Commit | Line | Data |
---|---|---|
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 | |
52 | struct 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 |
70 | struct kioctx_table { |
71 | struct rcu_head rcu; | |
72 | unsigned nr; | |
73 | struct kioctx *table[]; | |
74 | }; | |
75 | ||
e1bdd5f2 KO |
76 | struct kioctx_cpu { |
77 | unsigned reqs_available; | |
78 | }; | |
79 | ||
4e179bca | 80 | struct 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 KO |
114 | |
115 | struct { | |
34e83fc6 KO |
116 | /* |
117 | * This counts the number of available slots in the ringbuffer, | |
118 | * so we avoid overflowing it: it's decremented (if positive) | |
119 | * when allocating a kiocb and incremented when the resulting | |
120 | * io_event is pulled off the ringbuffer. | |
e1bdd5f2 KO |
121 | * |
122 | * We batch accesses to it with a percpu version. | |
34e83fc6 KO |
123 | */ |
124 | atomic_t reqs_available; | |
4e23bcae KO |
125 | } ____cacheline_aligned_in_smp; |
126 | ||
127 | struct { | |
128 | spinlock_t ctx_lock; | |
129 | struct list_head active_reqs; /* used for cancellation */ | |
130 | } ____cacheline_aligned_in_smp; | |
131 | ||
58c85dc2 KO |
132 | struct { |
133 | struct mutex ring_lock; | |
4e23bcae KO |
134 | wait_queue_head_t wait; |
135 | } ____cacheline_aligned_in_smp; | |
58c85dc2 KO |
136 | |
137 | struct { | |
138 | unsigned tail; | |
139 | spinlock_t completion_lock; | |
4e23bcae | 140 | } ____cacheline_aligned_in_smp; |
58c85dc2 KO |
141 | |
142 | struct page *internal_pages[AIO_RING_PAGES]; | |
36bc08cc | 143 | struct file *aio_ring_file; |
db446a08 BL |
144 | |
145 | unsigned id; | |
4e179bca KO |
146 | }; |
147 | ||
1da177e4 | 148 | /*------ sysctl variables----*/ |
d55b5fda ZB |
149 | static DEFINE_SPINLOCK(aio_nr_lock); |
150 | unsigned long aio_nr; /* current system wide number of aio requests */ | |
151 | unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */ | |
1da177e4 LT |
152 | /*----end sysctl variables---*/ |
153 | ||
e18b890b CL |
154 | static struct kmem_cache *kiocb_cachep; |
155 | static struct kmem_cache *kioctx_cachep; | |
1da177e4 | 156 | |
71ad7490 BL |
157 | static struct vfsmount *aio_mnt; |
158 | ||
159 | static const struct file_operations aio_ring_fops; | |
160 | static const struct address_space_operations aio_ctx_aops; | |
161 | ||
162 | static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages) | |
163 | { | |
164 | struct qstr this = QSTR_INIT("[aio]", 5); | |
165 | struct file *file; | |
166 | struct path path; | |
167 | struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb); | |
7f62656b DC |
168 | if (IS_ERR(inode)) |
169 | return ERR_CAST(inode); | |
71ad7490 BL |
170 | |
171 | inode->i_mapping->a_ops = &aio_ctx_aops; | |
172 | inode->i_mapping->private_data = ctx; | |
173 | inode->i_size = PAGE_SIZE * nr_pages; | |
174 | ||
175 | path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this); | |
176 | if (!path.dentry) { | |
177 | iput(inode); | |
178 | return ERR_PTR(-ENOMEM); | |
179 | } | |
180 | path.mnt = mntget(aio_mnt); | |
181 | ||
182 | d_instantiate(path.dentry, inode); | |
183 | file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops); | |
184 | if (IS_ERR(file)) { | |
185 | path_put(&path); | |
186 | return file; | |
187 | } | |
188 | ||
189 | file->f_flags = O_RDWR; | |
190 | file->private_data = ctx; | |
191 | return file; | |
192 | } | |
193 | ||
194 | static struct dentry *aio_mount(struct file_system_type *fs_type, | |
195 | int flags, const char *dev_name, void *data) | |
196 | { | |
197 | static const struct dentry_operations ops = { | |
198 | .d_dname = simple_dname, | |
199 | }; | |
200 | return mount_pseudo(fs_type, "aio:", NULL, &ops, 0xa10a10a1); | |
201 | } | |
202 | ||
1da177e4 LT |
203 | /* aio_setup |
204 | * Creates the slab caches used by the aio routines, panic on | |
205 | * failure as this is done early during the boot sequence. | |
206 | */ | |
207 | static int __init aio_setup(void) | |
208 | { | |
71ad7490 BL |
209 | static struct file_system_type aio_fs = { |
210 | .name = "aio", | |
211 | .mount = aio_mount, | |
212 | .kill_sb = kill_anon_super, | |
213 | }; | |
214 | aio_mnt = kern_mount(&aio_fs); | |
215 | if (IS_ERR(aio_mnt)) | |
216 | panic("Failed to create aio fs mount."); | |
217 | ||
0a31bd5f CL |
218 | kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC); |
219 | kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC); | |
1da177e4 | 220 | |
caf4167a | 221 | pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page)); |
1da177e4 LT |
222 | |
223 | return 0; | |
224 | } | |
385773e0 | 225 | __initcall(aio_setup); |
1da177e4 | 226 | |
5e9ae2e5 BL |
227 | static void put_aio_ring_file(struct kioctx *ctx) |
228 | { | |
229 | struct file *aio_ring_file = ctx->aio_ring_file; | |
230 | if (aio_ring_file) { | |
231 | truncate_setsize(aio_ring_file->f_inode, 0); | |
232 | ||
233 | /* Prevent further access to the kioctx from migratepages */ | |
234 | spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock); | |
235 | aio_ring_file->f_inode->i_mapping->private_data = NULL; | |
236 | ctx->aio_ring_file = NULL; | |
237 | spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock); | |
238 | ||
239 | fput(aio_ring_file); | |
240 | } | |
241 | } | |
242 | ||
1da177e4 LT |
243 | static void aio_free_ring(struct kioctx *ctx) |
244 | { | |
36bc08cc | 245 | int i; |
1da177e4 | 246 | |
fa8a53c3 BL |
247 | /* Disconnect the kiotx from the ring file. This prevents future |
248 | * accesses to the kioctx from page migration. | |
249 | */ | |
250 | put_aio_ring_file(ctx); | |
251 | ||
36bc08cc | 252 | for (i = 0; i < ctx->nr_pages; i++) { |
8e321fef | 253 | struct page *page; |
36bc08cc GZ |
254 | pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i, |
255 | page_count(ctx->ring_pages[i])); | |
8e321fef BL |
256 | page = ctx->ring_pages[i]; |
257 | if (!page) | |
258 | continue; | |
259 | ctx->ring_pages[i] = NULL; | |
260 | put_page(page); | |
36bc08cc | 261 | } |
1da177e4 | 262 | |
ddb8c45b | 263 | if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) { |
58c85dc2 | 264 | kfree(ctx->ring_pages); |
ddb8c45b SL |
265 | ctx->ring_pages = NULL; |
266 | } | |
36bc08cc GZ |
267 | } |
268 | ||
269 | static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma) | |
270 | { | |
271 | vma->vm_ops = &generic_file_vm_ops; | |
272 | return 0; | |
273 | } | |
274 | ||
275 | static const struct file_operations aio_ring_fops = { | |
276 | .mmap = aio_ring_mmap, | |
277 | }; | |
278 | ||
279 | static int aio_set_page_dirty(struct page *page) | |
