]>
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 | 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 |
154 | static DEFINE_SPINLOCK(aio_nr_lock); |
155 | unsigned long aio_nr; /* current system wide number of aio requests */ | |
156 | unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */ | |
1da177e4 LT |
157 | /*----end sysctl variables---*/ |
158 | ||
e18b890b CL |
159 | static struct kmem_cache *kiocb_cachep; |
160 | static struct kmem_cache *kioctx_cachep; | |
1da177e4 | 161 | |
71ad7490 BL |
162 | static struct vfsmount *aio_mnt; |
163 | ||
164 | static const struct file_operations aio_ring_fops; | |
165 | static const struct address_space_operations aio_ctx_aops; | |
166 | ||
167 | static 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 | ||
198 | static 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 | */ | |
211 | static 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 |
231 | static 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 |
247 | static 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 | ||
273 | static 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 | ||
279 | static const struct file_operations aio_ring_fops = { | |
280 | .mmap = aio_ring_mmap, | |
281 | }; | |
282 | ||
283 | static int aio_set_page_dirty(struct page *page) | |
284 | { | |
285 | return 0; | |
286 | } | |
287 | ||
0c45355f | 288 | #if IS_ENABLED(CONFIG_MIGRATION) |
36bc08cc GZ |
289 | static 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 |
350 | out_unlock: |
351 | mutex_unlock(&ctx->ring_lock); | |
352 | out: | |
353 | spin_unlock(&mapping->private_lock); | |
36bc08cc | 354 | return rc; |
1da177e4 | 355 | } |
0c45355f | 356 | #endif |
1da177e4 | 357 | |
36bc08cc GZ |
358 | static 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 |
365 | static 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 |
463 | void 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 | } | |
477 | EXPORT_SYMBOL(kiocb_set_cancel_fn); | |
478 | ||
d52a8f9e | 479 | static 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 | 500 | static 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); |
36f55889 KO |
508 | kmem_cache_free(kioctx_cachep, ctx); |
509 | } | |
510 | ||
e34ecee2 KO |
511 | static void free_ioctx_reqs(struct percpu_ref *ref) |
512 | { | |
513 | struct kioctx *ctx = container_of(ref, struct kioctx, reqs); | |
514 | ||
e02ba72a AP |
515 | /* At this point we know that there are no any in-flight requests */ |
516 | if (ctx->requests_done) | |
517 | complete(ctx->requests_done); | |
518 | ||
e34ecee2 KO |
519 | INIT_WORK(&ctx->free_work, free_ioctx); |
520 | schedule_work(&ctx->free_work); | |
521 | } | |
522 | ||
36f55889 KO |
523 | /* |
524 | * When this function runs, the kioctx has been removed from the "hash table" | |
525 | * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted - | |
526 | * now it's safe to cancel any that need to be. | |
527 | */ | |
e34ecee2 | 528 | static void free_ioctx_users(struct percpu_ref *ref) |
36f55889 | 529 | { |
e34ecee2 | 530 | struct kioctx *ctx = container_of(ref, struct kioctx, users); |
36f55889 KO |
531 | struct kiocb *req; |
532 | ||
533 | spin_lock_irq(&ctx->ctx_lock); | |
534 | ||
535 | while (!list_empty(&ctx->active_reqs)) { | |
536 | req = list_first_entry(&ctx->active_reqs, | |
537 | struct kiocb, ki_list); | |
538 | ||
539 | list_del_init(&req->ki_list); | |
d52a8f9e | 540 | kiocb_cancel(req); |
36f55889 KO |
541 | } |
542 | ||
543 | spin_unlock_irq(&ctx->ctx_lock); | |
544 | ||
e34ecee2 KO |
545 | percpu_ref_kill(&ctx->reqs); |
546 | percpu_ref_put(&ctx->reqs); | |
36f55889 KO |
547 | } |
548 | ||
db446a08 BL |
549 | static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm) |
550 | { | |
551 | unsigned i, new_nr; | |
552 | struct kioctx_table *table, *old; | |
553 | struct aio_ring *ring; | |
554 | ||
555 | spin_lock(&mm->ioctx_lock); | |
855ef0de | 556 | table = rcu_dereference_raw(mm->ioctx_table); |
db446a08 BL |
557 | |
558 | while (1) { | |
559 | if (table) | |
560 | for (i = 0; i < table->nr; i++) | |
561 | if (!table->table[i]) { | |
562 | ctx->id = i; | |
563 | table->table[i] = ctx; | |
564 | spin_unlock(&mm->ioctx_lock); | |
565 | ||
fa8a53c3 BL |
566 | /* While kioctx setup is in progress, |
567 | * we are protected from page migration | |
568 | * changes ring_pages by ->ring_lock. | |
569 | */ | |
db446a08 BL |
570 | ring = kmap_atomic(ctx->ring_pages[0]); |
571 | ring->id = ctx->id; | |
572 | kunmap_atomic(ring); | |
573 | return 0; | |
574 | } | |
575 | ||
576 | new_nr = (table ? table->nr : 1) * 4; | |
db446a08 BL |
577 | spin_unlock(&mm->ioctx_lock); |
578 | ||
579 | table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) * | |
580 | new_nr, GFP_KERNEL); | |
581 | if (!table) | |
582 | return -ENOMEM; | |
583 | ||
584 | table->nr = new_nr; | |
585 | ||
586 | spin_lock(&mm->ioctx_lock); | |
855ef0de | 587 | old = rcu_dereference_raw(mm->ioctx_table); |
db446a08 BL |
588 | |
589 | if (!old) { | |
590 | rcu_assign_pointer(mm->ioctx_table, table); | |
591 | } else if (table->nr > old->nr) { | |
592 | memcpy(table->table, old->table, | |
593 | old->nr * sizeof(struct kioctx *)); | |
594 | ||
595 | rcu_assign_pointer(mm->ioctx_table, table); | |
596 | kfree_rcu(old, rcu); | |
597 | } else { | |
598 | kfree(table); | |
599 | table = old; | |
600 | } | |
601 | } | |
602 | } | |
603 | ||
e34ecee2 KO |
604 | static void aio_nr_sub(unsigned nr) |
605 | { | |
606 | spin_lock(&aio_nr_lock); | |
