]>
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. | |
bfe4037e | 8 | * Copyright 2018 Christoph Hellwig. |
1da177e4 LT |
9 | * |
10 | * See ../COPYING for licensing terms. | |
11 | */ | |
caf4167a KO |
12 | #define pr_fmt(fmt) "%s: " fmt, __func__ |
13 | ||
1da177e4 LT |
14 | #include <linux/kernel.h> |
15 | #include <linux/init.h> | |
16 | #include <linux/errno.h> | |
17 | #include <linux/time.h> | |
18 | #include <linux/aio_abi.h> | |
630d9c47 | 19 | #include <linux/export.h> |
1da177e4 | 20 | #include <linux/syscalls.h> |
b9d128f1 | 21 | #include <linux/backing-dev.h> |
9018ccc4 | 22 | #include <linux/refcount.h> |
027445c3 | 23 | #include <linux/uio.h> |
1da177e4 | 24 | |
174cd4b1 | 25 | #include <linux/sched/signal.h> |
1da177e4 LT |
26 | #include <linux/fs.h> |
27 | #include <linux/file.h> | |
28 | #include <linux/mm.h> | |
29 | #include <linux/mman.h> | |
3d2d827f | 30 | #include <linux/mmu_context.h> |
e1bdd5f2 | 31 | #include <linux/percpu.h> |
1da177e4 LT |
32 | #include <linux/slab.h> |
33 | #include <linux/timer.h> | |
34 | #include <linux/aio.h> | |
35 | #include <linux/highmem.h> | |
36 | #include <linux/workqueue.h> | |
37 | #include <linux/security.h> | |
9c3060be | 38 | #include <linux/eventfd.h> |
cfb1e33e | 39 | #include <linux/blkdev.h> |
9d85cba7 | 40 | #include <linux/compat.h> |
36bc08cc GZ |
41 | #include <linux/migrate.h> |
42 | #include <linux/ramfs.h> | |
723be6e3 | 43 | #include <linux/percpu-refcount.h> |
71ad7490 | 44 | #include <linux/mount.h> |
52db59df | 45 | #include <linux/pseudo_fs.h> |
1da177e4 LT |
46 | |
47 | #include <asm/kmap_types.h> | |
7c0f6ba6 | 48 | #include <linux/uaccess.h> |
a538e3ff | 49 | #include <linux/nospec.h> |
1da177e4 | 50 | |
68d70d03 AV |
51 | #include "internal.h" |
52 | ||
f3a2752a CH |
53 | #define KIOCB_KEY 0 |
54 | ||
4e179bca KO |
55 | #define AIO_RING_MAGIC 0xa10a10a1 |
56 | #define AIO_RING_COMPAT_FEATURES 1 | |
57 | #define AIO_RING_INCOMPAT_FEATURES 0 | |
58 | struct aio_ring { | |
59 | unsigned id; /* kernel internal index number */ | |
60 | unsigned nr; /* number of io_events */ | |
fa8a53c3 BL |
61 | unsigned head; /* Written to by userland or under ring_lock |
62 | * mutex by aio_read_events_ring(). */ | |
4e179bca KO |
63 | unsigned tail; |
64 | ||
65 | unsigned magic; | |
66 | unsigned compat_features; | |
67 | unsigned incompat_features; | |
68 | unsigned header_length; /* size of aio_ring */ | |
69 | ||
70 | ||
71 | struct io_event io_events[0]; | |
72 | }; /* 128 bytes + ring size */ | |
73 | ||
a79d40e9 JA |
74 | /* |
75 | * Plugging is meant to work with larger batches of IOs. If we don't | |
76 | * have more than the below, then don't bother setting up a plug. | |
77 | */ | |
78 | #define AIO_PLUG_THRESHOLD 2 | |
79 | ||
4e179bca | 80 | #define AIO_RING_PAGES 8 |
4e179bca | 81 | |
db446a08 | 82 | struct kioctx_table { |
d0264c01 TH |
83 | struct rcu_head rcu; |
84 | unsigned nr; | |
85 | struct kioctx __rcu *table[]; | |
db446a08 BL |
86 | }; |
87 | ||
e1bdd5f2 KO |
88 | struct kioctx_cpu { |
89 | unsigned reqs_available; | |
90 | }; | |
91 | ||
dc48e56d JA |
92 | struct ctx_rq_wait { |
93 | struct completion comp; | |
94 | atomic_t count; | |
95 | }; | |
96 | ||
4e179bca | 97 | struct kioctx { |
723be6e3 | 98 | struct percpu_ref users; |
36f55889 | 99 | atomic_t dead; |
4e179bca | 100 | |
e34ecee2 KO |
101 | struct percpu_ref reqs; |
102 | ||
4e179bca | 103 | unsigned long user_id; |
4e179bca | 104 | |
e1bdd5f2 KO |
105 | struct __percpu kioctx_cpu *cpu; |
106 | ||
107 | /* | |
108 | * For percpu reqs_available, number of slots we move to/from global | |
109 | * counter at a time: | |
110 | */ | |
111 | unsigned req_batch; | |
3e845ce0 KO |
112 | /* |
113 | * This is what userspace passed to io_setup(), it's not used for | |
114 | * anything but counting against the global max_reqs quota. | |
115 | * | |
58c85dc2 | 116 | * The real limit is nr_events - 1, which will be larger (see |
3e845ce0 KO |
117 | * aio_setup_ring()) |
118 | */ | |
4e179bca KO |
119 | unsigned max_reqs; |
120 | ||
58c85dc2 KO |
121 | /* Size of ringbuffer, in units of struct io_event */ |
122 | unsigned nr_events; | |
4e179bca | 123 | |
58c85dc2 KO |
124 | unsigned long mmap_base; |
125 | unsigned long mmap_size; | |
126 | ||
127 | struct page **ring_pages; | |
128 | long nr_pages; | |
129 | ||
f729863a | 130 | struct rcu_work free_rwork; /* see free_ioctx() */ |
4e23bcae | 131 | |
e02ba72a AP |
132 | /* |
133 | * signals when all in-flight requests are done | |
134 | */ | |
dc48e56d | 135 | struct ctx_rq_wait *rq_wait; |
e02ba72a | 136 | |
4e23bcae | 137 | struct { |
34e83fc6 KO |
138 | /* |
139 | * This counts the number of available slots in the ringbuffer, | |
140 | * so we avoid overflowing it: it's decremented (if positive) | |
141 | * when allocating a kiocb and incremented when the resulting | |
142 | * io_event is pulled off the ringbuffer. | |
e1bdd5f2 KO |
143 | * |
144 | * We batch accesses to it with a percpu version. | |
34e83fc6 KO |
145 | */ |
146 | atomic_t reqs_available; | |
4e23bcae KO |
147 | } ____cacheline_aligned_in_smp; |
148 | ||
149 | struct { | |
150 | spinlock_t ctx_lock; | |
151 | struct list_head active_reqs; /* used for cancellation */ | |
152 | } ____cacheline_aligned_in_smp; | |
153 | ||
58c85dc2 KO |
154 | struct { |
155 | struct mutex ring_lock; | |
4e23bcae KO |
156 | wait_queue_head_t wait; |
157 | } ____cacheline_aligned_in_smp; | |
58c85dc2 KO |
158 | |
159 | struct { | |
160 | unsigned tail; | |
d856f32a | 161 | unsigned completed_events; |
58c85dc2 | 162 | spinlock_t completion_lock; |
4e23bcae | 163 | } ____cacheline_aligned_in_smp; |
58c85dc2 KO |
164 | |
165 | struct page *internal_pages[AIO_RING_PAGES]; | |
36bc08cc | 166 | struct file *aio_ring_file; |
db446a08 BL |
167 | |
168 | unsigned id; | |
4e179bca KO |
169 | }; |
170 | ||
84c4e1f8 LT |
171 | /* |
172 | * First field must be the file pointer in all the | |
173 | * iocb unions! See also 'struct kiocb' in <linux/fs.h> | |
174 | */ | |
a3c0d439 | 175 | struct fsync_iocb { |
a3c0d439 | 176 | struct file *file; |
84c4e1f8 | 177 | struct work_struct work; |
a3c0d439 CH |
178 | bool datasync; |
179 | }; | |
180 | ||
bfe4037e CH |
181 | struct poll_iocb { |
182 | struct file *file; | |
183 | struct wait_queue_head *head; | |
184 | __poll_t events; | |
af5c72b1 | 185 | bool done; |
bfe4037e CH |
186 | bool cancelled; |
187 | struct wait_queue_entry wait; | |
188 | struct work_struct work; | |
189 | }; | |
190 | ||
84c4e1f8 LT |
191 | /* |
192 | * NOTE! Each of the iocb union members has the file pointer | |
193 | * as the first entry in their struct definition. So you can | |
194 | * access the file pointer through any of the sub-structs, | |
195 | * or directly as just 'ki_filp' in this struct. | |
196 | */ | |
04b2fa9f | 197 | struct aio_kiocb { |
54843f87 | 198 | union { |
84c4e1f8 | 199 | struct file *ki_filp; |
54843f87 | 200 | struct kiocb rw; |
a3c0d439 | 201 | struct fsync_iocb fsync; |
bfe4037e | 202 | struct poll_iocb poll; |
54843f87 | 203 | }; |
04b2fa9f CH |
204 | |
205 | struct kioctx *ki_ctx; | |
206 | kiocb_cancel_fn *ki_cancel; | |
207 | ||
a9339b78 | 208 | struct io_event ki_res; |
04b2fa9f CH |
209 | |
210 | struct list_head ki_list; /* the aio core uses this | |
211 | * for cancellation */ | |
9018ccc4 | 212 | refcount_t ki_refcnt; |
04b2fa9f CH |
213 | |
214 | /* | |
215 | * If the aio_resfd field of the userspace iocb is not zero, | |
216 | * this is the underlying eventfd context to deliver events to. | |
217 | */ | |
218 | struct eventfd_ctx *ki_eventfd; | |
219 | }; | |
220 | ||
1da177e4 | 221 | /*------ sysctl variables----*/ |
d55b5fda ZB |
222 | static DEFINE_SPINLOCK(aio_nr_lock); |
223 | unsigned long aio_nr; /* current system wide number of aio requests */ | |
224 | unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */ | |
1da177e4 LT |
225 | /*----end sysctl variables---*/ |
226 | ||
e18b890b CL |
227 | static struct kmem_cache *kiocb_cachep; |
228 | static struct kmem_cache *kioctx_cachep; | |
1da177e4 | 229 | |
71ad7490 BL |
230 | static struct vfsmount *aio_mnt; |
231 | ||
232 | static const struct file_operations aio_ring_fops; | |
233 | static const struct address_space_operations aio_ctx_aops; | |
234 | ||
235 | static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages) | |
236 | { | |
71ad7490 | 237 | struct file *file; |
71ad7490 | 238 | struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb); |
7f62656b DC |
239 | if (IS_ERR(inode)) |
240 | return ERR_CAST(inode); | |
71ad7490 BL |
241 | |
242 | inode->i_mapping->a_ops = &aio_ctx_aops; | |