280 | { | |
281 | return 0; | |
282 | } | |
283 | ||
0c45355f | 284 | #if IS_ENABLED(CONFIG_MIGRATION) |
36bc08cc GZ |
285 | static int aio_migratepage(struct address_space *mapping, struct page *new, |
286 | struct page *old, enum migrate_mode mode) | |
287 | { | |
5e9ae2e5 | 288 | struct kioctx *ctx; |
36bc08cc | 289 | unsigned long flags; |
fa8a53c3 | 290 | pgoff_t idx; |
36bc08cc GZ |
291 | int rc; |
292 | ||
8e321fef BL |
293 | rc = 0; |
294 | ||
fa8a53c3 | 295 | /* mapping->private_lock here protects against the kioctx teardown. */ |
8e321fef BL |
296 | spin_lock(&mapping->private_lock); |
297 | ctx = mapping->private_data; | |
fa8a53c3 BL |
298 | if (!ctx) { |
299 | rc = -EINVAL; | |
300 | goto out; | |
301 | } | |
302 | ||
303 | /* The ring_lock mutex. The prevents aio_read_events() from writing | |
304 | * to the ring's head, and prevents page migration from mucking in | |
305 | * a partially initialized kiotx. | |
306 | */ | |
307 | if (!mutex_trylock(&ctx->ring_lock)) { | |
308 | rc = -EAGAIN; | |
309 | goto out; | |
310 | } | |
311 | ||
312 | idx = old->index; | |
313 | if (idx < (pgoff_t)ctx->nr_pages) { | |
314 | /* Make sure the old page hasn't already been changed */ | |
315 | if (ctx->ring_pages[idx] != old) | |
316 | rc = -EAGAIN; | |
8e321fef BL |
317 | } else |
318 | rc = -EINVAL; | |
8e321fef BL |
319 | |
320 | if (rc != 0) | |
fa8a53c3 | 321 | goto out_unlock; |
8e321fef | 322 | |
36bc08cc GZ |
323 | /* Writeback must be complete */ |
324 | BUG_ON(PageWriteback(old)); | |
8e321fef | 325 | get_page(new); |
36bc08cc | 326 | |
8e321fef | 327 | rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1); |
36bc08cc | 328 | if (rc != MIGRATEPAGE_SUCCESS) { |
8e321fef | 329 | put_page(new); |
fa8a53c3 | 330 | goto out_unlock; |
36bc08cc GZ |
331 | } |
332 | ||
fa8a53c3 BL |
333 | /* Take completion_lock to prevent other writes to the ring buffer |
334 | * while the old page is copied to the new. This prevents new | |
335 | * events from being lost. | |
5e9ae2e5 | 336 | */ |
fa8a53c3 BL |
337 | spin_lock_irqsave(&ctx->completion_lock, flags); |
338 | migrate_page_copy(new, old); | |
339 | BUG_ON(ctx->ring_pages[idx] != old); | |
340 | ctx->ring_pages[idx] = new; | |
341 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
36bc08cc | 342 | |
fa8a53c3 BL |
343 | /* The old page is no longer accessible. */ |
344 | put_page(old); | |
8e321fef | 345 | |
fa8a53c3 BL |
346 | out_unlock: |
347 | mutex_unlock(&ctx->ring_lock); | |
348 | out: | |
349 | spin_unlock(&mapping->private_lock); | |
36bc08cc | 350 | return rc; |
1da177e4 | 351 | } |
0c45355f | 352 | #endif |
1da177e4 | 353 | |
36bc08cc GZ |
354 | static const struct address_space_operations aio_ctx_aops = { |
355 | .set_page_dirty = aio_set_page_dirty, | |
0c45355f | 356 | #if IS_ENABLED(CONFIG_MIGRATION) |
36bc08cc | 357 | .migratepage = aio_migratepage, |
0c45355f | 358 | #endif |
36bc08cc GZ |
359 | }; |
360 | ||
1da177e4 LT |
361 | static int aio_setup_ring(struct kioctx *ctx) |
362 | { | |
363 | struct aio_ring *ring; | |
1da177e4 | 364 | unsigned nr_events = ctx->max_reqs; |
41003a7b | 365 | struct mm_struct *mm = current->mm; |
3dc9acb6 | 366 | unsigned long size, unused; |
1da177e4 | 367 | int nr_pages; |
36bc08cc GZ |
368 | int i; |
369 | struct file *file; | |
1da177e4 LT |
370 | |
371 | /* Compensate for the ring buffer's head/tail overlap entry */ | |
372 | nr_events += 2; /* 1 is required, 2 for good luck */ | |
373 | ||
374 | size = sizeof(struct aio_ring); | |
375 | size += sizeof(struct io_event) * nr_events; | |
1da177e4 | 376 | |
36bc08cc | 377 | nr_pages = PFN_UP(size); |
1da177e4 LT |
378 | if (nr_pages < 0) |
379 | return -EINVAL; | |
380 | ||
71ad7490 | 381 | file = aio_private_file(ctx, nr_pages); |
36bc08cc GZ |
382 | if (IS_ERR(file)) { |
383 | ctx->aio_ring_file = NULL; | |
fa8a53c3 | 384 | return -ENOMEM; |
36bc08cc GZ |
385 | } |
386 | ||
3dc9acb6 LT |
387 | ctx->aio_ring_file = file; |
388 | nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) | |
389 | / sizeof(struct io_event); | |
390 | ||
391 | ctx->ring_pages = ctx->internal_pages; | |
392 | if (nr_pages > AIO_RING_PAGES) { | |
393 | ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *), | |
394 | GFP_KERNEL); | |
395 | if (!ctx->ring_pages) { | |
396 | put_aio_ring_file(ctx); | |
397 | return -ENOMEM; | |
398 | } | |
399 | } | |
400 | ||
36bc08cc GZ |
401 | for (i = 0; i < nr_pages; i++) { |
402 | struct page *page; | |
403 | page = find_or_create_page(file->f_inode->i_mapping, | |
404 | i, GFP_HIGHUSER | __GFP_ZERO); | |
405 | if (!page) | |
406 | break; | |
407 | pr_debug("pid(%d) page[%d]->count=%d\n", | |
408 | current->pid, i, page_count(page)); | |
409 | SetPageUptodate(page); | |
410 | SetPageDirty(page); | |
411 | unlock_page(page); | |
3dc9acb6 LT |
412 | |
413 | ctx->ring_pages[i] = page; | |
36bc08cc | 414 | } |
3dc9acb6 | 415 | ctx->nr_pages = i; |
1da177e4 | 416 | |
3dc9acb6 LT |
417 | if (unlikely(i != nr_pages)) { |
418 | aio_free_ring(ctx); | |
fa8a53c3 | 419 | return -ENOMEM; |
1da177e4 LT |
420 | } |
421 | ||
58c85dc2 KO |
422 | ctx->mmap_size = nr_pages * PAGE_SIZE; |
423 | pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size); | |
36bc08cc | 424 | |
41003a7b | 425 | down_write(&mm->mmap_sem); |
36bc08cc GZ |
426 | ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size, |
427 | PROT_READ | PROT_WRITE, | |
3dc9acb6 LT |
428 | MAP_SHARED, 0, &unused); |
429 | up_write(&mm->mmap_sem); | |
58c85dc2 | 430 | if (IS_ERR((void *)ctx->mmap_base)) { |
58c85dc2 | 431 | ctx->mmap_size = 0; |
1da177e4 | 432 | aio_free_ring(ctx); |
fa8a53c3 | 433 | return -ENOMEM; |
1da177e4 LT |
434 | } |
435 | ||
58c85dc2 | 436 | pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base); |
d6c355c7 | 437 | |
58c85dc2 KO |
438 | ctx->user_id = ctx->mmap_base; |
439 | ctx->nr_events = nr_events; /* trusted copy */ | |
1da177e4 | 440 | |
58c85dc2 | 441 | ring = kmap_atomic(ctx->ring_pages[0]); |
1da177e4 | 442 | ring->nr = nr_events; /* user copy */ |
db446a08 | 443 | ring->id = ~0U; |
1da177e4 LT |
444 | ring->head = ring->tail = 0; |
445 | ring->magic = AIO_RING_MAGIC; | |
446 | ring->compat_features = AIO_RING_COMPAT_FEATURES; | |
447 | ring->incompat_features = AIO_RING_INCOMPAT_FEATURES; | |
448 | ring->header_length = sizeof(struct aio_ring); | |
e8e3c3d6 | 449 | kunmap_atomic(ring); |
58c85dc2 | 450 | flush_dcache_page(ctx->ring_pages[0]); |
1da177e4 LT |
451 | |
452 | return 0; | |
453 | } | |
454 | ||
1da177e4 LT |
455 | #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event)) |
