Merge tag 'pwm/for-3.17-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/thierry...
[deliverable/linux.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;
71ad7490
BL
195 return file;
196}
197
198static struct dentry *aio_mount(struct file_system_type *fs_type,
199 int flags, const char *dev_name, void *data)
200{
201 static const struct dentry_operations ops = {
202 .d_dname = simple_dname,
203 };
8dc4379e 204 return mount_pseudo(fs_type, "aio:", NULL, &ops, AIO_RING_MAGIC);
71ad7490
BL
205}
206
1da177e4
LT
207/* aio_setup
208 * Creates the slab caches used by the aio routines, panic on
209 * failure as this is done early during the boot sequence.
210 */
211static int __init aio_setup(void)
212{
71ad7490
BL
213 static struct file_system_type aio_fs = {
214 .name = "aio",
215 .mount = aio_mount,
216 .kill_sb = kill_anon_super,
217 };
218 aio_mnt = kern_mount(&aio_fs);
219 if (IS_ERR(aio_mnt))
220 panic("Failed to create aio fs mount.");
221
0a31bd5f
CL
222 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
223 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
1da177e4 224
caf4167a 225 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
1da177e4
LT
226
227 return 0;
228}
385773e0 229__initcall(aio_setup);
1da177e4 230
5e9ae2e5
BL
231static void put_aio_ring_file(struct kioctx *ctx)
232{
233 struct file *aio_ring_file = ctx->aio_ring_file;
234 if (aio_ring_file) {
235 truncate_setsize(aio_ring_file->f_inode, 0);
236
237 /* Prevent further access to the kioctx from migratepages */
238 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
239 aio_ring_file->f_inode->i_mapping->private_data = NULL;
240 ctx->aio_ring_file = NULL;
241 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
242
243 fput(aio_ring_file);
244 }
245}
246
1da177e4
LT
247static void aio_free_ring(struct kioctx *ctx)
248{
36bc08cc 249 int i;
1da177e4 250
fa8a53c3
BL
251 /* Disconnect the kiotx from the ring file. This prevents future
252 * accesses to the kioctx from page migration.
253 */
254 put_aio_ring_file(ctx);
255
36bc08cc 256 for (i = 0; i < ctx->nr_pages; i++) {
8e321fef 257 struct page *page;
36bc08cc
GZ
258 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
259 page_count(ctx->ring_pages[i]));
8e321fef
BL
260 page = ctx->ring_pages[i];
261 if (!page)
262 continue;
263 ctx->ring_pages[i] = NULL;
264 put_page(page);
36bc08cc 265 }
1da177e4 266
ddb8c45b 267 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
58c85dc2 268 kfree(ctx->ring_pages);
ddb8c45b
SL
269 ctx->ring_pages = NULL;
270 }
36bc08cc
GZ
271}
272
273static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
274{
275 vma->vm_ops = &generic_file_vm_ops;
276 return 0;
277}
278
279static const struct file_operations aio_ring_fops = {
280 .mmap = aio_ring_mmap,
281};
282
283static int aio_set_page_dirty(struct page *page)
284{
285 return 0;
286}
287
0c45355f 288#if IS_ENABLED(CONFIG_MIGRATION)
36bc08cc
GZ
289static int aio_migratepage(struct address_space *mapping, struct page *new,
290 struct page *old, enum migrate_mode mode)
291{
5e9ae2e5 292 struct kioctx *ctx;
36bc08cc 293 unsigned long flags;
fa8a53c3 294 pgoff_t idx;
36bc08cc
GZ
295 int rc;
296
8e321fef
BL
297 rc = 0;
298
fa8a53c3 299 /* mapping->private_lock here protects against the kioctx teardown. */
8e321fef
BL
300 spin_lock(&mapping->private_lock);
301 ctx = mapping->private_data;
fa8a53c3
BL
302 if (!ctx) {
303 rc = -EINVAL;
304 goto out;
305 }
306
307 /* The ring_lock mutex. The prevents aio_read_events() from writing
308 * to the ring's head, and prevents page migration from mucking in
309 * a partially initialized kiotx.
310 */
311 if (!mutex_trylock(&ctx->ring_lock)) {
312 rc = -EAGAIN;
313 goto out;
314 }
315
316 idx = old->index;
317 if (idx < (pgoff_t)ctx->nr_pages) {
318 /* Make sure the old page hasn't already been changed */
319 if (ctx->ring_pages[idx] != old)
320 rc = -EAGAIN;
8e321fef
BL
321 } else
322 rc = -EINVAL;
8e321fef
BL
323
324 if (rc != 0)
fa8a53c3 325 goto out_unlock;
8e321fef 326
36bc08cc
GZ
327 /* Writeback must be complete */
328 BUG_ON(PageWriteback(old));
8e321fef 329 get_page(new);
36bc08cc 330
8e321fef 331 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
36bc08cc 332 if (rc != MIGRATEPAGE_SUCCESS) {
8e321fef 333 put_page(new);
fa8a53c3 334 goto out_unlock;
36bc08cc
GZ
335 }
336
fa8a53c3
BL
337 /* Take completion_lock to prevent other writes to the ring buffer
338 * while the old page is copied to the new. This prevents new
339 * events from being lost.
