]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blame - fs/aio.c
percpu-refcount: add @gfp to percpu_ref_init()
[mirror_ubuntu-artful-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
4e179bca 80struct kioctx {
723be6e3 81 struct percpu_ref users;
36f55889 82 atomic_t dead;
4e179bca 83
e34ecee2
KO
84 struct percpu_ref reqs;
85
4e179bca 86 unsigned long user_id;
4e179bca 87
e1bdd5f2
KO
88 struct __percpu kioctx_cpu *cpu;
89
90 /*
91 * For percpu reqs_available, number of slots we move to/from global
92 * counter at a time:
93 */
94 unsigned req_batch;
3e845ce0
KO
95 /*
96 * This is what userspace passed to io_setup(), it's not used for
97 * anything but counting against the global max_reqs quota.
98 *
58c85dc2 99 * The real limit is nr_events - 1, which will be larger (see
3e845ce0
KO
100 * aio_setup_ring())
101 */
4e179bca
KO
102 unsigned max_reqs;
103
58c85dc2
KO
104 /* Size of ringbuffer, in units of struct io_event */
105 unsigned nr_events;
4e179bca 106
58c85dc2
KO
107 unsigned long mmap_base;
108 unsigned long mmap_size;
109
110 struct page **ring_pages;
111 long nr_pages;
112
723be6e3 113 struct work_struct free_work;
4e23bcae 114
e02ba72a
AP
115 /*
116 * signals when all in-flight requests are done
117 */
118 struct completion *requests_done;
119
4e23bcae 120 struct {
34e83fc6
KO
121 /*
122 * This counts the number of available slots in the ringbuffer,
123 * so we avoid overflowing it: it's decremented (if positive)
124 * when allocating a kiocb and incremented when the resulting
125 * io_event is pulled off the ringbuffer.
e1bdd5f2
KO
126 *
127 * We batch accesses to it with a percpu version.
34e83fc6
KO
128 */
129 atomic_t reqs_available;
4e23bcae
KO
130 } ____cacheline_aligned_in_smp;
131
132 struct {
133 spinlock_t ctx_lock;
134 struct list_head active_reqs; /* used for cancellation */
135 } ____cacheline_aligned_in_smp;
136
58c85dc2
KO
137 struct {
138 struct mutex ring_lock;
4e23bcae
KO
139 wait_queue_head_t wait;
140 } ____cacheline_aligned_in_smp;
58c85dc2
KO
141
142 struct {
143 unsigned tail;
144 spinlock_t completion_lock;
4e23bcae 145 } ____cacheline_aligned_in_smp;
58c85dc2
KO
146
147 struct page *internal_pages[AIO_RING_PAGES];
36bc08cc 148 struct file *aio_ring_file;
db446a08
BL
149
150 unsigned id;
4e179bca
KO
151};
152
1da177e4 153/*------ sysctl variables----*/
d55b5fda
ZB
154static DEFINE_SPINLOCK(aio_nr_lock);
155unsigned long aio_nr; /* current system wide number of aio requests */
156unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
1da177e4
LT
157/*----end sysctl variables---*/
158
e18b890b
CL
159static struct kmem_cache *kiocb_cachep;
160static struct kmem_cache *kioctx_cachep;
1da177e4 161
71ad7490
BL
162static struct vfsmount *aio_mnt;
163
164static const struct file_operations aio_ring_fops;
165static const struct address_space_operations aio_ctx_aops;
166
167static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
168{
169 struct qstr this = QSTR_INIT("[aio]", 5);
170 struct file *file;
171 struct path path;
172 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
7f62656b
DC
173 if (IS_ERR(inode))
174 return ERR_CAST(inode);
71ad7490
BL
175
176 inode->i_mapping->a_ops = &aio_ctx_aops;
177 inode->i_mapping->private_data = ctx;
178 inode->i_size = PAGE_SIZE * nr_pages;
179
180 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
181 if (!path.dentry) {
182 iput(inode);
183 return ERR_PTR(-ENOMEM);
184 }
185 path.mnt = mntget(aio_mnt);
186
187 d_instantiate(path.dentry, inode);
188 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
189 if (IS_ERR(file)) {
190 path_put(&path);
191 return file;
192 }
193
194 file->f_flags = O_RDWR;
195 file->private_data = ctx;
196 return file;
197}
198
199static struct dentry *aio_mount(struct file_system_type *fs_type,
200 int flags, const char *dev_name, void *data)
201{
202 static const struct dentry_operations ops = {
203 .d_dname = simple_dname,
204 };
205 return mount_pseudo(fs_type, "aio:", NULL, &ops, 0xa10a10a1);
206}
207
1da177e4
LT
208/* aio_setup
209 * Creates the slab caches used by the aio routines, panic on
210 * failure as this is done early during the boot sequence.
211 */
212static int __init aio_setup(void)
213{
71ad7490
BL
214 static struct file_system_type aio_fs = {
215 .name = "aio",
216 .mount = aio_mount,
217 .kill_sb = kill_anon_super,
218 };
219 aio_mnt = kern_mount(&aio_fs);
220 if (IS_ERR(aio_mnt))
221 panic("Failed to create aio fs mount.");
222
0a31bd5f
CL
223 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
224 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
1da177e4 225
caf4167a 226 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
1da177e4
LT
227
228 return 0;
229}
385773e0 230__initcall(aio_setup);
1da177e4 231
5e9ae2e5
BL
232static void put_aio_ring_file(struct kioctx *ctx)
233{
234 struct file *aio_ring_file = ctx->aio_ring_file;
235 if (aio_ring_file) {
236 truncate_setsize(aio_ring_file->f_inode, 0);
237
238 /* Prevent further access to the kioctx from migratepages */
239 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
240 aio_ring_file->f_inode->i_mapping->private_data = NULL;
241 ctx->aio_ring_file = NULL;
242 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
243
244 fput(aio_ring_file);
245 }
246}
247
1da177e4
LT
248static void aio_free_ring(struct kioctx *ctx)
249{
36bc08cc 250 int i;
1da177e4 251
fa8a53c3
BL
252 /* Disconnect the kiotx from the ring file. This prevents future
253 * accesses to the kioctx from page migration.
254 */
255 put_aio_ring_file(ctx);
256
36bc08cc 257 for (i = 0; i < ctx->nr_pages; i++) {
8e321fef 258 struct page *page;
36bc08cc
GZ
259 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
260 page_count(ctx->ring_pages[i]));
8e321fef
BL
261 page = ctx->ring_pages[i];
262 if (!page)
263 continue;
264 ctx->ring_pages[i] = NULL;
265 put_page(page);
36bc08cc 266 }
1da177e4 267
ddb8c45b 268 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
58c85dc2 269 kfree(ctx->ring_pages);
ddb8c45b
SL
270 ctx->ring_pages = NULL;
271 }
36bc08cc
GZ
272}
273
274static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
275{
276 vma->vm_ops = &generic_file_vm_ops;
277 return 0;
278}
279
280static const struct file_operations aio_ring_fops = {
281 .mmap = aio_ring_mmap,
282};
283
284static int aio_set_page_dirty(struct page *page)
285{
286 return 0;
287}
288
0c45355f 289#if IS_ENABLED(CONFIG_MIGRATION)
36bc08cc
GZ
290static int aio_migratepage(struct address_space *mapping, struct page *new,
291 struct page *old, enum migrate_mode mode)
292{
5e9ae2e5 293 struct kioctx *ctx;
36bc08cc 294 unsigned long flags;
fa8a53c3 295 pgoff_t idx;
36bc08cc
GZ
296 int rc;
297
8e321fef
BL
298 rc = 0;
299
fa8a53c3 300 /* mapping->private_lock here protects against the kioctx teardown. */
8e321fef
BL
301 spin_lock(&mapping->private_lock);
302 ctx = mapping->private_data;
fa8a53c3
BL
303 if (!ctx) {
304 rc = -EINVAL;
305 goto out;
306 }
307
308 /* The ring_lock mutex. The prevents aio_read_events() from writing
309 * to the ring's head, and prevents page migration from mucking in
310 * a partially initialized kiotx.