243 | inode->i_mapping->private_data = ctx; | |
244 | inode->i_size = PAGE_SIZE * nr_pages; | |
245 | ||
d93aa9d8 AV |
246 | file = alloc_file_pseudo(inode, aio_mnt, "[aio]", |
247 | O_RDWR, &aio_ring_fops); | |
c9c554f2 | 248 | if (IS_ERR(file)) |
71ad7490 | 249 | iput(inode); |
71ad7490 BL |
250 | return file; |
251 | } | |
252 | ||
52db59df | 253 | static int aio_init_fs_context(struct fs_context *fc) |
71ad7490 | 254 | { |
52db59df DH |
255 | if (!init_pseudo(fc, AIO_RING_MAGIC)) |
256 | return -ENOMEM; | |
257 | fc->s_iflags |= SB_I_NOEXEC; | |
258 | return 0; | |
71ad7490 BL |
259 | } |
260 | ||
1da177e4 LT |
261 | /* aio_setup |
262 | * Creates the slab caches used by the aio routines, panic on | |
263 | * failure as this is done early during the boot sequence. | |
264 | */ | |
265 | static int __init aio_setup(void) | |
266 | { | |
71ad7490 BL |
267 | static struct file_system_type aio_fs = { |
268 | .name = "aio", | |
52db59df | 269 | .init_fs_context = aio_init_fs_context, |
71ad7490 BL |
270 | .kill_sb = kill_anon_super, |
271 | }; | |
272 | aio_mnt = kern_mount(&aio_fs); | |
273 | if (IS_ERR(aio_mnt)) | |
274 | panic("Failed to create aio fs mount."); | |
275 | ||
04b2fa9f | 276 | kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC); |
0a31bd5f | 277 | kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC); |
1da177e4 LT |
278 | return 0; |
279 | } | |
385773e0 | 280 | __initcall(aio_setup); |
1da177e4 | 281 | |
5e9ae2e5 BL |
282 | static void put_aio_ring_file(struct kioctx *ctx) |
283 | { | |
284 | struct file *aio_ring_file = ctx->aio_ring_file; | |
de04e769 RV |
285 | struct address_space *i_mapping; |
286 | ||
5e9ae2e5 | 287 | if (aio_ring_file) { |
45063097 | 288 | truncate_setsize(file_inode(aio_ring_file), 0); |
5e9ae2e5 BL |
289 | |
290 | /* Prevent further access to the kioctx from migratepages */ | |
45063097 | 291 | i_mapping = aio_ring_file->f_mapping; |
de04e769 RV |
292 | spin_lock(&i_mapping->private_lock); |
293 | i_mapping->private_data = NULL; | |
5e9ae2e5 | 294 | ctx->aio_ring_file = NULL; |
de04e769 | 295 | spin_unlock(&i_mapping->private_lock); |
5e9ae2e5 BL |
296 | |
297 | fput(aio_ring_file); | |
298 | } | |
299 | } | |
300 | ||
1da177e4 LT |
301 | static void aio_free_ring(struct kioctx *ctx) |
302 | { | |
36bc08cc | 303 | int i; |
1da177e4 | 304 | |
fa8a53c3 BL |
305 | /* Disconnect the kiotx from the ring file. This prevents future |
306 | * accesses to the kioctx from page migration. | |
307 | */ | |
308 | put_aio_ring_file(ctx); | |
309 | ||
36bc08cc | 310 | for (i = 0; i < ctx->nr_pages; i++) { |
8e321fef | 311 | struct page *page; |
36bc08cc GZ |
312 | pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i, |
313 | page_count(ctx->ring_pages[i])); | |
8e321fef BL |
314 | page = ctx->ring_pages[i]; |
315 | if (!page) | |
316 | continue; | |
317 | ctx->ring_pages[i] = NULL; | |
318 | put_page(page); | |
36bc08cc | 319 | } |
1da177e4 | 320 | |
ddb8c45b | 321 | if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) { |
58c85dc2 | 322 | kfree(ctx->ring_pages); |
ddb8c45b SL |
323 | ctx->ring_pages = NULL; |
324 | } | |
36bc08cc GZ |
325 | } |
326 | ||
5477e70a | 327 | static int aio_ring_mremap(struct vm_area_struct *vma) |
e4a0d3e7 | 328 | { |
5477e70a | 329 | struct file *file = vma->vm_file; |
e4a0d3e7 PE |
330 | struct mm_struct *mm = vma->vm_mm; |
331 | struct kioctx_table *table; | |
b2edffdd | 332 | int i, res = -EINVAL; |
e4a0d3e7 PE |
333 | |
334 | spin_lock(&mm->ioctx_lock); | |
335 | rcu_read_lock(); | |
336 | table = rcu_dereference(mm->ioctx_table); | |
337 | for (i = 0; i < table->nr; i++) { | |
338 | struct kioctx *ctx; | |
339 | ||
d0264c01 | 340 | ctx = rcu_dereference(table->table[i]); |
e4a0d3e7 | 341 | if (ctx && ctx->aio_ring_file == file) { |
b2edffdd AV |
342 | if (!atomic_read(&ctx->dead)) { |
343 | ctx->user_id = ctx->mmap_base = vma->vm_start; | |
344 | res = 0; | |
345 | } | |
e4a0d3e7 PE |
346 | break; |
347 | } | |
348 | } | |
349 | ||
350 | rcu_read_unlock(); | |
351 | spin_unlock(&mm->ioctx_lock); | |
b2edffdd | 352 | return res; |
e4a0d3e7 PE |
353 | } |
354 | ||
5477e70a ON |
355 | static const struct vm_operations_struct aio_ring_vm_ops = { |
356 | .mremap = aio_ring_mremap, | |
357 | #if IS_ENABLED(CONFIG_MMU) | |
358 | .fault = filemap_fault, | |
359 | .map_pages = filemap_map_pages, | |
360 | .page_mkwrite = filemap_page_mkwrite, | |
361 | #endif | |
362 | }; | |
363 | ||
364 | static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma) | |
365 | { | |
366 | vma->vm_flags |= VM_DONTEXPAND; | |
367 | vma->vm_ops = &aio_ring_vm_ops; | |
368 | return 0; | |
369 | } | |
370 | ||
36bc08cc GZ |
371 | static const struct file_operations aio_ring_fops = { |
372 | .mmap = aio_ring_mmap, | |
373 | }; | |
374 | ||
0c45355f | 375 | #if IS_ENABLED(CONFIG_MIGRATION) |
36bc08cc GZ |
376 | static int aio_migratepage(struct address_space *mapping, struct page *new, |
377 | struct page *old, enum migrate_mode mode) | |
378 | { | |
5e9ae2e5 | 379 | struct kioctx *ctx; |
36bc08cc | 380 | unsigned long flags; |
fa8a53c3 | 381 | pgoff_t idx; |
36bc08cc GZ |
382 | int rc; |
383 | ||
2916ecc0 JG |
384 | /* |
385 | * We cannot support the _NO_COPY case here, because copy needs to | |
386 | * happen under the ctx->completion_lock. That does not work with the | |
387 | * migration workflow of MIGRATE_SYNC_NO_COPY. | |
388 | */ | |
389 | if (mode == MIGRATE_SYNC_NO_COPY) | |
390 | return -EINVAL; | |
391 | ||
8e321fef BL |
392 | rc = 0; |
393 | ||
fa8a53c3 | 394 | /* mapping->private_lock here protects against the kioctx teardown. */ |
8e321fef BL |
395 | spin_lock(&mapping->private_lock); |
396 | ctx = mapping->private_data; | |
fa8a53c3 BL |
397 | if (!ctx) { |
398 | rc = -EINVAL; | |
399 | goto out; | |
400 | } | |
401 | ||
402 | /* The ring_lock mutex. The prevents aio_read_events() from writing | |
403 | * to the ring's head, and prevents page migration from mucking in | |
404 | * a partially initialized kiotx. | |
405 | */ | |
406 | if (!mutex_trylock(&ctx->ring_lock)) { | |
407 | rc = -EAGAIN; | |
408 | goto out; | |
409 | } | |
410 | ||
411 | idx = old->index; | |
412 | if (idx < (pgoff_t)ctx->nr_pages) { | |
413 | /* Make sure the old page hasn't already been changed */ | |
414 | if (ctx->ring_pages[idx] != old) | |
415 | rc = -EAGAIN; | |
8e321fef BL |
416 | } else |
417 | rc = -EINVAL; | |
8e321fef BL |
418 | |
419 | if (rc != 0) | |
fa8a53c3 | 420 | goto out_unlock; |
8e321fef | 421 | |
36bc08cc GZ |
422 | /* Writeback must be complete */ |
423 | BUG_ON(PageWriteback(old)); | |
8e321fef | 424 | get_page(new); |
36bc08cc | 425 | |
37109694 | 426 | rc = migrate_page_move_mapping(mapping, new, old, 1); |
36bc08cc | 427 | if (rc != MIGRATEPAGE_SUCCESS) { |
8e321fef | 428 | put_page(new); |
fa8a53c3 | 429 | goto out_unlock; |
36bc08cc GZ |
430 | } |
431 | ||
fa8a53c3 BL |
432 | /* Take completion_lock to prevent other writes to the ring buffer |
433 | * while the old page is copied to the new. This prevents new | |
434 | * events from being lost. | |
5e9ae2e5 | 435 | */ |
fa8a53c3 BL |
436 | spin_lock_irqsave(&ctx->completion_lock, flags); |
437 | migrate_page_copy(new, old); | |
438 | BUG_ON(ctx->ring_pages[idx] != old); | |
439 | ctx->ring_pages[idx] = new; | |
440 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
36bc08cc | 441 | |
fa8a53c3 BL |
442 | /* The old page is no longer accessible. */ |
443 | put_page(old); | |
8e321fef | 444 | |
fa8a53c3 BL |
445 | out_unlock: |
446 | mutex_unlock(&ctx->ring_lock); | |
447 | out: | |
448 | spin_unlock(&mapping->private_lock); | |
36bc08cc | 449 | return rc; |
1da177e4 | 450 | } |
0c45355f | 451 | #endif |
1da177e4 | 452 | |
36bc08cc | 453 | static const struct address_space_operations aio_ctx_aops = { |
835f252c | 454 | .set_page_dirty = __set_page_dirty_no_writeback, |
0c45355f | 455 | #if IS_ENABLED(CONFIG_MIGRATION) |
36bc08cc | 456 | .migratepage = aio_migratepage, |
0c45355f | 457 | #endif |
36bc08cc GZ |
458 | }; |
459 | ||
2a8a9867 | 460 | static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events) |
1da177e4 LT |
461 | { |
462 | struct aio_ring *ring; | |
41003a7b | 463 | struct mm_struct *mm = current->mm; |
3dc9acb6 | 464 | unsigned long size, unused; |
1da177e4 | 465 | int nr_pages; |
36bc08cc GZ |
466 | int i; |
467 | struct file *file; | |
1da177e4 LT |
468 | |
469 | /* Compensate for the ring buffer's head/tail overlap entry */ | |
470 | nr_events += 2; /* 1 is required, 2 for good luck */ | |
471 | ||
472 | size = sizeof(struct aio_ring); | |
473 | size += sizeof(struct io_event) * nr_events; | |
1da177e4 | 474 | |
36bc08cc | 475 | nr_pages = PFN_UP(size); |