607 | if (WARN_ON(aio_nr - nr > aio_nr)) | |
608 | aio_nr = 0; | |
609 | else | |
610 | aio_nr -= nr; | |
611 | spin_unlock(&aio_nr_lock); | |
612 | } | |
613 | ||
1da177e4 LT |
614 | /* ioctx_alloc |
615 | * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed. | |
616 | */ | |
617 | static struct kioctx *ioctx_alloc(unsigned nr_events) | |
618 | { | |
41003a7b | 619 | struct mm_struct *mm = current->mm; |
1da177e4 | 620 | struct kioctx *ctx; |
e23754f8 | 621 | int err = -ENOMEM; |
1da177e4 | 622 | |
e1bdd5f2 KO |
623 | /* |
624 | * We keep track of the number of available ringbuffer slots, to prevent | |
625 | * overflow (reqs_available), and we also use percpu counters for this. | |
626 | * | |
627 | * So since up to half the slots might be on other cpu's percpu counters | |
628 | * and unavailable, double nr_events so userspace sees what they | |
629 | * expected: additionally, we move req_batch slots to/from percpu | |
630 | * counters at a time, so make sure that isn't 0: | |
631 | */ | |
632 | nr_events = max(nr_events, num_possible_cpus() * 4); | |
633 | nr_events *= 2; | |
634 | ||
1da177e4 LT |
635 | /* Prevent overflows */ |
636 | if ((nr_events > (0x10000000U / sizeof(struct io_event))) || | |
637 | (nr_events > (0x10000000U / sizeof(struct kiocb)))) { | |
638 | pr_debug("ENOMEM: nr_events too high\n"); | |
639 | return ERR_PTR(-EINVAL); | |
640 | } | |
641 | ||
4cd81c3d | 642 | if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL)) |
1da177e4 LT |
643 | return ERR_PTR(-EAGAIN); |
644 | ||
c3762229 | 645 | ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL); |
1da177e4 LT |
646 | if (!ctx) |
647 | return ERR_PTR(-ENOMEM); | |
648 | ||
1da177e4 | 649 | ctx->max_reqs = nr_events; |
1da177e4 | 650 | |
1da177e4 | 651 | spin_lock_init(&ctx->ctx_lock); |
0460fef2 | 652 | spin_lock_init(&ctx->completion_lock); |
58c85dc2 | 653 | mutex_init(&ctx->ring_lock); |
fa8a53c3 BL |
654 | /* Protect against page migration throughout kiotx setup by keeping |
655 | * the ring_lock mutex held until setup is complete. */ | |
656 | mutex_lock(&ctx->ring_lock); | |
1da177e4 LT |
657 | init_waitqueue_head(&ctx->wait); |
658 | ||
659 | INIT_LIST_HEAD(&ctx->active_reqs); | |
1da177e4 | 660 | |
fa8a53c3 BL |
661 | if (percpu_ref_init(&ctx->users, free_ioctx_users)) |
662 | goto err; | |
663 | ||
664 | if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs)) | |
665 | goto err; | |
666 | ||
e1bdd5f2 KO |
667 | ctx->cpu = alloc_percpu(struct kioctx_cpu); |
668 | if (!ctx->cpu) | |
e34ecee2 | 669 | goto err; |
1da177e4 | 670 | |
fa8a53c3 BL |
671 | err = aio_setup_ring(ctx); |
672 | if (err < 0) | |
e34ecee2 | 673 | goto err; |
e1bdd5f2 | 674 | |
34e83fc6 | 675 | atomic_set(&ctx->reqs_available, ctx->nr_events - 1); |
e1bdd5f2 | 676 | ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4); |
6878ea72 BL |
677 | if (ctx->req_batch < 1) |
678 | ctx->req_batch = 1; | |
34e83fc6 | 679 | |
1da177e4 | 680 | /* limit the number of system wide aios */ |
9fa1cb39 | 681 | spin_lock(&aio_nr_lock); |
4cd81c3d | 682 | if (aio_nr + nr_events > (aio_max_nr * 2UL) || |
2dd542b7 | 683 | aio_nr + nr_events < aio_nr) { |
9fa1cb39 | 684 | spin_unlock(&aio_nr_lock); |
e34ecee2 | 685 | err = -EAGAIN; |
d1b94327 | 686 | goto err_ctx; |
2dd542b7 AV |
687 | } |
688 | aio_nr += ctx->max_reqs; | |
9fa1cb39 | 689 | spin_unlock(&aio_nr_lock); |
1da177e4 | 690 | |
1881686f BL |
691 | percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */ |
692 | percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */ | |
723be6e3 | 693 | |
da90382c BL |
694 | err = ioctx_add_table(ctx, mm); |
695 | if (err) | |
e34ecee2 | 696 | goto err_cleanup; |
da90382c | 697 | |
fa8a53c3 BL |
698 | /* Release the ring_lock mutex now that all setup is complete. */ |
699 | mutex_unlock(&ctx->ring_lock); | |
700 | ||
caf4167a | 701 | pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n", |
58c85dc2 | 702 | ctx, ctx->user_id, mm, ctx->nr_events); |
1da177e4 LT |
703 | return ctx; |
704 | ||
e34ecee2 KO |
705 | err_cleanup: |
706 | aio_nr_sub(ctx->max_reqs); | |
d1b94327 GZ |
707 | err_ctx: |
708 | aio_free_ring(ctx); | |
e34ecee2 | 709 | err: |
fa8a53c3 | 710 | mutex_unlock(&ctx->ring_lock); |
e1bdd5f2 | 711 | free_percpu(ctx->cpu); |
e34ecee2 | 712 | free_percpu(ctx->reqs.pcpu_count); |
723be6e3 | 713 | free_percpu(ctx->users.pcpu_count); |
1da177e4 | 714 | kmem_cache_free(kioctx_cachep, ctx); |
caf4167a | 715 | pr_debug("error allocating ioctx %d\n", err); |
e23754f8 | 716 | return ERR_PTR(err); |
1da177e4 LT |
717 | } |
718 | ||
36f55889 KO |
719 | /* kill_ioctx |
720 | * Cancels all outstanding aio requests on an aio context. Used | |
721 | * when the processes owning a context have all exited to encourage | |
722 | * the rapid destruction of the kioctx. | |
723 | */ | |
fb2d4483 | 724 | static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx, |
e02ba72a | 725 | struct completion *requests_done) |
36f55889 | 726 | { |
fa88b6f8 | 727 | struct kioctx_table *table; |
db446a08 | 728 | |
fa88b6f8 BL |
729 | if (atomic_xchg(&ctx->dead, 1)) |
730 | return -EINVAL; | |
db446a08 | 731 | |
db446a08 | 732 | |
fa88b6f8 | 733 | spin_lock(&mm->ioctx_lock); |
855ef0de | 734 | table = rcu_dereference_raw(mm->ioctx_table); |
fa88b6f8 BL |
735 | WARN_ON(ctx != table->table[ctx->id]); |
736 | table->table[ctx->id] = NULL; | |
fa88b6f8 | 737 | spin_unlock(&mm->ioctx_lock); |
4fcc712f | 738 | |
fa88b6f8 BL |
739 | /* percpu_ref_kill() will do the necessary call_rcu() */ |
