]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blame - fs/aio.c
Merge branch 'for-linus-4.1' of git://git.kernel.org/pub/scm/linux/kernel/git/mason...
[mirror_ubuntu-bionic-kernel.git] / fs / aio.c
CommitLineData
1da177e4
LT
1/*
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 */
caf4167a
KO
11#define pr_fmt(fmt) "%s: " fmt, __func__
12
1da177e4
LT
13#include <linux/kernel.h>
14#include <linux/init.h>
15#include <linux/errno.h>
16#include <linux/time.h>
17#include <linux/aio_abi.h>
630d9c47 18#include <linux/export.h>
1da177e4 19#include <linux/syscalls.h>
b9d128f1 20#include <linux/backing-dev.h>
027445c3 21#include <linux/uio.h>
1da177e4 22
1da177e4
LT
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>
3d2d827f 28#include <linux/mmu_context.h>
e1bdd5f2 29#include <linux/percpu.h>
1da177e4
LT
30#include <linux/slab.h>
31#include <linux/timer.h>
32#include <linux/aio.h>
33#include <linux/highmem.h>
34#include <linux/workqueue.h>
35#include <linux/security.h>
9c3060be 36#include <linux/eventfd.h>
cfb1e33e 37#include <linux/blkdev.h>
9d85cba7 38#include <linux/compat.h>
36bc08cc
GZ
39#include <linux/migrate.h>
40#include <linux/ramfs.h>
723be6e3 41#include <linux/percpu-refcount.h>
71ad7490 42#include <linux/mount.h>
1da177e4
LT
43
44#include <asm/kmap_types.h>
45#include <asm/uaccess.h>
1da177e4 46
68d70d03
AV
47#include "internal.h"
48
4e179bca
KO
49#define AIO_RING_MAGIC 0xa10a10a1
50#define AIO_RING_COMPAT_FEATURES 1
51#define AIO_RING_INCOMPAT_FEATURES 0
52struct aio_ring {
53 unsigned id; /* kernel internal index number */
54 unsigned nr; /* number of io_events */
fa8a53c3
BL
55 unsigned head; /* Written to by userland or under ring_lock
56 * mutex by aio_read_events_ring(). */
4e179bca
KO
57 unsigned tail;
58
59 unsigned magic;
60 unsigned compat_features;
61 unsigned incompat_features;
62 unsigned header_length; /* size of aio_ring */
63
64
65 struct io_event io_events[0];
66}; /* 128 bytes + ring size */
67
68#define AIO_RING_PAGES 8
4e179bca 69
db446a08
BL
70struct kioctx_table {
71 struct rcu_head rcu;
72 unsigned nr;
73 struct kioctx *table[];
74};
75
e1bdd5f2
KO
76struct kioctx_cpu {
77 unsigned reqs_available;
78};
79
dc48e56d
JA
80struct ctx_rq_wait {
81 struct completion comp;
82 atomic_t count;
83};
84
4e179bca 85struct kioctx {
723be6e3 86 struct percpu_ref users;
36f55889 87 atomic_t dead;
4e179bca 88
e34ecee2
KO
89 struct percpu_ref reqs;
90
4e179bca 91 unsigned long user_id;
4e179bca 92
e1bdd5f2
KO
93 struct __percpu kioctx_cpu *cpu;
94
95 /*
96 * For percpu reqs_available, number of slots we move to/from global
97 * counter at a time:
98 */
99 unsigned req_batch;
3e845ce0
KO
100 /*
101 * This is what userspace passed to io_setup(), it's not used for
102 * anything but counting against the global max_reqs quota.
103 *
58c85dc2 104 * The real limit is nr_events - 1, which will be larger (see
3e845ce0
KO
105 * aio_setup_ring())
106 */
4e179bca
KO
107 unsigned max_reqs;
108
58c85dc2
KO
109 /* Size of ringbuffer, in units of struct io_event */
110 unsigned nr_events;
4e179bca 111
58c85dc2
KO
112 unsigned long mmap_base;
113 unsigned long mmap_size;
114
115 struct page **ring_pages;
116 long nr_pages;
117
723be6e3 118 struct work_struct free_work;
4e23bcae 119
e02ba72a
AP
120 /*
121 * signals when all in-flight requests are done
122 */
dc48e56d 123 struct ctx_rq_wait *rq_wait;
e02ba72a 124
4e23bcae 125 struct {
34e83fc6
KO
126 /*
127 * This counts the number of available slots in the ringbuffer,
128 * so we avoid overflowing it: it's decremented (if positive)
129 * when allocating a kiocb and incremented when the resulting
130 * io_event is pulled off the ringbuffer.
e1bdd5f2
KO
131 *
132 * We batch accesses to it with a percpu version.
34e83fc6
KO
133 */
134 atomic_t reqs_available;
4e23bcae
KO
135 } ____cacheline_aligned_in_smp;
136
137 struct {
138 spinlock_t ctx_lock;
139 struct list_head active_reqs; /* used for cancellation */
140 } ____cacheline_aligned_in_smp;
141
58c85dc2
KO
142 struct {
143 struct mutex ring_lock;
4e23bcae
KO
144 wait_queue_head_t wait;
145 } ____cacheline_aligned_in_smp;
58c85dc2
KO
146
147 struct {
148 unsigned tail;
d856f32a 149 unsigned completed_events;
58c85dc2 150 spinlock_t completion_lock;
4e23bcae 151 } ____cacheline_aligned_in_smp;
58c85dc2
KO
152
153 struct page *internal_pages[AIO_RING_PAGES];
36bc08cc 154 struct file *aio_ring_file;
db446a08
BL
155
156 unsigned id;
4e179bca
KO
157};
158
04b2fa9f
CH
159/*
160 * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either
161 * cancelled or completed (this makes a certain amount of sense because
162 * successful cancellation - io_cancel() - does deliver the completion to
163 * userspace).
164 *
165 * And since most things don't implement kiocb cancellation and we'd really like
166 * kiocb completion to be lockless when possible, we use ki_cancel to
167 * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED
168 * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel().
169 */
170#define KIOCB_CANCELLED ((void *) (~0ULL))
171
172struct aio_kiocb {
173 struct kiocb common;
174
175 struct kioctx *ki_ctx;
176 kiocb_cancel_fn *ki_cancel;
177
178 struct iocb __user *ki_user_iocb; /* user's aiocb */
179 __u64 ki_user_data; /* user's data for completion */
180
181 struct list_head ki_list; /* the aio core uses this
182 * for cancellation */
183
184 /*
185 * If the aio_resfd field of the userspace iocb is not zero,
186 * this is the underlying eventfd context to deliver events to.
187 */
188 struct eventfd_ctx *ki_eventfd;
189};
190
1da177e4 191/*------ sysctl variables----*/
d55b5fda
ZB
192static DEFINE_SPINLOCK(aio_nr_lock);
193unsigned long aio_nr; /* current system wide number of aio requests */
194unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
1da177e4
LT
195/*----end sysctl variables---*/
196
e18b890b
CL
197static struct kmem_cache *kiocb_cachep;
198static struct kmem_cache *kioctx_cachep;
1da177e4 199
71ad7490
BL
200static struct vfsmount *aio_mnt;
201
202static const struct file_operations aio_ring_fops;
203static const struct address_space_operations aio_ctx_aops;
204
205static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
206{
207 struct qstr this = QSTR_INIT("[aio]", 5);
208 struct file *file;
209 struct path path;
210 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
7f62656b
DC
211 if (IS_ERR(inode))
212 return ERR_CAST(inode);
71ad7490
BL
213
214 inode->i_mapping->a_ops = &aio_ctx_aops;
215 inode->i_mapping->private_data = ctx;
216 inode->i_size = PAGE_SIZE * nr_pages;
217
218 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
219 if (!path.dentry) {
220 iput(inode);
221 return ERR_PTR(-ENOMEM);
222 }
223 path.mnt = mntget(aio_mnt);
224
225 d_instantiate(path.dentry, inode);
226 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
227 if (IS_ERR(file)) {
228 path_put(&path);
229 return file;
230 }
231
232 file->f_flags = O_RDWR;
71ad7490
BL
233 return file;
234}
235
236static struct dentry *aio_mount(struct file_system_type *fs_type,
237 int flags, const char *dev_name, void *data)
238{
239 static const struct dentry_operations ops = {
240 .d_dname = simple_dname,
241 };
8dc4379e 242 return mount_pseudo(fs_type, "aio:", NULL, &ops, AIO_RING_MAGIC);
71ad7490
BL
243}
244
1da177e4
LT
245/* aio_setup
246 * Creates the slab caches used by the aio routines, panic on
247 * failure as this is done early during the boot sequence.