1da177e4 LT |
476 | if (nr_pages < 0) |
477 | return -EINVAL; | |
478 | ||
71ad7490 | 479 | file = aio_private_file(ctx, nr_pages); |
36bc08cc GZ |
480 | if (IS_ERR(file)) { |
481 | ctx->aio_ring_file = NULL; | |
fa8a53c3 | 482 | return -ENOMEM; |
36bc08cc GZ |
483 | } |
484 | ||
3dc9acb6 LT |
485 | ctx->aio_ring_file = file; |
486 | nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) | |
487 | / sizeof(struct io_event); | |
488 | ||
489 | ctx->ring_pages = ctx->internal_pages; | |
490 | if (nr_pages > AIO_RING_PAGES) { | |
491 | ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *), | |
492 | GFP_KERNEL); | |
493 | if (!ctx->ring_pages) { | |
494 | put_aio_ring_file(ctx); | |
495 | return -ENOMEM; | |
496 | } | |
497 | } | |
498 | ||
36bc08cc GZ |
499 | for (i = 0; i < nr_pages; i++) { |
500 | struct page *page; | |
45063097 | 501 | page = find_or_create_page(file->f_mapping, |
36bc08cc GZ |
502 | i, GFP_HIGHUSER | __GFP_ZERO); |
503 | if (!page) | |
504 | break; | |
505 | pr_debug("pid(%d) page[%d]->count=%d\n", | |
506 | current->pid, i, page_count(page)); | |
507 | SetPageUptodate(page); | |
36bc08cc | 508 | unlock_page(page); |
3dc9acb6 LT |
509 | |
510 | ctx->ring_pages[i] = page; | |
36bc08cc | 511 | } |
3dc9acb6 | 512 | ctx->nr_pages = i; |
1da177e4 | 513 | |
3dc9acb6 LT |
514 | if (unlikely(i != nr_pages)) { |
515 | aio_free_ring(ctx); | |
fa8a53c3 | 516 | return -ENOMEM; |
1da177e4 LT |
517 | } |
518 | ||
58c85dc2 KO |
519 | ctx->mmap_size = nr_pages * PAGE_SIZE; |
520 | pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size); | |
36bc08cc | 521 | |
013373e8 MH |
522 | if (down_write_killable(&mm->mmap_sem)) { |
523 | ctx->mmap_size = 0; | |
524 | aio_free_ring(ctx); | |
525 | return -EINTR; | |
526 | } | |
527 | ||
36bc08cc GZ |
528 | ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size, |
529 | PROT_READ | PROT_WRITE, | |
897ab3e0 | 530 | MAP_SHARED, 0, &unused, NULL); |
3dc9acb6 | 531 | up_write(&mm->mmap_sem); |
58c85dc2 | 532 | if (IS_ERR((void *)ctx->mmap_base)) { |
58c85dc2 | 533 | ctx->mmap_size = 0; |
1da177e4 | 534 | aio_free_ring(ctx); |
fa8a53c3 | 535 | return -ENOMEM; |
1da177e4 LT |
536 | } |
537 | ||
58c85dc2 | 538 | pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base); |
d6c355c7 | 539 | |
58c85dc2 KO |
540 | ctx->user_id = ctx->mmap_base; |
541 | ctx->nr_events = nr_events; /* trusted copy */ | |
1da177e4 | 542 | |
58c85dc2 | 543 | ring = kmap_atomic(ctx->ring_pages[0]); |
1da177e4 | 544 | ring->nr = nr_events; /* user copy */ |
db446a08 | 545 | ring->id = ~0U; |
1da177e4 LT |
546 | ring->head = ring->tail = 0; |
547 | ring->magic = AIO_RING_MAGIC; | |
548 | ring->compat_features = AIO_RING_COMPAT_FEATURES; | |
549 | ring->incompat_features = AIO_RING_INCOMPAT_FEATURES; | |
550 | ring->header_length = sizeof(struct aio_ring); | |
e8e3c3d6 | 551 | kunmap_atomic(ring); |
58c85dc2 | 552 | flush_dcache_page(ctx->ring_pages[0]); |
1da177e4 LT |
553 | |
554 | return 0; | |
555 | } | |
556 | ||
1da177e4 LT |
557 | #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event)) |
558 | #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event)) | |
559 | #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE) | |
560 | ||
04b2fa9f | 561 | void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel) |
0460fef2 | 562 | { |
54843f87 | 563 | struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw); |
0460fef2 KO |
564 | struct kioctx *ctx = req->ki_ctx; |
565 | unsigned long flags; | |
566 | ||
75321b50 CH |
567 | if (WARN_ON_ONCE(!list_empty(&req->ki_list))) |
568 | return; | |
0460fef2 | 569 | |
75321b50 CH |
570 | spin_lock_irqsave(&ctx->ctx_lock, flags); |
571 | list_add_tail(&req->ki_list, &ctx->active_reqs); | |
0460fef2 | 572 | req->ki_cancel = cancel; |
0460fef2 KO |
573 | spin_unlock_irqrestore(&ctx->ctx_lock, flags); |
574 | } | |
575 | EXPORT_SYMBOL(kiocb_set_cancel_fn); | |
576 | ||
a6d7cff4 TH |
577 | /* |
578 | * free_ioctx() should be RCU delayed to synchronize against the RCU | |
579 | * protected lookup_ioctx() and also needs process context to call | |
f729863a | 580 | * aio_free_ring(). Use rcu_work. |
a6d7cff4 | 581 | */ |
e34ecee2 | 582 | static void free_ioctx(struct work_struct *work) |
36f55889 | 583 | { |
f729863a TH |
584 | struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx, |
585 | free_rwork); | |
e34ecee2 | 586 | pr_debug("freeing %p\n", ctx); |
e1bdd5f2 | 587 | |
e34ecee2 | 588 | aio_free_ring(ctx); |
e1bdd5f2 | 589 | free_percpu(ctx->cpu); |
9a1049da TH |
590 | percpu_ref_exit(&ctx->reqs); |
591 | percpu_ref_exit(&ctx->users); | |
36f55889 KO |
592 | kmem_cache_free(kioctx_cachep, ctx); |
593 | } | |
594 | ||
e34ecee2 KO |
595 | static void free_ioctx_reqs(struct percpu_ref *ref) |
596 | { | |
597 | struct kioctx *ctx = container_of(ref, struct kioctx, reqs); | |
598 | ||
e02ba72a | 599 | /* At this point we know that there are no any in-flight requests */ |
dc48e56d JA |
600 | if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count)) |
601 | complete(&ctx->rq_wait->comp); | |
e02ba72a | 602 | |
a6d7cff4 | 603 | /* Synchronize against RCU protected table->table[] dereferences */ |
f729863a TH |
604 | INIT_RCU_WORK(&ctx->free_rwork, free_ioctx); |
605 | queue_rcu_work(system_wq, &ctx->free_rwork); | |
e34ecee2 KO |
606 | } |
607 | ||
36f55889 KO |
608 | /* |
609 | * When this function runs, the kioctx has been removed from the "hash table" | |
610 | * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted - | |
611 | * now it's safe to cancel any that need to be. | |
612 | */ | |
e34ecee2 | 613 | static void free_ioctx_users(struct percpu_ref *ref) |
36f55889 | 614 | { |
e34ecee2 | 615 | struct kioctx *ctx = container_of(ref, struct kioctx, users); |
04b2fa9f | 616 | struct aio_kiocb *req; |
36f55889 KO |
617 | |
618 | spin_lock_irq(&ctx->ctx_lock); | |
619 | ||
620 | while (!list_empty(&ctx->active_reqs)) { | |
621 | req = list_first_entry(&ctx->active_reqs, | |
04b2fa9f | 622 | struct aio_kiocb, ki_list); |
888933f8 | 623 | req->ki_cancel(&req->rw); |
4faa9996 | 624 | list_del_init(&req->ki_list); |
36f55889 KO |
625 | } |
626 | ||
627 | spin_unlock_irq(&ctx->ctx_lock); | |
628 | ||
e34ecee2 KO |
629 | percpu_ref_kill(&ctx->reqs); |
630 | percpu_ref_put(&ctx->reqs); | |
36f55889 KO |
631 | } |
632 | ||
db446a08 BL |
633 | static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm) |
634 | { | |
635 | unsigned i, new_nr; | |
636 | struct kioctx_table *table, *old; | |
637 | struct aio_ring *ring; | |
638 | ||
639 | spin_lock(&mm->ioctx_lock); | |
855ef0de | 640 | table = rcu_dereference_raw(mm->ioctx_table); |
db446a08 BL |
641 | |
642 | while (1) { | |
643 | if (table) | |
644 | for (i = 0; i < table->nr; i++) | |
d0264c01 | 645 | if (!rcu_access_pointer(table->table[i])) { |
db446a08 | 646 | ctx->id = i; |
d0264c01 | 647 | rcu_assign_pointer(table->table[i], ctx); |
db446a08 BL |
648 | spin_unlock(&mm->ioctx_lock); |
649 | ||
fa8a53c3 BL |
650 | /* While kioctx setup is in progress, |
651 | * we are protected from page migration | |
652 | * changes ring_pages by ->ring_lock. | |
653 | */ | |
db446a08 BL |
654 | ring = kmap_atomic(ctx->ring_pages[0]); |
655 | ring->id = ctx->id; | |
656 | kunmap_atomic(ring); | |
657 | return 0; | |
658 | } | |
659 | ||
660 | new_nr = (table ? table->nr : 1) * 4; | |
db446a08 BL |
661 | spin_unlock(&mm->ioctx_lock); |
662 | ||
663 | table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) * | |
664 | new_nr, GFP_KERNEL); | |
665 | if (!table) | |
666 | return -ENOMEM; | |
667 | ||
668 | table->nr = new_nr; | |
669 | ||
670 | spin_lock(&mm->ioctx_lock); | |
855ef0de | 671 | old = rcu_dereference_raw(mm->ioctx_table); |
db446a08 BL |
672 | |
673 | if (!old) { | |
674 | rcu_assign_pointer(mm->ioctx_table, table); | |
675 | } else if (table->nr > old->nr) { | |
676 | memcpy(table->table, old->table, | |
677 | old->nr * sizeof(struct kioctx *)); | |
678 | ||
679 | rcu_assign_pointer(mm->ioctx_table, table); | |
680 | kfree_rcu(old, rcu); | |
681 | } else { | |
682 | kfree(table); | |
683 | table = old; | |
684 | } | |
685 | } | |
686 | } | |
687 | ||
e34ecee2 KO |
688 | static void aio_nr_sub(unsigned nr) |
689 | { | |
690 | spin_lock(&aio_nr_lock); | |
691 | if (WARN_ON(aio_nr - nr > aio_nr)) | |
692 | aio_nr = 0; | |
693 | else | |
694 | aio_nr -= nr; | |
695 | spin_unlock(&aio_nr_lock); | |
696 | } | |
697 | ||
1da177e4 LT |
698 | /* ioctx_alloc |
699 | * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed. | |
700 | */ | |
701 | static struct kioctx *ioctx_alloc(unsigned nr_events) | |
702 | { | |
41003a7b | 703 | struct mm_struct *mm = current->mm; |