740 | wake_up_all(&ctx->wait); | |
4fcc712f | 741 | |
fa88b6f8 BL |
742 | /* |
743 | * It'd be more correct to do this in free_ioctx(), after all | |
744 | * the outstanding kiocbs have finished - but by then io_destroy | |
745 | * has already returned, so io_setup() could potentially return | |
746 | * -EAGAIN with no ioctxs actually in use (as far as userspace | |
747 | * could tell). | |
748 | */ | |
749 | aio_nr_sub(ctx->max_reqs); | |
4fcc712f | 750 | |
fa88b6f8 BL |
751 | if (ctx->mmap_size) |
752 | vm_munmap(ctx->mmap_base, ctx->mmap_size); | |
fb2d4483 | 753 | |
fa88b6f8 BL |
754 | ctx->requests_done = requests_done; |
755 | percpu_ref_kill(&ctx->users); | |
756 | return 0; | |
1da177e4 LT |
757 | } |
758 | ||
759 | /* wait_on_sync_kiocb: | |
760 | * Waits on the given sync kiocb to complete. | |
761 | */ | |
57282d8f | 762 | ssize_t wait_on_sync_kiocb(struct kiocb *req) |
1da177e4 | 763 | { |
57282d8f | 764 | while (!req->ki_ctx) { |
1da177e4 | 765 | set_current_state(TASK_UNINTERRUPTIBLE); |
57282d8f | 766 | if (req->ki_ctx) |
1da177e4 | 767 | break; |
41d10da3 | 768 | io_schedule(); |
1da177e4 LT |
769 | } |
770 | __set_current_state(TASK_RUNNING); | |
57282d8f | 771 | return req->ki_user_data; |
1da177e4 | 772 | } |
385773e0 | 773 | EXPORT_SYMBOL(wait_on_sync_kiocb); |
1da177e4 | 774 | |
36f55889 KO |
775 | /* |
776 | * exit_aio: called when the last user of mm goes away. At this point, there is | |
777 | * no way for any new requests to be submited or any of the io_* syscalls to be | |
778 | * called on the context. | |
779 | * | |
780 | * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on | |
781 | * them. | |
1da177e4 | 782 | */ |
fc9b52cd | 783 | void exit_aio(struct mm_struct *mm) |
1da177e4 | 784 | { |
4b70ac5f ON |
785 | struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table); |
786 | int i; | |
db446a08 | 787 | |
4b70ac5f ON |
788 | if (!table) |
789 | return; | |
db446a08 | 790 | |
4b70ac5f ON |
791 | for (i = 0; i < table->nr; ++i) { |
792 | struct kioctx *ctx = table->table[i]; | |
abf137dd | 793 | |
4b70ac5f ON |
794 | if (!ctx) |
795 | continue; | |
936af157 | 796 | /* |
4b70ac5f ON |
797 | * We don't need to bother with munmap() here - exit_mmap(mm) |
798 | * is coming and it'll unmap everything. And we simply can't, | |
799 | * this is not necessarily our ->mm. | |
800 | * Since kill_ioctx() uses non-zero ->mmap_size as indicator | |
801 | * that it needs to unmap the area, just set it to 0. | |
936af157 | 802 | */ |
58c85dc2 | 803 | ctx->mmap_size = 0; |
e02ba72a | 804 | kill_ioctx(mm, ctx, NULL); |
1da177e4 | 805 | } |
4b70ac5f ON |
806 | |
807 | RCU_INIT_POINTER(mm->ioctx_table, NULL); | |
808 | kfree(table); | |
1da177e4 LT |
809 | } |
810 | ||
e1bdd5f2 KO |
811 | static void put_reqs_available(struct kioctx *ctx, unsigned nr) |
812 | { | |
813 | struct kioctx_cpu *kcpu; | |
263782c1 | 814 | unsigned long flags; |
e1bdd5f2 | 815 | |
263782c1 | 816 | local_irq_save(flags); |
be6fb451 | 817 | kcpu = this_cpu_ptr(ctx->cpu); |
e1bdd5f2 | 818 | kcpu->reqs_available += nr; |
263782c1 | 819 | |
e1bdd5f2 KO |
820 | while (kcpu->reqs_available >= ctx->req_batch * 2) { |
821 | kcpu->reqs_available -= ctx->req_batch; | |
822 | atomic_add(ctx->req_batch, &ctx->reqs_available); | |
823 | } | |
824 | ||
263782c1 | 825 | local_irq_restore(flags); |
e1bdd5f2 KO |
826 | } |
827 | ||
828 | static bool get_reqs_available(struct kioctx *ctx) | |
829 | { | |
830 | struct kioctx_cpu *kcpu; | |
831 | bool ret = false; | |
263782c1 | 832 | unsigned long flags; |
e1bdd5f2 | 833 | |
263782c1 | 834 | local_irq_save(flags); |
be6fb451 | 835 | kcpu = this_cpu_ptr(ctx->cpu); |
e1bdd5f2 KO |
836 | if (!kcpu->reqs_available) { |
837 | int old, avail = atomic_read(&ctx->reqs_available); | |
838 | ||
839 | do { | |
840 | if (avail < ctx->req_batch) | |
841 | goto out; | |
842 | ||
843 | old = avail; | |
844 | avail = atomic_cmpxchg(&ctx->reqs_available, | |
845 | avail, avail - ctx->req_batch); | |
846 | } while (avail != old); | |
847 | ||
848 | kcpu->reqs_available += ctx->req_batch; | |
849 | } | |
850 | ||
851 | ret = true; | |
852 | kcpu->reqs_available--; | |
853 | out: | |
263782c1 | 854 | local_irq_restore(flags); |
e1bdd5f2 KO |
855 | return ret; |
856 | } | |
857 | ||
1da177e4 | 858 | /* aio_get_req |
57282d8f KO |
859 | * Allocate a slot for an aio request. |
860 | * Returns NULL if no requests are free. | |
1da177e4 | 861 | */ |
a1c8eae7 | 862 | static inline struct kiocb *aio_get_req(struct kioctx *ctx) |
1da177e4 | 863 | { |
a1c8eae7 KO |
864 | struct kiocb *req; |
865 | ||
e1bdd5f2 | 866 | if (!get_reqs_available(ctx)) |
a1c8eae7 KO |
867 | return NULL; |
868 | ||
0460fef2 | 869 | req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO); |
1da177e4 | 870 | if (unlikely(!req)) |
a1c8eae7 | 871 | goto out_put; |
1da177e4 | 872 | |
e34ecee2 KO |
873 | percpu_ref_get(&ctx->reqs); |
874 | ||
1da177e4 | 875 | req->ki_ctx = ctx; |
080d676d | 876 | return req; |
a1c8eae7 | 877 | out_put: |
e1bdd5f2 | 878 | put_reqs_available(ctx, 1); |
a1c8eae7 | 879 | return NULL; |
1da177e4 LT |
880 | } |
881 | ||
11599eba | 882 | static void kiocb_free(struct kiocb *req) |
1da177e4 | 883 | { |
1d98ebfc KO |
884 | if (req->ki_filp) |
885 | fput(req->ki_filp); | |
13389010 DL |
886 | if (req->ki_eventfd != NULL) |
887 | eventfd_ctx_put(req->ki_eventfd); | |
1da177e4 | 888 | kmem_cache_free(kiocb_cachep, req); |
1da177e4 LT |
889 | } |
890 | ||