5e9ae2e5 340 */
fa8a53c3
BL
341 spin_lock_irqsave(&ctx->completion_lock, flags);
342 migrate_page_copy(new, old);
343 BUG_ON(ctx->ring_pages[idx] != old);
344 ctx->ring_pages[idx] = new;
345 spin_unlock_irqrestore(&ctx->completion_lock, flags);
36bc08cc 346
fa8a53c3
BL
347 /* The old page is no longer accessible. */
348 put_page(old);
8e321fef 349
fa8a53c3
BL
350out_unlock:
351 mutex_unlock(&ctx->ring_lock);
352out:
353 spin_unlock(&mapping->private_lock);
36bc08cc 354 return rc;
1da177e4 355}
0c45355f 356#endif
1da177e4 357
36bc08cc
GZ
358static const struct address_space_operations aio_ctx_aops = {
359 .set_page_dirty = aio_set_page_dirty,
0c45355f 360#if IS_ENABLED(CONFIG_MIGRATION)
36bc08cc 361 .migratepage = aio_migratepage,
0c45355f 362#endif
36bc08cc
GZ
363};
364
1da177e4
LT
365static int aio_setup_ring(struct kioctx *ctx)
366{
367 struct aio_ring *ring;
1da177e4 368 unsigned nr_events = ctx->max_reqs;
41003a7b 369 struct mm_struct *mm = current->mm;
3dc9acb6 370 unsigned long size, unused;
1da177e4 371 int nr_pages;
36bc08cc
GZ
372 int i;
373 struct file *file;
1da177e4
LT
374
375 /* Compensate for the ring buffer's head/tail overlap entry */
376 nr_events += 2; /* 1 is required, 2 for good luck */
377
378 size = sizeof(struct aio_ring);
379 size += sizeof(struct io_event) * nr_events;
1da177e4 380
36bc08cc 381 nr_pages = PFN_UP(size);
1da177e4
LT
382 if (nr_pages < 0)
383 return -EINVAL;
384
71ad7490 385 file = aio_private_file(ctx, nr_pages);
36bc08cc
GZ
386 if (IS_ERR(file)) {
387 ctx->aio_ring_file = NULL;
fa8a53c3 388 return -ENOMEM;
36bc08cc
GZ
389 }
390
3dc9acb6
LT
391 ctx->aio_ring_file = file;
392 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
393 / sizeof(struct io_event);
394
395 ctx->ring_pages = ctx->internal_pages;
396 if (nr_pages > AIO_RING_PAGES) {
397 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
398 GFP_KERNEL);
399 if (!ctx->ring_pages) {
400 put_aio_ring_file(ctx);
401 return -ENOMEM;
402 }
403 }
404
36bc08cc
GZ
405 for (i = 0; i < nr_pages; i++) {
406 struct page *page;
407 page = find_or_create_page(file->f_inode->i_mapping,
408 i, GFP_HIGHUSER | __GFP_ZERO);
409 if (!page)
410 break;
411 pr_debug("pid(%d) page[%d]->count=%d\n",
412 current->pid, i, page_count(page));
413 SetPageUptodate(page);
414 SetPageDirty(page);
415 unlock_page(page);
3dc9acb6
LT
416
417 ctx->ring_pages[i] = page;
36bc08cc 418 }
3dc9acb6 419 ctx->nr_pages = i;
1da177e4 420
3dc9acb6
LT
421 if (unlikely(i != nr_pages)) {
422 aio_free_ring(ctx);
fa8a53c3 423 return -ENOMEM;
1da177e4
LT
424 }
425
58c85dc2
KO
426 ctx->mmap_size = nr_pages * PAGE_SIZE;
427 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
36bc08cc 428
41003a7b 429 down_write(&mm->mmap_sem);
36bc08cc
GZ
430 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
431 PROT_READ | PROT_WRITE,
3dc9acb6
LT
432 MAP_SHARED, 0, &unused);
433 up_write(&mm->mmap_sem);
58c85dc2 434 if (IS_ERR((void *)ctx->mmap_base)) {
58c85dc2 435 ctx->mmap_size = 0;
1da177e4 436 aio_free_ring(ctx);
fa8a53c3 437 return -ENOMEM;
1da177e4
LT
438 }
439
58c85dc2 440 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
d6c355c7 441
58c85dc2
KO
442 ctx->user_id = ctx->mmap_base;
443 ctx->nr_events = nr_events; /* trusted copy */
1da177e4 444
58c85dc2 445 ring = kmap_atomic(ctx->ring_pages[0]);
1da177e4 446 ring->nr = nr_events; /* user copy */
db446a08 447 ring->id = ~0U;
1da177e4
LT
448 ring->head = ring->tail = 0;
449 ring->magic = AIO_RING_MAGIC;
450 ring->compat_features = AIO_RING_COMPAT_FEATURES;
451 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
452 ring->header_length = sizeof(struct aio_ring);
e8e3c3d6 453 kunmap_atomic(ring);
58c85dc2 454 flush_dcache_page(ctx->ring_pages[0]);
1da177e4
LT
455
456 return 0;
457}
458
1da177e4
LT
459#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
460#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
461#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
462
0460fef2
KO
463void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
464{
465 struct kioctx *ctx = req->ki_ctx;
466 unsigned long flags;
467
468 spin_lock_irqsave(&ctx->ctx_lock, flags);
469
470 if (!req->ki_list.next)
471 list_add(&req->ki_list, &ctx->active_reqs);
472
473 req->ki_cancel = cancel;
474
475 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
476}
477EXPORT_SYMBOL(kiocb_set_cancel_fn);
478
d52a8f9e 479static int kiocb_cancel(struct kiocb *kiocb)
906b973c 480{
0460fef2 481 kiocb_cancel_fn *old, *cancel;
906b973c 482
0460fef2
KO
483 /*
484 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
485 * actually has a cancel function, hence the cmpxchg()
486 */
487
488 cancel = ACCESS_ONCE(kiocb->ki_cancel);
489 do {
490 if (!cancel || cancel == KIOCB_CANCELLED)
57282d8f 491 return -EINVAL;
906b973c 492
0460fef2
KO
493 old = cancel;
494 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
495 } while (cancel != old);
906b973c 496
57282d8f 497 return cancel(kiocb);
906b973c
KO
498}
499
e34ecee2 500static void free_ioctx(struct work_struct *work)
36f55889 501{
e34ecee2 502 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
e1bdd5f2 503
e34ecee2 504 pr_debug("freeing %p\n", ctx);
e1bdd5f2 505
e34ecee2 506 aio_free_ring(ctx);
e1bdd5f2 507 free_percpu(ctx->cpu);
9a1049da
TH
508 percpu_ref_exit(&ctx->reqs);
509 percpu_ref_exit(&ctx->users);
36f55889
KO
510 kmem_cache_free(kioctx_cachep, ctx);
511}
512
e34ecee2
KO
513static void free_ioctx_reqs(struct percpu_ref *ref)
514{
515 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
516
e02ba72a
AP
517 /* At this point we know that there are no any in-flight requests */
518 if (ctx->requests_done)
519 complete(ctx->requests_done);
520
e34ecee2
KO
521 INIT_WORK(&ctx->free_work, free_ioctx);
522 schedule_work(&ctx->free_work);
523}
524
36f55889
KO
525/*
526 * When this function runs, the kioctx has been removed from the "hash table"
527 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
528 * now it's safe to cancel any that need to be.
529 */
e34ecee2 530static void free_ioctx_users(struct percpu_ref *ref)
36f55889 531{
e34ecee2 532 struct kioctx *ctx = container_of(ref, struct kioctx, users);
36f55889
KO
533 struct kiocb *req;
534
535 spin_lock_irq(&ctx->ctx_lock);
536
537 while (!list_empty(&ctx->active_reqs)) {
538 req = list_first_entry(&ctx->active_reqs,
539 struct kiocb, ki_list);
540
541 list_del_init(&req->ki_list);
d52a8f9e 542 kiocb_cancel(req);
36f55889
KO
543 }
544
545 spin_unlock_irq(&ctx->ctx_lock);
546
e34ecee2
KO
547 percpu_ref_kill(&ctx->reqs);
548 percpu_ref_put(&ctx->reqs);
36f55889
KO
549}
550
db446a08
BL
551static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
552{
553 unsigned i, new_nr;
554 struct kioctx_table *table, *old;
555 struct aio_ring *ring;
556
557 spin_lock(&mm->ioctx_lock);
855ef0de 558 table = rcu_dereference_raw(mm->ioctx_table);
db446a08
BL
559
560 while (1) {
561 if (table)
562 for (i = 0; i < table->nr; i++)
563 if (!table->table[i]) {
564 ctx->id = i;
565 table->table[i] = ctx;
566 spin_unlock(&mm->ioctx_lock);
567
fa8a53c3
BL
568 /* While kioctx setup is in progress,
569 * we are protected from page migration
570 * changes ring_pages by ->ring_lock.