456 | #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event)) | |
457 | #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE) | |
458 | ||
0460fef2 KO |
459 | void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel) |
460 | { | |
461 | struct kioctx *ctx = req->ki_ctx; | |
462 | unsigned long flags; | |
463 | ||
464 | spin_lock_irqsave(&ctx->ctx_lock, flags); | |
465 | ||
466 | if (!req->ki_list.next) | |
467 | list_add(&req->ki_list, &ctx->active_reqs); | |
468 | ||
469 | req->ki_cancel = cancel; | |
470 | ||
471 | spin_unlock_irqrestore(&ctx->ctx_lock, flags); | |
472 | } | |
473 | EXPORT_SYMBOL(kiocb_set_cancel_fn); | |
474 | ||
bec68faa | 475 | static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb) |
906b973c | 476 | { |
0460fef2 | 477 | kiocb_cancel_fn *old, *cancel; |
906b973c | 478 | |
0460fef2 KO |
479 | /* |
480 | * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it | |
481 | * actually has a cancel function, hence the cmpxchg() | |
482 | */ | |
483 | ||
484 | cancel = ACCESS_ONCE(kiocb->ki_cancel); | |
485 | do { | |
486 | if (!cancel || cancel == KIOCB_CANCELLED) | |
57282d8f | 487 | return -EINVAL; |
906b973c | 488 | |
0460fef2 KO |
489 | old = cancel; |
490 | cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED); | |
491 | } while (cancel != old); | |
906b973c | 492 | |
57282d8f | 493 | return cancel(kiocb); |
906b973c KO |
494 | } |
495 | ||
e34ecee2 | 496 | static void free_ioctx(struct work_struct *work) |
36f55889 | 497 | { |
e34ecee2 | 498 | struct kioctx *ctx = container_of(work, struct kioctx, free_work); |
e1bdd5f2 | 499 | |
e34ecee2 | 500 | pr_debug("freeing %p\n", ctx); |
e1bdd5f2 | 501 | |
e34ecee2 | 502 | aio_free_ring(ctx); |
e1bdd5f2 | 503 | free_percpu(ctx->cpu); |
36f55889 KO |
504 | kmem_cache_free(kioctx_cachep, ctx); |
505 | } | |
506 | ||
e34ecee2 KO |
507 | static void free_ioctx_reqs(struct percpu_ref *ref) |
508 | { | |
509 | struct kioctx *ctx = container_of(ref, struct kioctx, reqs); | |
510 | ||
511 | INIT_WORK(&ctx->free_work, free_ioctx); | |
512 | schedule_work(&ctx->free_work); | |
513 | } | |
514 | ||
36f55889 KO |
515 | /* |
516 | * When this function runs, the kioctx has been removed from the "hash table" | |
517 | * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted - | |
518 | * now it's safe to cancel any that need to be. | |
519 | */ | |
e34ecee2 | 520 | static void free_ioctx_users(struct percpu_ref *ref) |
36f55889 | 521 | { |
e34ecee2 | 522 | struct kioctx *ctx = container_of(ref, struct kioctx, users); |
36f55889 KO |
523 | struct kiocb *req; |
524 | ||
525 | spin_lock_irq(&ctx->ctx_lock); | |
526 | ||
527 | while (!list_empty(&ctx->active_reqs)) { | |
528 | req = list_first_entry(&ctx->active_reqs, | |
529 | struct kiocb, ki_list); | |
530 | ||
531 | list_del_init(&req->ki_list); | |
bec68faa | 532 | kiocb_cancel(ctx, req); |
36f55889 KO |
533 | } |
534 | ||
535 | spin_unlock_irq(&ctx->ctx_lock); | |
536 | ||
e34ecee2 KO |
537 | percpu_ref_kill(&ctx->reqs); |
538 | percpu_ref_put(&ctx->reqs); | |
36f55889 KO |
539 | } |
540 | ||
db446a08 BL |
541 | static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm) |
542 | { | |
543 | unsigned i, new_nr; | |
544 | struct kioctx_table *table, *old; | |
545 | struct aio_ring *ring; | |
546 | ||
547 | spin_lock(&mm->ioctx_lock); | |
d9b2c871 | 548 | rcu_read_lock(); |
77d30b14 | 549 | table = rcu_dereference(mm->ioctx_table); |
db446a08 BL |
550 | |
551 | while (1) { | |
552 | if (table) | |
553 | for (i = 0; i < table->nr; i++) | |
554 | if (!table->table[i]) { | |
555 | ctx->id = i; | |
556 | table->table[i] = ctx; | |
d9b2c871 | 557 | rcu_read_unlock(); |
db446a08 BL |
558 | spin_unlock(&mm->ioctx_lock); |
559 | ||
fa8a53c3 BL |
560 | /* While kioctx setup is in progress, |
561 | * we are protected from page migration | |
562 | * changes ring_pages by ->ring_lock. | |
563 | */ | |
db446a08 BL |
564 | ring = kmap_atomic(ctx->ring_pages[0]); |
565 | ring->id = ctx->id; | |
566 | kunmap_atomic(ring); | |
567 | return 0; | |
568 | } | |
569 | ||
570 | new_nr = (table ? table->nr : 1) * 4; | |
571 | ||
d9b2c871 | 572 | rcu_read_unlock(); |
db446a08 BL |
573 | spin_unlock(&mm->ioctx_lock); |
574 | ||
575 | table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) * | |
576 | new_nr, GFP_KERNEL); | |
577 | if (!table) | |
578 | return -ENOMEM; | |
579 | ||
580 | table->nr = new_nr; | |
581 | ||
582 | spin_lock(&mm->ioctx_lock); | |
d9b2c871 | 583 | rcu_read_lock(); |
77d30b14 | 584 | old = rcu_dereference(mm->ioctx_table); |
db446a08 BL |
585 | |
586 | if (!old) { | |
587 | rcu_assign_pointer(mm->ioctx_table, table); | |
588 | } else if (table->nr > old->nr) { | |
589 | memcpy(table->table, old->table, | |
590 | old->nr * sizeof(struct kioctx *)); | |
591 | ||
592 | rcu_assign_pointer(mm->ioctx_table, table); | |
593 | kfree_rcu(old, rcu); | |
594 | } else { | |
595 | kfree(table); | |
596 | table = old; | |
597 | } | |
598 | } | |
599 | } | |
600 | ||
e34ecee2 KO |
601 | static void aio_nr_sub(unsigned nr) |
602 | { | |
603 | spin_lock(&aio_nr_lock); | |
604 | if (WARN_ON(aio_nr - nr > aio_nr)) | |
605 | aio_nr = 0; | |
606 | else | |
607 | aio_nr -= nr; | |
608 | spin_unlock(&aio_nr_lock); | |
609 | } | |
610 | ||
1da177e4 LT |
611 | /* ioctx_alloc |
612 | * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed. | |
613 | */ | |
614 | static struct kioctx *ioctx_alloc(unsigned nr_events) | |
615 | { | |
41003a7b | 616 | struct mm_struct *mm = current->mm; |
1da177e4 | 617 | struct kioctx *ctx; |
e23754f8 | 618 | int err = -ENOMEM; |
1da177e4 | 619 | |
e1bdd5f2 KO |
620 | /* |
621 | * We keep track of the number of available ringbuffer slots, to prevent | |
622 | * overflow (reqs_available), and we also use percpu counters for this. | |
623 | * | |
624 | * So since up to half the slots might be on other cpu's percpu counters | |
625 | * and unavailable, double nr_events so userspace sees what they | |
626 | * expected: additionally, we move req_batch slots to/from percpu | |
627 | * counters at a time, so make sure that isn't 0: | |
628 | */ | |
629 | nr_events = max(nr_events, num_possible_cpus() * 4); | |
630 | nr_events *= 2; | |
631 | ||
1da177e4 LT |
632 | /* Prevent overflows */ |
633 | if ((nr_events > (0x10000000U / sizeof(struct io_event))) || | |
634 | (nr_events > (0x10000000U / sizeof(struct kiocb)))) { | |
635 | pr_debug("ENOMEM: nr_events too high\n"); | |
636 | return ERR_PTR(-EINVAL); | |
637 | } | |
638 | ||
4cd81c3d | 639 | if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL)) |
1da177e4 LT |