d5470b59 | 891 | static struct kioctx *lookup_ioctx(unsigned long ctx_id) |
1da177e4 | 892 | { |
db446a08 | 893 | struct aio_ring __user *ring = (void __user *)ctx_id; |
abf137dd | 894 | struct mm_struct *mm = current->mm; |
65c24491 | 895 | struct kioctx *ctx, *ret = NULL; |
db446a08 BL |
896 | struct kioctx_table *table; |
897 | unsigned id; | |
898 | ||
899 | if (get_user(id, &ring->id)) | |
900 | return NULL; | |
1da177e4 | 901 | |
abf137dd | 902 | rcu_read_lock(); |
db446a08 | 903 | table = rcu_dereference(mm->ioctx_table); |
abf137dd | 904 | |
db446a08 BL |
905 | if (!table || id >= table->nr) |
906 | goto out; | |
1da177e4 | 907 | |
db446a08 | 908 | ctx = table->table[id]; |
f30d704f | 909 | if (ctx && ctx->user_id == ctx_id) { |
db446a08 BL |
910 | percpu_ref_get(&ctx->users); |
911 | ret = ctx; | |
912 | } | |
913 | out: | |
abf137dd | 914 | rcu_read_unlock(); |
65c24491 | 915 | return ret; |
1da177e4 LT |
916 | } |
917 | ||
1da177e4 LT |
918 | /* aio_complete |
919 | * Called when the io request on the given iocb is complete. | |
1da177e4 | 920 | */ |
2d68449e | 921 | void aio_complete(struct kiocb *iocb, long res, long res2) |
1da177e4 LT |
922 | { |
923 | struct kioctx *ctx = iocb->ki_ctx; | |
1da177e4 | 924 | struct aio_ring *ring; |
21b40200 | 925 | struct io_event *ev_page, *event; |
1da177e4 | 926 | unsigned long flags; |
21b40200 | 927 | unsigned tail, pos; |
1da177e4 | 928 | |
20dcae32 ZB |
929 | /* |
930 | * Special case handling for sync iocbs: | |
931 | * - events go directly into the iocb for fast handling | |
932 | * - the sync task with the iocb in its stack holds the single iocb | |
933 | * ref, no other paths have a way to get another ref | |
934 | * - the sync task helpfully left a reference to itself in the iocb | |
1da177e4 LT |
935 | */ |
936 | if (is_sync_kiocb(iocb)) { | |
1da177e4 | 937 | iocb->ki_user_data = res; |
57282d8f KO |
938 | smp_wmb(); |
939 | iocb->ki_ctx = ERR_PTR(-EXDEV); | |
1da177e4 | 940 | wake_up_process(iocb->ki_obj.tsk); |
2d68449e | 941 | return; |
1da177e4 LT |
942 | } |
943 | ||
0460fef2 KO |
944 | if (iocb->ki_list.next) { |
945 | unsigned long flags; | |
946 | ||
947 | spin_lock_irqsave(&ctx->ctx_lock, flags); | |
948 | list_del(&iocb->ki_list); | |
949 | spin_unlock_irqrestore(&ctx->ctx_lock, flags); | |
950 | } | |
11599eba | 951 | |
0460fef2 KO |
952 | /* |
953 | * Add a completion event to the ring buffer. Must be done holding | |
4b30f07e | 954 | * ctx->completion_lock to prevent other code from messing with the tail |
0460fef2 KO |
955 | * pointer since we might be called from irq context. |
956 | */ | |
957 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
958 | ||
58c85dc2 | 959 | tail = ctx->tail; |
21b40200 KO |
960 | pos = tail + AIO_EVENTS_OFFSET; |
961 | ||
58c85dc2 | 962 | if (++tail >= ctx->nr_events) |
4bf69b2a | 963 | tail = 0; |
1da177e4 | 964 | |
58c85dc2 | 965 | ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); |
21b40200 KO |
966 | event = ev_page + pos % AIO_EVENTS_PER_PAGE; |
967 | ||
1da177e4 LT |
968 | event->obj = (u64)(unsigned long)iocb->ki_obj.user; |
969 | event->data = iocb->ki_user_data; | |
970 | event->res = res; | |
971 | event->res2 = res2; | |
972 | ||
21b40200 | 973 | kunmap_atomic(ev_page); |
58c85dc2 | 974 | flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); |
21b40200 KO |
975 | |
976 | pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n", | |
caf4167a KO |
977 | ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data, |
978 | res, res2); | |
1da177e4 LT |
979 | |
980 | /* after flagging the request as done, we | |
981 | * must never even look at it again | |
982 | */ | |
983 | smp_wmb(); /* make event visible before updating tail */ | |
984 | ||
58c85dc2 | 985 | ctx->tail = tail; |
1da177e4 | 986 | |
58c85dc2 | 987 | ring = kmap_atomic(ctx->ring_pages[0]); |
21b40200 | 988 | ring->tail = tail; |
e8e3c3d6 | 989 | kunmap_atomic(ring); |
58c85dc2 | 990 | flush_dcache_page(ctx->ring_pages[0]); |
1da177e4 | 991 | |
0460fef2 KO |
992 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
993 | ||
21b40200 | 994 | pr_debug("added to ring %p at [%u]\n", iocb, tail); |
8d1c98b0 DL |
995 | |
996 | /* | |
997 | * Check if the user asked us to deliver the result through an | |
998 | * eventfd. The eventfd_signal() function is safe to be called | |
999 | * from IRQ context. | |
1000 | */ | |
87c3a86e | 1001 | if (iocb->ki_eventfd != NULL) |
8d1c98b0 DL |
1002 | eventfd_signal(iocb->ki_eventfd, 1); |
1003 | ||
1da177e4 | 1004 | /* everything turned out well, dispose of the aiocb. */ |
57282d8f | 1005 | kiocb_free(iocb); |
f8567a38 | 1006 | put_reqs_available(ctx, 1); |
1da177e4 | 1007 | |
6cb2a210 QB |
1008 | /* |
1009 | * We have to order our ring_info tail store above and test | |
1010 | * of the wait list below outside the wait lock. This is | |
1011 | * like in wake_up_bit() where clearing a bit has to be | |
1012 | * ordered with the unlocked test. | |
1013 | */ | |
1014 | smp_mb(); | |
1015 | ||
1da177e4 LT |
1016 | if (waitqueue_active(&ctx->wait)) |
1017 | wake_up(&ctx->wait); | |
1018 | ||
e34ecee2 | 1019 | percpu_ref_put(&ctx->reqs); |
1da177e4 | 1020 | } |
385773e0 | 1021 | EXPORT_SYMBOL(aio_complete); |
1da177e4 | 1022 | |
2be4e7de | 1023 | /* aio_read_events_ring |
a31ad380 KO |
1024 | * Pull an event off of the ioctx's event ring. Returns the number of |
1025 | * events fetched | |
1da177e4 | 1026 | */ |
a31ad380 KO |
1027 | static long aio_read_events_ring(struct kioctx *ctx, |