311 */
312 if (!mutex_trylock(&ctx->ring_lock)) {
313 rc = -EAGAIN;
314 goto out;
315 }
316
317 idx = old->index;
318 if (idx < (pgoff_t)ctx->nr_pages) {
319 /* Make sure the old page hasn't already been changed */
320 if (ctx->ring_pages[idx] != old)
321 rc = -EAGAIN;
8e321fef
BL
322 } else
323 rc = -EINVAL;
8e321fef
BL
324
325 if (rc != 0)
fa8a53c3 326 goto out_unlock;
8e321fef 327
36bc08cc
GZ
328 /* Writeback must be complete */
329 BUG_ON(PageWriteback(old));
8e321fef 330 get_page(new);
36bc08cc 331
8e321fef 332 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
36bc08cc 333 if (rc != MIGRATEPAGE_SUCCESS) {
8e321fef 334 put_page(new);
fa8a53c3 335 goto out_unlock;
36bc08cc
GZ
336 }
337
fa8a53c3
BL
338 /* Take completion_lock to prevent other writes to the ring buffer
339 * while the old page is copied to the new. This prevents new
340 * events from being lost.
5e9ae2e5 341 */
fa8a53c3
BL
342 spin_lock_irqsave(&ctx->completion_lock, flags);
343 migrate_page_copy(new, old);
344 BUG_ON(ctx->ring_pages[idx] != old);
345 ctx->ring_pages[idx] = new;
346 spin_unlock_irqrestore(&ctx->completion_lock, flags);
36bc08cc 347
fa8a53c3
BL
348 /* The old page is no longer accessible. */
349 put_page(old);
8e321fef 350
fa8a53c3
BL
351out_unlock:
352 mutex_unlock(&ctx->ring_lock);
353out:
354 spin_unlock(&mapping->private_lock);
36bc08cc 355 return rc;
1da177e4 356}
0c45355f 357#endif
1da177e4 358
36bc08cc
GZ
359static const struct address_space_operations aio_ctx_aops = {
360 .set_page_dirty = aio_set_page_dirty,
0c45355f 361#if IS_ENABLED(CONFIG_MIGRATION)
36bc08cc 362 .migratepage = aio_migratepage,
0c45355f 363#endif
36bc08cc
GZ
364};
365
1da177e4
LT
366static int aio_setup_ring(struct kioctx *ctx)
367{
368 struct aio_ring *ring;
1da177e4 369 unsigned nr_events = ctx->max_reqs;
41003a7b 370 struct mm_struct *mm = current->mm;
3dc9acb6 371 unsigned long size, unused;
1da177e4 372 int nr_pages;
36bc08cc
GZ
373 int i;
374 struct file *file;
1da177e4
LT
375
376 /* Compensate for the ring buffer's head/tail overlap entry */
377 nr_events += 2; /* 1 is required, 2 for good luck */
378
379 size = sizeof(struct aio_ring);
380 size += sizeof(struct io_event) * nr_events;
1da177e4 381
36bc08cc 382 nr_pages = PFN_UP(size);
1da177e4
LT
383 if (nr_pages < 0)
384 return -EINVAL;
385
71ad7490 386 file = aio_private_file(ctx, nr_pages);
36bc08cc
GZ
387 if (IS_ERR(file)) {
388 ctx->aio_ring_file = NULL;
fa8a53c3 389 return -ENOMEM;
36bc08cc
GZ
390 }
391
3dc9acb6
LT
392 ctx->aio_ring_file = file;
393 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
394 / sizeof(struct io_event);
395
396 ctx->ring_pages = ctx->internal_pages;
397 if (nr_pages > AIO_RING_PAGES) {
398 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
399 GFP_KERNEL);
400 if (!ctx->ring_pages) {
401 put_aio_ring_file(ctx);
402 return -ENOMEM;
403 }
404 }
405
36bc08cc
GZ
406 for (i = 0; i < nr_pages; i++) {
407 struct page *page;
408 page = find_or_create_page(file->f_inode->i_mapping,
409 i, GFP_HIGHUSER | __GFP_ZERO);
410 if (!page)
411 break;
412 pr_debug("pid(%d) page[%d]->count=%d\n",
413 current->pid, i, page_count(page));
414 SetPageUptodate(page);
415 SetPageDirty(page);
416 unlock_page(page);
3dc9acb6
LT
417
418 ctx->ring_pages[i] = page;
36bc08cc 419 }
3dc9acb6 420 ctx->nr_pages = i;
1da177e4 421
3dc9acb6
LT
422 if (unlikely(i != nr_pages)) {
423 aio_free_ring(ctx);
fa8a53c3 424 return -ENOMEM;
1da177e4
LT
425 }
426
58c85dc2
KO
427 ctx->mmap_size = nr_pages * PAGE_SIZE;
428 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
36bc08cc 429
41003a7b 430 down_write(&mm->mmap_sem);
36bc08cc
GZ
431 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
432 PROT_READ | PROT_WRITE,
3dc9acb6
LT
433 MAP_SHARED, 0, &unused);
434 up_write(&mm->mmap_sem);
58c85dc2 435 if (IS_ERR((void *)ctx->mmap_base)) {
58c85dc2 436 ctx->mmap_size = 0;
1da177e4 437 aio_free_ring(ctx);
fa8a53c3 438 return -ENOMEM;
1da177e4
LT
439 }
440
58c85dc2 441 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
d6c355c7 442
58c85dc2
KO
443 ctx->user_id = ctx->mmap_base;
444 ctx->nr_events = nr_events; /* trusted copy */
1da177e4 445
58c85dc2 446 ring = kmap_atomic(ctx->ring_pages[0]);
1da177e4 447 ring->nr = nr_events; /* user copy */
db446a08 448 ring->id = ~0U;
1da177e4
LT
449 ring->head = ring->tail = 0;
450 ring->magic = AIO_RING_MAGIC;
451 ring->compat_features = AIO_RING_COMPAT_FEATURES;
452 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
453 ring->header_length = sizeof(struct aio_ring);
e8e3c3d6 454 kunmap_atomic(ring);
58c85dc2 455 flush_dcache_page(ctx->ring_pages[0]);
1da177e4
LT
456
457 return 0;
458}
459
1da177e4
LT
460#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
461#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
462#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
463
0460fef2
KO
464void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
465{
466 struct kioctx *ctx = req->ki_ctx;
467 unsigned long flags;
468
469 spin_lock_irqsave(&ctx->ctx_lock, flags);
470
471 if (!req->ki_list.next)
472 list_add(&req->ki_list, &ctx->active_reqs);
473
474 req->ki_cancel = cancel;
475
476 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
477}
478EXPORT_SYMBOL(kiocb_set_cancel_fn);
479
d52a8f9e 480static int kiocb_cancel(struct kiocb *kiocb)
906b973c 481{
0460fef2 482 kiocb_cancel_fn *old, *cancel;
906b973c 483
0460fef2
KO
484 /*
485 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
486 * actually has a cancel function, hence the cmpxchg()
487 */
488
489 cancel = ACCESS_ONCE(kiocb->ki_cancel);
490 do {
491 if (!cancel || cancel == KIOCB_CANCELLED)
57282d8f 492 return -EINVAL;
906b973c 493
0460fef2
KO
494 old = cancel;
495 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
496 } while (cancel != old);
906b973c 497
57282d8f 498 return cancel(kiocb);
906b973c
KO
499}
500
e34ecee2 501static void free_ioctx(struct work_struct *work)
36f55889 502{
e34ecee2 503 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
e1bdd5f2 504
e34ecee2 505 pr_debug("freeing %p\n", ctx);
e1bdd5f2 506
e34ecee2 507 aio_free_ring(ctx);
e1bdd5f2 508 free_percpu(ctx->cpu);
9a1049da
TH
509 percpu_ref_exit(&ctx->reqs);
510 percpu_ref_exit(&ctx->users);
36f55889
KO
511 kmem_cache_free(kioctx_cachep, ctx);
512}
513
e34ecee2
KO
514static void free_ioctx_reqs(struct percpu_ref *ref)
515{
516 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
517
e02ba72a
AP
518 /* At this point we know that there are no any in-flight requests */
519 if (ctx->requests_done)
520 complete(ctx->requests_done);
521
e34ecee2
KO
522 INIT_WORK(&ctx->free_work, free_ioctx);
523 schedule_work(&ctx->free_work);
524}
525
36f55889
KO
526/*
527 * When this function runs, the kioctx has been removed from the "hash table"
528 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
529 * now it's safe to cancel any that need to be.