248 */
249static int __init aio_setup(void)
250{
71ad7490
BL
251 static struct file_system_type aio_fs = {
252 .name = "aio",
253 .mount = aio_mount,
254 .kill_sb = kill_anon_super,
255 };
256 aio_mnt = kern_mount(&aio_fs);
257 if (IS_ERR(aio_mnt))
258 panic("Failed to create aio fs mount.");
259
04b2fa9f 260 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
0a31bd5f 261 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
1da177e4 262
caf4167a 263 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
1da177e4
LT
264
265 return 0;
266}
385773e0 267__initcall(aio_setup);
1da177e4 268
5e9ae2e5
BL
269static void put_aio_ring_file(struct kioctx *ctx)
270{
271 struct file *aio_ring_file = ctx->aio_ring_file;
272 if (aio_ring_file) {
273 truncate_setsize(aio_ring_file->f_inode, 0);
274
275 /* Prevent further access to the kioctx from migratepages */
276 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
277 aio_ring_file->f_inode->i_mapping->private_data = NULL;
278 ctx->aio_ring_file = NULL;
279 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
280
281 fput(aio_ring_file);
282 }
283}
284
1da177e4
LT
285static void aio_free_ring(struct kioctx *ctx)
286{
36bc08cc 287 int i;
1da177e4 288
fa8a53c3
BL
289 /* Disconnect the kiotx from the ring file. This prevents future
290 * accesses to the kioctx from page migration.
291 */
292 put_aio_ring_file(ctx);
293
36bc08cc 294 for (i = 0; i < ctx->nr_pages; i++) {
8e321fef 295 struct page *page;
36bc08cc
GZ
296 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
297 page_count(ctx->ring_pages[i]));
8e321fef
BL
298 page = ctx->ring_pages[i];
299 if (!page)
300 continue;
301 ctx->ring_pages[i] = NULL;
302 put_page(page);
36bc08cc 303 }
1da177e4 304
ddb8c45b 305 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
58c85dc2 306 kfree(ctx->ring_pages);
ddb8c45b
SL
307 ctx->ring_pages = NULL;
308 }
36bc08cc
GZ
309}
310
311static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
312{
e4a0d3e7 313 vma->vm_flags |= VM_DONTEXPAND;
36bc08cc
GZ
314 vma->vm_ops = &generic_file_vm_ops;
315 return 0;
316}
317
b2edffdd 318static int aio_ring_remap(struct file *file, struct vm_area_struct *vma)
e4a0d3e7
PE
319{
320 struct mm_struct *mm = vma->vm_mm;
321 struct kioctx_table *table;
b2edffdd 322 int i, res = -EINVAL;
e4a0d3e7
PE
323
324 spin_lock(&mm->ioctx_lock);
325 rcu_read_lock();
326 table = rcu_dereference(mm->ioctx_table);
327 for (i = 0; i < table->nr; i++) {
328 struct kioctx *ctx;
329
330 ctx = table->table[i];
331 if (ctx && ctx->aio_ring_file == file) {
b2edffdd
AV
332 if (!atomic_read(&ctx->dead)) {
333 ctx->user_id = ctx->mmap_base = vma->vm_start;
334 res = 0;
335 }
e4a0d3e7
PE
336 break;
337 }
338 }
339
340 rcu_read_unlock();
341 spin_unlock(&mm->ioctx_lock);
b2edffdd 342 return res;
e4a0d3e7
PE
343}
344
36bc08cc
GZ
345static const struct file_operations aio_ring_fops = {
346 .mmap = aio_ring_mmap,
e4a0d3e7 347 .mremap = aio_ring_remap,
36bc08cc
GZ
348};
349
0c45355f 350#if IS_ENABLED(CONFIG_MIGRATION)
36bc08cc
GZ
351static int aio_migratepage(struct address_space *mapping, struct page *new,
352 struct page *old, enum migrate_mode mode)
353{
5e9ae2e5 354 struct kioctx *ctx;
36bc08cc 355 unsigned long flags;
fa8a53c3 356 pgoff_t idx;
36bc08cc
GZ
357 int rc;
358
8e321fef
BL
359 rc = 0;
360
fa8a53c3 361 /* mapping->private_lock here protects against the kioctx teardown. */
8e321fef
BL
362 spin_lock(&mapping->private_lock);
363 ctx = mapping->private_data;
fa8a53c3
BL
364 if (!ctx) {
365 rc = -EINVAL;
366 goto out;
367 }
368
369 /* The ring_lock mutex. The prevents aio_read_events() from writing
370 * to the ring's head, and prevents page migration from mucking in
371 * a partially initialized kiotx.
372 */
373 if (!mutex_trylock(&ctx->ring_lock)) {
374 rc = -EAGAIN;
375 goto out;
376 }
377
378 idx = old->index;
379 if (idx < (pgoff_t)ctx->nr_pages) {
380 /* Make sure the old page hasn't already been changed */
381 if (ctx->ring_pages[idx] != old)
382 rc = -EAGAIN;
8e321fef
BL
383 } else
384 rc = -EINVAL;
8e321fef
BL
385
386 if (rc != 0)
fa8a53c3 387 goto out_unlock;
8e321fef 388
36bc08cc
GZ
389 /* Writeback must be complete */
390 BUG_ON(PageWriteback(old));
8e321fef 391 get_page(new);
36bc08cc 392
8e321fef 393 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
36bc08cc 394 if (rc != MIGRATEPAGE_SUCCESS) {
8e321fef 395 put_page(new);
fa8a53c3 396 goto out_unlock;
36bc08cc
GZ
397 }
398
fa8a53c3
BL
399 /* Take completion_lock to prevent other writes to the ring buffer
400 * while the old page is copied to the new. This prevents new
401 * events from being lost.