640 | return ERR_PTR(-EAGAIN); |
641 | ||
c3762229 | 642 | ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL); |
1da177e4 LT |
643 | if (!ctx) |
644 | return ERR_PTR(-ENOMEM); | |
645 | ||
1da177e4 | 646 | ctx->max_reqs = nr_events; |
1da177e4 | 647 | |
1da177e4 | 648 | spin_lock_init(&ctx->ctx_lock); |
0460fef2 | 649 | spin_lock_init(&ctx->completion_lock); |
58c85dc2 | 650 | mutex_init(&ctx->ring_lock); |
fa8a53c3 BL |
651 | /* Protect against page migration throughout kiotx setup by keeping |
652 | * the ring_lock mutex held until setup is complete. */ | |
653 | mutex_lock(&ctx->ring_lock); | |
1da177e4 LT |
654 | init_waitqueue_head(&ctx->wait); |
655 | ||
656 | INIT_LIST_HEAD(&ctx->active_reqs); | |
1da177e4 | 657 | |
fa8a53c3 BL |
658 | if (percpu_ref_init(&ctx->users, free_ioctx_users)) |
659 | goto err; | |
660 | ||
661 | if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs)) | |
662 | goto err; | |
663 | ||
e1bdd5f2 KO |
664 | ctx->cpu = alloc_percpu(struct kioctx_cpu); |
665 | if (!ctx->cpu) | |
e34ecee2 | 666 | goto err; |
1da177e4 | 667 | |
fa8a53c3 BL |
668 | err = aio_setup_ring(ctx); |
669 | if (err < 0) | |
e34ecee2 | 670 | goto err; |
e1bdd5f2 | 671 | |
34e83fc6 | 672 | atomic_set(&ctx->reqs_available, ctx->nr_events - 1); |
e1bdd5f2 | 673 | ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4); |
6878ea72 BL |
674 | if (ctx->req_batch < 1) |
675 | ctx->req_batch = 1; | |
34e83fc6 | 676 | |
1da177e4 | 677 | /* limit the number of system wide aios */ |
9fa1cb39 | 678 | spin_lock(&aio_nr_lock); |
4cd81c3d | 679 | if (aio_nr + nr_events > (aio_max_nr * 2UL) || |
2dd542b7 | 680 | aio_nr + nr_events < aio_nr) { |
9fa1cb39 | 681 | spin_unlock(&aio_nr_lock); |
e34ecee2 | 682 | err = -EAGAIN; |
d1b94327 | 683 | goto err_ctx; |
2dd542b7 AV |
684 | } |
685 | aio_nr += ctx->max_reqs; | |
9fa1cb39 | 686 | spin_unlock(&aio_nr_lock); |
1da177e4 | 687 | |
1881686f BL |
688 | percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */ |
689 | percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */ | |
723be6e3 | 690 | |
da90382c BL |
691 | err = ioctx_add_table(ctx, mm); |
692 | if (err) | |
e34ecee2 | 693 | goto err_cleanup; |
da90382c | 694 | |
fa8a53c3 BL |
695 | /* Release the ring_lock mutex now that all setup is complete. */ |
696 | mutex_unlock(&ctx->ring_lock); | |
697 | ||
caf4167a | 698 | pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n", |
58c85dc2 | 699 | ctx, ctx->user_id, mm, ctx->nr_events); |
1da177e4 LT |
700 | return ctx; |
701 | ||
e34ecee2 KO |
702 | err_cleanup: |
703 | aio_nr_sub(ctx->max_reqs); | |
d1b94327 GZ |
704 | err_ctx: |
705 | aio_free_ring(ctx); | |
e34ecee2 | 706 | err: |
fa8a53c3 | 707 | mutex_unlock(&ctx->ring_lock); |
e1bdd5f2 | 708 | free_percpu(ctx->cpu); |
e34ecee2 | 709 | free_percpu(ctx->reqs.pcpu_count); |
723be6e3 | 710 | free_percpu(ctx->users.pcpu_count); |
1da177e4 | 711 | kmem_cache_free(kioctx_cachep, ctx); |
caf4167a | 712 | pr_debug("error allocating ioctx %d\n", err); |
e23754f8 | 713 | return ERR_PTR(err); |
1da177e4 LT |
714 | } |
715 | ||
36f55889 KO |
716 | /* kill_ioctx |
717 | * Cancels all outstanding aio requests on an aio context. Used | |
718 | * when the processes owning a context have all exited to encourage | |
719 | * the rapid destruction of the kioctx. | |
720 | */ | |
db446a08 | 721 | static void kill_ioctx(struct mm_struct *mm, struct kioctx *ctx) |
36f55889 KO |
722 | { |
723 | if (!atomic_xchg(&ctx->dead, 1)) { | |
db446a08 BL |
724 | struct kioctx_table *table; |
725 | ||
726 | spin_lock(&mm->ioctx_lock); | |
d9b2c871 | 727 | rcu_read_lock(); |
77d30b14 | 728 | table = rcu_dereference(mm->ioctx_table); |
db446a08 BL |
729 | |
730 | WARN_ON(ctx != table->table[ctx->id]); | |
731 | table->table[ctx->id] = NULL; | |
d9b2c871 | 732 | rcu_read_unlock(); |
db446a08 BL |
733 | spin_unlock(&mm->ioctx_lock); |
734 | ||
723be6e3 KO |
735 | /* percpu_ref_kill() will do the necessary call_rcu() */ |
736 | wake_up_all(&ctx->wait); | |
dee11c23 | 737 | |
36f55889 | 738 | /* |
4fcc712f KO |
739 | * It'd be more correct to do this in free_ioctx(), after all |
740 | * the outstanding kiocbs have finished - but by then io_destroy | |
741 | * has already returned, so io_setup() could potentially return | |
742 | * -EAGAIN with no ioctxs actually in use (as far as userspace | |
743 | * could tell). | |
36f55889 | 744 | */ |
e34ecee2 | 745 | aio_nr_sub(ctx->max_reqs); |
4fcc712f KO |
746 | |
747 | if (ctx->mmap_size) | |
748 | vm_munmap(ctx->mmap_base, ctx->mmap_size); | |
749 | ||
723be6e3 | 750 | percpu_ref_kill(&ctx->users); |
36f55889 | 751 | } |
1da177e4 LT |
752 | } |
753 | ||
754 | /* wait_on_sync_kiocb: | |
755 | * Waits on the given sync kiocb to complete. | |
756 | */ | |
57282d8f | 757 | ssize_t wait_on_sync_kiocb(struct kiocb *req) |
1da177e4 | 758 | { |
57282d8f | 759 | while (!req->ki_ctx) { |
1da177e4 | 760 | set_current_state(TASK_UNINTERRUPTIBLE); |
57282d8f | 761 | if (req->ki_ctx) |
1da177e4 | 762 | break; |
41d10da3 | 763 | io_schedule(); |
1da177e4 LT |
764 | } |
765 | __set_current_state(TASK_RUNNING); | |
57282d8f | 766 | return req->ki_user_data; |
1da177e4 | 767 | } |
385773e0 | 768 | EXPORT_SYMBOL(wait_on_sync_kiocb); |
1da177e4 | 769 | |
36f55889 KO |
770 | /* |
771 | * exit_aio: called when the last user of mm goes away. At this point, there is | |
772 | * no way for any new requests to be submited or any of the io_* syscalls to be | |
773 | * called on the context. | |
774 | * | |
775 | * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on | |
776 | * them. | |
1da177e4 | 777 | */ |
fc9b52cd | 778 | void exit_aio(struct mm_struct *mm) |
1da177e4 | 779 | { |
db446a08 | 780 | struct kioctx_table *table; |
abf137dd | 781 | struct kioctx *ctx; |
db446a08 BL |
782 | unsigned i = 0; |
783 | ||
784 | while (1) { | |
785 | rcu_read_lock(); | |
786 | table = rcu_dereference(mm->ioctx_table); | |
787 | ||
788 | do { | |
789 | if (!table || i >= table->nr) { | |
790 | rcu_read_unlock(); | |
791 | rcu_assign_pointer(mm->ioctx_table, NULL); | |
792 | if (table) | |
793 | kfree(table); | |
794 | return; | |
795 | } | |
796 | ||
797 | ctx = table->table[i++]; | |
798 | } while (!ctx); | |
799 | ||
800 | rcu_read_unlock(); | |
abf137dd | 801 | |
936af157 AV |
802 | /* |
803 | * We don't need to bother with munmap() here - | |
804 | * exit_mmap(mm) is coming and it'll unmap everything. | |
805 | * Since aio_free_ring() uses non-zero ->mmap_size | |
806 | * as indicator that it needs to unmap the area, | |
807 | * just set it to 0; aio_free_ring() is the only | |