530 */
e34ecee2 531static void free_ioctx_users(struct percpu_ref *ref)
36f55889 532{
e34ecee2 533 struct kioctx *ctx = container_of(ref, struct kioctx, users);
36f55889
KO
534 struct kiocb *req;
535
536 spin_lock_irq(&ctx->ctx_lock);
537
538 while (!list_empty(&ctx->active_reqs)) {
539 req = list_first_entry(&ctx->active_reqs,
540 struct kiocb, ki_list);
541
542 list_del_init(&req->ki_list);
d52a8f9e 543 kiocb_cancel(req);
36f55889
KO
544 }
545
546 spin_unlock_irq(&ctx->ctx_lock);
547
e34ecee2
KO
548 percpu_ref_kill(&ctx->reqs);
549 percpu_ref_put(&ctx->reqs);
36f55889
KO
550}
551
db446a08
BL
552static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
553{
554 unsigned i, new_nr;
555 struct kioctx_table *table, *old;
556 struct aio_ring *ring;
557
558 spin_lock(&mm->ioctx_lock);
d9b2c871 559 rcu_read_lock();
77d30b14 560 table = rcu_dereference(mm->ioctx_table);
db446a08
BL
561
562 while (1) {
563 if (table)
564 for (i = 0; i < table->nr; i++)
565 if (!table->table[i]) {
566 ctx->id = i;
567 table->table[i] = ctx;
d9b2c871 568 rcu_read_unlock();
db446a08
BL
569 spin_unlock(&mm->ioctx_lock);
570
fa8a53c3
BL
571 /* While kioctx setup is in progress,
572 * we are protected from page migration
573 * changes ring_pages by ->ring_lock.
574 */
db446a08
BL
575 ring = kmap_atomic(ctx->ring_pages[0]);
576 ring->id = ctx->id;
577 kunmap_atomic(ring);
578 return 0;
579 }
580
581 new_nr = (table ? table->nr : 1) * 4;
582
d9b2c871 583 rcu_read_unlock();
db446a08
BL
584 spin_unlock(&mm->ioctx_lock);
585
586 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
587 new_nr, GFP_KERNEL);
588 if (!table)
589 return -ENOMEM;
590
591 table->nr = new_nr;
592
593 spin_lock(&mm->ioctx_lock);
d9b2c871 594 rcu_read_lock();
77d30b14 595 old = rcu_dereference(mm->ioctx_table);
db446a08
BL
596
597 if (!old) {
598 rcu_assign_pointer(mm->ioctx_table, table);
599 } else if (table->nr > old->nr) {
600 memcpy(table->table, old->table,
601 old->nr * sizeof(struct kioctx *));
602
603 rcu_assign_pointer(mm->ioctx_table, table);
604 kfree_rcu(old, rcu);
605 } else {
606 kfree(table);
607 table = old;
608 }
609 }
610}
611
e34ecee2
KO
612static void aio_nr_sub(unsigned nr)
613{
614 spin_lock(&aio_nr_lock);
615 if (WARN_ON(aio_nr - nr > aio_nr))
616 aio_nr = 0;
617 else
618 aio_nr -= nr;
619 spin_unlock(&aio_nr_lock);
620}
621
1da177e4
LT
622/* ioctx_alloc
623 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
624 */
625static struct kioctx *ioctx_alloc(unsigned nr_events)
626{
41003a7b 627 struct mm_struct *mm = current->mm;
1da177e4 628 struct kioctx *ctx;
e23754f8 629 int err = -ENOMEM;
1da177e4 630
e1bdd5f2
KO
631 /*
632 * We keep track of the number of available ringbuffer slots, to prevent
633 * overflow (reqs_available), and we also use percpu counters for this.
634 *
635 * So since up to half the slots might be on other cpu's percpu counters
636 * and unavailable, double nr_events so userspace sees what they
637 * expected: additionally, we move req_batch slots to/from percpu
638 * counters at a time, so make sure that isn't 0:
639 */
640 nr_events = max(nr_events, num_possible_cpus() * 4);
641 nr_events *= 2;
642
1da177e4
LT
643 /* Prevent overflows */
644 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
645 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
646 pr_debug("ENOMEM: nr_events too high\n");
647 return ERR_PTR(-EINVAL);
648 }
649
4cd81c3d 650 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
1da177e4
LT
651 return ERR_PTR(-EAGAIN);
652
c3762229 653 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
1da177e4
LT
654 if (!ctx)
655 return ERR_PTR(-ENOMEM);
656
1da177e4 657 ctx->max_reqs = nr_events;
1da177e4 658
1da177e4 659 spin_lock_init(&ctx->ctx_lock);
0460fef2 660 spin_lock_init(&ctx->completion_lock);
58c85dc2 661 mutex_init(&ctx->ring_lock);
fa8a53c3
BL
662 /* Protect against page migration throughout kiotx setup by keeping
663 * the ring_lock mutex held until setup is complete. */
664 mutex_lock(&ctx->ring_lock);
1da177e4
LT
665 init_waitqueue_head(&ctx->wait);
666
667 INIT_LIST_HEAD(&ctx->active_reqs);
1da177e4 668
a34375ef 669 if (percpu_ref_init(&ctx->users, free_ioctx_users, GFP_KERNEL))
fa8a53c3
BL
670 goto err;
671
a34375ef 672 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, GFP_KERNEL))
fa8a53c3
BL
673 goto err;
674
e1bdd5f2
KO
675 ctx->cpu = alloc_percpu(struct kioctx_cpu);
676 if (!ctx->cpu)
e34ecee2 677 goto err;
1da177e4 678
fa8a53c3
BL
679 err = aio_setup_ring(ctx);
680 if (err < 0)
e34ecee2 681 goto err;
e1bdd5f2 682
34e83fc6 683 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
e1bdd5f2 684 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
6878ea72
BL
685 if (ctx->req_batch < 1)
686 ctx->req_batch = 1;
34e83fc6 687
1da177e4 688 /* limit the number of system wide aios */
9fa1cb39 689 spin_lock(&aio_nr_lock);
4cd81c3d 690 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
2dd542b7 691 aio_nr + nr_events < aio_nr) {
9fa1cb39 692 spin_unlock(&aio_nr_lock);
e34ecee2 693 err = -EAGAIN;
d1b94327 694 goto err_ctx;
2dd542b7
AV
695 }
696 aio_nr += ctx->max_reqs;
9fa1cb39 697 spin_unlock(&aio_nr_lock);
1da177e4 698
1881686f
BL
699 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
700 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
723be6e3 701
da90382c
BL
702 err = ioctx_add_table(ctx, mm);
703 if (err)
e34ecee2 704 goto err_cleanup;
da90382c 705
fa8a53c3
BL
706 /* Release the ring_lock mutex now that all setup is complete. */
707 mutex_unlock(&ctx->ring_lock);
708
caf4167a 709 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
58c85dc2 710 ctx, ctx->user_id, mm, ctx->nr_events);
1da177e4
LT
711 return ctx;
712
e34ecee2
KO
713err_cleanup:
714 aio_nr_sub(ctx->max_reqs);
d1b94327
GZ
715err_ctx:
716 aio_free_ring(ctx);
e34ecee2 717err:
fa8a53c3 718 mutex_unlock(&ctx->ring_lock);
e1bdd5f2 719 free_percpu(ctx->cpu);
9a1049da
TH
720 percpu_ref_exit(&ctx->reqs);
721 percpu_ref_exit(&ctx->users);
1da177e4 722 kmem_cache_free(kioctx_cachep, ctx);
caf4167a 723 pr_debug("error allocating ioctx %d\n", err);
e23754f8 724 return ERR_PTR(err);
1da177e4
LT
725}
726
36f55889
KO
727/* kill_ioctx
728 * Cancels all outstanding aio requests on an aio context. Used
729 * when the processes owning a context have all exited to encourage
730 * the rapid destruction of the kioctx.