5e9ae2e5 402 */
fa8a53c3
BL
403 spin_lock_irqsave(&ctx->completion_lock, flags);
404 migrate_page_copy(new, old);
405 BUG_ON(ctx->ring_pages[idx] != old);
406 ctx->ring_pages[idx] = new;
407 spin_unlock_irqrestore(&ctx->completion_lock, flags);
36bc08cc 408
fa8a53c3
BL
409 /* The old page is no longer accessible. */
410 put_page(old);
8e321fef 411
fa8a53c3
BL
412out_unlock:
413 mutex_unlock(&ctx->ring_lock);
414out:
415 spin_unlock(&mapping->private_lock);
36bc08cc 416 return rc;
1da177e4 417}
0c45355f 418#endif
1da177e4 419
36bc08cc 420static const struct address_space_operations aio_ctx_aops = {
835f252c 421 .set_page_dirty = __set_page_dirty_no_writeback,
0c45355f 422#if IS_ENABLED(CONFIG_MIGRATION)
36bc08cc 423 .migratepage = aio_migratepage,
0c45355f 424#endif
36bc08cc
GZ
425};
426
1da177e4
LT
427static int aio_setup_ring(struct kioctx *ctx)
428{
429 struct aio_ring *ring;
1da177e4 430 unsigned nr_events = ctx->max_reqs;
41003a7b 431 struct mm_struct *mm = current->mm;
3dc9acb6 432 unsigned long size, unused;
1da177e4 433 int nr_pages;
36bc08cc
GZ
434 int i;
435 struct file *file;
1da177e4
LT
436
437 /* Compensate for the ring buffer's head/tail overlap entry */
438 nr_events += 2; /* 1 is required, 2 for good luck */
439
440 size = sizeof(struct aio_ring);
441 size += sizeof(struct io_event) * nr_events;
1da177e4 442
36bc08cc 443 nr_pages = PFN_UP(size);
1da177e4
LT
444 if (nr_pages < 0)
445 return -EINVAL;
446
71ad7490 447 file = aio_private_file(ctx, nr_pages);
36bc08cc
GZ
448 if (IS_ERR(file)) {
449 ctx->aio_ring_file = NULL;
fa8a53c3 450 return -ENOMEM;
36bc08cc
GZ
451 }
452
3dc9acb6
LT
453 ctx->aio_ring_file = file;
454 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
455 / sizeof(struct io_event);
456
457 ctx->ring_pages = ctx->internal_pages;
458 if (nr_pages > AIO_RING_PAGES) {
459 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
460 GFP_KERNEL);
461 if (!ctx->ring_pages) {
462 put_aio_ring_file(ctx);
463 return -ENOMEM;
464 }
465 }
466
36bc08cc
GZ
467 for (i = 0; i < nr_pages; i++) {
468 struct page *page;
469 page = find_or_create_page(file->f_inode->i_mapping,
470 i, GFP_HIGHUSER | __GFP_ZERO);
471 if (!page)
472 break;
473 pr_debug("pid(%d) page[%d]->count=%d\n",
474 current->pid, i, page_count(page));
475 SetPageUptodate(page);
36bc08cc 476 unlock_page(page);
3dc9acb6
LT
477
478 ctx->ring_pages[i] = page;
36bc08cc 479 }
3dc9acb6 480 ctx->nr_pages = i;
1da177e4 481
3dc9acb6
LT
482 if (unlikely(i != nr_pages)) {
483 aio_free_ring(ctx);
fa8a53c3 484 return -ENOMEM;
1da177e4
LT
485 }
486
58c85dc2
KO
487 ctx->mmap_size = nr_pages * PAGE_SIZE;
488 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
36bc08cc 489
41003a7b 490 down_write(&mm->mmap_sem);
36bc08cc
GZ
491 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
492 PROT_READ | PROT_WRITE,
3dc9acb6
LT
493 MAP_SHARED, 0, &unused);
494 up_write(&mm->mmap_sem);
58c85dc2 495 if (IS_ERR((void *)ctx->mmap_base)) {
58c85dc2 496 ctx->mmap_size = 0;
1da177e4 497 aio_free_ring(ctx);
fa8a53c3 498 return -ENOMEM;
1da177e4
LT
499 }
500
58c85dc2 501 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
d6c355c7 502
58c85dc2
KO
503 ctx->user_id = ctx->mmap_base;
504 ctx->nr_events = nr_events; /* trusted copy */
1da177e4 505
58c85dc2 506 ring = kmap_atomic(ctx->ring_pages[0]);
1da177e4 507 ring->nr = nr_events; /* user copy */
db446a08 508 ring->id = ~0U;
1da177e4
LT
509 ring->head = ring->tail = 0;
510 ring->magic = AIO_RING_MAGIC;
511 ring->compat_features = AIO_RING_COMPAT_FEATURES;
512 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
513 ring->header_length = sizeof(struct aio_ring);
e8e3c3d6 514 kunmap_atomic(ring);
58c85dc2 515 flush_dcache_page(ctx->ring_pages[0]);
1da177e4
LT
516
517 return 0;
518}
519
1da177e4
LT
520#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
521#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
522#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
523
04b2fa9f 524void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
0460fef2 525{
04b2fa9f 526 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common);
0460fef2
KO
527 struct kioctx *ctx = req->ki_ctx;
528 unsigned long flags;
529
530 spin_lock_irqsave(&ctx->ctx_lock, flags);
531
532 if (!req->ki_list.next)
533 list_add(&req->ki_list, &ctx->active_reqs);
534
535 req->ki_cancel = cancel;
536
537 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
538}
539EXPORT_SYMBOL(kiocb_set_cancel_fn);
540
04b2fa9f 541static int kiocb_cancel(struct aio_kiocb *kiocb)
906b973c 542{
0460fef2 543 kiocb_cancel_fn *old, *cancel;
906b973c 544
0460fef2
KO
545 /*
546 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
547 * actually has a cancel function, hence the cmpxchg()
548 */
549
550 cancel = ACCESS_ONCE(kiocb->ki_cancel);
551 do {
552 if (!cancel || cancel == KIOCB_CANCELLED)
57282d8f 553 return -EINVAL;
906b973c 554
0460fef2
KO
555 old = cancel;
556 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
557 } while (cancel != old);
906b973c 558
04b2fa9f 559 return cancel(&kiocb->common);
906b973c
KO
560}
561
e34ecee2 562static void free_ioctx(struct work_struct *work)
36f55889 563{
e34ecee2 564 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
e1bdd5f2 565
e34ecee2 566 pr_debug("freeing %p\n", ctx);
e1bdd5f2 567
e34ecee2 568 aio_free_ring(ctx);
e1bdd5f2 569 free_percpu(ctx->cpu);
9a1049da
TH
570 percpu_ref_exit(&ctx->reqs);
571 percpu_ref_exit(&ctx->users);
36f55889
KO
572 kmem_cache_free(kioctx_cachep, ctx);
573}
574
e34ecee2
KO
575static void free_ioctx_reqs(struct percpu_ref *ref)
576{
577 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
578
e02ba72a 579 /* At this point we know that there are no any in-flight requests */
dc48e56d
JA
580 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
581 complete(&ctx->rq_wait->comp);
e02ba72a 582
e34ecee2
KO
583 INIT_WORK(&ctx->free_work, free_ioctx);
584 schedule_work(&ctx->free_work);
585}
586
36f55889
KO
587/*
588 * When this function runs, the kioctx has been removed from the "hash table"
589 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
590 * now it's safe to cancel any that need to be.
591 */
e34ecee2 592static void free_ioctx_users(struct percpu_ref *ref)
36f55889 593{
e34ecee2 594 struct kioctx *ctx = container_of(ref, struct kioctx, users);
04b2fa9f 595 struct aio_kiocb *req;
36f55889
KO
596
597 spin_lock_irq(&ctx->ctx_lock);
598
599 while (!list_empty(&ctx->active_reqs)) {
600 req = list_first_entry(&ctx->active_reqs,
04b2fa9f 601 struct aio_kiocb, ki_list);
36f55889
KO
602
603 list_del_init(&req->ki_list);
d52a8f9e 604 kiocb_cancel(req);
36f55889
KO
605 }
606
607 spin_unlock_irq(&ctx->ctx_lock);
608
e34ecee2
KO
609 percpu_ref_kill(&ctx->reqs);
610 percpu_ref_put(&ctx->reqs);
36f55889
KO
611}
612
db446a08
BL
613static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
614{
615 unsigned i, new_nr;
616 struct kioctx_table *table, *old;
617 struct aio_ring *ring;
618
619 spin_lock(&mm->ioctx_lock);
855ef0de 620 table = rcu_dereference_raw(mm->ioctx_table);
db446a08
BL
621
622 while (1) {
623 if (table)
624 for (i = 0; i < table->nr; i++)
625 if (!table->table[i]) {
626 ctx->id = i;
627 table->table[i] = ctx;
628 spin_unlock(&mm->ioctx_lock);
629
fa8a53c3
BL
630 /* While kioctx setup is in progress,
631 * we are protected from page migration
632 * changes ring_pages by ->ring_lock.
633 */
db446a08
BL
634 ring = kmap_atomic(ctx->ring_pages[0]);
635 ring->id = ctx->id;
636 kunmap_atomic(ring);
637 return 0;
638 }
639
640 new_nr = (table ? table->nr : 1) * 4;
db446a08
BL
641 spin_unlock(&mm->ioctx_lock);
642
643 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
644 new_nr, GFP_KERNEL);
645 if (!table)
646 return -ENOMEM;
647
648 table->nr = new_nr;
649
650 spin_lock(&mm->ioctx_lock);
855ef0de 651 old = rcu_dereference_raw(mm->ioctx_table);
db446a08
BL
652
653 if (!old) {
654 rcu_assign_pointer(mm->ioctx_table, table);
655 } else if (table->nr > old->nr) {
656 memcpy(table->table, old->table,
657 old->nr * sizeof(struct kioctx *));
658
659 rcu_assign_pointer(mm->ioctx_table, table);
660 kfree_rcu(old, rcu);
661 } else {
662 kfree(table);
663 table = old;
664 }
665 }
666}
667
e34ecee2
KO
668static void aio_nr_sub(unsigned nr)
669{
670 spin_lock(&aio_nr_lock);
671 if (WARN_ON(aio_nr - nr > aio_nr))
672 aio_nr = 0;
673 else
674 aio_nr -= nr;
675 spin_unlock(&aio_nr_lock);
676}
677
1da177e4
LT
678/* ioctx_alloc
679 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
680 */
681static struct kioctx *ioctx_alloc(unsigned nr_events)
682{
41003a7b 683 struct mm_struct *mm = current->mm;
1da177e4 684 struct kioctx *ctx;
e23754f8 685 int err = -ENOMEM;
1da177e4 686
e1bdd5f2
KO
687 /*
688 * We keep track of the number of available ringbuffer slots, to prevent
689 * overflow (reqs_available), and we also use percpu counters for this.