571 */
db446a08
BL
572 ring = kmap_atomic(ctx->ring_pages[0]);
573 ring->id = ctx->id;
574 kunmap_atomic(ring);
575 return 0;
576 }
577
578 new_nr = (table ? table->nr : 1) * 4;
db446a08
BL
579 spin_unlock(&mm->ioctx_lock);
580
581 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
582 new_nr, GFP_KERNEL);
583 if (!table)
584 return -ENOMEM;
585
586 table->nr = new_nr;
587
588 spin_lock(&mm->ioctx_lock);
855ef0de 589 old = rcu_dereference_raw(mm->ioctx_table);
db446a08
BL
590
591 if (!old) {
592 rcu_assign_pointer(mm->ioctx_table, table);
593 } else if (table->nr > old->nr) {
594 memcpy(table->table, old->table,
595 old->nr * sizeof(struct kioctx *));
596
597 rcu_assign_pointer(mm->ioctx_table, table);
598 kfree_rcu(old, rcu);
599 } else {
600 kfree(table);
601 table = old;
602 }
603 }
604}
605
e34ecee2
KO
606static void aio_nr_sub(unsigned nr)
607{
608 spin_lock(&aio_nr_lock);
609 if (WARN_ON(aio_nr - nr > aio_nr))
610 aio_nr = 0;
611 else
612 aio_nr -= nr;
613 spin_unlock(&aio_nr_lock);
614}
615
1da177e4
LT
616/* ioctx_alloc
617 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
618 */
619static struct kioctx *ioctx_alloc(unsigned nr_events)
620{
41003a7b 621 struct mm_struct *mm = current->mm;
1da177e4 622 struct kioctx *ctx;
e23754f8 623 int err = -ENOMEM;
1da177e4 624
e1bdd5f2
KO
625 /*
626 * We keep track of the number of available ringbuffer slots, to prevent
627 * overflow (reqs_available), and we also use percpu counters for this.
628 *
629 * So since up to half the slots might be on other cpu's percpu counters
630 * and unavailable, double nr_events so userspace sees what they
631 * expected: additionally, we move req_batch slots to/from percpu
632 * counters at a time, so make sure that isn't 0:
633 */
634 nr_events = max(nr_events, num_possible_cpus() * 4);
635 nr_events *= 2;
636
1da177e4
LT
637 /* Prevent overflows */
638 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
639 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
640 pr_debug("ENOMEM: nr_events too high\n");
641 return ERR_PTR(-EINVAL);
642 }
643
4cd81c3d 644 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
1da177e4
LT
645 return ERR_PTR(-EAGAIN);
646
c3762229 647 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
1da177e4
LT
648 if (!ctx)
649 return ERR_PTR(-ENOMEM);
650
1da177e4 651 ctx->max_reqs = nr_events;
1da177e4 652
1da177e4 653 spin_lock_init(&ctx->ctx_lock);
0460fef2 654 spin_lock_init(&ctx->completion_lock);
58c85dc2 655 mutex_init(&ctx->ring_lock);
fa8a53c3
BL
656 /* Protect against page migration throughout kiotx setup by keeping
657 * the ring_lock mutex held until setup is complete. */
658 mutex_lock(&ctx->ring_lock);
1da177e4
LT
659 init_waitqueue_head(&ctx->wait);
660
661 INIT_LIST_HEAD(&ctx->active_reqs);
1da177e4 662
fa8a53c3
BL
663 if (percpu_ref_init(&ctx->users, free_ioctx_users))
664 goto err;
665
666 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs))
667 goto err;
668
e1bdd5f2
KO
669 ctx->cpu = alloc_percpu(struct kioctx_cpu);
670 if (!ctx->cpu)
e34ecee2 671 goto err;
1da177e4 672
fa8a53c3
BL
673 err = aio_setup_ring(ctx);
674 if (err < 0)
e34ecee2 675 goto err;
e1bdd5f2 676
34e83fc6 677 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
e1bdd5f2 678 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
6878ea72
BL
679 if (ctx->req_batch < 1)
680 ctx->req_batch = 1;
34e83fc6 681
1da177e4 682 /* limit the number of system wide aios */
9fa1cb39 683 spin_lock(&aio_nr_lock);
4cd81c3d 684 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
2dd542b7 685 aio_nr + nr_events < aio_nr) {
9fa1cb39 686 spin_unlock(&aio_nr_lock);
e34ecee2 687 err = -EAGAIN;
d1b94327 688 goto err_ctx;
2dd542b7
AV
689 }
690 aio_nr += ctx->max_reqs;
9fa1cb39 691 spin_unlock(&aio_nr_lock);
1da177e4 692
1881686f
BL
693 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
694 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
723be6e3 695
da90382c
BL
696 err = ioctx_add_table(ctx, mm);
697 if (err)
e34ecee2 698 goto err_cleanup;
da90382c 699
fa8a53c3
BL
700 /* Release the ring_lock mutex now that all setup is complete. */
701 mutex_unlock(&ctx->ring_lock);
702
caf4167a 703 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
58c85dc2 704 ctx, ctx->user_id, mm, ctx->nr_events);
1da177e4
LT
705 return ctx;
706
e34ecee2
KO
707err_cleanup:
708 aio_nr_sub(ctx->max_reqs);
d1b94327
GZ
709err_ctx:
710 aio_free_ring(ctx);
e34ecee2 711err:
fa8a53c3 712 mutex_unlock(&ctx->ring_lock);
e1bdd5f2 713 free_percpu(ctx->cpu);
9a1049da
TH
714 percpu_ref_exit(&ctx->reqs);
715 percpu_ref_exit(&ctx->users);
1da177e4 716 kmem_cache_free(kioctx_cachep, ctx);
caf4167a 717 pr_debug("error allocating ioctx %d\n", err);
e23754f8 718 return ERR_PTR(err);
1da177e4
LT
719}
720
36f55889
KO
721/* kill_ioctx
722 * Cancels all outstanding aio requests on an aio context. Used
723 * when the processes owning a context have all exited to encourage
724 * the rapid destruction of the kioctx.
725 */
fb2d4483 726static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
e02ba72a 727 struct completion *requests_done)
36f55889 728{
fa88b6f8 729 struct kioctx_table *table;
db446a08 730
fa88b6f8
BL
731 if (atomic_xchg(&ctx->dead, 1))
732 return -EINVAL;
db446a08 733
db446a08 734
fa88b6f8 735 spin_lock(&mm->ioctx_lock);
855ef0de 736 table = rcu_dereference_raw(mm->ioctx_table);
fa88b6f8
BL
737 WARN_ON(ctx != table->table[ctx->id]);
738 table->table[ctx->id] = NULL;
fa88b6f8 739 spin_unlock(&mm->ioctx_lock);
4fcc712f 740
fa88b6f8
BL
741 /* percpu_ref_kill() will do the necessary call_rcu() */
742 wake_up_all(&ctx->wait);
4fcc712f 743
fa88b6f8
BL
744 /*
745 * It'd be more correct to do this in free_ioctx(), after all
746 * the outstanding kiocbs have finished - but by then io_destroy
747 * has already returned, so io_setup() could potentially return
748 * -EAGAIN with no ioctxs actually in use (as far as userspace
749 * could tell).