1da177e4 | 704 | struct kioctx *ctx; |
e23754f8 | 705 | int err = -ENOMEM; |
1da177e4 | 706 | |
2a8a9867 MFO |
707 | /* |
708 | * Store the original nr_events -- what userspace passed to io_setup(), | |
709 | * for counting against the global limit -- before it changes. | |
710 | */ | |
711 | unsigned int max_reqs = nr_events; | |
712 | ||
e1bdd5f2 KO |
713 | /* |
714 | * We keep track of the number of available ringbuffer slots, to prevent | |
715 | * overflow (reqs_available), and we also use percpu counters for this. | |
716 | * | |
717 | * So since up to half the slots might be on other cpu's percpu counters | |
718 | * and unavailable, double nr_events so userspace sees what they | |
719 | * expected: additionally, we move req_batch slots to/from percpu | |
720 | * counters at a time, so make sure that isn't 0: | |
721 | */ | |
722 | nr_events = max(nr_events, num_possible_cpus() * 4); | |
723 | nr_events *= 2; | |
724 | ||
1da177e4 | 725 | /* Prevent overflows */ |
08397acd | 726 | if (nr_events > (0x10000000U / sizeof(struct io_event))) { |
1da177e4 LT |
727 | pr_debug("ENOMEM: nr_events too high\n"); |
728 | return ERR_PTR(-EINVAL); | |
729 | } | |
730 | ||
2a8a9867 | 731 | if (!nr_events || (unsigned long)max_reqs > aio_max_nr) |
1da177e4 LT |
732 | return ERR_PTR(-EAGAIN); |
733 | ||
c3762229 | 734 | ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL); |
1da177e4 LT |
735 | if (!ctx) |
736 | return ERR_PTR(-ENOMEM); | |
737 | ||
2a8a9867 | 738 | ctx->max_reqs = max_reqs; |
1da177e4 | 739 | |
1da177e4 | 740 | spin_lock_init(&ctx->ctx_lock); |
0460fef2 | 741 | spin_lock_init(&ctx->completion_lock); |
58c85dc2 | 742 | mutex_init(&ctx->ring_lock); |
fa8a53c3 BL |
743 | /* Protect against page migration throughout kiotx setup by keeping |
744 | * the ring_lock mutex held until setup is complete. */ | |
745 | mutex_lock(&ctx->ring_lock); | |
1da177e4 LT |
746 | init_waitqueue_head(&ctx->wait); |
747 | ||
748 | INIT_LIST_HEAD(&ctx->active_reqs); | |
1da177e4 | 749 | |
2aad2a86 | 750 | if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL)) |
fa8a53c3 BL |
751 | goto err; |
752 | ||
2aad2a86 | 753 | if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL)) |
fa8a53c3 BL |
754 | goto err; |
755 | ||
e1bdd5f2 KO |
756 | ctx->cpu = alloc_percpu(struct kioctx_cpu); |
757 | if (!ctx->cpu) | |
e34ecee2 | 758 | goto err; |
1da177e4 | 759 | |
2a8a9867 | 760 | err = aio_setup_ring(ctx, nr_events); |
fa8a53c3 | 761 | if (err < 0) |
e34ecee2 | 762 | goto err; |
e1bdd5f2 | 763 | |
34e83fc6 | 764 | atomic_set(&ctx->reqs_available, ctx->nr_events - 1); |
e1bdd5f2 | 765 | ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4); |
6878ea72 BL |
766 | if (ctx->req_batch < 1) |
767 | ctx->req_batch = 1; | |
34e83fc6 | 768 | |
1da177e4 | 769 | /* limit the number of system wide aios */ |
9fa1cb39 | 770 | spin_lock(&aio_nr_lock); |
2a8a9867 MFO |
771 | if (aio_nr + ctx->max_reqs > aio_max_nr || |
772 | aio_nr + ctx->max_reqs < aio_nr) { | |
9fa1cb39 | 773 | spin_unlock(&aio_nr_lock); |
e34ecee2 | 774 | err = -EAGAIN; |
d1b94327 | 775 | goto err_ctx; |
2dd542b7 AV |
776 | } |
777 | aio_nr += ctx->max_reqs; | |
9fa1cb39 | 778 | spin_unlock(&aio_nr_lock); |
1da177e4 | 779 | |
1881686f BL |
780 | percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */ |
781 | percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */ | |
723be6e3 | 782 | |
da90382c BL |
783 | err = ioctx_add_table(ctx, mm); |
784 | if (err) | |
e34ecee2 | 785 | goto err_cleanup; |
da90382c | 786 | |
fa8a53c3 BL |
787 | /* Release the ring_lock mutex now that all setup is complete. */ |
788 | mutex_unlock(&ctx->ring_lock); | |
789 | ||
caf4167a | 790 | pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n", |
58c85dc2 | 791 | ctx, ctx->user_id, mm, ctx->nr_events); |
1da177e4 LT |
792 | return ctx; |
793 | ||
e34ecee2 KO |
794 | err_cleanup: |
795 | aio_nr_sub(ctx->max_reqs); | |
d1b94327 | 796 | err_ctx: |
deeb8525 AV |
797 | atomic_set(&ctx->dead, 1); |
798 | if (ctx->mmap_size) | |
799 | vm_munmap(ctx->mmap_base, ctx->mmap_size); | |
d1b94327 | 800 | aio_free_ring(ctx); |
e34ecee2 | 801 | err: |
fa8a53c3 | 802 | mutex_unlock(&ctx->ring_lock); |
e1bdd5f2 | 803 | free_percpu(ctx->cpu); |
9a1049da TH |
804 | percpu_ref_exit(&ctx->reqs); |
805 | percpu_ref_exit(&ctx->users); | |
1da177e4 | 806 | kmem_cache_free(kioctx_cachep, ctx); |
caf4167a | 807 | pr_debug("error allocating ioctx %d\n", err); |
e23754f8 | 808 | return ERR_PTR(err); |
1da177e4 LT |
809 | } |
810 | ||
36f55889 KO |
811 | /* kill_ioctx |
812 | * Cancels all outstanding aio requests on an aio context. Used | |
813 | * when the processes owning a context have all exited to encourage | |
814 | * the rapid destruction of the kioctx. | |
815 | */ | |
fb2d4483 | 816 | static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx, |
dc48e56d | 817 | struct ctx_rq_wait *wait) |
36f55889 | 818 | { |
fa88b6f8 | 819 | struct kioctx_table *table; |
db446a08 | 820 | |
b2edffdd AV |
821 | spin_lock(&mm->ioctx_lock); |
822 | if (atomic_xchg(&ctx->dead, 1)) { | |
823 | spin_unlock(&mm->ioctx_lock); | |
fa88b6f8 | 824 | return -EINVAL; |
b2edffdd | 825 | } |
db446a08 | 826 | |
855ef0de | 827 | table = rcu_dereference_raw(mm->ioctx_table); |
d0264c01 TH |
828 | WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id])); |
829 | RCU_INIT_POINTER(table->table[ctx->id], NULL); | |
fa88b6f8 | 830 | spin_unlock(&mm->ioctx_lock); |
4fcc712f | 831 | |
a6d7cff4 | 832 | /* free_ioctx_reqs() will do the necessary RCU synchronization */ |
fa88b6f8 | 833 | wake_up_all(&ctx->wait); |
4fcc712f | 834 | |
fa88b6f8 BL |
835 | /* |
836 | * It'd be more correct to do this in free_ioctx(), after all | |
837 | * the outstanding kiocbs have finished - but by then io_destroy | |
838 | * has already returned, so io_setup() could potentially return | |
839 | * -EAGAIN with no ioctxs actually in use (as far as userspace | |
840 | * could tell). | |
841 | */ | |
842 | aio_nr_sub(ctx->max_reqs); | |
4fcc712f | 843 | |
fa88b6f8 BL |
844 | if (ctx->mmap_size) |
845 | vm_munmap(ctx->mmap_base, ctx->mmap_size); | |
fb2d4483 | 846 | |
dc48e56d | 847 | ctx->rq_wait = wait; |
fa88b6f8 BL |
848 | percpu_ref_kill(&ctx->users); |
849 | return 0; | |
1da177e4 LT |
850 | } |
851 | ||
36f55889 KO |
852 | /* |
853 | * exit_aio: called when the last user of mm goes away. At this point, there is | |
854 | * no way for any new requests to be submited or any of the io_* syscalls to be | |
855 | * called on the context. | |
856 | * | |
857 | * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on | |
858 | * them. | |
1da177e4 | 859 | */ |
fc9b52cd | 860 | void exit_aio(struct mm_struct *mm) |
1da177e4 | 861 | { |
4b70ac5f | 862 | struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table); |
dc48e56d JA |
863 | struct ctx_rq_wait wait; |
864 | int i, skipped; | |
db446a08 | 865 | |
4b70ac5f ON |
866 | if (!table) |
867 | return; | |
db446a08 | 868 | |
dc48e56d JA |
869 | atomic_set(&wait.count, table->nr); |
870 | init_completion(&wait.comp); | |
871 | ||
872 | skipped = 0; | |
4b70ac5f | 873 | for (i = 0; i < table->nr; ++i) { |
d0264c01 TH |
874 | struct kioctx *ctx = |
875 | rcu_dereference_protected(table->table[i], true); | |
abf137dd | 876 | |
dc48e56d JA |
877 | if (!ctx) { |
878 | skipped++; | |
4b70ac5f | 879 | continue; |
dc48e56d JA |
880 | } |
881 | ||
936af157 | 882 | /* |
4b70ac5f ON |
883 | * We don't need to bother with munmap() here - exit_mmap(mm) |
884 | * is coming and it'll unmap everything. And we simply can't, | |
885 | * this is not necessarily our ->mm. | |
886 | * Since kill_ioctx() uses non-zero ->mmap_size as indicator | |
887 | * that it needs to unmap the area, just set it to 0. | |
936af157 | 888 | */ |
58c85dc2 | 889 | ctx->mmap_size = 0; |
dc48e56d JA |
890 | kill_ioctx(mm, ctx, &wait); |
891 | } | |
36f55889 | 892 | |
dc48e56d | 893 | if (!atomic_sub_and_test(skipped, &wait.count)) { |
6098b45b | 894 | /* Wait until all IO for the context are done. */ |
dc48e56d | 895 | wait_for_completion(&wait.comp); |
1da177e4 | 896 | } |
4b70ac5f ON |
897 | |
898 | RCU_INIT_POINTER(mm->ioctx_table, NULL); | |
899 | kfree(table); | |
1da177e4 LT |
900 | } |
901 | ||
e1bdd5f2 KO |
902 | static void put_reqs_available(struct kioctx *ctx, unsigned nr) |
903 | { | |
904 | struct kioctx_cpu *kcpu; | |
263782c1 | 905 | unsigned long flags; |
e1bdd5f2 | 906 | |
263782c1 | 907 | local_irq_save(flags); |
be6fb451 | 908 | kcpu = this_cpu_ptr(ctx->cpu); |
e1bdd5f2 | 909 | kcpu->reqs_available += nr; |
263782c1 | 910 | |
e1bdd5f2 KO |