808 | * place that uses ->mmap_size, so it's safe. | |
936af157 | 809 | */ |
58c85dc2 | 810 | ctx->mmap_size = 0; |
36f55889 | 811 | |
db446a08 | 812 | kill_ioctx(mm, ctx); |
1da177e4 LT |
813 | } |
814 | } | |
815 | ||
e1bdd5f2 KO |
816 | static void put_reqs_available(struct kioctx *ctx, unsigned nr) |
817 | { | |
818 | struct kioctx_cpu *kcpu; | |
819 | ||
820 | preempt_disable(); | |
821 | kcpu = this_cpu_ptr(ctx->cpu); | |
822 | ||
823 | kcpu->reqs_available += nr; | |
824 | while (kcpu->reqs_available >= ctx->req_batch * 2) { | |
825 | kcpu->reqs_available -= ctx->req_batch; | |
826 | atomic_add(ctx->req_batch, &ctx->reqs_available); | |
827 | } | |
828 | ||
829 | preempt_enable(); | |
830 | } | |
831 | ||
832 | static bool get_reqs_available(struct kioctx *ctx) | |
833 | { | |
834 | struct kioctx_cpu *kcpu; | |
835 | bool ret = false; | |
836 | ||
837 | preempt_disable(); | |
838 | kcpu = this_cpu_ptr(ctx->cpu); | |
839 | ||
840 | if (!kcpu->reqs_available) { | |
841 | int old, avail = atomic_read(&ctx->reqs_available); | |
842 | ||
843 | do { | |
844 | if (avail < ctx->req_batch) | |
845 | goto out; | |
846 | ||
847 | old = avail; | |
848 | avail = atomic_cmpxchg(&ctx->reqs_available, | |
849 | avail, avail - ctx->req_batch); | |
850 | } while (avail != old); | |
851 | ||
852 | kcpu->reqs_available += ctx->req_batch; | |
853 | } | |
854 | ||
855 | ret = true; | |
856 | kcpu->reqs_available--; | |
857 | out: | |
858 | preempt_enable(); | |
859 | return ret; | |
860 | } | |
861 | ||
1da177e4 | 862 | /* aio_get_req |
57282d8f KO |
863 | * Allocate a slot for an aio request. |
864 | * Returns NULL if no requests are free. | |
1da177e4 | 865 | */ |
a1c8eae7 | 866 | static inline struct kiocb *aio_get_req(struct kioctx *ctx) |
1da177e4 | 867 | { |
a1c8eae7 KO |
868 | struct kiocb *req; |
869 | ||
e1bdd5f2 | 870 | if (!get_reqs_available(ctx)) |
a1c8eae7 KO |
871 | return NULL; |
872 | ||
0460fef2 | 873 | req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO); |
1da177e4 | 874 | if (unlikely(!req)) |
a1c8eae7 | 875 | goto out_put; |
1da177e4 | 876 | |
e34ecee2 KO |
877 | percpu_ref_get(&ctx->reqs); |
878 | ||
1da177e4 | 879 | req->ki_ctx = ctx; |
080d676d | 880 | return req; |
a1c8eae7 | 881 | out_put: |
e1bdd5f2 | 882 | put_reqs_available(ctx, 1); |
a1c8eae7 | 883 | return NULL; |
1da177e4 LT |
884 | } |
885 | ||
11599eba | 886 | static void kiocb_free(struct kiocb *req) |
1da177e4 | 887 | { |
1d98ebfc KO |
888 | if (req->ki_filp) |
889 | fput(req->ki_filp); | |
13389010 DL |
890 | if (req->ki_eventfd != NULL) |
891 | eventfd_ctx_put(req->ki_eventfd); | |
1da177e4 | 892 | kmem_cache_free(kiocb_cachep, req); |
1da177e4 LT |
893 | } |
894 | ||
d5470b59 | 895 | static struct kioctx *lookup_ioctx(unsigned long ctx_id) |
1da177e4 | 896 | { |
db446a08 | 897 | struct aio_ring __user *ring = (void __user *)ctx_id; |
abf137dd | 898 | struct mm_struct *mm = current->mm; |
65c24491 | 899 | struct kioctx *ctx, *ret = NULL; |
db446a08 BL |
900 | struct kioctx_table *table; |
901 | unsigned id; | |
902 | ||
903 | if (get_user(id, &ring->id)) | |
904 | return NULL; | |
1da177e4 | 905 | |
abf137dd | 906 | rcu_read_lock(); |
db446a08 | 907 | table = rcu_dereference(mm->ioctx_table); |
abf137dd | 908 | |
db446a08 BL |
909 | if (!table || id >= table->nr) |
910 | goto out; | |
1da177e4 | 911 | |
db446a08 | 912 | ctx = table->table[id]; |
f30d704f | 913 | if (ctx && ctx->user_id == ctx_id) { |
db446a08 BL |
914 | percpu_ref_get(&ctx->users); |
915 | ret = ctx; | |
916 | } | |
917 | out: | |
abf137dd | 918 | rcu_read_unlock(); |
65c24491 | 919 | return ret; |
1da177e4 LT |
920 | } |
921 | ||
1da177e4 LT |
922 | /* aio_complete |
923 | * Called when the io request on the given iocb is complete. | |
1da177e4 | 924 | */ |
2d68449e | 925 | void aio_complete(struct kiocb *iocb, long res, long res2) |
1da177e4 LT |
926 | { |
927 | struct kioctx *ctx = iocb->ki_ctx; | |
1da177e4 | 928 | struct aio_ring *ring; |
21b40200 | 929 | struct io_event *ev_page, *event; |
1da177e4 | 930 | unsigned long flags; |
21b40200 | 931 | unsigned tail, pos; |
1da177e4 | 932 | |
20dcae32 ZB |
933 | /* |
934 | * Special case handling for sync iocbs: | |
935 | * - events go directly into the iocb for fast handling | |
936 | * - the sync task with the iocb in its stack holds the single iocb | |
937 | * ref, no other paths have a way to get another ref | |
938 | * - the sync task helpfully left a reference to itself in the iocb | |
1da177e4 LT |
939 | */ |
940 | if (is_sync_kiocb(iocb)) { | |
1da177e4 | 941 | iocb->ki_user_data = res; |
57282d8f KO |
942 | smp_wmb(); |
943 | iocb->ki_ctx = ERR_PTR(-EXDEV); | |
1da177e4 | 944 | wake_up_process(iocb->ki_obj.tsk); |
2d68449e | 945 | return; |
1da177e4 LT |
946 | } |
947 | ||
0460fef2 KO |
948 | if (iocb->ki_list.next) { |
949 | unsigned long flags; | |
950 | ||
951 | spin_lock_irqsave(&ctx->ctx_lock, flags); | |
952 | list_del(&iocb->ki_list); | |
953 | spin_unlock_irqrestore(&ctx->ctx_lock, flags); | |
954 | } | |
11599eba | 955 | |
0460fef2 KO |
956 | /* |
957 | * Add a completion event to the ring buffer. Must be done holding | |
4b30f07e | 958 | * ctx->completion_lock to prevent other code from messing with the tail |
0460fef2 KO |
959 | * pointer since we might be called from irq context. |
960 | */ | |
961 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
962 | ||
58c85dc2 | 963 | tail = ctx->tail; |
21b40200 KO |
964 | pos = tail + AIO_EVENTS_OFFSET; |
965 | ||
58c85dc2 | 966 | if (++tail >= ctx->nr_events) |
4bf69b2a | 967 | tail = 0; |
1da177e4 | 968 | |
58c85dc2 | 969 | ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); |
21b40200 KO |
970 | event = ev_page + pos % AIO_EVENTS_PER_PAGE; |
971 | ||
1da177e4 LT |
972 | event->obj = (u64)(unsigned long)iocb->ki_obj.user; |
973 | event->data = iocb->ki_user_data; | |
974 | event->res = res; | |
975 | event->res2 = res2; | |
976 | ||
21b40200 | 977 | kunmap_atomic(ev_page); |
58c85dc2 | 978 | flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); |
21b40200 KO |
979 | |
980 | pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n", | |
caf4167a KO |
981 | ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data, |
982 | res, res2); | |
1da177e4 LT |
983 | |
984 | /* after flagging the request as done, we | |
985 | * must never even look at it again | |
986 | */ | |
987 | smp_wmb(); /* make event visible before updating tail */ | |
988 | ||
58c85dc2 | 989 | ctx->tail = tail; |
1da177e4 | 990 | |
58c85dc2 | 991 | ring = kmap_atomic(ctx->ring_pages[0]); |