750 */
751 aio_nr_sub(ctx->max_reqs);
4fcc712f 752
fa88b6f8
BL
753 if (ctx->mmap_size)
754 vm_munmap(ctx->mmap_base, ctx->mmap_size);
fb2d4483 755
fa88b6f8
BL
756 ctx->requests_done = requests_done;
757 percpu_ref_kill(&ctx->users);
758 return 0;
1da177e4
LT
759}
760
761/* wait_on_sync_kiocb:
762 * Waits on the given sync kiocb to complete.
763 */
57282d8f 764ssize_t wait_on_sync_kiocb(struct kiocb *req)
1da177e4 765{
57282d8f 766 while (!req->ki_ctx) {
1da177e4 767 set_current_state(TASK_UNINTERRUPTIBLE);
57282d8f 768 if (req->ki_ctx)
1da177e4 769 break;
41d10da3 770 io_schedule();
1da177e4
LT
771 }
772 __set_current_state(TASK_RUNNING);
57282d8f 773 return req->ki_user_data;
1da177e4 774}
385773e0 775EXPORT_SYMBOL(wait_on_sync_kiocb);
1da177e4 776
36f55889
KO
777/*
778 * exit_aio: called when the last user of mm goes away. At this point, there is
779 * no way for any new requests to be submited or any of the io_* syscalls to be
780 * called on the context.
781 *
782 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
783 * them.
1da177e4 784 */
fc9b52cd 785void exit_aio(struct mm_struct *mm)
1da177e4 786{
4b70ac5f
ON
787 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
788 int i;
db446a08 789
4b70ac5f
ON
790 if (!table)
791 return;
db446a08 792
4b70ac5f
ON
793 for (i = 0; i < table->nr; ++i) {
794 struct kioctx *ctx = table->table[i];
abf137dd 795
4b70ac5f
ON
796 if (!ctx)
797 continue;
936af157 798 /*
4b70ac5f
ON
799 * We don't need to bother with munmap() here - exit_mmap(mm)
800 * is coming and it'll unmap everything. And we simply can't,
801 * this is not necessarily our ->mm.
802 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
803 * that it needs to unmap the area, just set it to 0.
936af157 804 */
58c85dc2 805 ctx->mmap_size = 0;
e02ba72a 806 kill_ioctx(mm, ctx, NULL);
1da177e4 807 }
4b70ac5f
ON
808
809 RCU_INIT_POINTER(mm->ioctx_table, NULL);
810 kfree(table);
1da177e4
LT
811}
812
e1bdd5f2
KO
813static void put_reqs_available(struct kioctx *ctx, unsigned nr)
814{
815 struct kioctx_cpu *kcpu;
263782c1 816 unsigned long flags;
e1bdd5f2 817
263782c1 818 local_irq_save(flags);
be6fb451 819 kcpu = this_cpu_ptr(ctx->cpu);
e1bdd5f2 820 kcpu->reqs_available += nr;
263782c1 821
e1bdd5f2
KO
822 while (kcpu->reqs_available >= ctx->req_batch * 2) {
823 kcpu->reqs_available -= ctx->req_batch;
824 atomic_add(ctx->req_batch, &ctx->reqs_available);
825 }
826
263782c1 827 local_irq_restore(flags);
e1bdd5f2
KO
828}
829
830static bool get_reqs_available(struct kioctx *ctx)
831{
832 struct kioctx_cpu *kcpu;
833 bool ret = false;
263782c1 834 unsigned long flags;
e1bdd5f2 835
263782c1 836 local_irq_save(flags);
be6fb451 837 kcpu = this_cpu_ptr(ctx->cpu);
e1bdd5f2
KO
838 if (!kcpu->reqs_available) {
839 int old, avail = atomic_read(&ctx->reqs_available);
840
841 do {
842 if (avail < ctx->req_batch)
843 goto out;
844
845 old = avail;
846 avail = atomic_cmpxchg(&ctx->reqs_available,
847 avail, avail - ctx->req_batch);
848 } while (avail != old);
849
850 kcpu->reqs_available += ctx->req_batch;
851 }
852
853 ret = true;
854 kcpu->reqs_available--;
855out:
263782c1 856 local_irq_restore(flags);
e1bdd5f2
KO
857 return ret;
858}
859
1da177e4 860/* aio_get_req
57282d8f
KO
861 * Allocate a slot for an aio request.
862 * Returns NULL if no requests are free.
1da177e4 863 */
a1c8eae7 864static inline struct kiocb *aio_get_req(struct kioctx *ctx)
1da177e4 865{
a1c8eae7
KO
866 struct kiocb *req;
867
e1bdd5f2 868 if (!get_reqs_available(ctx))
a1c8eae7
KO
869 return NULL;
870
0460fef2 871 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
1da177e4 872 if (unlikely(!req))
a1c8eae7 873 goto out_put;
1da177e4 874
e34ecee2
KO
875 percpu_ref_get(&ctx->reqs);
876
1da177e4 877 req->ki_ctx = ctx;
080d676d 878 return req;
a1c8eae7 879out_put:
e1bdd5f2 880 put_reqs_available(ctx, 1);
a1c8eae7 881 return NULL;
1da177e4
LT
882}
883
11599eba 884static void kiocb_free(struct kiocb *req)
1da177e4 885{
1d98ebfc
KO
886 if (req->ki_filp)
887 fput(req->ki_filp);
13389010
DL
888 if (req->ki_eventfd != NULL)
889 eventfd_ctx_put(req->ki_eventfd);
1da177e4 890 kmem_cache_free(kiocb_cachep, req);
1da177e4
LT
891}
892
d5470b59 893static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1da177e4 894{
db446a08 895 struct aio_ring __user *ring = (void __user *)ctx_id;
abf137dd 896 struct mm_struct *mm = current->mm;
65c24491 897 struct kioctx *ctx, *ret = NULL;
db446a08
BL
898 struct kioctx_table *table;
899 unsigned id;
900
901 if (get_user(id, &ring->id))
902 return NULL;
1da177e4 903
abf137dd 904 rcu_read_lock();
db446a08 905 table = rcu_dereference(mm->ioctx_table);
abf137dd 906
db446a08
BL
907 if (!table || id >= table->nr)
908 goto out;
1da177e4 909
db446a08 910 ctx = table->table[id];
f30d704f 911 if (ctx && ctx->user_id == ctx_id) {
db446a08
BL
912 percpu_ref_get(&ctx->users);
913 ret = ctx;
914 }
915out:
abf137dd 916 rcu_read_unlock();
65c24491 917 return ret;
1da177e4
LT
918}
919
1da177e4
LT
920/* aio_complete
921 * Called when the io request on the given iocb is complete.