690 *
691 * So since up to half the slots might be on other cpu's percpu counters
692 * and unavailable, double nr_events so userspace sees what they
693 * expected: additionally, we move req_batch slots to/from percpu
694 * counters at a time, so make sure that isn't 0:
695 */
696 nr_events = max(nr_events, num_possible_cpus() * 4);
697 nr_events *= 2;
698
1da177e4 699 /* Prevent overflows */
08397acd 700 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
1da177e4
LT
701 pr_debug("ENOMEM: nr_events too high\n");
702 return ERR_PTR(-EINVAL);
703 }
704
4cd81c3d 705 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
1da177e4
LT
706 return ERR_PTR(-EAGAIN);
707
c3762229 708 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
1da177e4
LT
709 if (!ctx)
710 return ERR_PTR(-ENOMEM);
711
1da177e4 712 ctx->max_reqs = nr_events;
1da177e4 713
1da177e4 714 spin_lock_init(&ctx->ctx_lock);
0460fef2 715 spin_lock_init(&ctx->completion_lock);
58c85dc2 716 mutex_init(&ctx->ring_lock);
fa8a53c3
BL
717 /* Protect against page migration throughout kiotx setup by keeping
718 * the ring_lock mutex held until setup is complete. */
719 mutex_lock(&ctx->ring_lock);
1da177e4
LT
720 init_waitqueue_head(&ctx->wait);
721
722 INIT_LIST_HEAD(&ctx->active_reqs);
1da177e4 723
2aad2a86 724 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
fa8a53c3
BL
725 goto err;
726
2aad2a86 727 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
fa8a53c3
BL
728 goto err;
729
e1bdd5f2
KO
730 ctx->cpu = alloc_percpu(struct kioctx_cpu);
731 if (!ctx->cpu)
e34ecee2 732 goto err;
1da177e4 733
fa8a53c3
BL
734 err = aio_setup_ring(ctx);
735 if (err < 0)
e34ecee2 736 goto err;
e1bdd5f2 737
34e83fc6 738 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
e1bdd5f2 739 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
6878ea72
BL
740 if (ctx->req_batch < 1)
741 ctx->req_batch = 1;
34e83fc6 742
1da177e4 743 /* limit the number of system wide aios */
9fa1cb39 744 spin_lock(&aio_nr_lock);
4cd81c3d 745 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
2dd542b7 746 aio_nr + nr_events < aio_nr) {
9fa1cb39 747 spin_unlock(&aio_nr_lock);
e34ecee2 748 err = -EAGAIN;
d1b94327 749 goto err_ctx;
2dd542b7
AV
750 }
751 aio_nr += ctx->max_reqs;
9fa1cb39 752 spin_unlock(&aio_nr_lock);
1da177e4 753
1881686f
BL
754 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
755 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
723be6e3 756
da90382c
BL
757 err = ioctx_add_table(ctx, mm);
758 if (err)
e34ecee2 759 goto err_cleanup;
da90382c 760
fa8a53c3
BL
761 /* Release the ring_lock mutex now that all setup is complete. */
762 mutex_unlock(&ctx->ring_lock);
763
caf4167a 764 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
58c85dc2 765 ctx, ctx->user_id, mm, ctx->nr_events);
1da177e4
LT
766 return ctx;
767
e34ecee2
KO
768err_cleanup:
769 aio_nr_sub(ctx->max_reqs);
d1b94327 770err_ctx:
deeb8525
AV
771 atomic_set(&ctx->dead, 1);
772 if (ctx->mmap_size)
773 vm_munmap(ctx->mmap_base, ctx->mmap_size);
d1b94327 774 aio_free_ring(ctx);
e34ecee2 775err:
fa8a53c3 776 mutex_unlock(&ctx->ring_lock);
e1bdd5f2 777 free_percpu(ctx->cpu);
9a1049da
TH
778 percpu_ref_exit(&ctx->reqs);
779 percpu_ref_exit(&ctx->users);
1da177e4 780 kmem_cache_free(kioctx_cachep, ctx);
caf4167a 781 pr_debug("error allocating ioctx %d\n", err);
e23754f8 782 return ERR_PTR(err);
1da177e4
LT
783}
784
36f55889
KO
785/* kill_ioctx
786 * Cancels all outstanding aio requests on an aio context. Used
787 * when the processes owning a context have all exited to encourage
788 * the rapid destruction of the kioctx.
789 */
fb2d4483 790static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
dc48e56d 791 struct ctx_rq_wait *wait)
36f55889 792{
fa88b6f8 793 struct kioctx_table *table;
db446a08 794
b2edffdd
AV
795 spin_lock(&mm->ioctx_lock);
796 if (atomic_xchg(&ctx->dead, 1)) {
797 spin_unlock(&mm->ioctx_lock);
fa88b6f8 798 return -EINVAL;
b2edffdd 799 }
db446a08 800
855ef0de 801 table = rcu_dereference_raw(mm->ioctx_table);
fa88b6f8
BL
802 WARN_ON(ctx != table->table[ctx->id]);
803 table->table[ctx->id] = NULL;
fa88b6f8 804 spin_unlock(&mm->ioctx_lock);
4fcc712f 805
fa88b6f8
BL
806 /* percpu_ref_kill() will do the necessary call_rcu() */
807 wake_up_all(&ctx->wait);
4fcc712f 808
fa88b6f8
BL
809 /*
810 * It'd be more correct to do this in free_ioctx(), after all
811 * the outstanding kiocbs have finished - but by then io_destroy
812 * has already returned, so io_setup() could potentially return
813 * -EAGAIN with no ioctxs actually in use (as far as userspace
814 * could tell).
815 */
816 aio_nr_sub(ctx->max_reqs);
4fcc712f 817
fa88b6f8
BL
818 if (ctx->mmap_size)
819 vm_munmap(ctx->mmap_base, ctx->mmap_size);
fb2d4483 820
dc48e56d 821 ctx->rq_wait = wait;
fa88b6f8
BL
822 percpu_ref_kill(&ctx->users);
823 return 0;
1da177e4
LT
824}
825
36f55889
KO
826/*
827 * exit_aio: called when the last user of mm goes away. At this point, there is
828 * no way for any new requests to be submited or any of the io_* syscalls to be
829 * called on the context.
830 *
831 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
832 * them.
1da177e4 833 */
fc9b52cd 834void exit_aio(struct mm_struct *mm)
1da177e4 835{
4b70ac5f 836 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
dc48e56d
JA
837 struct ctx_rq_wait wait;
838 int i, skipped;
db446a08 839
4b70ac5f
ON
840 if (!table)
841 return;
db446a08 842
dc48e56d
JA
843 atomic_set(&wait.count, table->nr);
844 init_completion(&wait.comp);
845
846 skipped = 0;
4b70ac5f
ON
847 for (i = 0; i < table->nr; ++i) {
848 struct kioctx *ctx = table->table[i];
abf137dd 849
dc48e56d
JA
850 if (!ctx) {
851 skipped++;
4b70ac5f 852 continue;
dc48e56d
JA
853 }
854
936af157 855 /*
4b70ac5f
ON
856 * We don't need to bother with munmap() here - exit_mmap(mm)
857 * is coming and it'll unmap everything. And we simply can't,
858 * this is not necessarily our ->mm.
859 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
860 * that it needs to unmap the area, just set it to 0.