21b40200 | 992 | ring->tail = tail; |
e8e3c3d6 | 993 | kunmap_atomic(ring); |
58c85dc2 | 994 | flush_dcache_page(ctx->ring_pages[0]); |
1da177e4 | 995 | |
0460fef2 KO |
996 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
997 | ||
21b40200 | 998 | pr_debug("added to ring %p at [%u]\n", iocb, tail); |
8d1c98b0 DL |
999 | |
1000 | /* | |
1001 | * Check if the user asked us to deliver the result through an | |
1002 | * eventfd. The eventfd_signal() function is safe to be called | |
1003 | * from IRQ context. | |
1004 | */ | |
87c3a86e | 1005 | if (iocb->ki_eventfd != NULL) |
8d1c98b0 DL |
1006 | eventfd_signal(iocb->ki_eventfd, 1); |
1007 | ||
1da177e4 | 1008 | /* everything turned out well, dispose of the aiocb. */ |
57282d8f | 1009 | kiocb_free(iocb); |
1da177e4 | 1010 | |
6cb2a210 QB |
1011 | /* |
1012 | * We have to order our ring_info tail store above and test | |
1013 | * of the wait list below outside the wait lock. This is | |
1014 | * like in wake_up_bit() where clearing a bit has to be | |
1015 | * ordered with the unlocked test. | |
1016 | */ | |
1017 | smp_mb(); | |
1018 | ||
1da177e4 LT |
1019 | if (waitqueue_active(&ctx->wait)) |
1020 | wake_up(&ctx->wait); | |
1021 | ||
e34ecee2 | 1022 | percpu_ref_put(&ctx->reqs); |
1da177e4 | 1023 | } |
385773e0 | 1024 | EXPORT_SYMBOL(aio_complete); |
1da177e4 | 1025 | |
a31ad380 KO |
1026 | /* aio_read_events |
1027 | * Pull an event off of the ioctx's event ring. Returns the number of | |
1028 | * events fetched | |
1da177e4 | 1029 | */ |
a31ad380 KO |
1030 | static long aio_read_events_ring(struct kioctx *ctx, |
1031 | struct io_event __user *event, long nr) | |
1da177e4 | 1032 | { |
1da177e4 | 1033 | struct aio_ring *ring; |
5ffac122 | 1034 | unsigned head, tail, pos; |
a31ad380 KO |
1035 | long ret = 0; |
1036 | int copy_ret; | |
1037 | ||
58c85dc2 | 1038 | mutex_lock(&ctx->ring_lock); |
1da177e4 | 1039 | |
fa8a53c3 | 1040 | /* Access to ->ring_pages here is protected by ctx->ring_lock. */ |
58c85dc2 | 1041 | ring = kmap_atomic(ctx->ring_pages[0]); |
a31ad380 | 1042 | head = ring->head; |
5ffac122 | 1043 | tail = ring->tail; |
a31ad380 KO |
1044 | kunmap_atomic(ring); |
1045 | ||
5ffac122 | 1046 | pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events); |
1da177e4 | 1047 | |
5ffac122 | 1048 | if (head == tail) |
1da177e4 LT |
1049 | goto out; |
1050 | ||
a31ad380 KO |
1051 | while (ret < nr) { |
1052 | long avail; | |
1053 | struct io_event *ev; | |
1054 | struct page *page; | |
1055 | ||
5ffac122 KO |
1056 | avail = (head <= tail ? tail : ctx->nr_events) - head; |
1057 | if (head == tail) | |
a31ad380 KO |
1058 | break; |
1059 | ||
1060 | avail = min(avail, nr - ret); | |
1061 | avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - | |
1062 | ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE)); | |
1063 | ||
1064 | pos = head + AIO_EVENTS_OFFSET; | |
58c85dc2 | 1065 | page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]; |
a31ad380 KO |
1066 | pos %= AIO_EVENTS_PER_PAGE; |
1067 | ||
1068 | ev = kmap(page); | |
1069 | copy_ret = copy_to_user(event + ret, ev + pos, | |
1070 | sizeof(*ev) * avail); | |
1071 | kunmap(page); | |
1072 | ||
1073 | if (unlikely(copy_ret)) { | |
1074 | ret = -EFAULT; | |
1075 | goto out; | |
1076 | } | |
1077 | ||
1078 | ret += avail; | |
1079 | head += avail; | |
58c85dc2 | 1080 | head %= ctx->nr_events; |
1da177e4 | 1081 | } |
1da177e4 | 1082 | |
58c85dc2 | 1083 | ring = kmap_atomic(ctx->ring_pages[0]); |
a31ad380 | 1084 | ring->head = head; |
91d80a84 | 1085 | kunmap_atomic(ring); |
58c85dc2 | 1086 | flush_dcache_page(ctx->ring_pages[0]); |
a31ad380 | 1087 | |
5ffac122 | 1088 | pr_debug("%li h%u t%u\n", ret, head, tail); |
3e845ce0 | 1089 | |
e1bdd5f2 | 1090 | put_reqs_available(ctx, ret); |
a31ad380 | 1091 | out: |
58c85dc2 | 1092 | mutex_unlock(&ctx->ring_lock); |
a31ad380 | 1093 | |
1da177e4 LT |
1094 | return ret; |
1095 | } | |
1096 | ||
a31ad380 KO |
1097 | static 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 | 1114 | static 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 | 1166 | SYSCALL_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) |
db446a08 | 1188 | kill_ioctx(current->mm, ioctx); |
723be6e3 | 1189 | percpu_ref_put(&ioctx->users); |
1da177e4 LT |
1190 | } |
1191 | ||
1192 | out: | |
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 | 1202 | SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx) |
1da177e4 LT |
1203 | { |
1204 | struct kioctx *ioctx = lookup_ioctx(ctx); | |
1205 | if (likely(NULL != ioctx)) { | |
db446a08 | 1206 | kill_ioctx(current->mm, ioctx); |
723be6e3 | 1207 | percpu_ref_put(&ioctx->users); |
1da177e4 LT |
1208 | return 0; |
1209 | } | |
1210 | pr_debug("EINVAL: io_destroy: invalid context id\n"); | |
1211 | return -EINVAL; | |
1212 | } | |
1213 | ||
41ef4eb8 KO |
1214 | typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *, |
1215 | unsigned long, loff_t); | |
1216 | ||
8bc92afc KO |
1217 | static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb, |
1218 | int rw, char __user *buf, | |
1219 | unsigned long *nr_segs, | |
1220 | struct iovec **iovec, | |
1221 | bool compat) | |
eed4e51f BP |
1222 | { |
1223 | ssize_t ret; | |
1224 | ||
8bc92afc | 1225 | *nr_segs = kiocb->ki_nbytes; |
41ef4eb8 | 1226 | |
9d85cba7 JM |
1227 | #ifdef CONFIG_COMPAT |
1228 | if (compat) | |
41ef4eb8 | 1229 | ret = compat_rw_copy_check_uvector(rw, |
8bc92afc KO |
1230 | (struct compat_iovec __user *)buf, |
1231 | *nr_segs, 1, *iovec, iovec); | |
9d85cba7 JM |
1232 | else |
1233 | #endif | |
41ef4eb8 | 1234 | ret = rw_copy_check_uvector(rw, |
8bc92afc KO |
1235 | (struct iovec __user *)buf, |
1236 | *nr_segs, 1, *iovec, iovec); | |
eed4e51f | 1237 | if (ret < 0) |
41ef4eb8 | 1238 | return ret; |
a70b52ec | 1239 | |
41ef4eb8 | 1240 | /* ki_nbytes now reflect bytes instead of segs */ |
eed4e51f | 1241 | kiocb->ki_nbytes = ret; |
41ef4eb8 | 1242 | return 0; |
eed4e51f BP |
1243 | } |
1244 | ||
8bc92afc KO |
1245 | static ssize_t aio_setup_single_vector(struct kiocb *kiocb, |
1246 | int rw, char __user *buf, | |
1247 | unsigned long *nr_segs, | |
1248 | struct iovec *iovec) | |
eed4e51f | 1249 | { |
8bc92afc | 1250 | if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes))) |
41ef4eb8 | 1251 | return -EFAULT; |
a70b52ec | 1252 | |
8bc92afc KO |
1253 | iovec->iov_base = buf; |
1254 | iovec->iov_len = kiocb->ki_nbytes; | |
1255 | *nr_segs = 1; | |
eed4e51f BP |
1256 | return 0; |
1257 | } | |
1258 | ||
1da177e4 LT |
1259 | /* |
1260 | * aio_setup_iocb: | |
1261 | * Performs the initial checks and aio retry method | |