1da177e4 922 */
2d68449e 923void aio_complete(struct kiocb *iocb, long res, long res2)
1da177e4
LT
924{
925 struct kioctx *ctx = iocb->ki_ctx;
1da177e4 926 struct aio_ring *ring;
21b40200 927 struct io_event *ev_page, *event;
1da177e4 928 unsigned long flags;
21b40200 929 unsigned tail, pos;
1da177e4 930
20dcae32
ZB
931 /*
932 * Special case handling for sync iocbs:
933 * - events go directly into the iocb for fast handling
934 * - the sync task with the iocb in its stack holds the single iocb
935 * ref, no other paths have a way to get another ref
936 * - the sync task helpfully left a reference to itself in the iocb
1da177e4
LT
937 */
938 if (is_sync_kiocb(iocb)) {
1da177e4 939 iocb->ki_user_data = res;
57282d8f
KO
940 smp_wmb();
941 iocb->ki_ctx = ERR_PTR(-EXDEV);
1da177e4 942 wake_up_process(iocb->ki_obj.tsk);
2d68449e 943 return;
1da177e4
LT
944 }
945
0460fef2
KO
946 if (iocb->ki_list.next) {
947 unsigned long flags;
948
949 spin_lock_irqsave(&ctx->ctx_lock, flags);
950 list_del(&iocb->ki_list);
951 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
952 }
11599eba 953
0460fef2
KO
954 /*
955 * Add a completion event to the ring buffer. Must be done holding
4b30f07e 956 * ctx->completion_lock to prevent other code from messing with the tail
0460fef2
KO
957 * pointer since we might be called from irq context.
958 */
959 spin_lock_irqsave(&ctx->completion_lock, flags);
960
58c85dc2 961 tail = ctx->tail;
21b40200
KO
962 pos = tail + AIO_EVENTS_OFFSET;
963
58c85dc2 964 if (++tail >= ctx->nr_events)
4bf69b2a 965 tail = 0;
1da177e4 966
58c85dc2 967 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
21b40200
KO
968 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
969
1da177e4
LT
970 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
971 event->data = iocb->ki_user_data;
972 event->res = res;
973 event->res2 = res2;
974
21b40200 975 kunmap_atomic(ev_page);
58c85dc2 976 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
21b40200
KO
977
978 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
caf4167a
KO
979 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
980 res, res2);
1da177e4
LT
981
982 /* after flagging the request as done, we
983 * must never even look at it again
984 */
985 smp_wmb(); /* make event visible before updating tail */
986
58c85dc2 987 ctx->tail = tail;
1da177e4 988
58c85dc2 989 ring = kmap_atomic(ctx->ring_pages[0]);
21b40200 990 ring->tail = tail;
e8e3c3d6 991 kunmap_atomic(ring);
58c85dc2 992 flush_dcache_page(ctx->ring_pages[0]);
1da177e4 993
0460fef2
KO
994 spin_unlock_irqrestore(&ctx->completion_lock, flags);
995
21b40200 996 pr_debug("added to ring %p at [%u]\n", iocb, tail);
8d1c98b0
DL
997
998 /*
999 * Check if the user asked us to deliver the result through an
1000 * eventfd. The eventfd_signal() function is safe to be called
1001 * from IRQ context.
1002 */
87c3a86e 1003 if (iocb->ki_eventfd != NULL)
8d1c98b0
DL
1004 eventfd_signal(iocb->ki_eventfd, 1);
1005
1da177e4 1006 /* everything turned out well, dispose of the aiocb. */
57282d8f 1007 kiocb_free(iocb);
f8567a38 1008 put_reqs_available(ctx, 1);
1da177e4 1009
6cb2a210
QB
1010 /*
1011 * We have to order our ring_info tail store above and test
1012 * of the wait list below outside the wait lock. This is
1013 * like in wake_up_bit() where clearing a bit has to be
1014 * ordered with the unlocked test.
1015 */
1016 smp_mb();
1017
1da177e4
LT
1018 if (waitqueue_active(&ctx->wait))
1019 wake_up(&ctx->wait);
1020
e34ecee2 1021 percpu_ref_put(&ctx->reqs);
1da177e4 1022}
385773e0 1023EXPORT_SYMBOL(aio_complete);
1da177e4 1024
2be4e7de 1025/* aio_read_events_ring
a31ad380
KO
1026 * Pull an event off of the ioctx's event ring. Returns the number of
1027 * events fetched
1da177e4 1028 */
a31ad380
KO
1029static long aio_read_events_ring(struct kioctx *ctx,
1030 struct io_event __user *event, long nr)
1da177e4 1031{
1da177e4 1032 struct aio_ring *ring;
5ffac122 1033 unsigned head, tail, pos;
a31ad380
KO
1034 long ret = 0;
1035 int copy_ret;
1036
58c85dc2 1037 mutex_lock(&ctx->ring_lock);
1da177e4 1038
fa8a53c3 1039 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
58c85dc2 1040 ring = kmap_atomic(ctx->ring_pages[0]);
a31ad380 1041 head = ring->head;
5ffac122 1042 tail = ring->tail;
a31ad380
KO
1043 kunmap_atomic(ring);
1044
5ffac122 1045 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1da177e4 1046
5ffac122 1047 if (head == tail)
1da177e4
LT
1048 goto out;
1049
edfbbf38
BL
1050 head %= ctx->nr_events;
1051 tail %= ctx->nr_events;
1052
a31ad380
KO
1053 while (ret < nr) {
1054 long avail;
1055 struct io_event *ev;
1056 struct page *page;
1057
5ffac122
KO
1058 avail = (head <= tail ? tail : ctx->nr_events) - head;
1059 if (head == tail)
a31ad380
KO
1060 break;
1061
1062 avail = min(avail, nr - ret);
1063 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1064 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1065
1066 pos = head + AIO_EVENTS_OFFSET;
58c85dc2 1067 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
a31ad380
KO
1068 pos %= AIO_EVENTS_PER_PAGE;
1069
1070 ev = kmap(page);
1071 copy_ret = copy_to_user(event + ret, ev + pos,
1072 sizeof(*ev) * avail);
1073 kunmap(page);
1074
1075 if (unlikely(copy_ret)) {
1076 ret = -EFAULT;
1077 goto out;
1078 }
1079
1080 ret += avail;
1081 head += avail;
58c85dc2 1082 head %= ctx->nr_events;
1da177e4 1083 }
1da177e4 1084
58c85dc2 1085 ring = kmap_atomic(ctx->ring_pages[0]);
a31ad380 1086 ring->head = head;
91d80a84 1087 kunmap_atomic(ring);
58c85dc2 1088 flush_dcache_page(ctx->ring_pages[0]);
a31ad380 1089
5ffac122 1090 pr_debug("%li h%u t%u\n", ret, head, tail);
a31ad380 1091out:
58c85dc2 1092 mutex_unlock(&ctx->ring_lock);
a31ad380 1093
1da177e4
LT
1094 return ret;
1095}
1096
a31ad380
KO
1097static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1098 struct io_event __user *event, long *i)
1da177e4 1099{
a31ad380 1100 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1da177e4 1101
a31ad380
KO
1102 if (ret > 0)
1103 *i += ret;
1da177e4 1104
a31ad380
KO
1105 if (unlikely(atomic_read(&ctx->dead)))
1106 ret = -EINVAL;
1da177e4 1107
a31ad380
KO
1108 if (!*i)
1109 *i = ret;
1da177e4 1110
a31ad380 1111 return ret < 0 || *i >= min_nr;
1da177e4
LT
1112}
1113
a31ad380 1114static long read_events(struct kioctx *ctx, long min_nr, long nr,
1da177e4
LT
1115 struct io_event __user *event,
1116 struct timespec __user *timeout)
1117{
a31ad380
KO
1118 ktime_t until = { .tv64 = KTIME_MAX };
1119 long ret = 0;
1da177e4 1120
1da177e4
LT
1121 if (timeout) {
1122 struct timespec ts;
a31ad380 1123
1da177e4 1124 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
a31ad380 1125 return -EFAULT;
1da177e4 1126
a31ad380 1127 until = timespec_to_ktime(ts);
1da177e4
LT
1128 }
1129
a31ad380
KO
1130 /*
1131 * Note that aio_read_events() is being called as the conditional - i.e.