936af157 861 */
58c85dc2 862 ctx->mmap_size = 0;
dc48e56d
JA
863 kill_ioctx(mm, ctx, &wait);
864 }
36f55889 865
dc48e56d 866 if (!atomic_sub_and_test(skipped, &wait.count)) {
6098b45b 867 /* Wait until all IO for the context are done. */
dc48e56d 868 wait_for_completion(&wait.comp);
1da177e4 869 }
4b70ac5f
ON
870
871 RCU_INIT_POINTER(mm->ioctx_table, NULL);
872 kfree(table);
1da177e4
LT
873}
874
e1bdd5f2
KO
875static void put_reqs_available(struct kioctx *ctx, unsigned nr)
876{
877 struct kioctx_cpu *kcpu;
263782c1 878 unsigned long flags;
e1bdd5f2 879
263782c1 880 local_irq_save(flags);
be6fb451 881 kcpu = this_cpu_ptr(ctx->cpu);
e1bdd5f2 882 kcpu->reqs_available += nr;
263782c1 883
e1bdd5f2
KO
884 while (kcpu->reqs_available >= ctx->req_batch * 2) {
885 kcpu->reqs_available -= ctx->req_batch;
886 atomic_add(ctx->req_batch, &ctx->reqs_available);
887 }
888
263782c1 889 local_irq_restore(flags);
e1bdd5f2
KO
890}
891
892static bool get_reqs_available(struct kioctx *ctx)
893{
894 struct kioctx_cpu *kcpu;
895 bool ret = false;
263782c1 896 unsigned long flags;
e1bdd5f2 897
263782c1 898 local_irq_save(flags);
be6fb451 899 kcpu = this_cpu_ptr(ctx->cpu);
e1bdd5f2
KO
900 if (!kcpu->reqs_available) {
901 int old, avail = atomic_read(&ctx->reqs_available);
902
903 do {
904 if (avail < ctx->req_batch)
905 goto out;
906
907 old = avail;
908 avail = atomic_cmpxchg(&ctx->reqs_available,
909 avail, avail - ctx->req_batch);
910 } while (avail != old);
911
912 kcpu->reqs_available += ctx->req_batch;
913 }
914
915 ret = true;
916 kcpu->reqs_available--;
917out:
263782c1 918 local_irq_restore(flags);
e1bdd5f2
KO
919 return ret;
920}
921
d856f32a
BL
922/* refill_reqs_available
923 * Updates the reqs_available reference counts used for tracking the
924 * number of free slots in the completion ring. This can be called
925 * from aio_complete() (to optimistically update reqs_available) or
926 * from aio_get_req() (the we're out of events case). It must be
927 * called holding ctx->completion_lock.
928 */
929static void refill_reqs_available(struct kioctx *ctx, unsigned head,
930 unsigned tail)
931{
932 unsigned events_in_ring, completed;
933
934 /* Clamp head since userland can write to it. */
935 head %= ctx->nr_events;
936 if (head <= tail)
937 events_in_ring = tail - head;
938 else
939 events_in_ring = ctx->nr_events - (head - tail);
940
941 completed = ctx->completed_events;
942 if (events_in_ring < completed)
943 completed -= events_in_ring;
944 else
945 completed = 0;
946
947 if (!completed)
948 return;
949
950 ctx->completed_events -= completed;
951 put_reqs_available(ctx, completed);
952}
953
954/* user_refill_reqs_available
955 * Called to refill reqs_available when aio_get_req() encounters an
956 * out of space in the completion ring.
957 */
958static void user_refill_reqs_available(struct kioctx *ctx)
959{
960 spin_lock_irq(&ctx->completion_lock);
961 if (ctx->completed_events) {
962 struct aio_ring *ring;
963 unsigned head;
964
965 /* Access of ring->head may race with aio_read_events_ring()
966 * here, but that's okay since whether we read the old version
967 * or the new version, and either will be valid. The important
968 * part is that head cannot pass tail since we prevent
969 * aio_complete() from updating tail by holding
970 * ctx->completion_lock. Even if head is invalid, the check
971 * against ctx->completed_events below will make sure we do the
972 * safe/right thing.
973 */
974 ring = kmap_atomic(ctx->ring_pages[0]);
975 head = ring->head;
976 kunmap_atomic(ring);
977
978 refill_reqs_available(ctx, head, ctx->tail);
979 }
980
981 spin_unlock_irq(&ctx->completion_lock);
982}
983
1da177e4 984/* aio_get_req
57282d8f
KO
985 * Allocate a slot for an aio request.
986 * Returns NULL if no requests are free.
1da177e4 987 */
04b2fa9f 988static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1da177e4 989{
04b2fa9f 990 struct aio_kiocb *req;
a1c8eae7 991
d856f32a
BL
992 if (!get_reqs_available(ctx)) {
993 user_refill_reqs_available(ctx);
994 if (!get_reqs_available(ctx))
995 return NULL;
996 }
a1c8eae7 997
0460fef2 998 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
1da177e4 999 if (unlikely(!req))
a1c8eae7 1000 goto out_put;
1da177e4 1001
e34ecee2
KO
1002 percpu_ref_get(&ctx->reqs);
1003
1da177e4 1004 req->ki_ctx = ctx;
080d676d 1005 return req;
a1c8eae7 1006out_put:
e1bdd5f2 1007 put_reqs_available(ctx, 1);
a1c8eae7 1008 return NULL;
1da177e4
LT
1009}
1010
04b2fa9f 1011static void kiocb_free(struct aio_kiocb *req)
1da177e4 1012{
04b2fa9f
CH
1013 if (req->common.ki_filp)
1014 fput(req->common.ki_filp);
13389010
DL
1015 if (req->ki_eventfd != NULL)
1016 eventfd_ctx_put(req->ki_eventfd);
1da177e4 1017 kmem_cache_free(kiocb_cachep, req);
1da177e4
LT
1018}
1019
d5470b59 1020static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1da177e4 1021{
db446a08 1022 struct aio_ring __user *ring = (void __user *)ctx_id;
abf137dd 1023 struct mm_struct *mm = current->mm;
65c24491 1024 struct kioctx *ctx, *ret = NULL;
db446a08
BL
1025 struct kioctx_table *table;
1026 unsigned id;
1027
1028 if (get_user(id, &ring->id))
1029 return NULL;
1da177e4 1030
abf137dd 1031 rcu_read_lock();
db446a08 1032 table = rcu_dereference(mm->ioctx_table);
abf137dd 1033
db446a08
BL
1034 if (!table || id >= table->nr)
1035 goto out;
1da177e4 1036
db446a08 1037 ctx = table->table[id];
f30d704f 1038 if (ctx && ctx->user_id == ctx_id) {
db446a08
BL
1039 percpu_ref_get(&ctx->users);
1040 ret = ctx;
1041 }
1042out:
abf137dd 1043 rcu_read_unlock();
65c24491 1044 return ret;
1da177e4
LT
1045}
1046
1da177e4
LT
1047/* aio_complete
1048 * Called when the io request on the given iocb is complete.
1da177e4 1049 */
04b2fa9f 1050static void aio_complete(struct kiocb *kiocb, long res, long res2)
1da177e4 1051{
04b2fa9f 1052 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common);
1da177e4 1053 struct kioctx *ctx = iocb->ki_ctx;
1da177e4 1054 struct aio_ring *ring;
21b40200 1055 struct io_event *ev_page, *event;
d856f32a 1056 unsigned tail, pos, head;
1da177e4 1057 unsigned long flags;
1da177e4 1058
20dcae32
ZB
1059 /*
1060 * Special case handling for sync iocbs:
1061 * - events go directly into the iocb for fast handling
1062 * - the sync task with the iocb in its stack holds the single iocb
1063 * ref, no other paths have a way to get another ref
1064 * - the sync task helpfully left a reference to itself in the iocb
1da177e4 1065 */
04b2fa9f 1066 BUG_ON(is_sync_kiocb(kiocb));
1da177e4 1067
0460fef2
KO
1068 if (iocb->ki_list.next) {
1069 unsigned long flags;
1070
1071 spin_lock_irqsave(&ctx->ctx_lock, flags);
1072 list_del(&iocb->ki_list);
1073 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1074 }
11599eba 1075
0460fef2
KO
1076 /*
1077 * Add a completion event to the ring buffer. Must be done holding
4b30f07e 1078 * ctx->completion_lock to prevent other code from messing with the tail
0460fef2
KO
1079 * pointer since we might be called from irq context.
1080 */
1081 spin_lock_irqsave(&ctx->completion_lock, flags);
1082
58c85dc2 1083 tail = ctx->tail;
21b40200
KO
1084 pos = tail + AIO_EVENTS_OFFSET;
1085
58c85dc2 1086 if (++tail >= ctx->nr_events)
4bf69b2a 1087 tail = 0;
1da177e4 1088
58c85dc2 1089 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
21b40200
KO
1090 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1091
04b2fa9f 1092 event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
1da177e4
LT
1093 event->data = iocb->ki_user_data;
1094 event->res = res;
1095 event->res2 = res2;
1096
21b40200 1097 kunmap_atomic(ev_page);
58c85dc2 1098 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
21b40200
KO
1099
1100 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
04b2fa9f 1101 ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
caf4167a 1102 res, res2);
1da177e4
LT
1103
1104 /* after flagging the request as done, we
1105 * must never even look at it again
1106 */
1107 smp_wmb(); /* make event visible before updating tail */
1108
58c85dc2 1109 ctx->tail = tail;
1da177e4 1110
58c85dc2 1111 ring = kmap_atomic(ctx->ring_pages[0]);
d856f32a 1112 head = ring->head;
21b40200 1113 ring->tail = tail;
e8e3c3d6 1114 kunmap_atomic(ring);
58c85dc2 1115 flush_dcache_page(ctx->ring_pages[0]);
1da177e4 1116
d856f32a
BL
1117 ctx->completed_events++;
1118 if (ctx->completed_events > 1)
1119 refill_reqs_available(ctx, head, tail);
0460fef2
KO
1120 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1121
21b40200 1122 pr_debug("added to ring %p at [%u]\n", iocb, tail);
8d1c98b0
DL
1123
1124 /*
1125 * Check if the user asked us to deliver the result through an
1126 * eventfd. The eventfd_signal() function is safe to be called
1127 * from IRQ context.