911 | while (kcpu->reqs_available >= ctx->req_batch * 2) { |
912 | kcpu->reqs_available -= ctx->req_batch; | |
913 | atomic_add(ctx->req_batch, &ctx->reqs_available); | |
914 | } | |
915 | ||
263782c1 | 916 | local_irq_restore(flags); |
e1bdd5f2 KO |
917 | } |
918 | ||
432c7997 | 919 | static bool __get_reqs_available(struct kioctx *ctx) |
e1bdd5f2 KO |
920 | { |
921 | struct kioctx_cpu *kcpu; | |
922 | bool ret = false; | |
263782c1 | 923 | unsigned long flags; |
e1bdd5f2 | 924 | |
263782c1 | 925 | local_irq_save(flags); |
be6fb451 | 926 | kcpu = this_cpu_ptr(ctx->cpu); |
e1bdd5f2 KO |
927 | if (!kcpu->reqs_available) { |
928 | int old, avail = atomic_read(&ctx->reqs_available); | |
929 | ||
930 | do { | |
931 | if (avail < ctx->req_batch) | |
932 | goto out; | |
933 | ||
934 | old = avail; | |
935 | avail = atomic_cmpxchg(&ctx->reqs_available, | |
936 | avail, avail - ctx->req_batch); | |
937 | } while (avail != old); | |
938 | ||
939 | kcpu->reqs_available += ctx->req_batch; | |
940 | } | |
941 | ||
942 | ret = true; | |
943 | kcpu->reqs_available--; | |
944 | out: | |
263782c1 | 945 | local_irq_restore(flags); |
e1bdd5f2 KO |
946 | return ret; |
947 | } | |
948 | ||
d856f32a BL |
949 | /* refill_reqs_available |
950 | * Updates the reqs_available reference counts used for tracking the | |
951 | * number of free slots in the completion ring. This can be called | |
952 | * from aio_complete() (to optimistically update reqs_available) or | |
953 | * from aio_get_req() (the we're out of events case). It must be | |
954 | * called holding ctx->completion_lock. | |
955 | */ | |
956 | static void refill_reqs_available(struct kioctx *ctx, unsigned head, | |
957 | unsigned tail) | |
958 | { | |
959 | unsigned events_in_ring, completed; | |
960 | ||
961 | /* Clamp head since userland can write to it. */ | |
962 | head %= ctx->nr_events; | |
963 | if (head <= tail) | |
964 | events_in_ring = tail - head; | |
965 | else | |
966 | events_in_ring = ctx->nr_events - (head - tail); | |
967 | ||
968 | completed = ctx->completed_events; | |
969 | if (events_in_ring < completed) | |
970 | completed -= events_in_ring; | |
971 | else | |
972 | completed = 0; | |
973 | ||
974 | if (!completed) | |
975 | return; | |
976 | ||
977 | ctx->completed_events -= completed; | |
978 | put_reqs_available(ctx, completed); | |
979 | } | |
980 | ||
981 | /* user_refill_reqs_available | |
982 | * Called to refill reqs_available when aio_get_req() encounters an | |
983 | * out of space in the completion ring. | |
984 | */ | |
985 | static void user_refill_reqs_available(struct kioctx *ctx) | |
986 | { | |
987 | spin_lock_irq(&ctx->completion_lock); | |
988 | if (ctx->completed_events) { | |
989 | struct aio_ring *ring; | |
990 | unsigned head; | |
991 | ||
992 | /* Access of ring->head may race with aio_read_events_ring() | |
993 | * here, but that's okay since whether we read the old version | |
994 | * or the new version, and either will be valid. The important | |
995 | * part is that head cannot pass tail since we prevent | |
996 | * aio_complete() from updating tail by holding | |
997 | * ctx->completion_lock. Even if head is invalid, the check | |
998 | * against ctx->completed_events below will make sure we do the | |
999 | * safe/right thing. | |
1000 | */ | |
1001 | ring = kmap_atomic(ctx->ring_pages[0]); | |
1002 | head = ring->head; | |
1003 | kunmap_atomic(ring); | |
1004 | ||
1005 | refill_reqs_available(ctx, head, ctx->tail); | |
1006 | } | |
1007 | ||
1008 | spin_unlock_irq(&ctx->completion_lock); | |
1009 | } | |
1010 | ||
432c7997 CH |
1011 | static bool get_reqs_available(struct kioctx *ctx) |
1012 | { | |
1013 | if (__get_reqs_available(ctx)) | |
1014 | return true; | |
1015 | user_refill_reqs_available(ctx); | |
1016 | return __get_reqs_available(ctx); | |
1017 | } | |
1018 | ||
1da177e4 | 1019 | /* aio_get_req |
57282d8f KO |
1020 | * Allocate a slot for an aio request. |
1021 | * Returns NULL if no requests are free. | |
b53119f1 LT |
1022 | * |
1023 | * The refcount is initialized to 2 - one for the async op completion, | |
1024 | * one for the synchronous code that does this. | |
1da177e4 | 1025 | */ |
04b2fa9f | 1026 | static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx) |
1da177e4 | 1027 | { |
04b2fa9f | 1028 | struct aio_kiocb *req; |
a1c8eae7 | 1029 | |
2bc4ca9b | 1030 | req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL); |
1da177e4 | 1031 | if (unlikely(!req)) |
432c7997 | 1032 | return NULL; |
1da177e4 | 1033 | |
fa0ca2ae | 1034 | if (unlikely(!get_reqs_available(ctx))) { |
6af1c849 | 1035 | kmem_cache_free(kiocb_cachep, req); |
fa0ca2ae AV |
1036 | return NULL; |
1037 | } | |
1038 | ||
e34ecee2 | 1039 | percpu_ref_get(&ctx->reqs); |
2bc4ca9b | 1040 | req->ki_ctx = ctx; |
75321b50 | 1041 | INIT_LIST_HEAD(&req->ki_list); |
b53119f1 | 1042 | refcount_set(&req->ki_refcnt, 2); |
2bc4ca9b | 1043 | req->ki_eventfd = NULL; |
080d676d | 1044 | return req; |
1da177e4 LT |
1045 | } |
1046 | ||
d5470b59 | 1047 | static struct kioctx *lookup_ioctx(unsigned long ctx_id) |
1da177e4 | 1048 | { |
db446a08 | 1049 | struct aio_ring __user *ring = (void __user *)ctx_id; |
abf137dd | 1050 | struct mm_struct *mm = current->mm; |
65c24491 | 1051 | struct kioctx *ctx, *ret = NULL; |
db446a08 BL |
1052 | struct kioctx_table *table; |
1053 | unsigned id; | |
1054 | ||
1055 | if (get_user(id, &ring->id)) | |
1056 | return NULL; | |
1da177e4 | 1057 | |
abf137dd | 1058 | rcu_read_lock(); |
db446a08 | 1059 | table = rcu_dereference(mm->ioctx_table); |
abf137dd | 1060 | |
db446a08 BL |
1061 | if (!table || id >= table->nr) |
1062 | goto out; | |
1da177e4 | 1063 | |
a538e3ff | 1064 | id = array_index_nospec(id, table->nr); |
d0264c01 | 1065 | ctx = rcu_dereference(table->table[id]); |
f30d704f | 1066 | if (ctx && ctx->user_id == ctx_id) { |
baf10564 AV |
1067 | if (percpu_ref_tryget_live(&ctx->users)) |
1068 | ret = ctx; | |
db446a08 BL |
1069 | } |
1070 | out: | |
abf137dd | 1071 | rcu_read_unlock(); |
65c24491 | 1072 | return ret; |
1da177e4 LT |
1073 | } |
1074 | ||
b53119f1 LT |
1075 | static inline void iocb_destroy(struct aio_kiocb *iocb) |
1076 | { | |
74259703 AV |
1077 | if (iocb->ki_eventfd) |
1078 | eventfd_ctx_put(iocb->ki_eventfd); | |
b53119f1 LT |
1079 | if (iocb->ki_filp) |
1080 | fput(iocb->ki_filp); | |
1081 | percpu_ref_put(&iocb->ki_ctx->reqs); | |
1082 | kmem_cache_free(kiocb_cachep, iocb); | |
1083 | } | |
1084 | ||
1da177e4 LT |
1085 | /* aio_complete |
1086 | * Called when the io request on the given iocb is complete. | |
1da177e4 | 1087 | */ |
2bb874c0 | 1088 | static void aio_complete(struct aio_kiocb *iocb) |
1da177e4 LT |
1089 | { |
1090 | struct kioctx *ctx = iocb->ki_ctx; | |
1da177e4 | 1091 | struct aio_ring *ring; |
21b40200 | 1092 | struct io_event *ev_page, *event; |
d856f32a | 1093 | unsigned tail, pos, head; |
1da177e4 | 1094 | unsigned long flags; |
1da177e4 | 1095 | |
0460fef2 KO |
1096 | /* |
1097 | * Add a completion event to the ring buffer. Must be done holding | |
4b30f07e | 1098 | * ctx->completion_lock to prevent other code from messing with the tail |
0460fef2 KO |
1099 | * pointer since we might be called from irq context. |
1100 | */ | |
1101 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
1102 | ||
58c85dc2 | 1103 | tail = ctx->tail; |
21b40200 KO |
1104 | pos = tail + AIO_EVENTS_OFFSET; |
1105 | ||
58c85dc2 | 1106 | if (++tail >= ctx->nr_events) |
4bf69b2a | 1107 | tail = 0; |
1da177e4 | 1108 | |
58c85dc2 | 1109 | ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); |
21b40200 KO |
1110 | event = ev_page + pos % AIO_EVENTS_PER_PAGE; |
1111 | ||
a9339b78 | 1112 | *event = iocb->ki_res; |
1da177e4 | 1113 | |
21b40200 | 1114 | kunmap_atomic(ev_page); |
58c85dc2 | 1115 | flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]); |
21b40200 | 1116 | |
a9339b78 AV |
1117 | pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb, |
1118 | (void __user *)(unsigned long)iocb->ki_res.obj, | |
1119 | iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2); | |
1da177e4 LT |
1120 | |
1121 | /* after flagging the request as done, we | |
1122 | * must never even look at it again | |
1123 | */ | |
1124 | smp_wmb(); /* make event visible before updating tail */ | |
1125 | ||
58c85dc2 | 1126 | ctx->tail = tail; |
1da177e4 | 1127 | |
58c85dc2 | 1128 | ring = kmap_atomic(ctx->ring_pages[0]); |
d856f32a | 1129 | head = ring->head; |
21b40200 | 1130 | ring->tail = tail; |
e8e3c3d6 | 1131 | kunmap_atomic(ring); |
58c85dc2 | 1132 | flush_dcache_page(ctx->ring_pages[0]); |
1da177e4 | 1133 | |
d856f32a BL |
1134 | ctx->completed_events++; |
1135 | if (ctx->completed_events > 1) | |