1028 | struct io_event __user *event, long nr) | |
1da177e4 | 1029 | { |
1da177e4 | 1030 | struct aio_ring *ring; |
5ffac122 | 1031 | unsigned head, tail, pos; |
a31ad380 KO |
1032 | long ret = 0; |
1033 | int copy_ret; | |
1034 | ||
58c85dc2 | 1035 | mutex_lock(&ctx->ring_lock); |
1da177e4 | 1036 | |
fa8a53c3 | 1037 | /* Access to ->ring_pages here is protected by ctx->ring_lock. */ |
58c85dc2 | 1038 | ring = kmap_atomic(ctx->ring_pages[0]); |
a31ad380 | 1039 | head = ring->head; |
5ffac122 | 1040 | tail = ring->tail; |
a31ad380 KO |
1041 | kunmap_atomic(ring); |
1042 | ||
5ffac122 | 1043 | pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events); |
1da177e4 | 1044 | |
5ffac122 | 1045 | if (head == tail) |
1da177e4 LT |
1046 | goto out; |
1047 | ||
edfbbf38 BL |
1048 | head %= ctx->nr_events; |
1049 | tail %= ctx->nr_events; | |
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); |
a31ad380 | 1089 | out: |
58c85dc2 | 1090 | mutex_unlock(&ctx->ring_lock); |
a31ad380 | 1091 | |
1da177e4 LT |
1092 | return ret; |
1093 | } | |
1094 | ||
a31ad380 KO |
1095 | static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr, |
1096 | struct io_event __user *event, long *i) | |
1da177e4 | 1097 | { |
a31ad380 | 1098 | long ret = aio_read_events_ring(ctx, event + *i, nr - *i); |
1da177e4 | 1099 | |
a31ad380 KO |
1100 | if (ret > 0) |
1101 | *i += ret; | |
1da177e4 | 1102 | |
a31ad380 KO |
1103 | if (unlikely(atomic_read(&ctx->dead))) |
1104 | ret = -EINVAL; | |
1da177e4 | 1105 | |
a31ad380 KO |
1106 | if (!*i) |
1107 | *i = ret; | |
1da177e4 | 1108 | |
a31ad380 | 1109 | return ret < 0 || *i >= min_nr; |
1da177e4 LT |
1110 | } |
1111 | ||
a31ad380 | 1112 | static long read_events(struct kioctx *ctx, long min_nr, long nr, |
1da177e4 LT |
1113 | struct io_event __user *event, |
1114 | struct timespec __user *timeout) | |
1115 | { | |
a31ad380 KO |
1116 | ktime_t until = { .tv64 = KTIME_MAX }; |
1117 | long ret = 0; | |
1da177e4 | 1118 | |
1da177e4 LT |
1119 | if (timeout) { |
1120 | struct timespec ts; | |
a31ad380 | 1121 | |
1da177e4 | 1122 | if (unlikely(copy_from_user(&ts, timeout, sizeof(ts)))) |
a31ad380 | 1123 | return -EFAULT; |
1da177e4 | 1124 | |
a31ad380 | 1125 | until = timespec_to_ktime(ts); |
1da177e4 LT |
1126 | } |
1127 | ||
a31ad380 KO |
1128 | /* |
1129 | * Note that aio_read_events() is being called as the conditional - i.e. | |
1130 | * we're calling it after prepare_to_wait() has set task state to | |
1131 | * TASK_INTERRUPTIBLE. | |
1132 | * | |
1133 | * But aio_read_events() can block, and if it blocks it's going to flip | |
1134 | * the task state back to TASK_RUNNING. | |
1135 | * | |
1136 | * This should be ok, provided it doesn't flip the state back to | |
1137 | * TASK_RUNNING and return 0 too much - that causes us to spin. That | |
1138 | * will only happen if the mutex_lock() call blocks, and we then find | |
1139 | * the ringbuffer empty. So in practice we should be ok, but it's | |
1140 | * something to be aware of when touching this code. | |
1141 | */ | |
1142 | wait_event_interruptible_hrtimeout(ctx->wait, | |
1143 | aio_read_events(ctx, min_nr, nr, event, &ret), until); | |
1da177e4 | 1144 | |
a31ad380 KO |
1145 | if (!ret && signal_pending(current)) |
1146 | ret = -EINTR; | |
1da177e4 | 1147 | |
a31ad380 | 1148 | return ret; |
1da177e4 LT |
1149 | } |
1150 | ||
1da177e4 LT |
1151 | /* sys_io_setup: |
1152 | * Create an aio_context capable of receiving at least nr_events. | |
1153 | * ctxp must not point to an aio_context that already exists, and | |
1154 | * must be initialized to 0 prior to the call. On successful | |
1155 | * creation of the aio_context, *ctxp is filled in with the resulting | |
1156 | * handle. May fail with -EINVAL if *ctxp is not initialized, | |
1157 | * if the specified nr_events exceeds internal limits. May fail | |
1158 | * with -EAGAIN if the specified nr_events exceeds the user's limit | |
1159 | * of available events. May fail with -ENOMEM if insufficient kernel | |
1160 | * resources are available. May fail with -EFAULT if an invalid | |
1161 | * pointer is passed for ctxp. Will fail with -ENOSYS if not | |
1162 | * implemented. | |
1163 | */ | |
002c8976 | 1164 | SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp) |
1da177e4 LT |
1165 | { |
1166 | struct kioctx *ioctx = NULL; | |
1167 | unsigned long ctx; | |
1168 | long ret; | |
1169 | ||
1170 | ret = get_user(ctx, ctxp); | |
1171 | if (unlikely(ret)) | |
1172 | goto out; | |
1173 | ||
1174 | ret = -EINVAL; | |
d55b5fda ZB |
1175 | if (unlikely(ctx || nr_events == 0)) { |
1176 | pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n", | |
1177 | ctx, nr_events); | |
1da177e4 LT |
1178 | goto out; |
1179 | } | |
1180 | ||
1181 | ioctx = ioctx_alloc(nr_events); | |
1182 | ret = PTR_ERR(ioctx); | |
1183 | if (!IS_ERR(ioctx)) { | |
1184 | ret = put_user(ioctx->user_id, ctxp); | |
a2e1859a | 1185 | if (ret) |
e02ba72a | 1186 | kill_ioctx(current->mm, ioctx, NULL); |
723be6e3 | 1187 | percpu_ref_put(&ioctx->users); |
1da177e4 LT |
1188 | } |
1189 | ||
1190 | out: | |
1191 | return ret; | |
1192 | } | |
1193 | ||
1194 | /* sys_io_destroy: | |
1195 | * Destroy the aio_context specified. May cancel any outstanding | |
1196 | * AIOs and block on completion. Will fail with -ENOSYS if not | |
642b5123 | 1197 | * implemented. May fail with -EINVAL if the context pointed to |
1da177e4 LT |
1198 | * is invalid. |
1199 | */ | |
002c8976 | 1200 | SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx) |
