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1745 * fail with -ENOSYS if not implemented.
1746 */
1747SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1748 struct iocb __user * __user *, iocbpp)
1749{
1750 return do_io_submit(ctx_id, nr, iocbpp, 0);
1751}
1752
Linus Torvalds1da177e2005-04-16 15:20:36 -07001753/* lookup_kiocb
1754 * Finds a given iocb for cancellation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001755 */
Adrian Bunk25ee7e32005-04-25 08:18:14 -07001756static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1757 u32 key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001758{
1759 struct list_head *pos;
Zach Brownd00689a2005-11-13 16:07:34 -08001760
1761 assert_spin_locked(&ctx->ctx_lock);
1762
Linus Torvalds1da177e2005-04-16 15:20:36 -07001763 /* TODO: use a hash or array, this sucks. */
1764 list_for_each(pos, &ctx->active_reqs) {
1765 struct kiocb *kiocb = list_kiocb(pos);
1766 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1767 return kiocb;
1768 }
1769 return NULL;
1770}
1771
1772/* sys_io_cancel:
1773 * Attempts to cancel an iocb previously passed to io_submit. If
1774 * the operation is successfully cancelled, the resulting event is
1775 * copied into the memory pointed to by result without being placed
1776 * into the completion queue and 0 is returned. May fail with
1777 * -EFAULT if any of the data structures pointed to are invalid.
1778 * May fail with -EINVAL if aio_context specified by ctx_id is
1779 * invalid. May fail with -EAGAIN if the iocb specified was not
1780 * cancelled. Will fail with -ENOSYS if not implemented.
1781 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001782SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1783 struct io_event __user *, result)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784{
1785 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1786 struct kioctx *ctx;
1787 struct kiocb *kiocb;
1788 u32 key;
1789 int ret;
1790
1791 ret = get_user(key, &iocb->aio_key);
1792 if (unlikely(ret))
1793 return -EFAULT;
1794
1795 ctx = lookup_ioctx(ctx_id);
1796 if (unlikely(!ctx))
1797 return -EINVAL;
1798
1799 spin_lock_irq(&ctx->ctx_lock);
1800 ret = -EAGAIN;
1801 kiocb = lookup_kiocb(ctx, iocb, key);
1802 if (kiocb && kiocb->ki_cancel) {
1803 cancel = kiocb->ki_cancel;
1804 kiocb->ki_users ++;
1805 kiocbSetCancelled(kiocb);
1806 } else
1807 cancel = NULL;
1808 spin_unlock_irq(&ctx->ctx_lock);
1809
1810 if (NULL != cancel) {
1811 struct io_event tmp;
1812 pr_debug("calling cancel\n");
1813 memset(&tmp, 0, sizeof(tmp));
1814 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1815 tmp.data = kiocb->ki_user_data;
1816 ret = cancel(kiocb, &tmp);
1817 if (!ret) {
1818 /* Cancellation succeeded -- copy the result
1819 * into the user's buffer.
1820 */
1821 if (copy_to_user(result, &tmp, sizeof(tmp)))
1822 ret = -EFAULT;
1823 }
1824 } else
Wendy Cheng8f582022005-09-09 13:02:08 -07001825 ret = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001826
1827 put_ioctx(ctx);
1828
1829 return ret;
1830}
1831
1832/* io_getevents:
1833 * Attempts to read at least min_nr events and up to nr events from
Satoru Takeuchi642b5122010-08-05 11:23:11 -07001834 * the completion queue for the aio_context specified by ctx_id. If
1835 * it succeeds, the number of read events is returned. May fail with
1836 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1837 * out of range, if timeout is out of range. May fail with -EFAULT
1838 * if any of the memory specified is invalid. May return 0 or
1839 * < min_nr if the timeout specified by timeout has elapsed
1840 * before sufficient events are available, where timeout == NULL
1841 * specifies an infinite timeout. Note that the timeout pointed to by
1842 * timeout is relative and will be updated if not NULL and the
1843 * operation blocks. Will fail with -ENOSYS if not implemented.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001844 */
Heiko Carstens002c8972009-01-14 14:14:18 +01001845SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1846 long, min_nr,
1847 long, nr,
1848 struct io_event __user *, events,
1849 struct timespec __user *, timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850{
1851 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1852 long ret = -EINVAL;
1853
1854 if (likely(ioctx)) {
Namhyung Kim2e410252011-01-12 17:01:08 -08001855 if (likely(min_nr <= nr && min_nr >= 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001856 ret = read_events(ioctx, min_nr, nr, events, timeout);
1857 put_ioctx(ioctx);
1858 }
1859
Roland McGrath598af052008-04-10 15:38:45 -07001860 asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001861 return ret;
1862}