731 */
fb2d4483 732static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
e02ba72a 733 struct completion *requests_done)
36f55889 734{
fa88b6f8 735 struct kioctx_table *table;
db446a08 736
fa88b6f8
BL
737 if (atomic_xchg(&ctx->dead, 1))
738 return -EINVAL;
db446a08 739
db446a08 740
fa88b6f8
BL
741 spin_lock(&mm->ioctx_lock);
742 rcu_read_lock();
743 table = rcu_dereference(mm->ioctx_table);
dee11c23 744
fa88b6f8
BL
745 WARN_ON(ctx != table->table[ctx->id]);
746 table->table[ctx->id] = NULL;
747 rcu_read_unlock();
748 spin_unlock(&mm->ioctx_lock);
4fcc712f 749
fa88b6f8
BL
750 /* percpu_ref_kill() will do the necessary call_rcu() */
751 wake_up_all(&ctx->wait);
4fcc712f 752
fa88b6f8
BL
753 /*
754 * It'd be more correct to do this in free_ioctx(), after all
755 * the outstanding kiocbs have finished - but by then io_destroy
756 * has already returned, so io_setup() could potentially return
757 * -EAGAIN with no ioctxs actually in use (as far as userspace
758 * could tell).
759 */
760 aio_nr_sub(ctx->max_reqs);
4fcc712f 761
fa88b6f8
BL
762 if (ctx->mmap_size)
763 vm_munmap(ctx->mmap_base, ctx->mmap_size);
fb2d4483 764
fa88b6f8
BL
765 ctx->requests_done = requests_done;
766 percpu_ref_kill(&ctx->users);
767 return 0;
1da177e4
LT
768}
769
770/* wait_on_sync_kiocb:
771 * Waits on the given sync kiocb to complete.
772 */
57282d8f 773ssize_t wait_on_sync_kiocb(struct kiocb *req)
1da177e4 774{
57282d8f 775 while (!req->ki_ctx) {
1da177e4 776 set_current_state(TASK_UNINTERRUPTIBLE);
57282d8f 777 if (req->ki_ctx)
1da177e4 778 break;
41d10da3 779 io_schedule();
1da177e4
LT
780 }
781 __set_current_state(TASK_RUNNING);
57282d8f 782 return req->ki_user_data;
1da177e4 783}
385773e0 784EXPORT_SYMBOL(wait_on_sync_kiocb);
1da177e4 785
36f55889
KO
786/*
787 * exit_aio: called when the last user of mm goes away. At this point, there is
788 * no way for any new requests to be submited or any of the io_* syscalls to be
789 * called on the context.
790 *
791 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
792 * them.
1da177e4 793 */
fc9b52cd 794void exit_aio(struct mm_struct *mm)
1da177e4 795{
db446a08 796 struct kioctx_table *table;
abf137dd 797 struct kioctx *ctx;
db446a08
BL
798 unsigned i = 0;
799
800 while (1) {
801 rcu_read_lock();
802 table = rcu_dereference(mm->ioctx_table);
803
804 do {
805 if (!table || i >= table->nr) {
806 rcu_read_unlock();
807 rcu_assign_pointer(mm->ioctx_table, NULL);
808 if (table)
809 kfree(table);
810 return;
811 }
812
813 ctx = table->table[i++];
814 } while (!ctx);
815
816 rcu_read_unlock();
abf137dd 817
936af157
AV
818 /*
819 * We don't need to bother with munmap() here -
820 * exit_mmap(mm) is coming and it'll unmap everything.
821 * Since aio_free_ring() uses non-zero ->mmap_size
822 * as indicator that it needs to unmap the area,
823 * just set it to 0; aio_free_ring() is the only
824 * place that uses ->mmap_size, so it's safe.
936af157 825 */
58c85dc2 826 ctx->mmap_size = 0;
36f55889 827
e02ba72a 828 kill_ioctx(mm, ctx, NULL);
1da177e4
LT
829 }
830}
831
e1bdd5f2
KO
832static void put_reqs_available(struct kioctx *ctx, unsigned nr)
833{
834 struct kioctx_cpu *kcpu;
263782c1 835 unsigned long flags;
e1bdd5f2
KO
836
837 preempt_disable();
838 kcpu = this_cpu_ptr(ctx->cpu);
839
263782c1 840 local_irq_save(flags);
e1bdd5f2 841 kcpu->reqs_available += nr;
263782c1 842
e1bdd5f2
KO
843 while (kcpu->reqs_available >= ctx->req_batch * 2) {
844 kcpu->reqs_available -= ctx->req_batch;
845 atomic_add(ctx->req_batch, &ctx->reqs_available);
846 }
847
263782c1 848 local_irq_restore(flags);
e1bdd5f2
KO
849 preempt_enable();
850}
851
852static bool get_reqs_available(struct kioctx *ctx)
853{
854 struct kioctx_cpu *kcpu;
855 bool ret = false;
263782c1 856 unsigned long flags;
e1bdd5f2
KO
857
858 preempt_disable();
859 kcpu = this_cpu_ptr(ctx->cpu);
860
263782c1 861 local_irq_save(flags);
e1bdd5f2
KO
862 if (!kcpu->reqs_available) {
863 int old, avail = atomic_read(&ctx->reqs_available);
864
865 do {
866 if (avail < ctx->req_batch)
867 goto out;
868
869 old = avail;
870 avail = atomic_cmpxchg(&ctx->reqs_available,
871 avail, avail - ctx->req_batch);
872 } while (avail != old);
873
874 kcpu->reqs_available += ctx->req_batch;
875 }
876
877 ret = true;
878 kcpu->reqs_available--;
879out:
263782c1 880 local_irq_restore(flags);
e1bdd5f2
KO
881 preempt_enable();
882 return ret;
883}
884
1da177e4 885/* aio_get_req
57282d8f
KO
886 * Allocate a slot for an aio request.
887 * Returns NULL if no requests are free.