1da177e4 LT |
1201 | { |
1202 | struct kioctx *ioctx = lookup_ioctx(ctx); | |
1203 | if (likely(NULL != ioctx)) { | |
e02ba72a AP |
1204 | struct completion requests_done = |
1205 | COMPLETION_INITIALIZER_ONSTACK(requests_done); | |
fb2d4483 | 1206 | int ret; |
e02ba72a AP |
1207 | |
1208 | /* Pass requests_done to kill_ioctx() where it can be set | |
1209 | * in a thread-safe way. If we try to set it here then we have | |
1210 | * a race condition if two io_destroy() called simultaneously. | |
1211 | */ | |
fb2d4483 | 1212 | ret = kill_ioctx(current->mm, ioctx, &requests_done); |
723be6e3 | 1213 | percpu_ref_put(&ioctx->users); |
e02ba72a AP |
1214 | |
1215 | /* Wait until all IO for the context are done. Otherwise kernel | |
1216 | * keep using user-space buffers even if user thinks the context | |
1217 | * is destroyed. | |
1218 | */ | |
fb2d4483 BL |
1219 | if (!ret) |
1220 | wait_for_completion(&requests_done); | |
e02ba72a | 1221 | |
fb2d4483 | 1222 | return ret; |
1da177e4 LT |
1223 | } |
1224 | pr_debug("EINVAL: io_destroy: invalid context id\n"); | |
1225 | return -EINVAL; | |
1226 | } | |
1227 | ||
41ef4eb8 KO |
1228 | typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *, |
1229 | unsigned long, loff_t); | |
293bc982 | 1230 | typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *); |
41ef4eb8 | 1231 | |
8bc92afc KO |
1232 | static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb, |
1233 | int rw, char __user *buf, | |
1234 | unsigned long *nr_segs, | |
1235 | struct iovec **iovec, | |
1236 | bool compat) | |
eed4e51f BP |
1237 | { |
1238 | ssize_t ret; | |
1239 | ||
8bc92afc | 1240 | *nr_segs = kiocb->ki_nbytes; |
41ef4eb8 | 1241 | |
9d85cba7 JM |
1242 | #ifdef CONFIG_COMPAT |
1243 | if (compat) | |
41ef4eb8 | 1244 | ret = compat_rw_copy_check_uvector(rw, |
8bc92afc KO |
1245 | (struct compat_iovec __user *)buf, |
1246 | *nr_segs, 1, *iovec, iovec); | |
9d85cba7 JM |
1247 | else |
1248 | #endif | |
41ef4eb8 | 1249 | ret = rw_copy_check_uvector(rw, |
8bc92afc KO |
1250 | (struct iovec __user *)buf, |
1251 | *nr_segs, 1, *iovec, iovec); | |
eed4e51f | 1252 | if (ret < 0) |
41ef4eb8 | 1253 | return ret; |
a70b52ec | 1254 | |
41ef4eb8 | 1255 | /* ki_nbytes now reflect bytes instead of segs */ |
eed4e51f | 1256 | kiocb->ki_nbytes = ret; |
41ef4eb8 | 1257 | return 0; |
eed4e51f BP |
1258 | } |
1259 | ||
8bc92afc KO |
1260 | static ssize_t aio_setup_single_vector(struct kiocb *kiocb, |
1261 | int rw, char __user *buf, | |
1262 | unsigned long *nr_segs, | |
1263 | struct iovec *iovec) | |
eed4e51f | 1264 | { |
8bc92afc | 1265 | if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes))) |
41ef4eb8 | 1266 | return -EFAULT; |
a70b52ec | 1267 | |
8bc92afc KO |
1268 | iovec->iov_base = buf; |
1269 | iovec->iov_len = kiocb->ki_nbytes; | |
1270 | *nr_segs = 1; | |
eed4e51f BP |
1271 | return 0; |
1272 | } | |
1273 | ||
1da177e4 | 1274 | /* |
2be4e7de GZ |
1275 | * aio_run_iocb: |
1276 | * Performs the initial checks and io submission. | |
1da177e4 | 1277 | */ |
8bc92afc KO |
1278 | static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode, |
1279 | char __user *buf, bool compat) | |
1da177e4 | 1280 | { |
41ef4eb8 KO |
1281 | struct file *file = req->ki_filp; |
1282 | ssize_t ret; | |
8bc92afc | 1283 | unsigned long nr_segs; |
41ef4eb8 KO |
1284 | int rw; |
1285 | fmode_t mode; | |
1286 | aio_rw_op *rw_op; | |
293bc982 | 1287 | rw_iter_op *iter_op; |
8bc92afc | 1288 | struct iovec inline_vec, *iovec = &inline_vec; |
293bc982 | 1289 | struct iov_iter iter; |
1da177e4 | 1290 | |
8bc92afc | 1291 | switch (opcode) { |
1da177e4 | 1292 | case IOCB_CMD_PREAD: |
eed4e51f | 1293 | case IOCB_CMD_PREADV: |
41ef4eb8 KO |
1294 | mode = FMODE_READ; |
1295 | rw = READ; | |
1296 | rw_op = file->f_op->aio_read; | |
293bc982 | 1297 | iter_op = file->f_op->read_iter; |
41ef4eb8 KO |
1298 | goto rw_common; |
1299 | ||
1300 | case IOCB_CMD_PWRITE: | |
eed4e51f | 1301 | case IOCB_CMD_PWRITEV: |
41ef4eb8 KO |
1302 | mode = FMODE_WRITE; |
1303 | rw = WRITE; | |
1304 | rw_op = file->f_op->aio_write; | |
293bc982 | 1305 | iter_op = file->f_op->write_iter; |
41ef4eb8 KO |
1306 | goto rw_common; |
1307 | rw_common: | |
1308 | if (unlikely(!(file->f_mode & mode))) | |
1309 | return -EBADF; | |
1310 | ||
293bc982 | 1311 | if (!rw_op && !iter_op) |
41ef4eb8 KO |
1312 | return -EINVAL; |
1313 | ||
8bc92afc KO |
1314 | ret = (opcode == IOCB_CMD_PREADV || |
1315 | opcode == IOCB_CMD_PWRITEV) | |
1316 | ? aio_setup_vectored_rw(req, rw, buf, &nr_segs, | |
1317 | &iovec, compat) | |
1318 | : aio_setup_single_vector(req, rw, buf, &nr_segs, | |
1319 | iovec); | |
754320d6 LY |
1320 | if (!ret) |
1321 | ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes); | |
8bc92afc KO |
1322 | if (ret < 0) { |
1323 | if (iovec != &inline_vec) | |
1324 | kfree(iovec); | |
41ef4eb8 | 1325 | return ret; |
8bc92afc | 1326 | } |
41ef4eb8 KO |
1327 | |
1328 | req->ki_nbytes = ret; | |
41ef4eb8 | 1329 | |
73a7075e KO |
1330 | /* XXX: move/kill - rw_verify_area()? */ |
1331 | /* This matches the pread()/pwrite() logic */ | |
1332 | if (req->ki_pos < 0) { | |
1333 | ret = -EINVAL; | |
1334 | break; | |
1335 | } | |
1336 | ||
1337 | if (rw == WRITE) | |
1338 | file_start_write(file); | |
1339 | ||
293bc982 AV |
1340 | if (iter_op) { |
1341 | iov_iter_init(&iter, rw, iovec, nr_segs, req->ki_nbytes); | |
1342 | ret = iter_op(req, &iter); | |
1343 | } else { | |
1344 | ret = rw_op(req, iovec, nr_segs, req->ki_pos); | |