1262 | * setup for the kiocb at the time of io submission. | |
1263 | */ | |
8bc92afc KO |
1264 | static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode, |
1265 | char __user *buf, bool compat) | |
1da177e4 | 1266 | { |
41ef4eb8 KO |
1267 | struct file *file = req->ki_filp; |
1268 | ssize_t ret; | |
8bc92afc | 1269 | unsigned long nr_segs; |
41ef4eb8 KO |
1270 | int rw; |
1271 | fmode_t mode; | |
1272 | aio_rw_op *rw_op; | |
8bc92afc | 1273 | struct iovec inline_vec, *iovec = &inline_vec; |
1da177e4 | 1274 | |
8bc92afc | 1275 | switch (opcode) { |
1da177e4 | 1276 | case IOCB_CMD_PREAD: |
eed4e51f | 1277 | case IOCB_CMD_PREADV: |
41ef4eb8 KO |
1278 | mode = FMODE_READ; |
1279 | rw = READ; | |
1280 | rw_op = file->f_op->aio_read; | |
1281 | goto rw_common; | |
1282 | ||
1283 | case IOCB_CMD_PWRITE: | |
eed4e51f | 1284 | case IOCB_CMD_PWRITEV: |
41ef4eb8 KO |
1285 | mode = FMODE_WRITE; |
1286 | rw = WRITE; | |
1287 | rw_op = file->f_op->aio_write; | |
1288 | goto rw_common; | |
1289 | rw_common: | |
1290 | if (unlikely(!(file->f_mode & mode))) | |
1291 | return -EBADF; | |
1292 | ||
1293 | if (!rw_op) | |
1294 | return -EINVAL; | |
1295 | ||
8bc92afc KO |
1296 | ret = (opcode == IOCB_CMD_PREADV || |
1297 | opcode == IOCB_CMD_PWRITEV) | |
1298 | ? aio_setup_vectored_rw(req, rw, buf, &nr_segs, | |
1299 | &iovec, compat) | |
1300 | : aio_setup_single_vector(req, rw, buf, &nr_segs, | |
1301 | iovec); | |
eed4e51f | 1302 | if (ret) |
41ef4eb8 KO |
1303 | return ret; |
1304 | ||
1305 | ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes); | |
8bc92afc KO |
1306 | if (ret < 0) { |
1307 | if (iovec != &inline_vec) | |
1308 | kfree(iovec); | |
41ef4eb8 | 1309 | return ret; |
8bc92afc | 1310 | } |
41ef4eb8 KO |
1311 | |
1312 | req->ki_nbytes = ret; | |
41ef4eb8 | 1313 | |
73a7075e KO |
1314 | /* XXX: move/kill - rw_verify_area()? */ |
1315 | /* This matches the pread()/pwrite() logic */ | |
1316 | if (req->ki_pos < 0) { | |
1317 | ret = -EINVAL; | |
1318 | break; | |
1319 | } | |
1320 | ||
1321 | if (rw == WRITE) | |
1322 | file_start_write(file); | |
1323 | ||
8bc92afc | 1324 | ret = rw_op(req, iovec, nr_segs, req->ki_pos); |
73a7075e KO |
1325 | |
1326 | if (rw == WRITE) | |
1327 | file_end_write(file); | |
1da177e4 | 1328 | break; |
41ef4eb8 | 1329 | |
1da177e4 | 1330 | case IOCB_CMD_FDSYNC: |
41ef4eb8 KO |
1331 | if (!file->f_op->aio_fsync) |
1332 | return -EINVAL; | |
1333 | ||
1334 | ret = file->f_op->aio_fsync(req, 1); | |
1da177e4 | 1335 | break; |
41ef4eb8 | 1336 | |
1da177e4 | 1337 | case IOCB_CMD_FSYNC: |
41ef4eb8 KO |
1338 | if (!file->f_op->aio_fsync) |
1339 | return -EINVAL; | |
1340 | ||
1341 | ret = file->f_op->aio_fsync(req, 0); | |
1da177e4 | 1342 | break; |
41ef4eb8 | 1343 | |
1da177e4 | 1344 | default: |
caf4167a | 1345 | pr_debug("EINVAL: no operation provided\n"); |
41ef4eb8 | 1346 | return -EINVAL; |
1da177e4 LT |
1347 | } |
1348 | ||
8bc92afc KO |
1349 | if (iovec != &inline_vec) |
1350 | kfree(iovec); | |
1351 | ||
41ef4eb8 KO |
1352 | if (ret != -EIOCBQUEUED) { |
1353 | /* | |
1354 | * There's no easy way to restart the syscall since other AIO's | |
1355 | * may be already running. Just fail this IO with EINTR. | |
1356 | */ | |
1357 | if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR || | |
1358 | ret == -ERESTARTNOHAND || | |
1359 | ret == -ERESTART_RESTARTBLOCK)) | |
1360 | ret = -EINTR; | |
1361 | aio_complete(req, ret, 0); | |
1362 | } | |
1da177e4 LT |
1363 | |
1364 | return 0; | |
1365 | } | |
1366 | ||
d5470b59 | 1367 | static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb, |
a1c8eae7 | 1368 | struct iocb *iocb, bool compat) |
1da177e4 LT |
1369 | { |
1370 | struct kiocb *req; | |
1da177e4 LT |
1371 | ssize_t ret; |
1372 | ||
1373 | /* enforce forwards compatibility on users */ | |
9c3060be | 1374 | if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) { |
caf4167a | 1375 | pr_debug("EINVAL: reserve field set\n"); |
1da177e4 LT |
1376 | return -EINVAL; |
1377 | } | |
1378 | ||
1379 | /* prevent overflows */ | |
1380 | if (unlikely( | |
1381 | (iocb->aio_buf != (unsigned long)iocb->aio_buf) || | |
1382 | (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) || | |
1383 | ((ssize_t)iocb->aio_nbytes < 0) | |
1384 | )) { | |
1385 | pr_debug("EINVAL: io_submit: overflow check\n"); | |
1386 | return -EINVAL; | |
1387 | } | |
1388 | ||
41ef4eb8 | 1389 | req = aio_get_req(ctx); |
1d98ebfc | 1390 | if (unlikely(!req)) |
1da177e4 | 1391 | return -EAGAIN; |
1d98ebfc KO |
1392 | |
1393 | req->ki_filp = fget(iocb->aio_fildes); | |
1394 | if (unlikely(!req->ki_filp)) { | |
1395 | ret = -EBADF; | |
1396 | goto out_put_req; | |
1da177e4 | 1397 | } |
1d98ebfc | 1398 | |
9c3060be DL |
1399 | if (iocb->aio_flags & IOCB_FLAG_RESFD) { |
1400 | /* | |
1401 | * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an | |
1402 | * instance of the file* now. The file descriptor must be | |
1403 | * an eventfd() fd, and will be signaled for each completed | |
1404 | * event using the eventfd_signal() function. | |
1405 | */ | |
13389010 | 1406 | req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd); |
801678c5 | 1407 | if (IS_ERR(req->ki_eventfd)) { |
9c3060be | 1408 | ret = PTR_ERR(req->ki_eventfd); |
87c3a86e | 1409 | req->ki_eventfd = NULL; |
9c3060be DL |
1410 | goto out_put_req; |
1411 | } | |
1412 | } | |
1da177e4 | 1413 | |
8a660890 | 1414 | ret = put_user(KIOCB_KEY, &user_iocb->aio_key); |
1da177e4 | 1415 | if (unlikely(ret)) { |
caf4167a | 1416 | pr_debug("EFAULT: aio_key\n"); |
1da177e4 LT |
1417 | goto out_put_req; |
1418 | } | |
1419 | ||
1420 | req->ki_obj.user = user_iocb; | |
1421 | req->ki_user_data = iocb->aio_data; | |
1422 | req->ki_pos = iocb->aio_offset; | |
73a7075e | 1423 | req->ki_nbytes = iocb->aio_nbytes; |
1da177e4 | 1424 | |
8bc92afc KO |
1425 | ret = aio_run_iocb(req, iocb->aio_lio_opcode, |
1426 | (char __user *)(unsigned long)iocb->aio_buf, | |
1427 | compat); | |
41003a7b | 1428 | if (ret) |
7137c6bd | 1429 | goto out_put_req; |
41003a7b | 1430 | |
1da177e4 | 1431 | return 0; |
1da177e4 | 1432 | out_put_req: |
e1bdd5f2 | 1433 | put_reqs_available(ctx, 1); |
e34ecee2 | 1434 | percpu_ref_put(&ctx->reqs); |
57282d8f | 1435 | kiocb_free(req); |
1da177e4 LT |
1436 | return ret; |
1437 | } | |
1438 | ||
9d85cba7 JM |
1439 | long do_io_submit(aio_context_t ctx_id, long nr, |
1440 | struct iocb __user *__user *iocbpp, bool compat) | |
1da177e4 LT |
1441 | { |
1442 | struct kioctx *ctx; | |
1443 | long ret = 0; | |
080d676d | 1444 | int i = 0; |
9f5b9425 | 1445 | struct blk_plug plug; |