1132 * we're calling it after prepare_to_wait() has set task state to
1133 * TASK_INTERRUPTIBLE.
1134 *
1135 * But aio_read_events() can block, and if it blocks it's going to flip
1136 * the task state back to TASK_RUNNING.
1137 *
1138 * This should be ok, provided it doesn't flip the state back to
1139 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1140 * will only happen if the mutex_lock() call blocks, and we then find
1141 * the ringbuffer empty. So in practice we should be ok, but it's
1142 * something to be aware of when touching this code.
1143 */
1144 wait_event_interruptible_hrtimeout(ctx->wait,
1145 aio_read_events(ctx, min_nr, nr, event, &ret), until);
1da177e4 1146
a31ad380
KO
1147 if (!ret && signal_pending(current))
1148 ret = -EINTR;
1da177e4 1149
a31ad380 1150 return ret;
1da177e4
LT
1151}
1152
1da177e4
LT
1153/* sys_io_setup:
1154 * Create an aio_context capable of receiving at least nr_events.
1155 * ctxp must not point to an aio_context that already exists, and
1156 * must be initialized to 0 prior to the call. On successful
1157 * creation of the aio_context, *ctxp is filled in with the resulting
1158 * handle. May fail with -EINVAL if *ctxp is not initialized,
1159 * if the specified nr_events exceeds internal limits. May fail
1160 * with -EAGAIN if the specified nr_events exceeds the user's limit
1161 * of available events. May fail with -ENOMEM if insufficient kernel
1162 * resources are available. May fail with -EFAULT if an invalid
1163 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1164 * implemented.
1165 */
002c8976 1166SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1da177e4
LT
1167{
1168 struct kioctx *ioctx = NULL;
1169 unsigned long ctx;
1170 long ret;
1171
1172 ret = get_user(ctx, ctxp);
1173 if (unlikely(ret))
1174 goto out;
1175
1176 ret = -EINVAL;
d55b5fda
ZB
1177 if (unlikely(ctx || nr_events == 0)) {
1178 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1179 ctx, nr_events);
1da177e4
LT
1180 goto out;
1181 }
1182
1183 ioctx = ioctx_alloc(nr_events);
1184 ret = PTR_ERR(ioctx);
1185 if (!IS_ERR(ioctx)) {
1186 ret = put_user(ioctx->user_id, ctxp);
a2e1859a 1187 if (ret)
e02ba72a 1188 kill_ioctx(current->mm, ioctx, NULL);
723be6e3 1189 percpu_ref_put(&ioctx->users);
1da177e4
LT
1190 }
1191
1192out:
1193 return ret;
1194}
1195
1196/* sys_io_destroy:
1197 * Destroy the aio_context specified. May cancel any outstanding
1198 * AIOs and block on completion. Will fail with -ENOSYS if not
642b5123 1199 * implemented. May fail with -EINVAL if the context pointed to
1da177e4
LT
1200 * is invalid.
1201 */
002c8976 1202SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1da177e4
LT
1203{
1204 struct kioctx *ioctx = lookup_ioctx(ctx);
1205 if (likely(NULL != ioctx)) {
e02ba72a
AP
1206 struct completion requests_done =
1207 COMPLETION_INITIALIZER_ONSTACK(requests_done);
fb2d4483 1208 int ret;
e02ba72a
AP
1209
1210 /* Pass requests_done to kill_ioctx() where it can be set
1211 * in a thread-safe way. If we try to set it here then we have
1212 * a race condition if two io_destroy() called simultaneously.
1213 */
fb2d4483 1214 ret = kill_ioctx(current->mm, ioctx, &requests_done);
723be6e3 1215 percpu_ref_put(&ioctx->users);
e02ba72a
AP
1216
1217 /* Wait until all IO for the context are done. Otherwise kernel
1218 * keep using user-space buffers even if user thinks the context
1219 * is destroyed.
1220 */
fb2d4483
BL
1221 if (!ret)
1222 wait_for_completion(&requests_done);
e02ba72a 1223
fb2d4483 1224 return ret;
1da177e4
LT
1225 }
1226 pr_debug("EINVAL: io_destroy: invalid context id\n");
1227 return -EINVAL;
1228}
1229
41ef4eb8
KO
1230typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1231 unsigned long, loff_t);
293bc982 1232typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
41ef4eb8 1233
8bc92afc
KO
1234static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1235 int rw, char __user *buf,
1236 unsigned long *nr_segs,
1237 struct iovec **iovec,
1238 bool compat)
eed4e51f
BP
1239{
1240 ssize_t ret;
1241
8bc92afc 1242 *nr_segs = kiocb->ki_nbytes;
41ef4eb8 1243
9d85cba7
JM
1244#ifdef CONFIG_COMPAT
1245 if (compat)
41ef4eb8 1246 ret = compat_rw_copy_check_uvector(rw,
8bc92afc 1247 (struct compat_iovec __user *)buf,
00fefb9c 1248 *nr_segs, UIO_FASTIOV, *iovec, iovec);
9d85cba7
JM
1249 else
1250#endif
41ef4eb8 1251 ret = rw_copy_check_uvector(rw,
8bc92afc 1252 (struct iovec __user *)buf,
00fefb9c 1253 *nr_segs, UIO_FASTIOV, *iovec, iovec);
eed4e51f 1254 if (ret < 0)
41ef4eb8 1255 return ret;
a70b52ec 1256
41ef4eb8 1257 /* ki_nbytes now reflect bytes instead of segs */
eed4e51f 1258 kiocb->ki_nbytes = ret;
41ef4eb8 1259 return 0;
eed4e51f
BP
1260}
1261
8bc92afc
KO
1262static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1263 int rw, char __user *buf,
1264 unsigned long *nr_segs,
1265 struct iovec *iovec)
eed4e51f 1266{
8bc92afc 1267 if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
41ef4eb8 1268 return -EFAULT;
a70b52ec 1269
8bc92afc
KO
1270 iovec->iov_base = buf;
1271 iovec->iov_len = kiocb->ki_nbytes;
1272 *nr_segs = 1;
eed4e51f
BP
1273 return 0;
1274}
1275
1da177e4 1276/*
2be4e7de
GZ
1277 * aio_run_iocb:
1278 * Performs the initial checks and io submission.