1128 */
87c3a86e 1129 if (iocb->ki_eventfd != NULL)
8d1c98b0
DL
1130 eventfd_signal(iocb->ki_eventfd, 1);
1131
1da177e4 1132 /* everything turned out well, dispose of the aiocb. */
57282d8f 1133 kiocb_free(iocb);
1da177e4 1134
6cb2a210
QB
1135 /*
1136 * We have to order our ring_info tail store above and test
1137 * of the wait list below outside the wait lock. This is
1138 * like in wake_up_bit() where clearing a bit has to be
1139 * ordered with the unlocked test.
1140 */
1141 smp_mb();
1142
1da177e4
LT
1143 if (waitqueue_active(&ctx->wait))
1144 wake_up(&ctx->wait);
1145
e34ecee2 1146 percpu_ref_put(&ctx->reqs);
1da177e4
LT
1147}
1148
2be4e7de 1149/* aio_read_events_ring
a31ad380
KO
1150 * Pull an event off of the ioctx's event ring. Returns the number of
1151 * events fetched
1da177e4 1152 */
a31ad380
KO
1153static long aio_read_events_ring(struct kioctx *ctx,
1154 struct io_event __user *event, long nr)
1da177e4 1155{
1da177e4 1156 struct aio_ring *ring;
5ffac122 1157 unsigned head, tail, pos;
a31ad380
KO
1158 long ret = 0;
1159 int copy_ret;
1160
9c9ce763
DC
1161 /*
1162 * The mutex can block and wake us up and that will cause
1163 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1164 * and repeat. This should be rare enough that it doesn't cause
1165 * peformance issues. See the comment in read_events() for more detail.
1166 */
1167 sched_annotate_sleep();
58c85dc2 1168 mutex_lock(&ctx->ring_lock);
1da177e4 1169
fa8a53c3 1170 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
58c85dc2 1171 ring = kmap_atomic(ctx->ring_pages[0]);
a31ad380 1172 head = ring->head;
5ffac122 1173 tail = ring->tail;
a31ad380
KO
1174 kunmap_atomic(ring);
1175
2ff396be
JM
1176 /*
1177 * Ensure that once we've read the current tail pointer, that
1178 * we also see the events that were stored up to the tail.
1179 */
1180 smp_rmb();
1181
5ffac122 1182 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1da177e4 1183
5ffac122 1184 if (head == tail)
1da177e4
LT
1185 goto out;
1186
edfbbf38
BL
1187 head %= ctx->nr_events;
1188 tail %= ctx->nr_events;
1189
a31ad380
KO
1190 while (ret < nr) {
1191 long avail;
1192 struct io_event *ev;
1193 struct page *page;
1194
5ffac122
KO
1195 avail = (head <= tail ? tail : ctx->nr_events) - head;
1196 if (head == tail)
a31ad380
KO
1197 break;
1198
1199 avail = min(avail, nr - ret);
1200 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1201 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1202
1203 pos = head + AIO_EVENTS_OFFSET;
58c85dc2 1204 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
a31ad380
KO
1205 pos %= AIO_EVENTS_PER_PAGE;
1206
1207 ev = kmap(page);
1208 copy_ret = copy_to_user(event + ret, ev + pos,
1209 sizeof(*ev) * avail);
1210 kunmap(page);
1211
1212 if (unlikely(copy_ret)) {
1213 ret = -EFAULT;
1214 goto out;
1215 }
1216
1217 ret += avail;
1218 head += avail;
58c85dc2 1219 head %= ctx->nr_events;
1da177e4 1220 }
1da177e4 1221
58c85dc2 1222 ring = kmap_atomic(ctx->ring_pages[0]);
a31ad380 1223 ring->head = head;
91d80a84 1224 kunmap_atomic(ring);
58c85dc2 1225 flush_dcache_page(ctx->ring_pages[0]);
a31ad380 1226
5ffac122 1227 pr_debug("%li h%u t%u\n", ret, head, tail);
a31ad380 1228out:
58c85dc2 1229 mutex_unlock(&ctx->ring_lock);
a31ad380 1230
1da177e4
LT
1231 return ret;
1232}
1233
a31ad380
KO
1234static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1235 struct io_event __user *event, long *i)
1da177e4 1236{
a31ad380 1237 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1da177e4 1238
a31ad380
KO
1239 if (ret > 0)
1240 *i += ret;
1da177e4 1241
a31ad380
KO
1242 if (unlikely(atomic_read(&ctx->dead)))
1243 ret = -EINVAL;
1da177e4 1244
a31ad380
KO
1245 if (!*i)
1246 *i = ret;
1da177e4 1247
a31ad380 1248 return ret < 0 || *i >= min_nr;
1da177e4
LT
1249}
1250
a31ad380 1251static long read_events(struct kioctx *ctx, long min_nr, long nr,
1da177e4
LT
1252 struct io_event __user *event,
1253 struct timespec __user *timeout)
1254{
a31ad380
KO
1255 ktime_t until = { .tv64 = KTIME_MAX };
1256 long ret = 0;
1da177e4 1257
1da177e4
LT
1258 if (timeout) {
1259 struct timespec ts;
a31ad380 1260
1da177e4 1261 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
a31ad380 1262 return -EFAULT;
1da177e4 1263
a31ad380 1264 until = timespec_to_ktime(ts);
1da177e4
LT
1265 }
1266
a31ad380
KO
1267 /*
1268 * Note that aio_read_events() is being called as the conditional - i.e.
1269 * we're calling it after prepare_to_wait() has set task state to
1270 * TASK_INTERRUPTIBLE.
1271 *
1272 * But aio_read_events() can block, and if it blocks it's going to flip
1273 * the task state back to TASK_RUNNING.
1274 *
1275 * This should be ok, provided it doesn't flip the state back to
1276 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1277 * will only happen if the mutex_lock() call blocks, and we then find
1278 * the ringbuffer empty. So in practice we should be ok, but it's
1279 * something to be aware of when touching this code.
1280 */
5f785de5
FZ
1281 if (until.tv64 == 0)
1282 aio_read_events(ctx, min_nr, nr, event, &ret);
1283 else
1284 wait_event_interruptible_hrtimeout(ctx->wait,
1285 aio_read_events(ctx, min_nr, nr, event, &ret),
1286 until);
1da177e4 1287
a31ad380
KO
1288 if (!ret && signal_pending(current))
1289 ret = -EINTR;
1da177e4 1290
a31ad380 1291 return ret;
1da177e4
LT
1292}
1293
1da177e4
LT
1294/* sys_io_setup:
1295 * Create an aio_context capable of receiving at least nr_events.
1296 * ctxp must not point to an aio_context that already exists, and
1297 * must be initialized to 0 prior to the call. On successful
1298 * creation of the aio_context, *ctxp is filled in with the resulting
1299 * handle. May fail with -EINVAL if *ctxp is not initialized,
1300 * if the specified nr_events exceeds internal limits. May fail
1301 * with -EAGAIN if the specified nr_events exceeds the user's limit
1302 * of available events. May fail with -ENOMEM if insufficient kernel
1303 * resources are available. May fail with -EFAULT if an invalid
1304 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1305 * implemented.