1da177e4 LT |
1446 | |
1447 | if (unlikely(nr < 0)) | |
1448 | return -EINVAL; | |
1449 | ||
75e1c70f JM |
1450 | if (unlikely(nr > LONG_MAX/sizeof(*iocbpp))) |
1451 | nr = LONG_MAX/sizeof(*iocbpp); | |
1452 | ||
1da177e4 LT |
1453 | if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp))))) |
1454 | return -EFAULT; | |
1455 | ||
1456 | ctx = lookup_ioctx(ctx_id); | |
1457 | if (unlikely(!ctx)) { | |
caf4167a | 1458 | pr_debug("EINVAL: invalid context id\n"); |
1da177e4 LT |
1459 | return -EINVAL; |
1460 | } | |
1461 | ||
9f5b9425 SL |
1462 | blk_start_plug(&plug); |
1463 | ||
1da177e4 LT |
1464 | /* |
1465 | * AKPM: should this return a partial result if some of the IOs were | |
1466 | * successfully submitted? | |
1467 | */ | |
1468 | for (i=0; i<nr; i++) { | |
1469 | struct iocb __user *user_iocb; | |
1470 | struct iocb tmp; | |
1471 | ||
1472 | if (unlikely(__get_user(user_iocb, iocbpp + i))) { | |
1473 | ret = -EFAULT; | |
1474 | break; | |
1475 | } | |
1476 | ||
1477 | if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) { | |
1478 | ret = -EFAULT; | |
1479 | break; | |
1480 | } | |
1481 | ||
a1c8eae7 | 1482 | ret = io_submit_one(ctx, user_iocb, &tmp, compat); |
1da177e4 LT |
1483 | if (ret) |
1484 | break; | |
1485 | } | |
9f5b9425 | 1486 | blk_finish_plug(&plug); |
1da177e4 | 1487 | |
723be6e3 | 1488 | percpu_ref_put(&ctx->users); |
1da177e4 LT |
1489 | return i ? i : ret; |
1490 | } | |
1491 | ||
9d85cba7 JM |
1492 | /* sys_io_submit: |
1493 | * Queue the nr iocbs pointed to by iocbpp for processing. Returns | |
1494 | * the number of iocbs queued. May return -EINVAL if the aio_context | |
1495 | * specified by ctx_id is invalid, if nr is < 0, if the iocb at | |
1496 | * *iocbpp[0] is not properly initialized, if the operation specified | |
1497 | * is invalid for the file descriptor in the iocb. May fail with | |
1498 | * -EFAULT if any of the data structures point to invalid data. May | |
1499 | * fail with -EBADF if the file descriptor specified in the first | |
1500 | * iocb is invalid. May fail with -EAGAIN if insufficient resources | |
1501 | * are available to queue any iocbs. Will return 0 if nr is 0. Will | |
1502 | * fail with -ENOSYS if not implemented. | |
1503 | */ | |
1504 | SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr, | |
1505 | struct iocb __user * __user *, iocbpp) | |
1506 | { | |
1507 | return do_io_submit(ctx_id, nr, iocbpp, 0); | |
1508 | } | |
1509 | ||
1da177e4 LT |
1510 | /* lookup_kiocb |
1511 | * Finds a given iocb for cancellation. | |
1da177e4 | 1512 | */ |
25ee7e38 AB |
1513 | static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, |
1514 | u32 key) | |
1da177e4 LT |
1515 | { |
1516 | struct list_head *pos; | |
d00689af ZB |
1517 | |
1518 | assert_spin_locked(&ctx->ctx_lock); | |
1519 | ||
8a660890 KO |
1520 | if (key != KIOCB_KEY) |
1521 | return NULL; | |
1522 | ||
1da177e4 LT |
1523 | /* TODO: use a hash or array, this sucks. */ |
1524 | list_for_each(pos, &ctx->active_reqs) { | |
1525 | struct kiocb *kiocb = list_kiocb(pos); | |
8a660890 | 1526 | if (kiocb->ki_obj.user == iocb) |
1da177e4 LT |
1527 | return kiocb; |
1528 | } | |
1529 | return NULL; | |
1530 | } | |
1531 | ||
1532 | /* sys_io_cancel: | |
1533 | * Attempts to cancel an iocb previously passed to io_submit. If | |
1534 | * the operation is successfully cancelled, the resulting event is | |
1535 | * copied into the memory pointed to by result without being placed | |
1536 | * into the completion queue and 0 is returned. May fail with | |
1537 | * -EFAULT if any of the data structures pointed to are invalid. | |
1538 | * May fail with -EINVAL if aio_context specified by ctx_id is | |
1539 | * invalid. May fail with -EAGAIN if the iocb specified was not | |
1540 | * cancelled. Will fail with -ENOSYS if not implemented. | |
1541 | */ | |
002c8976 HC |
1542 | SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb, |
1543 | struct io_event __user *, result) | |
1da177e4 | 1544 | { |
1da177e4 LT |
1545 | struct kioctx *ctx; |
1546 | struct kiocb *kiocb; | |
1547 | u32 key; | |
1548 | int ret; | |
1549 | ||
1550 | ret = get_user(key, &iocb->aio_key); | |
1551 | if (unlikely(ret)) | |
1552 | return -EFAULT; | |
1553 | ||
1554 | ctx = lookup_ioctx(ctx_id); | |
1555 | if (unlikely(!ctx)) | |
1556 | return -EINVAL; | |
1557 | ||
1558 | spin_lock_irq(&ctx->ctx_lock); | |
906b973c | 1559 | |
1da177e4 | 1560 | kiocb = lookup_kiocb(ctx, iocb, key); |
906b973c | 1561 | if (kiocb) |
bec68faa | 1562 | ret = kiocb_cancel(ctx, kiocb); |
906b973c KO |
1563 | else |
1564 | ret = -EINVAL; | |
1565 | ||
1da177e4 LT |
1566 | spin_unlock_irq(&ctx->ctx_lock); |
1567 | ||
906b973c | 1568 | if (!ret) { |
bec68faa KO |
1569 | /* |
1570 | * The result argument is no longer used - the io_event is | |
1571 | * always delivered via the ring buffer. -EINPROGRESS indicates | |
1572 | * cancellation is progress: | |
906b973c | 1573 | */ |
bec68faa | 1574 | ret = -EINPROGRESS; |
906b973c | 1575 | } |
1da177e4 | 1576 | |
723be6e3 | 1577 | percpu_ref_put(&ctx->users); |
1da177e4 LT |
1578 | |
1579 | return ret; | |
1580 | } | |
1581 | ||
1582 | /* io_getevents: | |
1583 | * Attempts to read at least min_nr events and up to nr events from | |
642b5123 ST |
1584 | * the completion queue for the aio_context specified by ctx_id. If |
1585 | * it succeeds, the number of read events is returned. May fail with | |
1586 | * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is | |
1587 | * out of range, if timeout is out of range. May fail with -EFAULT | |
1588 | * if any of the memory specified is invalid. May return 0 or | |
1589 | * < min_nr if the timeout specified by timeout has elapsed | |
1590 | * before sufficient events are available, where timeout == NULL | |
1591 | * specifies an infinite timeout. Note that the timeout pointed to by | |
6900807c | 1592 | * timeout is relative. Will fail with -ENOSYS if not implemented. |
1da177e4 | 1593 | */ |
002c8976 HC |
1594 | SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id, |
1595 | long, min_nr, | |
1596 | long, nr, | |
1597 | struct io_event __user *, events, | |
1598 | struct timespec __user *, timeout) | |
1da177e4 LT |
1599 | { |
1600 | struct kioctx *ioctx = lookup_ioctx(ctx_id); | |
1601 | long ret = -EINVAL; | |
1602 | ||
1603 | if (likely(ioctx)) { | |
2e410255 | 1604 | if (likely(min_nr <= nr && min_nr >= 0)) |
1da177e4 | 1605 | ret = read_events(ioctx, min_nr, nr, events, timeout); |
723be6e3 | 1606 | percpu_ref_put(&ioctx->users); |
1da177e4 | 1607 | } |
1da177e4 LT |
1608 | return ret; |
1609 | } |