1da177e4 1279 */
8bc92afc
KO
1280static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1281 char __user *buf, bool compat)
1da177e4 1282{
41ef4eb8
KO
1283 struct file *file = req->ki_filp;
1284 ssize_t ret;
8bc92afc 1285 unsigned long nr_segs;
41ef4eb8
KO
1286 int rw;
1287 fmode_t mode;
1288 aio_rw_op *rw_op;
293bc982 1289 rw_iter_op *iter_op;
00fefb9c 1290 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
293bc982 1291 struct iov_iter iter;
1da177e4 1292
8bc92afc 1293 switch (opcode) {
1da177e4 1294 case IOCB_CMD_PREAD:
eed4e51f 1295 case IOCB_CMD_PREADV:
41ef4eb8
KO
1296 mode = FMODE_READ;
1297 rw = READ;
1298 rw_op = file->f_op->aio_read;
293bc982 1299 iter_op = file->f_op->read_iter;
41ef4eb8
KO
1300 goto rw_common;
1301
1302 case IOCB_CMD_PWRITE:
eed4e51f 1303 case IOCB_CMD_PWRITEV:
41ef4eb8
KO
1304 mode = FMODE_WRITE;
1305 rw = WRITE;
1306 rw_op = file->f_op->aio_write;
293bc982 1307 iter_op = file->f_op->write_iter;
41ef4eb8
KO
1308 goto rw_common;
1309rw_common:
1310 if (unlikely(!(file->f_mode & mode)))
1311 return -EBADF;
1312
293bc982 1313 if (!rw_op && !iter_op)
41ef4eb8
KO
1314 return -EINVAL;
1315
8bc92afc
KO
1316 ret = (opcode == IOCB_CMD_PREADV ||
1317 opcode == IOCB_CMD_PWRITEV)
1318 ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1319 &iovec, compat)
1320 : aio_setup_single_vector(req, rw, buf, &nr_segs,
1321 iovec);
754320d6
LY
1322 if (!ret)
1323 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
8bc92afc 1324 if (ret < 0) {
00fefb9c 1325 if (iovec != inline_vecs)
8bc92afc 1326 kfree(iovec);
41ef4eb8 1327 return ret;
8bc92afc 1328 }
41ef4eb8
KO
1329
1330 req->ki_nbytes = ret;
41ef4eb8 1331
73a7075e
KO
1332 /* XXX: move/kill - rw_verify_area()? */
1333 /* This matches the pread()/pwrite() logic */
1334 if (req->ki_pos < 0) {
1335 ret = -EINVAL;
1336 break;
1337 }
1338
1339 if (rw == WRITE)
1340 file_start_write(file);
1341
293bc982
AV
1342 if (iter_op) {
1343 iov_iter_init(&iter, rw, iovec, nr_segs, req->ki_nbytes);
1344 ret = iter_op(req, &iter);
1345 } else {
1346 ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1347 }
73a7075e
KO
1348
1349 if (rw == WRITE)
1350 file_end_write(file);
1da177e4 1351 break;
41ef4eb8 1352
1da177e4 1353 case IOCB_CMD_FDSYNC:
41ef4eb8
KO
1354 if (!file->f_op->aio_fsync)
1355 return -EINVAL;
1356
1357 ret = file->f_op->aio_fsync(req, 1);
1da177e4 1358 break;
41ef4eb8 1359
1da177e4 1360 case IOCB_CMD_FSYNC:
41ef4eb8
KO
1361 if (!file->f_op->aio_fsync)
1362 return -EINVAL;
1363
1364 ret = file->f_op->aio_fsync(req, 0);
1da177e4 1365 break;
41ef4eb8 1366
1da177e4 1367 default:
caf4167a 1368 pr_debug("EINVAL: no operation provided\n");
41ef4eb8 1369 return -EINVAL;
1da177e4
LT
1370 }
1371
00fefb9c 1372 if (iovec != inline_vecs)
8bc92afc
KO
1373 kfree(iovec);
1374
41ef4eb8
KO
1375 if (ret != -EIOCBQUEUED) {
1376 /*
1377 * There's no easy way to restart the syscall since other AIO's
1378 * may be already running. Just fail this IO with EINTR.
1379 */
1380 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1381 ret == -ERESTARTNOHAND ||
1382 ret == -ERESTART_RESTARTBLOCK))
1383 ret = -EINTR;
1384 aio_complete(req, ret, 0);
1385 }
1da177e4
LT
1386
1387 return 0;
1388}
1389
d5470b59 1390static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
a1c8eae7 1391 struct iocb *iocb, bool compat)
1da177e4
LT
1392{
1393 struct kiocb *req;
1da177e4
LT
1394 ssize_t ret;
1395
1396 /* enforce forwards compatibility on users */
9c3060be 1397 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
caf4167a 1398 pr_debug("EINVAL: reserve field set\n");
1da177e4
LT
1399 return -EINVAL;
1400 }
1401
1402 /* prevent overflows */
1403 if (unlikely(
1404 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1405 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1406 ((ssize_t)iocb->aio_nbytes < 0)
1407 )) {
1408 pr_debug("EINVAL: io_submit: overflow check\n");
1409 return -EINVAL;
1410 }
1411
41ef4eb8 1412 req = aio_get_req(ctx);
1d98ebfc 1413 if (unlikely(!req))
1da177e4 1414 return -EAGAIN;
1d98ebfc
KO
1415
1416 req->ki_filp = fget(iocb->aio_fildes);
1417 if (unlikely(!req->ki_filp)) {
1418 ret = -EBADF;
1419 goto out_put_req;
1da177e4 1420 }
1d98ebfc 1421
9c3060be
DL
1422 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1423 /*
1424 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1425 * instance of the file* now. The file descriptor must be
1426 * an eventfd() fd, and will be signaled for each completed
1427 * event using the eventfd_signal() function.