1136 | refill_reqs_available(ctx, head, tail); | |
0460fef2 KO |
1137 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
1138 | ||
21b40200 | 1139 | pr_debug("added to ring %p at [%u]\n", iocb, tail); |
8d1c98b0 DL |
1140 | |
1141 | /* | |
1142 | * Check if the user asked us to deliver the result through an | |
1143 | * eventfd. The eventfd_signal() function is safe to be called | |
1144 | * from IRQ context. | |
1145 | */ | |
74259703 | 1146 | if (iocb->ki_eventfd) |
8d1c98b0 DL |
1147 | eventfd_signal(iocb->ki_eventfd, 1); |
1148 | ||
6cb2a210 QB |
1149 | /* |
1150 | * We have to order our ring_info tail store above and test | |
1151 | * of the wait list below outside the wait lock. This is | |
1152 | * like in wake_up_bit() where clearing a bit has to be | |
1153 | * ordered with the unlocked test. | |
1154 | */ | |
1155 | smp_mb(); | |
1156 | ||
1da177e4 LT |
1157 | if (waitqueue_active(&ctx->wait)) |
1158 | wake_up(&ctx->wait); | |
2bb874c0 AV |
1159 | } |
1160 | ||
1161 | static inline void iocb_put(struct aio_kiocb *iocb) | |
1162 | { | |
1163 | if (refcount_dec_and_test(&iocb->ki_refcnt)) { | |
1164 | aio_complete(iocb); | |
1165 | iocb_destroy(iocb); | |
1166 | } | |
1da177e4 LT |
1167 | } |
1168 | ||
2be4e7de | 1169 | /* aio_read_events_ring |
a31ad380 KO |
1170 | * Pull an event off of the ioctx's event ring. Returns the number of |
1171 | * events fetched | |
1da177e4 | 1172 | */ |
a31ad380 KO |
1173 | static long aio_read_events_ring(struct kioctx *ctx, |
1174 | struct io_event __user *event, long nr) | |
1da177e4 | 1175 | { |
1da177e4 | 1176 | struct aio_ring *ring; |
5ffac122 | 1177 | unsigned head, tail, pos; |
a31ad380 KO |
1178 | long ret = 0; |
1179 | int copy_ret; | |
1180 | ||
9c9ce763 DC |
1181 | /* |
1182 | * The mutex can block and wake us up and that will cause | |
1183 | * wait_event_interruptible_hrtimeout() to schedule without sleeping | |
1184 | * and repeat. This should be rare enough that it doesn't cause | |
1185 | * peformance issues. See the comment in read_events() for more detail. | |
1186 | */ | |
1187 | sched_annotate_sleep(); | |
58c85dc2 | 1188 | mutex_lock(&ctx->ring_lock); |
1da177e4 | 1189 | |
fa8a53c3 | 1190 | /* Access to ->ring_pages here is protected by ctx->ring_lock. */ |
58c85dc2 | 1191 | ring = kmap_atomic(ctx->ring_pages[0]); |
a31ad380 | 1192 | head = ring->head; |
5ffac122 | 1193 | tail = ring->tail; |
a31ad380 KO |
1194 | kunmap_atomic(ring); |
1195 | ||
2ff396be JM |
1196 | /* |
1197 | * Ensure that once we've read the current tail pointer, that | |
1198 | * we also see the events that were stored up to the tail. | |
1199 | */ | |
1200 | smp_rmb(); | |
1201 | ||
5ffac122 | 1202 | pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events); |
1da177e4 | 1203 | |
5ffac122 | 1204 | if (head == tail) |
1da177e4 LT |
1205 | goto out; |
1206 | ||
edfbbf38 BL |
1207 | head %= ctx->nr_events; |
1208 | tail %= ctx->nr_events; | |
1209 | ||
a31ad380 KO |
1210 | while (ret < nr) { |
1211 | long avail; | |
1212 | struct io_event *ev; | |
1213 | struct page *page; | |
1214 | ||
5ffac122 KO |
1215 | avail = (head <= tail ? tail : ctx->nr_events) - head; |
1216 | if (head == tail) | |
a31ad380 KO |
1217 | break; |
1218 | ||
a31ad380 | 1219 | pos = head + AIO_EVENTS_OFFSET; |
58c85dc2 | 1220 | page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]; |
a31ad380 KO |
1221 | pos %= AIO_EVENTS_PER_PAGE; |
1222 | ||
d2988bd4 AV |
1223 | avail = min(avail, nr - ret); |
1224 | avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos); | |
1225 | ||
a31ad380 KO |
1226 | ev = kmap(page); |
1227 | copy_ret = copy_to_user(event + ret, ev + pos, | |
1228 | sizeof(*ev) * avail); | |
1229 | kunmap(page); | |
1230 | ||
1231 | if (unlikely(copy_ret)) { | |
1232 | ret = -EFAULT; | |
1233 | goto out; | |
1234 | } | |
1235 | ||
1236 | ret += avail; | |
1237 | head += avail; | |
58c85dc2 | 1238 | head %= ctx->nr_events; |
1da177e4 | 1239 | } |
1da177e4 | 1240 | |
58c85dc2 | 1241 | ring = kmap_atomic(ctx->ring_pages[0]); |
a31ad380 | 1242 | ring->head = head; |
91d80a84 | 1243 | kunmap_atomic(ring); |
58c85dc2 | 1244 | flush_dcache_page(ctx->ring_pages[0]); |
a31ad380 | 1245 | |
5ffac122 | 1246 | pr_debug("%li h%u t%u\n", ret, head, tail); |
a31ad380 | 1247 | out: |
58c85dc2 | 1248 | mutex_unlock(&ctx->ring_lock); |
a31ad380 | 1249 | |
1da177e4 LT |
1250 | return ret; |
1251 | } | |
1252 | ||
a31ad380 KO |
1253 | static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr, |
1254 | struct io_event __user *event, long *i) | |
1da177e4 | 1255 | { |
a31ad380 | 1256 | long ret = aio_read_events_ring(ctx, event + *i, nr - *i); |
1da177e4 | 1257 | |
a31ad380 KO |
1258 | if (ret > 0) |
1259 | *i += ret; | |
1da177e4 | 1260 | |
a31ad380 KO |
1261 | if (unlikely(atomic_read(&ctx->dead))) |
1262 | ret = -EINVAL; | |
1da177e4 | 1263 | |
a31ad380 KO |
1264 | if (!*i) |
1265 | *i = ret; | |
1da177e4 | 1266 | |
a31ad380 | 1267 | return ret < 0 || *i >= min_nr; |
1da177e4 LT |
1268 | } |
1269 | ||
a31ad380 | 1270 | static long read_events(struct kioctx *ctx, long min_nr, long nr, |
1da177e4 | 1271 | struct io_event __user *event, |
fa2e62a5 | 1272 | ktime_t until) |
1da177e4 | 1273 | { |
a31ad380 | 1274 | long ret = 0; |
1da177e4 | 1275 | |
a31ad380 KO |
1276 | /* |
1277 | * Note that aio_read_events() is being called as the conditional - i.e. | |
1278 | * we're calling it after prepare_to_wait() has set task state to | |
1279 | * TASK_INTERRUPTIBLE. | |
1280 | * | |
1281 | * But aio_read_events() can block, and if it blocks it's going to flip | |
1282 | * the task state back to TASK_RUNNING. | |
1283 | * | |
1284 | * This should be ok, provided it doesn't flip the state back to | |
1285 | * TASK_RUNNING and return 0 too much - that causes us to spin. That | |
1286 | * will only happen if the mutex_lock() call blocks, and we then find | |
1287 | * the ringbuffer empty. So in practice we should be ok, but it's | |
1288 | * something to be aware of when touching this code. | |
1289 | */ | |
2456e855 | 1290 | if (until == 0) |
5f785de5 FZ |
1291 | aio_read_events(ctx, min_nr, nr, event, &ret); |
1292 | else | |
1293 | wait_event_interruptible_hrtimeout(ctx->wait, | |
1294 | aio_read_events(ctx, min_nr, nr, event, &ret), | |
1295 | until); | |
a31ad380 | 1296 | return ret; |
1da177e4 LT |
1297 | } |
1298 | ||
1da177e4 LT |
1299 | /* sys_io_setup: |
1300 | * Create an aio_context capable of receiving at least nr_events. | |
1301 | * ctxp must not point to an aio_context that already exists, and | |
1302 | * must be initialized to 0 prior to the call. On successful | |
1303 | * creation of the aio_context, *ctxp is filled in with the resulting | |
1304 | * handle. May fail with -EINVAL if *ctxp is not initialized, | |
1305 | * if the specified nr_events exceeds internal limits. May fail | |
1306 | * with -EAGAIN if the specified nr_events exceeds the user's limit | |
1307 | * of available events. May fail with -ENOMEM if insufficient kernel | |
1308 | * resources are available. May fail with -EFAULT if an invalid | |
1309 | * pointer is passed for ctxp. Will fail with -ENOSYS if not | |
1310 | * implemented. | |
1311 | */ | |
002c8976 | 1312 | SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp) |
1da177e4 LT |
1313 | { |
1314 | struct kioctx *ioctx = NULL; | |
1315 | unsigned long ctx; | |
1316 | long ret; | |
1317 | ||
1318 | ret = get_user(ctx, ctxp); | |
1319 | if (unlikely(ret)) | |
1320 | goto out; | |
1321 | ||
1322 | ret = -EINVAL; | |
d55b5fda | 1323 | if (unlikely(ctx || nr_events == 0)) { |
acd88d4e | 1324 | pr_debug("EINVAL: ctx %lu nr_events %u\n", |
d55b5fda | 1325 | ctx, nr_events); |
1da177e4 LT |
1326 | goto out; |
1327 | } | |
1328 | ||
1329 | ioctx = ioctx_alloc(nr_events); | |
1330 | ret = PTR_ERR(ioctx); | |
1331 | if (!IS_ERR(ioctx)) { | |
1332 | ret = put_user(ioctx->user_id, ctxp); | |
a2e1859a | 1333 | if (ret) |
e02ba72a | 1334 | kill_ioctx(current->mm, ioctx, NULL); |
723be6e3 | 1335 | percpu_ref_put(&ioctx->users); |
1da177e4 LT |
1336 | } |
1337 | ||
1338 | out: | |
1339 | return ret; | |
1340 | } | |
1341 | ||
c00d2c7e AV |
1342 | #ifdef CONFIG_COMPAT |
1343 | COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p) | |
1344 | { | |
1345 | struct kioctx *ioctx = NULL; | |
1346 | unsigned long ctx; | |
1347 | long ret; | |
1348 | ||
1349 | ret = get_user(ctx, ctx32p); | |
1350 | if (unlikely(ret)) | |
1351 | goto out; | |
1352 | ||
1353 | ret = -EINVAL; | |
1354 | if (unlikely(ctx || nr_events == 0)) { | |
1355 | pr_debug("EINVAL: ctx %lu nr_events %u\n", | |
1356 | ctx, nr_events); | |
1357 | goto out; | |
1358 | } | |