1345 | } | |
73a7075e KO |
1346 | |
1347 | if (rw == WRITE) | |
1348 | file_end_write(file); | |
1da177e4 | 1349 | break; |
41ef4eb8 | 1350 | |
1da177e4 | 1351 | case IOCB_CMD_FDSYNC: |
41ef4eb8 KO |
1352 | if (!file->f_op->aio_fsync) |
1353 | return -EINVAL; | |
1354 | ||
1355 | ret = file->f_op->aio_fsync(req, 1); | |
1da177e4 | 1356 | break; |
41ef4eb8 | 1357 | |
1da177e4 | 1358 | case IOCB_CMD_FSYNC: |
41ef4eb8 KO |
1359 | if (!file->f_op->aio_fsync) |
1360 | return -EINVAL; | |
1361 | ||
1362 | ret = file->f_op->aio_fsync(req, 0); | |
1da177e4 | 1363 | break; |
41ef4eb8 | 1364 | |
1da177e4 | 1365 | default: |
caf4167a | 1366 | pr_debug("EINVAL: no operation provided\n"); |
41ef4eb8 | 1367 | return -EINVAL; |
1da177e4 LT |
1368 | } |
1369 | ||
8bc92afc KO |
1370 | if (iovec != &inline_vec) |
1371 | kfree(iovec); | |
1372 | ||
41ef4eb8 KO |
1373 | if (ret != -EIOCBQUEUED) { |
1374 | /* | |
1375 | * There's no easy way to restart the syscall since other AIO's | |
1376 | * may be already running. Just fail this IO with EINTR. | |
1377 | */ | |
1378 | if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR || | |
1379 | ret == -ERESTARTNOHAND || | |
1380 | ret == -ERESTART_RESTARTBLOCK)) | |
1381 | ret = -EINTR; | |
1382 | aio_complete(req, ret, 0); | |
1383 | } | |
1da177e4 LT |
1384 | |
1385 | return 0; | |
1386 | } | |
1387 | ||
d5470b59 | 1388 | static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb, |
a1c8eae7 | 1389 | struct iocb *iocb, bool compat) |
1da177e4 LT |
1390 | { |
1391 | struct kiocb *req; | |
1da177e4 LT |
1392 | ssize_t ret; |
1393 | ||
1394 | /* enforce forwards compatibility on users */ | |
9c3060be | 1395 | if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) { |
caf4167a | 1396 | pr_debug("EINVAL: reserve field set\n"); |
1da177e4 LT |
1397 | return -EINVAL; |
1398 | } | |
1399 | ||
1400 | /* prevent overflows */ | |
1401 | if (unlikely( | |
1402 | (iocb->aio_buf != (unsigned long)iocb->aio_buf) || | |
1403 | (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) || | |
1404 | ((ssize_t)iocb->aio_nbytes < 0) | |
1405 | )) { | |
1406 | pr_debug("EINVAL: io_submit: overflow check\n"); | |
1407 | return -EINVAL; | |
1408 | } | |
1409 | ||
41ef4eb8 | 1410 | req = aio_get_req(ctx); |
1d98ebfc | 1411 | if (unlikely(!req)) |
1da177e4 | 1412 | return -EAGAIN; |
1d98ebfc KO |
1413 | |
1414 | req->ki_filp = fget(iocb->aio_fildes); | |
1415 | if (unlikely(!req->ki_filp)) { | |
1416 | ret = -EBADF; | |
1417 | goto out_put_req; | |
1da177e4 | 1418 | } |
1d98ebfc | 1419 | |
9c3060be DL |
1420 | if (iocb->aio_flags & IOCB_FLAG_RESFD) { |
1421 | /* | |
1422 | * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an | |
1423 | * instance of the file* now. The file descriptor must be | |
1424 | * an eventfd() fd, and will be signaled for each completed | |
1425 | * event using the eventfd_signal() function. | |
1426 | */ | |
13389010 | 1427 | req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd); |
801678c5 | 1428 | if (IS_ERR(req->ki_eventfd)) { |
9c3060be | 1429 | ret = PTR_ERR(req->ki_eventfd); |
87c3a86e | 1430 | req->ki_eventfd = NULL; |
9c3060be DL |
1431 | goto out_put_req; |
1432 | } | |
1433 | } | |
1da177e4 | 1434 | |
8a660890 | 1435 | ret = put_user(KIOCB_KEY, &user_iocb->aio_key); |
1da177e4 | 1436 | if (unlikely(ret)) { |
caf4167a | 1437 | pr_debug("EFAULT: aio_key\n"); |
1da177e4 LT |
1438 | goto out_put_req; |
1439 | } | |
1440 | ||
1441 | req->ki_obj.user = user_iocb; | |
1442 | req->ki_user_data = iocb->aio_data; | |
1443 | req->ki_pos = iocb->aio_offset; | |
73a7075e | 1444 | req->ki_nbytes = iocb->aio_nbytes; |
1da177e4 | 1445 | |
8bc92afc KO |
1446 | ret = aio_run_iocb(req, iocb->aio_lio_opcode, |
1447 | (char __user *)(unsigned long)iocb->aio_buf, | |
1448 | compat); | |
41003a7b | 1449 | if (ret) |
7137c6bd | 1450 | goto out_put_req; |
41003a7b | 1451 | |
1da177e4 | 1452 | return 0; |
1da177e4 | 1453 | out_put_req: |
e1bdd5f2 | 1454 | put_reqs_available(ctx, 1); |
e34ecee2 | 1455 | percpu_ref_put(&ctx->reqs); |
57282d8f | 1456 | kiocb_free(req); |
1da177e4 LT |
1457 | return ret; |
1458 | } | |
1459 | ||
9d85cba7 JM |
1460 | long do_io_submit(aio_context_t ctx_id, long nr, |
1461 | struct iocb __user *__user *iocbpp, bool compat) | |
1da177e4 LT |
1462 | { |
1463 | struct kioctx *ctx; | |
1464 | long ret = 0; | |
080d676d | 1465 | int i = 0; |
9f5b9425 | 1466 | struct blk_plug plug; |
1da177e4 LT |
1467 | |
1468 | if (unlikely(nr < 0)) | |
1469 | return -EINVAL; | |
1470 | ||
75e1c70f JM |
1471 | if (unlikely(nr > LONG_MAX/sizeof(*iocbpp))) |
1472 | nr = LONG_MAX/sizeof(*iocbpp); | |
1473 | ||
1da177e4 LT |
1474 | if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp))))) |
1475 | return -EFAULT; | |
1476 | ||
1477 | ctx = lookup_ioctx(ctx_id); | |
1478 | if (unlikely(!ctx)) { | |
caf4167a | 1479 | pr_debug("EINVAL: invalid context id\n"); |
1da177e4 LT |
1480 | return -EINVAL; |
1481 | } | |
1482 | ||
9f5b9425 SL |
1483 | blk_start_plug(&plug); |
1484 | ||
1da177e4 LT |
1485 | /* |
1486 | * AKPM: should this return a partial result if some of the IOs were | |
1487 | * successfully submitted? | |
1488 | */ | |
1489 | for (i=0; i<nr; i++) { | |
1490 | struct iocb __user *user_iocb; | |
1491 | struct iocb tmp; | |
1492 | ||
1493 | if (unlikely(__get_user(user_iocb, iocbpp + i))) { | |
1494 | ret = -EFAULT; | |
1495 | break; | |
1496 | } | |
1497 | ||
1498 | if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) { | |
1499 | ret = -EFAULT; | |