1da177e4 888 */
a1c8eae7 889static inline struct kiocb *aio_get_req(struct kioctx *ctx)
1da177e4 890{
a1c8eae7
KO
891 struct kiocb *req;
892
e1bdd5f2 893 if (!get_reqs_available(ctx))
a1c8eae7
KO
894 return NULL;
895
0460fef2 896 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
1da177e4 897 if (unlikely(!req))
a1c8eae7 898 goto out_put;
1da177e4 899
e34ecee2
KO
900 percpu_ref_get(&ctx->reqs);
901
1da177e4 902 req->ki_ctx = ctx;
080d676d 903 return req;
a1c8eae7 904out_put:
e1bdd5f2 905 put_reqs_available(ctx, 1);
a1c8eae7 906 return NULL;
1da177e4
LT
907}
908
11599eba 909static void kiocb_free(struct kiocb *req)
1da177e4 910{
1d98ebfc
KO
911 if (req->ki_filp)
912 fput(req->ki_filp);
13389010
DL
913 if (req->ki_eventfd != NULL)
914 eventfd_ctx_put(req->ki_eventfd);
1da177e4 915 kmem_cache_free(kiocb_cachep, req);
1da177e4
LT
916}
917
d5470b59 918static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1da177e4 919{
db446a08 920 struct aio_ring __user *ring = (void __user *)ctx_id;
abf137dd 921 struct mm_struct *mm = current->mm;
65c24491 922 struct kioctx *ctx, *ret = NULL;
db446a08
BL
923 struct kioctx_table *table;
924 unsigned id;
925
926 if (get_user(id, &ring->id))
927 return NULL;
1da177e4 928
abf137dd 929 rcu_read_lock();
db446a08 930 table = rcu_dereference(mm->ioctx_table);
abf137dd 931
db446a08
BL
932 if (!table || id >= table->nr)
933 goto out;
1da177e4 934
db446a08 935 ctx = table->table[id];
f30d704f 936 if (ctx && ctx->user_id == ctx_id) {
db446a08
BL
937 percpu_ref_get(&ctx->users);
938 ret = ctx;
939 }
940out:
abf137dd 941 rcu_read_unlock();
65c24491 942 return ret;
1da177e4
LT
943}
944
1da177e4
LT
945/* aio_complete
946 * Called when the io request on the given iocb is complete.
1da177e4 947 */
2d68449e 948void aio_complete(struct kiocb *iocb, long res, long res2)
1da177e4
LT
949{
950 struct kioctx *ctx = iocb->ki_ctx;
1da177e4 951 struct aio_ring *ring;
21b40200 952 struct io_event *ev_page, *event;
1da177e4 953 unsigned long flags;
21b40200 954 unsigned tail, pos;
1da177e4 955
20dcae32
ZB
956 /*
957 * Special case handling for sync iocbs:
958 * - events go directly into the iocb for fast handling
959 * - the sync task with the iocb in its stack holds the single iocb
960 * ref, no other paths have a way to get another ref
961 * - the sync task helpfully left a reference to itself in the iocb
1da177e4
LT
962 */
963 if (is_sync_kiocb(iocb)) {
1da177e4 964 iocb->ki_user_data = res;
57282d8f
KO
965 smp_wmb();
966 iocb->ki_ctx = ERR_PTR(-EXDEV);
1da177e4 967 wake_up_process(iocb->ki_obj.tsk);
2d68449e 968 return;
1da177e4
LT
969 }
970
0460fef2
KO
971 if (iocb->ki_list.next) {
972 unsigned long flags;
973
974 spin_lock_irqsave(&ctx->ctx_lock, flags);
975 list_del(&iocb->ki_list);
976 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
977 }
11599eba 978
0460fef2
KO
979 /*
980 * Add a completion event to the ring buffer. Must be done holding
4b30f07e 981 * ctx->completion_lock to prevent other code from messing with the tail
0460fef2
KO
982 * pointer since we might be called from irq context.
983 */
984 spin_lock_irqsave(&ctx->completion_lock, flags);
985
58c85dc2 986 tail = ctx->tail;
21b40200
KO
987 pos = tail + AIO_EVENTS_OFFSET;
988
58c85dc2 989 if (++tail >= ctx->nr_events)
4bf69b2a 990 tail = 0;
1da177e4 991
58c85dc2 992 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
21b40200
KO
993 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
994
1da177e4
LT
995 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
996 event->data = iocb->ki_user_data;
997 event->res = res;
998 event->res2 = res2;
999
21b40200 1000 kunmap_atomic(ev_page);
58c85dc2 1001 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
21b40200
KO
1002
1003 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
caf4167a
KO
1004 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
1005 res, res2);
1da177e4
LT
1006
1007 /* after flagging the request as done, we
1008 * must never even look at it again
1009 */
1010 smp_wmb(); /* make event visible before updating tail */
1011
58c85dc2 1012 ctx->tail = tail;
1da177e4 1013
58c85dc2 1014 ring = kmap_atomic(ctx->ring_pages[0]);
21b40200 1015 ring->tail = tail;
e8e3c3d6 1016 kunmap_atomic(ring);
58c85dc2 1017 flush_dcache_page(ctx->ring_pages[0]);
1da177e4 1018
0460fef2
KO
1019 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1020
21b40200 1021 pr_debug("added to ring %p at [%u]\n", iocb, tail);
8d1c98b0
DL
1022
1023 /*
1024 * Check if the user asked us to deliver the result through an
1025 * eventfd. The eventfd_signal() function is safe to be called
1026 * from IRQ context.
1027 */
87c3a86e 1028 if (iocb->ki_eventfd != NULL)
8d1c98b0
DL
1029 eventfd_signal(iocb->ki_eventfd, 1);
1030
1da177e4 1031 /* everything turned out well, dispose of the aiocb. */
57282d8f 1032 kiocb_free(iocb);
f8567a38 1033 put_reqs_available(ctx, 1);
1da177e4 1034
6cb2a210
QB
1035 /*
1036 * We have to order our ring_info tail store above and test
1037 * of the wait list below outside the wait lock. This is
1038 * like in wake_up_bit() where clearing a bit has to be
1039 * ordered with the unlocked test.