1428 */
13389010 1429 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
801678c5 1430 if (IS_ERR(req->ki_eventfd)) {
9c3060be 1431 ret = PTR_ERR(req->ki_eventfd);
87c3a86e 1432 req->ki_eventfd = NULL;
9c3060be
DL
1433 goto out_put_req;
1434 }
1435 }
1da177e4 1436
8a660890 1437 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1da177e4 1438 if (unlikely(ret)) {
caf4167a 1439 pr_debug("EFAULT: aio_key\n");
1da177e4
LT
1440 goto out_put_req;
1441 }
1442
1443 req->ki_obj.user = user_iocb;
1444 req->ki_user_data = iocb->aio_data;
1445 req->ki_pos = iocb->aio_offset;
73a7075e 1446 req->ki_nbytes = iocb->aio_nbytes;
1da177e4 1447
8bc92afc
KO
1448 ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1449 (char __user *)(unsigned long)iocb->aio_buf,
1450 compat);
41003a7b 1451 if (ret)
7137c6bd 1452 goto out_put_req;
41003a7b 1453
1da177e4 1454 return 0;
1da177e4 1455out_put_req:
e1bdd5f2 1456 put_reqs_available(ctx, 1);
e34ecee2 1457 percpu_ref_put(&ctx->reqs);
57282d8f 1458 kiocb_free(req);
1da177e4
LT
1459 return ret;
1460}
1461
9d85cba7
JM
1462long do_io_submit(aio_context_t ctx_id, long nr,
1463 struct iocb __user *__user *iocbpp, bool compat)
1da177e4
LT
1464{
1465 struct kioctx *ctx;
1466 long ret = 0;
080d676d 1467 int i = 0;
9f5b9425 1468 struct blk_plug plug;
1da177e4
LT
1469
1470 if (unlikely(nr < 0))
1471 return -EINVAL;
1472
75e1c70f
JM
1473 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1474 nr = LONG_MAX/sizeof(*iocbpp);
1475
1da177e4
LT
1476 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1477 return -EFAULT;
1478
1479 ctx = lookup_ioctx(ctx_id);
1480 if (unlikely(!ctx)) {
caf4167a 1481 pr_debug("EINVAL: invalid context id\n");
1da177e4
LT
1482 return -EINVAL;
1483 }
1484
9f5b9425
SL
1485 blk_start_plug(&plug);
1486
1da177e4
LT
1487 /*
1488 * AKPM: should this return a partial result if some of the IOs were
1489 * successfully submitted?
1490 */
1491 for (i=0; i<nr; i++) {
1492 struct iocb __user *user_iocb;
1493 struct iocb tmp;
1494
1495 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1496 ret = -EFAULT;
1497 break;
1498 }
1499
1500 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1501 ret = -EFAULT;
1502 break;
1503 }
1504
a1c8eae7 1505 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1da177e4
LT
1506 if (ret)
1507 break;
1508 }
9f5b9425 1509 blk_finish_plug(&plug);
1da177e4 1510
723be6e3 1511 percpu_ref_put(&ctx->users);
1da177e4
LT
1512 return i ? i : ret;
1513}
1514
9d85cba7
JM
1515/* sys_io_submit:
1516 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1517 * the number of iocbs queued. May return -EINVAL if the aio_context
1518 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1519 * *iocbpp[0] is not properly initialized, if the operation specified
1520 * is invalid for the file descriptor in the iocb. May fail with
1521 * -EFAULT if any of the data structures point to invalid data. May
1522 * fail with -EBADF if the file descriptor specified in the first
1523 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1524 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1525 * fail with -ENOSYS if not implemented.
1526 */
1527SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1528 struct iocb __user * __user *, iocbpp)
1529{
1530 return do_io_submit(ctx_id, nr, iocbpp, 0);
1531}
1532
1da177e4
LT
1533/* lookup_kiocb
1534 * Finds a given iocb for cancellation.
1da177e4 1535 */
25ee7e38
AB
1536static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1537 u32 key)
1da177e4
LT
1538{
1539 struct list_head *pos;
d00689af
ZB
1540
1541 assert_spin_locked(&ctx->ctx_lock);
1542
8a660890
KO
1543 if (key != KIOCB_KEY)
1544 return NULL;
1545
1da177e4
LT
1546 /* TODO: use a hash or array, this sucks. */
1547 list_for_each(pos, &ctx->active_reqs) {
1548 struct kiocb *kiocb = list_kiocb(pos);
8a660890 1549 if (kiocb->ki_obj.user == iocb)
1da177e4
LT
1550 return kiocb;
1551 }
1552 return NULL;
1553}
1554
1555/* sys_io_cancel:
1556 * Attempts to cancel an iocb previously passed to io_submit. If
1557 * the operation is successfully cancelled, the resulting event is
1558 * copied into the memory pointed to by result without being placed
1559 * into the completion queue and 0 is returned. May fail with
1560 * -EFAULT if any of the data structures pointed to are invalid.
1561 * May fail with -EINVAL if aio_context specified by ctx_id is
1562 * invalid. May fail with -EAGAIN if the iocb specified was not
1563 * cancelled. Will fail with -ENOSYS if not implemented.
1564 */
002c8976
HC
1565SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1566 struct io_event __user *, result)
1da177e4 1567{
1da177e4
LT
1568 struct kioctx *ctx;
1569 struct kiocb *kiocb;
1570 u32 key;
1571 int ret;
1572
1573 ret = get_user(key, &iocb->aio_key);
1574 if (unlikely(ret))
1575 return -EFAULT;
1576
1577 ctx = lookup_ioctx(ctx_id);
1578 if (unlikely(!ctx))
1579 return -EINVAL;
1580
1581 spin_lock_irq(&ctx->ctx_lock);
906b973c 1582
1da177e4 1583 kiocb = lookup_kiocb(ctx, iocb, key);
906b973c 1584 if (kiocb)
d52a8f9e 1585 ret = kiocb_cancel(kiocb);
906b973c
KO
1586 else
1587 ret = -EINVAL;
1588
1da177e4
LT
1589 spin_unlock_irq(&ctx->ctx_lock);
1590
906b973c 1591 if (!ret) {
bec68faa
KO
1592 /*
1593 * The result argument is no longer used - the io_event is
1594 * always delivered via the ring buffer. -EINPROGRESS indicates
1595 * cancellation is progress:
906b973c 1596 */
bec68faa 1597 ret = -EINPROGRESS;
906b973c 1598 }
1da177e4 1599
723be6e3 1600 percpu_ref_put(&ctx->users);
1da177e4
LT
1601
1602 return ret;
1603}
1604
1605/* io_getevents:
1606 * Attempts to read at least min_nr events and up to nr events from
642b5123
ST
1607 * the completion queue for the aio_context specified by ctx_id. If
1608 * it succeeds, the number of read events is returned. May fail with
1609 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1610 * out of range, if timeout is out of range. May fail with -EFAULT
1611 * if any of the memory specified is invalid. May return 0 or
1612 * < min_nr if the timeout specified by timeout has elapsed
1613 * before sufficient events are available, where timeout == NULL
1614 * specifies an infinite timeout. Note that the timeout pointed to by
6900807c 1615 * timeout is relative. Will fail with -ENOSYS if not implemented.
1da177e4 1616 */
002c8976
HC
1617SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1618 long, min_nr,
1619 long, nr,
1620 struct io_event __user *, events,
1621 struct timespec __user *, timeout)
1da177e4
LT
1622{
1623 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1624 long ret = -EINVAL;
1625
1626 if (likely(ioctx)) {
2e410255 1627 if (likely(min_nr <= nr && min_nr >= 0))
1da177e4 1628 ret = read_events(ioctx, min_nr, nr, events, timeout);
723be6e3 1629 percpu_ref_put(&ioctx->users);
1da177e4 1630 }
1da177e4
LT
1631 return ret;
1632}
This page took 0.822273 seconds and 5 git commands to generate.