1306 */
002c8976 1307SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1da177e4
LT
1308{
1309 struct kioctx *ioctx = NULL;
1310 unsigned long ctx;
1311 long ret;
1312
1313 ret = get_user(ctx, ctxp);
1314 if (unlikely(ret))
1315 goto out;
1316
1317 ret = -EINVAL;
d55b5fda 1318 if (unlikely(ctx || nr_events == 0)) {
acd88d4e 1319 pr_debug("EINVAL: ctx %lu nr_events %u\n",
d55b5fda 1320 ctx, nr_events);
1da177e4
LT
1321 goto out;
1322 }
1323
1324 ioctx = ioctx_alloc(nr_events);
1325 ret = PTR_ERR(ioctx);
1326 if (!IS_ERR(ioctx)) {
1327 ret = put_user(ioctx->user_id, ctxp);
a2e1859a 1328 if (ret)
e02ba72a 1329 kill_ioctx(current->mm, ioctx, NULL);
723be6e3 1330 percpu_ref_put(&ioctx->users);
1da177e4
LT
1331 }
1332
1333out:
1334 return ret;
1335}
1336
1337/* sys_io_destroy:
1338 * Destroy the aio_context specified. May cancel any outstanding
1339 * AIOs and block on completion. Will fail with -ENOSYS if not
642b5123 1340 * implemented. May fail with -EINVAL if the context pointed to
1da177e4
LT
1341 * is invalid.
1342 */
002c8976 1343SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1da177e4
LT
1344{
1345 struct kioctx *ioctx = lookup_ioctx(ctx);
1346 if (likely(NULL != ioctx)) {
dc48e56d 1347 struct ctx_rq_wait wait;
fb2d4483 1348 int ret;
e02ba72a 1349
dc48e56d
JA
1350 init_completion(&wait.comp);
1351 atomic_set(&wait.count, 1);
1352
e02ba72a
AP
1353 /* Pass requests_done to kill_ioctx() where it can be set
1354 * in a thread-safe way. If we try to set it here then we have
1355 * a race condition if two io_destroy() called simultaneously.
1356 */
dc48e56d 1357 ret = kill_ioctx(current->mm, ioctx, &wait);
723be6e3 1358 percpu_ref_put(&ioctx->users);
e02ba72a
AP
1359
1360 /* Wait until all IO for the context are done. Otherwise kernel
1361 * keep using user-space buffers even if user thinks the context
1362 * is destroyed.
1363 */
fb2d4483 1364 if (!ret)
dc48e56d 1365 wait_for_completion(&wait.comp);
e02ba72a 1366
fb2d4483 1367 return ret;
1da177e4 1368 }
acd88d4e 1369 pr_debug("EINVAL: invalid context id\n");
1da177e4
LT
1370 return -EINVAL;
1371}
1372
293bc982 1373typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
41ef4eb8 1374
a96114fa
AV
1375static int aio_setup_vectored_rw(int rw, char __user *buf, size_t len,
1376 struct iovec **iovec,
1377 bool compat,
1378 struct iov_iter *iter)
eed4e51f 1379{
9d85cba7
JM
1380#ifdef CONFIG_COMPAT
1381 if (compat)
32a56afa 1382 return compat_import_iovec(rw,
8bc92afc 1383 (struct compat_iovec __user *)buf,
32a56afa 1384 len, UIO_FASTIOV, iovec, iter);
9d85cba7 1385#endif
32a56afa
AV
1386 return import_iovec(rw, (struct iovec __user *)buf,
1387 len, UIO_FASTIOV, iovec, iter);
eed4e51f
BP
1388}
1389
1da177e4 1390/*
2be4e7de
GZ
1391 * aio_run_iocb:
1392 * Performs the initial checks and io submission.
1da177e4 1393 */
8bc92afc 1394static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
66ee59af 1395 char __user *buf, size_t len, bool compat)
1da177e4 1396{
41ef4eb8
KO
1397 struct file *file = req->ki_filp;
1398 ssize_t ret;
1399 int rw;
1400 fmode_t mode;
293bc982 1401 rw_iter_op *iter_op;
00fefb9c 1402 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
293bc982 1403 struct iov_iter iter;
1da177e4 1404
8bc92afc 1405 switch (opcode) {
1da177e4 1406 case IOCB_CMD_PREAD:
eed4e51f 1407 case IOCB_CMD_PREADV:
41ef4eb8
KO
1408 mode = FMODE_READ;
1409 rw = READ;
293bc982 1410 iter_op = file->f_op->read_iter;
41ef4eb8
KO
1411 goto rw_common;
1412
1413 case IOCB_CMD_PWRITE:
eed4e51f 1414 case IOCB_CMD_PWRITEV:
41ef4eb8
KO
1415 mode = FMODE_WRITE;
1416 rw = WRITE;
293bc982 1417 iter_op = file->f_op->write_iter;
41ef4eb8
KO
1418 goto rw_common;
1419rw_common:
1420 if (unlikely(!(file->f_mode & mode)))
1421 return -EBADF;
1422
84363182 1423 if (!iter_op)
41ef4eb8
KO
1424 return -EINVAL;
1425
66ee59af 1426 if (opcode == IOCB_CMD_PREADV || opcode == IOCB_CMD_PWRITEV)
a96114fa
AV
1427 ret = aio_setup_vectored_rw(rw, buf, len,
1428 &iovec, compat, &iter);
32a56afa 1429 else {
d4fb392f 1430 ret = import_single_range(rw, buf, len, iovec, &iter);
32a56afa
AV
1431 iovec = NULL;
1432 }
754320d6 1433 if (!ret)
a96114fa
AV
1434 ret = rw_verify_area(rw, file, &req->ki_pos,
1435 iov_iter_count(&iter));
8bc92afc 1436 if (ret < 0) {
32a56afa 1437 kfree(iovec);
41ef4eb8 1438 return ret;
8bc92afc 1439 }
41ef4eb8 1440
66ee59af 1441 len = ret;
41ef4eb8 1442
73a7075e
KO
1443 if (rw == WRITE)
1444 file_start_write(file);
1445
84363182 1446 ret = iter_op(req, &iter);
73a7075e
KO
1447
1448 if (rw == WRITE)
1449 file_end_write(file);
32a56afa 1450 kfree(iovec);
1da177e4 1451 break;
41ef4eb8 1452
1da177e4 1453 case IOCB_CMD_FDSYNC:
41ef4eb8
KO
1454 if (!file->f_op->aio_fsync)
1455 return -EINVAL;
1456
1457 ret = file->f_op->aio_fsync(req, 1);
1da177e4 1458 break;
41ef4eb8 1459
1da177e4 1460 case IOCB_CMD_FSYNC:
41ef4eb8
KO
1461 if (!file->f_op->aio_fsync)
1462 return -EINVAL;
1463
1464 ret = file->f_op->aio_fsync(req, 0);
1da177e4 1465 break;
41ef4eb8 1466
1da177e4 1467 default:
caf4167a 1468 pr_debug("EINVAL: no operation provided\n");
41ef4eb8 1469 return -EINVAL;
1da177e4
LT
1470 }
1471
41ef4eb8
KO
1472 if (ret != -EIOCBQUEUED) {
1473 /*
1474 * There's no easy way to restart the syscall since other AIO's
1475 * may be already running. Just fail this IO with EINTR.
1476 */
1477 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1478 ret == -ERESTARTNOHAND ||
1479 ret == -ERESTART_RESTARTBLOCK))
1480 ret = -EINTR;
1481 aio_complete(req, ret, 0);
1482 }
1da177e4
LT
1483
1484 return 0;
1485}
1486
d5470b59 1487static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
a1c8eae7 1488 struct iocb *iocb, bool compat)
1da177e4 1489{
04b2fa9f 1490 struct aio_kiocb *req;
1da177e4
LT
1491 ssize_t ret;
1492
1493 /* enforce forwards compatibility on users */
9c3060be 1494 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
caf4167a 1495 pr_debug("EINVAL: reserve field set\n");
1da177e4
LT
1496 return -EINVAL;
1497 }
1498
1499 /* prevent overflows */
1500 if (unlikely(
1501 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1502 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1503 ((ssize_t)iocb->aio_nbytes < 0)
1504 )) {
acd88d4e 1505 pr_debug("EINVAL: overflow check\n");
1da177e4
LT
1506 return -EINVAL;
1507 }
1508
41ef4eb8 1509 req = aio_get_req(ctx);
1d98ebfc 1510 if (unlikely(!req))
1da177e4 1511 return -EAGAIN;
1d98ebfc 1512
04b2fa9f
CH
1513 req->common.ki_filp = fget(iocb->aio_fildes);
1514 if (unlikely(!req->common.ki_filp)) {
1d98ebfc
KO
1515 ret = -EBADF;
1516 goto out_put_req;
1da177e4 1517 }
04b2fa9f
CH
1518 req->common.ki_pos = iocb->aio_offset;
1519 req->common.ki_complete = aio_complete;
2ba48ce5 1520 req->common.ki_flags = iocb_flags(req->common.ki_filp);
1d98ebfc 1521
9c3060be
DL
1522 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1523 /*
1524 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1525 * instance of the file* now. The file descriptor must be
1526 * an eventfd() fd, and will be signaled for each completed
1527 * event using the eventfd_signal() function.