1040 */
1041 smp_mb();
1042
1da177e4
LT
1043 if (waitqueue_active(&ctx->wait))
1044 wake_up(&ctx->wait);
1045
e34ecee2 1046 percpu_ref_put(&ctx->reqs);
1da177e4 1047}
385773e0 1048EXPORT_SYMBOL(aio_complete);
1da177e4 1049
a31ad380
KO
1050/* aio_read_events
1051 * Pull an event off of the ioctx's event ring. Returns the number of
1052 * events fetched
1da177e4 1053 */
a31ad380
KO
1054static long aio_read_events_ring(struct kioctx *ctx,
1055 struct io_event __user *event, long nr)
1da177e4 1056{
1da177e4 1057 struct aio_ring *ring;
5ffac122 1058 unsigned head, tail, pos;
a31ad380
KO
1059 long ret = 0;
1060 int copy_ret;
1061
58c85dc2 1062 mutex_lock(&ctx->ring_lock);
1da177e4 1063
fa8a53c3 1064 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
58c85dc2 1065 ring = kmap_atomic(ctx->ring_pages[0]);
a31ad380 1066 head = ring->head;
5ffac122 1067 tail = ring->tail;
a31ad380
KO
1068 kunmap_atomic(ring);
1069
5ffac122 1070 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1da177e4 1071
5ffac122 1072 if (head == tail)
1da177e4
LT
1073 goto out;
1074
edfbbf38
BL
1075 head %= ctx->nr_events;
1076 tail %= ctx->nr_events;
1077
a31ad380
KO
1078 while (ret < nr) {
1079 long avail;
1080 struct io_event *ev;
1081 struct page *page;
1082
5ffac122
KO
1083 avail = (head <= tail ? tail : ctx->nr_events) - head;
1084 if (head == tail)
a31ad380
KO
1085 break;
1086
1087 avail = min(avail, nr - ret);
1088 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1089 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1090
1091 pos = head + AIO_EVENTS_OFFSET;
58c85dc2 1092 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
a31ad380
KO
1093 pos %= AIO_EVENTS_PER_PAGE;
1094
1095 ev = kmap(page);
1096 copy_ret = copy_to_user(event + ret, ev + pos,
1097 sizeof(*ev) * avail);
1098 kunmap(page);
1099
1100 if (unlikely(copy_ret)) {
1101 ret = -EFAULT;
1102 goto out;
1103 }
1104
1105 ret += avail;
1106 head += avail;
58c85dc2 1107 head %= ctx->nr_events;
1da177e4 1108 }
1da177e4 1109
58c85dc2 1110 ring = kmap_atomic(ctx->ring_pages[0]);
a31ad380 1111 ring->head = head;
91d80a84 1112 kunmap_atomic(ring);
58c85dc2 1113 flush_dcache_page(ctx->ring_pages[0]);
a31ad380 1114
5ffac122 1115 pr_debug("%li h%u t%u\n", ret, head, tail);
a31ad380 1116out:
58c85dc2 1117 mutex_unlock(&ctx->ring_lock);
a31ad380 1118
1da177e4
LT
1119 return ret;
1120}
1121
a31ad380
KO
1122static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1123 struct io_event __user *event, long *i)
1da177e4 1124{
a31ad380 1125 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1da177e4 1126
a31ad380
KO
1127 if (ret > 0)
1128 *i += ret;
1da177e4 1129
a31ad380
KO
1130 if (unlikely(atomic_read(&ctx->dead)))
1131 ret = -EINVAL;
1da177e4 1132
a31ad380
KO
1133 if (!*i)
1134 *i = ret;
1da177e4 1135
a31ad380 1136 return ret < 0 || *i >= min_nr;
1da177e4
LT
1137}
1138
a31ad380 1139static long read_events(struct kioctx *ctx, long min_nr, long nr,
1da177e4
LT
1140 struct io_event __user *event,
1141 struct timespec __user *timeout)
1142{
a31ad380
KO
1143 ktime_t until = { .tv64 = KTIME_MAX };
1144 long ret = 0;
1da177e4 1145
1da177e4
LT
1146 if (timeout) {
1147 struct timespec ts;
a31ad380 1148
1da177e4 1149 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
a31ad380 1150 return -EFAULT;
1da177e4 1151
a31ad380 1152 until = timespec_to_ktime(ts);
1da177e4
LT
1153 }
1154
a31ad380
KO
1155 /*
1156 * Note that aio_read_events() is being called as the conditional - i.e.
1157 * we're calling it after prepare_to_wait() has set task state to
1158 * TASK_INTERRUPTIBLE.
1159 *
1160 * But aio_read_events() can block, and if it blocks it's going to flip
1161 * the task state back to TASK_RUNNING.
1162 *
1163 * This should be ok, provided it doesn't flip the state back to
1164 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1165 * will only happen if the mutex_lock() call blocks, and we then find
1166 * the ringbuffer empty. So in practice we should be ok, but it's
1167 * something to be aware of when touching this code.
1168 */
1169 wait_event_interruptible_hrtimeout(ctx->wait,
1170 aio_read_events(ctx, min_nr, nr, event, &ret), until);
1da177e4 1171
a31ad380
KO
1172 if (!ret && signal_pending(current))
1173 ret = -EINTR;
1da177e4 1174
a31ad380 1175 return ret;
1da177e4
LT
1176}
1177
1da177e4
LT
1178/* sys_io_setup:
1179 * Create an aio_context capable of receiving at least nr_events.
1180 * ctxp must not point to an aio_context that already exists, and
1181 * must be initialized to 0 prior to the call. On successful
1182 * creation of the aio_context, *ctxp is filled in with the resulting
1183 * handle. May fail with -EINVAL if *ctxp is not initialized,
1184 * if the specified nr_events exceeds internal limits. May fail
1185 * with -EAGAIN if the specified nr_events exceeds the user's limit
1186 * of available events. May fail with -ENOMEM if insufficient kernel
1187 * resources are available. May fail with -EFAULT if an invalid
1188 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1189 * implemented.
1190 */
002c8976 1191SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1da177e4
LT
1192{
1193 struct kioctx *ioctx = NULL;
1194 unsigned long ctx;
1195 long ret;
1196
1197 ret = get_user(ctx, ctxp);
1198 if (unlikely(ret))
1199 goto out;
1200
1201 ret = -EINVAL;
d55b5fda
ZB
1202 if (unlikely(ctx || nr_events == 0)) {
1203 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1204 ctx, nr_events);
1da177e4
LT
1205 goto out;
1206 }
1207
1208 ioctx = ioctx_alloc(nr_events);
1209 ret = PTR_ERR(ioctx);
1210 if (!IS_ERR(ioctx)) {
1211 ret = put_user(ioctx->user_id, ctxp);
a2e1859a 1212 if (ret)
e02ba72a 1213 kill_ioctx(current->mm, ioctx, NULL);
723be6e3 1214 percpu_ref_put(&ioctx->users);
1da177e4
LT
1215 }
1216
1217out:
1218 return ret;
1219}
1220
1221/* sys_io_destroy:
1222 * Destroy the aio_context specified. May cancel any outstanding
1223 * AIOs and block on completion. Will fail with -ENOSYS if not
642b5123 1224 * implemented. May fail with -EINVAL if the context pointed to
1da177e4
LT
1225 * is invalid.
1226 */
002c8976 1227SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1da177e4
LT
1228{
1229 struct kioctx *ioctx = lookup_ioctx(ctx);
1230 if (likely(NULL != ioctx)) {
e02ba72a
AP
1231 struct completion requests_done =
1232 COMPLETION_INITIALIZER_ONSTACK(requests_done);
fb2d4483 1233 int ret;
e02ba72a
AP
1234
1235 /* Pass requests_done to kill_ioctx() where it can be set
1236 * in a thread-safe way. If we try to set it here then we have
1237 * a race condition if two io_destroy() called simultaneously.
1238 */
fb2d4483 1239 ret = kill_ioctx(current->mm, ioctx, &requests_done);
723be6e3 1240 percpu_ref_put(&ioctx->users);
e02ba72a
AP
1241
1242 /* Wait until all IO for the context are done. Otherwise kernel
1243 * keep using user-space buffers even if user thinks the context
1244 * is destroyed.