1359 | ||
1360 | ioctx = ioctx_alloc(nr_events); | |
1361 | ret = PTR_ERR(ioctx); | |
1362 | if (!IS_ERR(ioctx)) { | |
1363 | /* truncating is ok because it's a user address */ | |
1364 | ret = put_user((u32)ioctx->user_id, ctx32p); | |
1365 | if (ret) | |
1366 | kill_ioctx(current->mm, ioctx, NULL); | |
1367 | percpu_ref_put(&ioctx->users); | |
1368 | } | |
1369 | ||
1370 | out: | |
1371 | return ret; | |
1372 | } | |
1373 | #endif | |
1374 | ||
1da177e4 LT |
1375 | /* sys_io_destroy: |
1376 | * Destroy the aio_context specified. May cancel any outstanding | |
1377 | * AIOs and block on completion. Will fail with -ENOSYS if not | |
642b5123 | 1378 | * implemented. May fail with -EINVAL if the context pointed to |
1da177e4 LT |
1379 | * is invalid. |
1380 | */ | |
002c8976 | 1381 | SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx) |
1da177e4 LT |
1382 | { |
1383 | struct kioctx *ioctx = lookup_ioctx(ctx); | |
1384 | if (likely(NULL != ioctx)) { | |
dc48e56d | 1385 | struct ctx_rq_wait wait; |
fb2d4483 | 1386 | int ret; |
e02ba72a | 1387 | |
dc48e56d JA |
1388 | init_completion(&wait.comp); |
1389 | atomic_set(&wait.count, 1); | |
1390 | ||
e02ba72a AP |
1391 | /* Pass requests_done to kill_ioctx() where it can be set |
1392 | * in a thread-safe way. If we try to set it here then we have | |
1393 | * a race condition if two io_destroy() called simultaneously. | |
1394 | */ | |
dc48e56d | 1395 | ret = kill_ioctx(current->mm, ioctx, &wait); |
723be6e3 | 1396 | percpu_ref_put(&ioctx->users); |
e02ba72a AP |
1397 | |
1398 | /* Wait until all IO for the context are done. Otherwise kernel | |
1399 | * keep using user-space buffers even if user thinks the context | |
1400 | * is destroyed. | |
1401 | */ | |
fb2d4483 | 1402 | if (!ret) |
dc48e56d | 1403 | wait_for_completion(&wait.comp); |
e02ba72a | 1404 | |
fb2d4483 | 1405 | return ret; |
1da177e4 | 1406 | } |
acd88d4e | 1407 | pr_debug("EINVAL: invalid context id\n"); |
1da177e4 LT |
1408 | return -EINVAL; |
1409 | } | |
1410 | ||
3c96c7f4 AV |
1411 | static void aio_remove_iocb(struct aio_kiocb *iocb) |
1412 | { | |
1413 | struct kioctx *ctx = iocb->ki_ctx; | |
1414 | unsigned long flags; | |
1415 | ||
1416 | spin_lock_irqsave(&ctx->ctx_lock, flags); | |
1417 | list_del(&iocb->ki_list); | |
1418 | spin_unlock_irqrestore(&ctx->ctx_lock, flags); | |
1419 | } | |
1420 | ||
54843f87 CH |
1421 | static void aio_complete_rw(struct kiocb *kiocb, long res, long res2) |
1422 | { | |
1423 | struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw); | |
1424 | ||
3c96c7f4 AV |
1425 | if (!list_empty_careful(&iocb->ki_list)) |
1426 | aio_remove_iocb(iocb); | |
1427 | ||
54843f87 CH |
1428 | if (kiocb->ki_flags & IOCB_WRITE) { |
1429 | struct inode *inode = file_inode(kiocb->ki_filp); | |
1430 | ||
1431 | /* | |
1432 | * Tell lockdep we inherited freeze protection from submission | |
1433 | * thread. | |
1434 | */ | |
1435 | if (S_ISREG(inode->i_mode)) | |
1436 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); | |
1437 | file_end_write(kiocb->ki_filp); | |
1438 | } | |
1439 | ||
2bb874c0 AV |
1440 | iocb->ki_res.res = res; |
1441 | iocb->ki_res.res2 = res2; | |
1442 | iocb_put(iocb); | |
54843f87 CH |
1443 | } |
1444 | ||
88a6f18b | 1445 | static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb) |
54843f87 CH |
1446 | { |
1447 | int ret; | |
1448 | ||
54843f87 | 1449 | req->ki_complete = aio_complete_rw; |
ec51f8ee | 1450 | req->private = NULL; |
54843f87 CH |
1451 | req->ki_pos = iocb->aio_offset; |
1452 | req->ki_flags = iocb_flags(req->ki_filp); | |
1453 | if (iocb->aio_flags & IOCB_FLAG_RESFD) | |
1454 | req->ki_flags |= IOCB_EVENTFD; | |
fc28724d | 1455 | req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp)); |
d9a08a9e AM |
1456 | if (iocb->aio_flags & IOCB_FLAG_IOPRIO) { |
1457 | /* | |
1458 | * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then | |
1459 | * aio_reqprio is interpreted as an I/O scheduling | |
1460 | * class and priority. | |
1461 | */ | |
1462 | ret = ioprio_check_cap(iocb->aio_reqprio); | |
1463 | if (ret) { | |
9a6d9a62 | 1464 | pr_debug("aio ioprio check cap error: %d\n", ret); |
84c4e1f8 | 1465 | return ret; |
d9a08a9e AM |
1466 | } |
1467 | ||
1468 | req->ki_ioprio = iocb->aio_reqprio; | |
1469 | } else | |
76dc8913 | 1470 | req->ki_ioprio = get_current_ioprio(); |
d9a08a9e | 1471 | |
54843f87 CH |
1472 | ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags); |
1473 | if (unlikely(ret)) | |
84c4e1f8 | 1474 | return ret; |
154989e4 CH |
1475 | |
1476 | req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */ | |
1477 | return 0; | |
54843f87 CH |
1478 | } |
1479 | ||
87e5e6da JA |
1480 | static ssize_t aio_setup_rw(int rw, const struct iocb *iocb, |
1481 | struct iovec **iovec, bool vectored, bool compat, | |
1482 | struct iov_iter *iter) | |
eed4e51f | 1483 | { |
89319d31 CH |
1484 | void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf; |
1485 | size_t len = iocb->aio_nbytes; | |
1486 | ||
1487 | if (!vectored) { | |
1488 | ssize_t ret = import_single_range(rw, buf, len, *iovec, iter); | |
1489 | *iovec = NULL; | |
1490 | return ret; | |
1491 | } | |
9d85cba7 JM |
1492 | #ifdef CONFIG_COMPAT |
1493 | if (compat) | |
89319d31 CH |
1494 | return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec, |
1495 | iter); | |
9d85cba7 | 1496 | #endif |
89319d31 | 1497 | return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter); |
eed4e51f BP |
1498 | } |
1499 | ||
9061d14a | 1500 | static inline void aio_rw_done(struct kiocb *req, ssize_t ret) |
89319d31 CH |
1501 | { |
1502 | switch (ret) { | |
1503 | case -EIOCBQUEUED: | |
9061d14a | 1504 | break; |
89319d31 CH |
1505 | case -ERESTARTSYS: |
1506 | case -ERESTARTNOINTR: | |
1507 | case -ERESTARTNOHAND: | |
1508 | case -ERESTART_RESTARTBLOCK: | |
1509 | /* | |
1510 | * There's no easy way to restart the syscall since other AIO's | |
1511 | * may be already running. Just fail this IO with EINTR. | |
1512 | */ | |
1513 | ret = -EINTR; | |
1514 | /*FALLTHRU*/ | |
1515 | default: | |
bc9bff61 | 1516 | req->ki_complete(req, ret, 0); |
89319d31 CH |
1517 | } |
1518 | } | |
1519 | ||
958c13ce | 1520 | static int aio_read(struct kiocb *req, const struct iocb *iocb, |
88a6f18b | 1521 | bool vectored, bool compat) |
1da177e4 | 1522 | { |
00fefb9c | 1523 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; |
293bc982 | 1524 | struct iov_iter iter; |
54843f87 | 1525 | struct file *file; |
958c13ce | 1526 | int ret; |
1da177e4 | 1527 | |
54843f87 CH |
1528 | ret = aio_prep_rw(req, iocb); |
1529 | if (ret) | |
1530 | return ret; | |
1531 | file = req->ki_filp; | |
89319d31 | 1532 | if (unlikely(!(file->f_mode & FMODE_READ))) |
84c4e1f8 | 1533 | return -EBADF; |
54843f87 | 1534 | ret = -EINVAL; |
89319d31 | 1535 | if (unlikely(!file->f_op->read_iter)) |
84c4e1f8 | 1536 | return -EINVAL; |
73a7075e | 1537 | |
89319d31 | 1538 | ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter); |
87e5e6da | 1539 | if (ret < 0) |
84c4e1f8 | 1540 | return ret; |
89319d31 CH |
1541 | ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter)); |
1542 | if (!ret) | |
9061d14a | 1543 | aio_rw_done(req, call_read_iter(file, req, &iter)); |
89319d31 CH |
1544 | kfree(iovec); |
1545 | return ret; | |
1546 | } | |
73a7075e | 1547 | |
958c13ce | 1548 | static int aio_write(struct kiocb *req, const struct iocb *iocb, |
88a6f18b | 1549 | bool vectored, bool compat) |
89319d31 | 1550 | { |
89319d31 CH |
1551 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; |
1552 | struct iov_iter iter; | |
54843f87 | 1553 | struct file *file; |
958c13ce | 1554 | int ret; |
41ef4eb8 | 1555 | |
54843f87 CH |
1556 | ret = aio_prep_rw(req, iocb); |
1557 | if (ret) | |
1558 | return ret; | |
1559 | file = req->ki_filp; | |
1560 | ||
89319d31 | 1561 | if (unlikely(!(file->f_mode & FMODE_WRITE))) |
84c4e1f8 | 1562 | return -EBADF; |
89319d31 | 1563 | if (unlikely(!file->f_op->write_iter)) |
84c4e1f8 | 1564 | return -EINVAL; |
1da177e4 | 1565 | |
89319d31 | 1566 | ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter); |
87e5e6da | 1567 | if (ret < 0) |
84c4e1f8 | 1568 | return ret; |
89319d31 CH |
1569 | ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter)); |
1570 | if (!ret) { | |
70fe2f48 | 1571 | /* |
92ce4728 | 1572 | * Open-code file_start_write here to grab freeze protection, |
54843f87 CH |
1573 | * which will be released by another thread in |
1574 | * aio_complete_rw(). Fool lockdep by telling it the lock got | |
1575 | * released so that it doesn't complain about the held lock when | |
1576 | * we return to userspace. | |
70fe2f48 | 1577 | */ |