1528 */
13389010 1529 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
801678c5 1530 if (IS_ERR(req->ki_eventfd)) {
9c3060be 1531 ret = PTR_ERR(req->ki_eventfd);
87c3a86e 1532 req->ki_eventfd = NULL;
9c3060be
DL
1533 goto out_put_req;
1534 }
04b2fa9f
CH
1535
1536 req->common.ki_flags |= IOCB_EVENTFD;
9c3060be 1537 }
1da177e4 1538
8a660890 1539 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1da177e4 1540 if (unlikely(ret)) {
caf4167a 1541 pr_debug("EFAULT: aio_key\n");
1da177e4
LT
1542 goto out_put_req;
1543 }
1544
04b2fa9f 1545 req->ki_user_iocb = user_iocb;
1da177e4 1546 req->ki_user_data = iocb->aio_data;
1da177e4 1547
04b2fa9f 1548 ret = aio_run_iocb(&req->common, iocb->aio_lio_opcode,
8bc92afc 1549 (char __user *)(unsigned long)iocb->aio_buf,
66ee59af 1550 iocb->aio_nbytes,
8bc92afc 1551 compat);
41003a7b 1552 if (ret)
7137c6bd 1553 goto out_put_req;
41003a7b 1554
1da177e4 1555 return 0;
1da177e4 1556out_put_req:
e1bdd5f2 1557 put_reqs_available(ctx, 1);
e34ecee2 1558 percpu_ref_put(&ctx->reqs);
57282d8f 1559 kiocb_free(req);
1da177e4
LT
1560 return ret;
1561}
1562
9d85cba7
JM
1563long do_io_submit(aio_context_t ctx_id, long nr,
1564 struct iocb __user *__user *iocbpp, bool compat)
1da177e4
LT
1565{
1566 struct kioctx *ctx;
1567 long ret = 0;
080d676d 1568 int i = 0;
9f5b9425 1569 struct blk_plug plug;
1da177e4
LT
1570
1571 if (unlikely(nr < 0))
1572 return -EINVAL;
1573
75e1c70f
JM
1574 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1575 nr = LONG_MAX/sizeof(*iocbpp);
1576
1da177e4
LT
1577 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1578 return -EFAULT;
1579
1580 ctx = lookup_ioctx(ctx_id);
1581 if (unlikely(!ctx)) {
caf4167a 1582 pr_debug("EINVAL: invalid context id\n");
1da177e4
LT
1583 return -EINVAL;
1584 }
1585
9f5b9425
SL
1586 blk_start_plug(&plug);
1587
1da177e4
LT
1588 /*
1589 * AKPM: should this return a partial result if some of the IOs were
1590 * successfully submitted?
1591 */
1592 for (i=0; i<nr; i++) {
1593 struct iocb __user *user_iocb;
1594 struct iocb tmp;
1595
1596 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1597 ret = -EFAULT;
1598 break;
1599 }
1600
1601 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1602 ret = -EFAULT;
1603 break;
1604 }
1605
a1c8eae7 1606 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1da177e4
LT
1607 if (ret)
1608 break;
1609 }
9f5b9425 1610 blk_finish_plug(&plug);
1da177e4 1611
723be6e3 1612 percpu_ref_put(&ctx->users);
1da177e4
LT
1613 return i ? i : ret;
1614}
1615
9d85cba7
JM
1616/* sys_io_submit:
1617 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1618 * the number of iocbs queued. May return -EINVAL if the aio_context
1619 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1620 * *iocbpp[0] is not properly initialized, if the operation specified
1621 * is invalid for the file descriptor in the iocb. May fail with
1622 * -EFAULT if any of the data structures point to invalid data. May
1623 * fail with -EBADF if the file descriptor specified in the first
1624 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1625 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1626 * fail with -ENOSYS if not implemented.
1627 */
1628SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1629 struct iocb __user * __user *, iocbpp)
1630{
1631 return do_io_submit(ctx_id, nr, iocbpp, 0);
1632}
1633
1da177e4
LT
1634/* lookup_kiocb
1635 * Finds a given iocb for cancellation.
1da177e4 1636 */
04b2fa9f
CH
1637static struct aio_kiocb *
1638lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key)
1da177e4 1639{
04b2fa9f 1640 struct aio_kiocb *kiocb;
d00689af
ZB
1641
1642 assert_spin_locked(&ctx->ctx_lock);
1643
8a660890
KO
1644 if (key != KIOCB_KEY)
1645 return NULL;
1646
1da177e4 1647 /* TODO: use a hash or array, this sucks. */
04b2fa9f
CH
1648 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
1649 if (kiocb->ki_user_iocb == iocb)
1da177e4
LT
1650 return kiocb;
1651 }
1652 return NULL;
1653}
1654
1655/* sys_io_cancel:
1656 * Attempts to cancel an iocb previously passed to io_submit. If
1657 * the operation is successfully cancelled, the resulting event is
1658 * copied into the memory pointed to by result without being placed
1659 * into the completion queue and 0 is returned. May fail with
1660 * -EFAULT if any of the data structures pointed to are invalid.
1661 * May fail with -EINVAL if aio_context specified by ctx_id is
1662 * invalid. May fail with -EAGAIN if the iocb specified was not
1663 * cancelled. Will fail with -ENOSYS if not implemented.
1664 */
002c8976
HC
1665SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1666 struct io_event __user *, result)
1da177e4 1667{
1da177e4 1668 struct kioctx *ctx;
04b2fa9f 1669 struct aio_kiocb *kiocb;
1da177e4
LT
1670 u32 key;
1671 int ret;
1672
1673 ret = get_user(key, &iocb->aio_key);
1674 if (unlikely(ret))
1675 return -EFAULT;
1676
1677 ctx = lookup_ioctx(ctx_id);
1678 if (unlikely(!ctx))
1679 return -EINVAL;
1680
1681 spin_lock_irq(&ctx->ctx_lock);
906b973c 1682
1da177e4 1683 kiocb = lookup_kiocb(ctx, iocb, key);
906b973c 1684 if (kiocb)
d52a8f9e 1685 ret = kiocb_cancel(kiocb);
906b973c
KO
1686 else
1687 ret = -EINVAL;
1688
1da177e4
LT
1689 spin_unlock_irq(&ctx->ctx_lock);
1690
906b973c 1691 if (!ret) {
bec68faa
KO
1692 /*
1693 * The result argument is no longer used - the io_event is
1694 * always delivered via the ring buffer. -EINPROGRESS indicates
1695 * cancellation is progress:
906b973c 1696 */
bec68faa 1697 ret = -EINPROGRESS;
906b973c 1698 }
1da177e4 1699
723be6e3 1700 percpu_ref_put(&ctx->users);
1da177e4
LT
1701
1702 return ret;
1703}
1704
1705/* io_getevents:
1706 * Attempts to read at least min_nr events and up to nr events from
642b5123
ST
1707 * the completion queue for the aio_context specified by ctx_id. If
1708 * it succeeds, the number of read events is returned. May fail with
1709 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1710 * out of range, if timeout is out of range. May fail with -EFAULT
1711 * if any of the memory specified is invalid. May return 0 or
1712 * < min_nr if the timeout specified by timeout has elapsed
1713 * before sufficient events are available, where timeout == NULL
1714 * specifies an infinite timeout. Note that the timeout pointed to by
6900807c 1715 * timeout is relative. Will fail with -ENOSYS if not implemented.
1da177e4 1716 */
002c8976
HC
1717SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1718 long, min_nr,
1719 long, nr,
1720 struct io_event __user *, events,
1721 struct timespec __user *, timeout)
1da177e4
LT
1722{
1723 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1724 long ret = -EINVAL;
1725
1726 if (likely(ioctx)) {
2e410255 1727 if (likely(min_nr <= nr && min_nr >= 0))
1da177e4 1728 ret = read_events(ioctx, min_nr, nr, events, timeout);
723be6e3 1729 percpu_ref_put(&ioctx->users);
1da177e4 1730 }
1da177e4
LT
1731 return ret;
1732}