1245 */
fb2d4483
BL
1246 if (!ret)
1247 wait_for_completion(&requests_done);
e02ba72a 1248
fb2d4483 1249 return ret;
1da177e4
LT
1250 }
1251 pr_debug("EINVAL: io_destroy: invalid context id\n");
1252 return -EINVAL;
1253}
1254
41ef4eb8
KO
1255typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1256 unsigned long, loff_t);
293bc982 1257typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
41ef4eb8 1258
8bc92afc
KO
1259static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1260 int rw, char __user *buf,
1261 unsigned long *nr_segs,
1262 struct iovec **iovec,
1263 bool compat)
eed4e51f
BP
1264{
1265 ssize_t ret;
1266
8bc92afc 1267 *nr_segs = kiocb->ki_nbytes;
41ef4eb8 1268
9d85cba7
JM
1269#ifdef CONFIG_COMPAT
1270 if (compat)
41ef4eb8 1271 ret = compat_rw_copy_check_uvector(rw,
8bc92afc
KO
1272 (struct compat_iovec __user *)buf,
1273 *nr_segs, 1, *iovec, iovec);
9d85cba7
JM
1274 else
1275#endif
41ef4eb8 1276 ret = rw_copy_check_uvector(rw,
8bc92afc
KO
1277 (struct iovec __user *)buf,
1278 *nr_segs, 1, *iovec, iovec);
eed4e51f 1279 if (ret < 0)
41ef4eb8 1280 return ret;
a70b52ec 1281
41ef4eb8 1282 /* ki_nbytes now reflect bytes instead of segs */
eed4e51f 1283 kiocb->ki_nbytes = ret;
41ef4eb8 1284 return 0;
eed4e51f
BP
1285}
1286
8bc92afc
KO
1287static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1288 int rw, char __user *buf,
1289 unsigned long *nr_segs,
1290 struct iovec *iovec)
eed4e51f 1291{
8bc92afc 1292 if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
41ef4eb8 1293 return -EFAULT;
a70b52ec 1294
8bc92afc
KO
1295 iovec->iov_base = buf;
1296 iovec->iov_len = kiocb->ki_nbytes;
1297 *nr_segs = 1;
eed4e51f
BP
1298 return 0;
1299}
1300
1da177e4
LT
1301/*
1302 * aio_setup_iocb:
1303 * Performs the initial checks and aio retry method
1304 * setup for the kiocb at the time of io submission.
1305 */
8bc92afc
KO
1306static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1307 char __user *buf, bool compat)
1da177e4 1308{
41ef4eb8
KO
1309 struct file *file = req->ki_filp;
1310 ssize_t ret;
8bc92afc 1311 unsigned long nr_segs;
41ef4eb8
KO
1312 int rw;
1313 fmode_t mode;
1314 aio_rw_op *rw_op;
293bc982 1315 rw_iter_op *iter_op;
8bc92afc 1316 struct iovec inline_vec, *iovec = &inline_vec;
293bc982 1317 struct iov_iter iter;
1da177e4 1318
8bc92afc 1319 switch (opcode) {
1da177e4 1320 case IOCB_CMD_PREAD:
eed4e51f 1321 case IOCB_CMD_PREADV:
41ef4eb8
KO
1322 mode = FMODE_READ;
1323 rw = READ;
1324 rw_op = file->f_op->aio_read;
293bc982 1325 iter_op = file->f_op->read_iter;
41ef4eb8
KO
1326 goto rw_common;
1327
1328 case IOCB_CMD_PWRITE:
eed4e51f 1329 case IOCB_CMD_PWRITEV:
41ef4eb8
KO
1330 mode = FMODE_WRITE;
1331 rw = WRITE;
1332 rw_op = file->f_op->aio_write;
293bc982 1333 iter_op = file->f_op->write_iter;
41ef4eb8
KO
1334 goto rw_common;
1335rw_common:
1336 if (unlikely(!(file->f_mode & mode)))
1337 return -EBADF;
1338
293bc982 1339 if (!rw_op && !iter_op)
41ef4eb8
KO
1340 return -EINVAL;
1341
8bc92afc
KO
1342 ret = (opcode == IOCB_CMD_PREADV ||
1343 opcode == IOCB_CMD_PWRITEV)
1344 ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1345 &iovec, compat)
1346 : aio_setup_single_vector(req, rw, buf, &nr_segs,
1347 iovec);
754320d6
LY
1348 if (!ret)
1349 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
8bc92afc
KO
1350 if (ret < 0) {
1351 if (iovec != &inline_vec)
1352 kfree(iovec);
41ef4eb8 1353 return ret;
8bc92afc 1354 }
41ef4eb8
KO
1355
1356 req->ki_nbytes = ret;
41ef4eb8 1357
73a7075e
KO
1358 /* XXX: move/kill - rw_verify_area()? */
1359 /* This matches the pread()/pwrite() logic */
1360 if (req->ki_pos < 0) {
1361 ret = -EINVAL;
1362 break;
1363 }
1364
1365 if (rw == WRITE)
1366 file_start_write(file);
1367
293bc982
AV
1368 if (iter_op) {
1369 iov_iter_init(&iter, rw, iovec, nr_segs, req->ki_nbytes);
1370 ret = iter_op(req, &iter);
1371 } else {
1372 ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1373 }
73a7075e
KO
1374
1375 if (rw == WRITE)
1376 file_end_write(file);
1da177e4 1377 break;
41ef4eb8 1378
1da177e4 1379 case IOCB_CMD_FDSYNC:
41ef4eb8
KO
1380 if (!file->f_op->aio_fsync)
1381 return -EINVAL;
1382
1383 ret = file->f_op->aio_fsync(req, 1);
1da177e4 1384 break;
41ef4eb8 1385
1da177e4 1386 case IOCB_CMD_FSYNC:
41ef4eb8
KO
1387 if (!file->f_op->aio_fsync)
1388 return -EINVAL;
1389
1390 ret = file->f_op->aio_fsync(req, 0);
1da177e4 1391 break;
41ef4eb8 1392
1da177e4 1393 default:
caf4167a 1394 pr_debug("EINVAL: no operation provided\n");
41ef4eb8 1395 return -EINVAL;
1da177e4
LT
1396 }
1397
8bc92afc
KO
1398 if (iovec != &inline_vec)
1399 kfree(iovec);
1400
41ef4eb8
KO
1401 if (ret != -EIOCBQUEUED) {
1402 /*
1403 * There's no easy way to restart the syscall since other AIO's
1404 * may be already running. Just fail this IO with EINTR.
1405 */
1406 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1407 ret == -ERESTARTNOHAND ||
1408 ret == -ERESTART_RESTARTBLOCK))
1409 ret = -EINTR;
1410 aio_complete(req, ret, 0);
1411 }
1da177e4
LT
1412
1413 return 0;
1414}
1415
d5470b59 1416static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
a1c8eae7 1417 struct iocb *iocb, bool compat)
1da177e4
LT
1418{
1419 struct kiocb *req;
1da177e4
LT
1420 ssize_t ret;
1421
1422 /* enforce forwards compatibility on users */
9c3060be 1423 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
caf4167a 1424 pr_debug("EINVAL: reserve field set\n");
1da177e4
LT
1425 return -EINVAL;
1426 }
1427
1428 /* prevent overflows */
1429 if (unlikely(
1430 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1431 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1432 ((ssize_t)iocb->aio_nbytes < 0)
1433 )) {
1434 pr_debug("EINVAL: io_submit: overflow check\n");
1435 return -EINVAL;
1436 }
1437
41ef4eb8 1438 req = aio_get_req(ctx);
1d98ebfc 1439 if (unlikely(!req))
1da177e4 1440 return -EAGAIN;
1d98ebfc
KO
1441
1442 req->ki_filp = fget(iocb->aio_fildes);
1443 if (unlikely(!req->ki_filp)) {
1444 ret = -EBADF;
1445 goto out_put_req;
1da177e4 1446 }
1d98ebfc 1447
9c3060be
DL
1448 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1449 /*
1450 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1451 * instance of the file* now. The file descriptor must be
1452 * an eventfd() fd, and will be signaled for each completed
1453 * event using the eventfd_signal() function.