1500 | break; | |
1501 | } | |
1502 | ||
a1c8eae7 | 1503 | ret = io_submit_one(ctx, user_iocb, &tmp, compat); |
1da177e4 LT |
1504 | if (ret) |
1505 | break; | |
1506 | } | |
9f5b9425 | 1507 | blk_finish_plug(&plug); |
1da177e4 | 1508 | |
723be6e3 | 1509 | percpu_ref_put(&ctx->users); |
1da177e4 LT |
1510 | return i ? i : ret; |
1511 | } | |
1512 | ||
9d85cba7 JM |
1513 | /* sys_io_submit: |
1514 | * Queue the nr iocbs pointed to by iocbpp for processing. Returns | |
1515 | * the number of iocbs queued. May return -EINVAL if the aio_context | |
1516 | * specified by ctx_id is invalid, if nr is < 0, if the iocb at | |
1517 | * *iocbpp[0] is not properly initialized, if the operation specified | |
1518 | * is invalid for the file descriptor in the iocb. May fail with | |
1519 | * -EFAULT if any of the data structures point to invalid data. May | |
1520 | * fail with -EBADF if the file descriptor specified in the first | |
1521 | * iocb is invalid. May fail with -EAGAIN if insufficient resources | |
1522 | * are available to queue any iocbs. Will return 0 if nr is 0. Will | |
1523 | * fail with -ENOSYS if not implemented. | |
1524 | */ | |
1525 | SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr, | |
1526 | struct iocb __user * __user *, iocbpp) | |
1527 | { | |
1528 | return do_io_submit(ctx_id, nr, iocbpp, 0); | |
1529 | } | |
1530 | ||
1da177e4 LT |
1531 | /* lookup_kiocb |
1532 | * Finds a given iocb for cancellation. | |
1da177e4 | 1533 | */ |
25ee7e38 AB |
1534 | static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, |
1535 | u32 key) | |
1da177e4 LT |
1536 | { |
1537 | struct list_head *pos; | |
d00689af ZB |
1538 | |
1539 | assert_spin_locked(&ctx->ctx_lock); | |
1540 | ||
8a660890 KO |
1541 | if (key != KIOCB_KEY) |
1542 | return NULL; | |
1543 | ||
1da177e4 LT |
1544 | /* TODO: use a hash or array, this sucks. */ |
1545 | list_for_each(pos, &ctx->active_reqs) { | |
1546 | struct kiocb *kiocb = list_kiocb(pos); | |
8a660890 | 1547 | if (kiocb->ki_obj.user == iocb) |
1da177e4 LT |
1548 | return kiocb; |
1549 | } | |
1550 | return NULL; | |
1551 | } | |
1552 | ||
1553 | /* sys_io_cancel: | |
1554 | * Attempts to cancel an iocb previously passed to io_submit. If | |
1555 | * the operation is successfully cancelled, the resulting event is | |
1556 | * copied into the memory pointed to by result without being placed | |
1557 | * into the completion queue and 0 is returned. May fail with | |
1558 | * -EFAULT if any of the data structures pointed to are invalid. | |
1559 | * May fail with -EINVAL if aio_context specified by ctx_id is | |
1560 | * invalid. May fail with -EAGAIN if the iocb specified was not | |
1561 | * cancelled. Will fail with -ENOSYS if not implemented. | |
1562 | */ | |
002c8976 HC |
1563 | SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb, |
1564 | struct io_event __user *, result) | |
1da177e4 | 1565 | { |
1da177e4 LT |
1566 | struct kioctx *ctx; |
1567 | struct kiocb *kiocb; | |
1568 | u32 key; | |
1569 | int ret; | |
1570 | ||
1571 | ret = get_user(key, &iocb->aio_key); | |
1572 | if (unlikely(ret)) | |
1573 | return -EFAULT; | |
1574 | ||
1575 | ctx = lookup_ioctx(ctx_id); | |
1576 | if (unlikely(!ctx)) | |
1577 | return -EINVAL; | |
1578 | ||
1579 | spin_lock_irq(&ctx->ctx_lock); | |
906b973c | 1580 | |
1da177e4 | 1581 | kiocb = lookup_kiocb(ctx, iocb, key); |
906b973c | 1582 | if (kiocb) |
d52a8f9e | 1583 | ret = kiocb_cancel(kiocb); |
906b973c KO |
1584 | else |
1585 | ret = -EINVAL; | |
1586 | ||
1da177e4 LT |
1587 | spin_unlock_irq(&ctx->ctx_lock); |
1588 | ||
906b973c | 1589 | if (!ret) { |
bec68faa KO |
1590 | /* |
1591 | * The result argument is no longer used - the io_event is | |
1592 | * always delivered via the ring buffer. -EINPROGRESS indicates | |
1593 | * cancellation is progress: | |
906b973c | 1594 | */ |
bec68faa | 1595 | ret = -EINPROGRESS; |
906b973c | 1596 | } |
1da177e4 | 1597 | |
723be6e3 | 1598 | percpu_ref_put(&ctx->users); |
1da177e4 LT |
1599 | |
1600 | return ret; | |
1601 | } | |
1602 | ||
1603 | /* io_getevents: | |
1604 | * Attempts to read at least min_nr events and up to nr events from | |
642b5123 ST |
1605 | * the completion queue for the aio_context specified by ctx_id. If |
1606 | * it succeeds, the number of read events is returned. May fail with | |
1607 | * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is | |
1608 | * out of range, if timeout is out of range. May fail with -EFAULT | |
1609 | * if any of the memory specified is invalid. May return 0 or | |
1610 | * < min_nr if the timeout specified by timeout has elapsed | |
1611 | * before sufficient events are available, where timeout == NULL | |
1612 | * specifies an infinite timeout. Note that the timeout pointed to by | |
6900807c | 1613 | * timeout is relative. Will fail with -ENOSYS if not implemented. |
1da177e4 | 1614 | */ |
002c8976 HC |
1615 | SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id, |
1616 | long, min_nr, | |
1617 | long, nr, | |
1618 | struct io_event __user *, events, | |
1619 | struct timespec __user *, timeout) | |
1da177e4 LT |
1620 | { |
1621 | struct kioctx *ioctx = lookup_ioctx(ctx_id); | |
1622 | long ret = -EINVAL; | |
1623 | ||
1624 | if (likely(ioctx)) { | |
2e410255 | 1625 | if (likely(min_nr <= nr && min_nr >= 0)) |
1da177e4 | 1626 | ret = read_events(ioctx, min_nr, nr, events, timeout); |
723be6e3 | 1627 | percpu_ref_put(&ioctx->users); |
1da177e4 | 1628 | } |
1da177e4 LT |
1629 | return ret; |
1630 | } |