1454 */
13389010 1455 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
801678c5 1456 if (IS_ERR(req->ki_eventfd)) {
9c3060be 1457 ret = PTR_ERR(req->ki_eventfd);
87c3a86e 1458 req->ki_eventfd = NULL;
9c3060be
DL
1459 goto out_put_req;
1460 }
1461 }
1da177e4 1462
8a660890 1463 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1da177e4 1464 if (unlikely(ret)) {
caf4167a 1465 pr_debug("EFAULT: aio_key\n");
1da177e4
LT
1466 goto out_put_req;
1467 }
1468
1469 req->ki_obj.user = user_iocb;
1470 req->ki_user_data = iocb->aio_data;
1471 req->ki_pos = iocb->aio_offset;
73a7075e 1472 req->ki_nbytes = iocb->aio_nbytes;
1da177e4 1473
8bc92afc
KO
1474 ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1475 (char __user *)(unsigned long)iocb->aio_buf,
1476 compat);
41003a7b 1477 if (ret)
7137c6bd 1478 goto out_put_req;
41003a7b 1479
1da177e4 1480 return 0;
1da177e4 1481out_put_req:
e1bdd5f2 1482 put_reqs_available(ctx, 1);
e34ecee2 1483 percpu_ref_put(&ctx->reqs);
57282d8f 1484 kiocb_free(req);
1da177e4
LT
1485 return ret;
1486}
1487
9d85cba7
JM
1488long do_io_submit(aio_context_t ctx_id, long nr,
1489 struct iocb __user *__user *iocbpp, bool compat)
1da177e4
LT
1490{
1491 struct kioctx *ctx;
1492 long ret = 0;
080d676d 1493 int i = 0;
9f5b9425 1494 struct blk_plug plug;
1da177e4
LT
1495
1496 if (unlikely(nr < 0))
1497 return -EINVAL;
1498
75e1c70f
JM
1499 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1500 nr = LONG_MAX/sizeof(*iocbpp);
1501
1da177e4
LT
1502 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1503 return -EFAULT;
1504
1505 ctx = lookup_ioctx(ctx_id);
1506 if (unlikely(!ctx)) {
caf4167a 1507 pr_debug("EINVAL: invalid context id\n");
1da177e4
LT
1508 return -EINVAL;
1509 }
1510
9f5b9425
SL
1511 blk_start_plug(&plug);
1512
1da177e4
LT
1513 /*
1514 * AKPM: should this return a partial result if some of the IOs were
1515 * successfully submitted?
1516 */
1517 for (i=0; i<nr; i++) {
1518 struct iocb __user *user_iocb;
1519 struct iocb tmp;
1520
1521 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1522 ret = -EFAULT;
1523 break;
1524 }
1525
1526 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1527 ret = -EFAULT;
1528 break;
1529 }
1530
a1c8eae7 1531 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1da177e4
LT
1532 if (ret)
1533 break;
1534 }
9f5b9425 1535 blk_finish_plug(&plug);
1da177e4 1536
723be6e3 1537 percpu_ref_put(&ctx->users);
1da177e4
LT
1538 return i ? i : ret;
1539}
1540
9d85cba7
JM
1541/* sys_io_submit:
1542 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1543 * the number of iocbs queued. May return -EINVAL if the aio_context
1544 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1545 * *iocbpp[0] is not properly initialized, if the operation specified
1546 * is invalid for the file descriptor in the iocb. May fail with
1547 * -EFAULT if any of the data structures point to invalid data. May
1548 * fail with -EBADF if the file descriptor specified in the first
1549 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1550 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1551 * fail with -ENOSYS if not implemented.
1552 */
1553SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1554 struct iocb __user * __user *, iocbpp)
1555{
1556 return do_io_submit(ctx_id, nr, iocbpp, 0);
1557}
1558
1da177e4
LT
1559/* lookup_kiocb
1560 * Finds a given iocb for cancellation.
1da177e4 1561 */
25ee7e38
AB
1562static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1563 u32 key)
1da177e4
LT
1564{
1565 struct list_head *pos;
d00689af
ZB
1566
1567 assert_spin_locked(&ctx->ctx_lock);
1568
8a660890
KO
1569 if (key != KIOCB_KEY)
1570 return NULL;
1571
1da177e4
LT
1572 /* TODO: use a hash or array, this sucks. */
1573 list_for_each(pos, &ctx->active_reqs) {
1574 struct kiocb *kiocb = list_kiocb(pos);
8a660890 1575 if (kiocb->ki_obj.user == iocb)
1da177e4
LT
1576 return kiocb;
1577 }
1578 return NULL;
1579}
1580
1581/* sys_io_cancel:
1582 * Attempts to cancel an iocb previously passed to io_submit. If
1583 * the operation is successfully cancelled, the resulting event is
1584 * copied into the memory pointed to by result without being placed
1585 * into the completion queue and 0 is returned. May fail with
1586 * -EFAULT if any of the data structures pointed to are invalid.
1587 * May fail with -EINVAL if aio_context specified by ctx_id is
1588 * invalid. May fail with -EAGAIN if the iocb specified was not
1589 * cancelled. Will fail with -ENOSYS if not implemented.
1590 */
002c8976
HC
1591SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1592 struct io_event __user *, result)
1da177e4 1593{
1da177e4
LT
1594 struct kioctx *ctx;
1595 struct kiocb *kiocb;
1596 u32 key;
1597 int ret;
1598
1599 ret = get_user(key, &iocb->aio_key);
1600 if (unlikely(ret))
1601 return -EFAULT;
1602
1603 ctx = lookup_ioctx(ctx_id);
1604 if (unlikely(!ctx))
1605 return -EINVAL;
1606
1607 spin_lock_irq(&ctx->ctx_lock);
906b973c 1608
1da177e4 1609 kiocb = lookup_kiocb(ctx, iocb, key);
906b973c 1610 if (kiocb)
d52a8f9e 1611 ret = kiocb_cancel(kiocb);
906b973c
KO
1612 else
1613 ret = -EINVAL;
1614
1da177e4
LT
1615 spin_unlock_irq(&ctx->ctx_lock);
1616
906b973c 1617 if (!ret) {
bec68faa
KO
1618 /*
1619 * The result argument is no longer used - the io_event is
1620 * always delivered via the ring buffer. -EINPROGRESS indicates
1621 * cancellation is progress:
906b973c 1622 */
bec68faa 1623 ret = -EINPROGRESS;
906b973c 1624 }
1da177e4 1625
723be6e3 1626 percpu_ref_put(&ctx->users);
1da177e4
LT
1627
1628 return ret;
1629}
1630
1631/* io_getevents:
1632 * Attempts to read at least min_nr events and up to nr events from
642b5123
ST
1633 * the completion queue for the aio_context specified by ctx_id. If
1634 * it succeeds, the number of read events is returned. May fail with
1635 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1636 * out of range, if timeout is out of range. May fail with -EFAULT
1637 * if any of the memory specified is invalid. May return 0 or
1638 * < min_nr if the timeout specified by timeout has elapsed
1639 * before sufficient events are available, where timeout == NULL
1640 * specifies an infinite timeout. Note that the timeout pointed to by
6900807c 1641 * timeout is relative. Will fail with -ENOSYS if not implemented.
1da177e4 1642 */
002c8976
HC
1643SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1644 long, min_nr,
1645 long, nr,
1646 struct io_event __user *, events,
1647 struct timespec __user *, timeout)
1da177e4
LT
1648{
1649 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1650 long ret = -EINVAL;
1651
1652 if (likely(ioctx)) {
2e410255 1653 if (likely(min_nr <= nr && min_nr >= 0))
1da177e4 1654 ret = read_events(ioctx, min_nr, nr, events, timeout);
723be6e3 1655 percpu_ref_put(&ioctx->users);
1da177e4 1656 }
1da177e4
LT
1657 return ret;
1658}