92ce4728 CH |
1578 | if (S_ISREG(file_inode(file)->i_mode)) { |
1579 | __sb_start_write(file_inode(file)->i_sb, SB_FREEZE_WRITE, true); | |
a12f1ae6 | 1580 | __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE); |
92ce4728 CH |
1581 | } |
1582 | req->ki_flags |= IOCB_WRITE; | |
9061d14a | 1583 | aio_rw_done(req, call_write_iter(file, req, &iter)); |
41ef4eb8 | 1584 | } |
89319d31 CH |
1585 | kfree(iovec); |
1586 | return ret; | |
1da177e4 LT |
1587 | } |
1588 | ||
a3c0d439 CH |
1589 | static void aio_fsync_work(struct work_struct *work) |
1590 | { | |
2bb874c0 | 1591 | struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work); |
a3c0d439 | 1592 | |
2bb874c0 AV |
1593 | iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync); |
1594 | iocb_put(iocb); | |
a3c0d439 CH |
1595 | } |
1596 | ||
88a6f18b JA |
1597 | static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb, |
1598 | bool datasync) | |
a3c0d439 CH |
1599 | { |
1600 | if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes || | |
1601 | iocb->aio_rw_flags)) | |
1602 | return -EINVAL; | |
a11e1d43 | 1603 | |
84c4e1f8 | 1604 | if (unlikely(!req->file->f_op->fsync)) |
a3c0d439 | 1605 | return -EINVAL; |
a3c0d439 CH |
1606 | |
1607 | req->datasync = datasync; | |
1608 | INIT_WORK(&req->work, aio_fsync_work); | |
1609 | schedule_work(&req->work); | |
9061d14a | 1610 | return 0; |
a3c0d439 CH |
1611 | } |
1612 | ||
bfe4037e CH |
1613 | static void aio_poll_complete_work(struct work_struct *work) |
1614 | { | |
1615 | struct poll_iocb *req = container_of(work, struct poll_iocb, work); | |
1616 | struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll); | |
1617 | struct poll_table_struct pt = { ._key = req->events }; | |
1618 | struct kioctx *ctx = iocb->ki_ctx; | |
1619 | __poll_t mask = 0; | |
1620 | ||
1621 | if (!READ_ONCE(req->cancelled)) | |
1622 | mask = vfs_poll(req->file, &pt) & req->events; | |
1623 | ||
1624 | /* | |
1625 | * Note that ->ki_cancel callers also delete iocb from active_reqs after | |
1626 | * calling ->ki_cancel. We need the ctx_lock roundtrip here to | |
1627 | * synchronize with them. In the cancellation case the list_del_init | |
1628 | * itself is not actually needed, but harmless so we keep it in to | |
1629 | * avoid further branches in the fast path. | |
1630 | */ | |
1631 | spin_lock_irq(&ctx->ctx_lock); | |
1632 | if (!mask && !READ_ONCE(req->cancelled)) { | |
1633 | add_wait_queue(req->head, &req->wait); | |
1634 | spin_unlock_irq(&ctx->ctx_lock); | |
1635 | return; | |
1636 | } | |
1637 | list_del_init(&iocb->ki_list); | |
af5c72b1 AV |
1638 | iocb->ki_res.res = mangle_poll(mask); |
1639 | req->done = true; | |
bfe4037e CH |
1640 | spin_unlock_irq(&ctx->ctx_lock); |
1641 | ||
af5c72b1 | 1642 | iocb_put(iocb); |
bfe4037e CH |
1643 | } |
1644 | ||
1645 | /* assumes we are called with irqs disabled */ | |
1646 | static int aio_poll_cancel(struct kiocb *iocb) | |
1647 | { | |
1648 | struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw); | |
1649 | struct poll_iocb *req = &aiocb->poll; | |
1650 | ||
1651 | spin_lock(&req->head->lock); | |
1652 | WRITE_ONCE(req->cancelled, true); | |
1653 | if (!list_empty(&req->wait.entry)) { | |
1654 | list_del_init(&req->wait.entry); | |
1655 | schedule_work(&aiocb->poll.work); | |
1656 | } | |
1657 | spin_unlock(&req->head->lock); | |
1658 | ||
1659 | return 0; | |
1660 | } | |
1661 | ||
1662 | static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
1663 | void *key) | |
1664 | { | |
1665 | struct poll_iocb *req = container_of(wait, struct poll_iocb, wait); | |
e8693bcf | 1666 | struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll); |
bfe4037e | 1667 | __poll_t mask = key_to_poll(key); |
d3d6a18d | 1668 | unsigned long flags; |
bfe4037e | 1669 | |
bfe4037e | 1670 | /* for instances that support it check for an event match first: */ |
af5c72b1 AV |
1671 | if (mask && !(mask & req->events)) |
1672 | return 0; | |
e8693bcf | 1673 | |
af5c72b1 AV |
1674 | list_del_init(&req->wait.entry); |
1675 | ||
1676 | if (mask && spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) { | |
d3d6a18d BVA |
1677 | /* |
1678 | * Try to complete the iocb inline if we can. Use | |
1679 | * irqsave/irqrestore because not all filesystems (e.g. fuse) | |
1680 | * call this function with IRQs disabled and because IRQs | |
1681 | * have to be disabled before ctx_lock is obtained. | |
1682 | */ | |
af5c72b1 AV |
1683 | list_del(&iocb->ki_list); |
1684 | iocb->ki_res.res = mangle_poll(mask); | |
1685 | req->done = true; | |
1686 | spin_unlock_irqrestore(&iocb->ki_ctx->ctx_lock, flags); | |
1687 | iocb_put(iocb); | |
1688 | } else { | |
1689 | schedule_work(&req->work); | |
e8693bcf | 1690 | } |
bfe4037e CH |
1691 | return 1; |
1692 | } | |
1693 | ||
1694 | struct aio_poll_table { | |
1695 | struct poll_table_struct pt; | |
1696 | struct aio_kiocb *iocb; | |
1697 | int error; | |
1698 | }; | |
1699 | ||
1700 | static void | |
1701 | aio_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
1702 | struct poll_table_struct *p) | |
1703 | { | |
1704 | struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt); | |
1705 | ||
1706 | /* multiple wait queues per file are not supported */ | |
1707 | if (unlikely(pt->iocb->poll.head)) { | |
1708 | pt->error = -EINVAL; | |
1709 | return; | |
1710 | } | |
1711 | ||
1712 | pt->error = 0; | |
1713 | pt->iocb->poll.head = head; | |
1714 | add_wait_queue(head, &pt->iocb->poll.wait); | |
1715 | } | |
1716 | ||
958c13ce | 1717 | static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb) |
bfe4037e CH |
1718 | { |
1719 | struct kioctx *ctx = aiocb->ki_ctx; | |
1720 | struct poll_iocb *req = &aiocb->poll; | |
1721 | struct aio_poll_table apt; | |
af5c72b1 | 1722 | bool cancel = false; |
bfe4037e CH |
1723 | __poll_t mask; |
1724 | ||
1725 | /* reject any unknown events outside the normal event mask. */ | |
1726 | if ((u16)iocb->aio_buf != iocb->aio_buf) | |
1727 | return -EINVAL; | |
1728 | /* reject fields that are not defined for poll */ | |
1729 | if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags) | |
1730 | return -EINVAL; | |
1731 | ||
1732 | INIT_WORK(&req->work, aio_poll_complete_work); | |
1733 | req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP; | |
bfe4037e | 1734 | |
2bc4ca9b | 1735 | req->head = NULL; |
af5c72b1 | 1736 | req->done = false; |
2bc4ca9b JA |
1737 | req->cancelled = false; |
1738 | ||
bfe4037e CH |
1739 | apt.pt._qproc = aio_poll_queue_proc; |
1740 | apt.pt._key = req->events; | |
1741 | apt.iocb = aiocb; | |
1742 | apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */ | |
1743 | ||
1744 | /* initialized the list so that we can do list_empty checks */ | |
1745 | INIT_LIST_HEAD(&req->wait.entry); | |
1746 | init_waitqueue_func_entry(&req->wait, aio_poll_wake); | |
1747 | ||
bfe4037e | 1748 | mask = vfs_poll(req->file, &apt.pt) & req->events; |
bfe4037e | 1749 | spin_lock_irq(&ctx->ctx_lock); |
af5c72b1 AV |
1750 | if (likely(req->head)) { |
1751 | spin_lock(&req->head->lock); | |
1752 | if (unlikely(list_empty(&req->wait.entry))) { | |
1753 | if (apt.error) | |
1754 | cancel = true; | |
1755 | apt.error = 0; | |
1756 | mask = 0; | |
1757 | } | |
1758 | if (mask || apt.error) { | |
1759 | list_del_init(&req->wait.entry); | |
1760 | } else if (cancel) { | |
1761 | WRITE_ONCE(req->cancelled, true); | |
1762 | } else if (!req->done) { /* actually waiting for an event */ | |
1763 | list_add_tail(&aiocb->ki_list, &ctx->active_reqs); | |
1764 | aiocb->ki_cancel = aio_poll_cancel; | |
1765 | } | |
1766 | spin_unlock(&req->head->lock); | |
1767 | } | |
1768 | if (mask) { /* no async, we'd stolen it */ | |
1769 | aiocb->ki_res.res = mangle_poll(mask); | |
bfe4037e | 1770 | apt.error = 0; |
bfe4037e | 1771 | } |
bfe4037e | 1772 | spin_unlock_irq(&ctx->ctx_lock); |
bfe4037e | 1773 | if (mask) |
af5c72b1 AV |
1774 | iocb_put(aiocb); |
1775 | return apt.error; | |
bfe4037e CH |
1776 | } |
1777 | ||
88a6f18b | 1778 | static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb, |
7316b49c AV |
1779 | struct iocb __user *user_iocb, struct aio_kiocb *req, |
1780 | bool compat) | |
1da177e4 | 1781 | { |
84c4e1f8 | 1782 | req->ki_filp = fget(iocb->aio_fildes); |
84c4e1f8 | 1783 | if (unlikely(!req->ki_filp)) |
7316b49c | 1784 | return -EBADF; |
84c4e1f8 | 1785 | |
88a6f18b | 1786 | if (iocb->aio_flags & IOCB_FLAG_RESFD) { |
74259703 | 1787 | struct eventfd_ctx *eventfd; |
9c3060be DL |
1788 | /* |
1789 | * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an | |
1790 | * instance of the file* now. The file descriptor must be | |
1791 | * an eventfd() fd, and will be signaled for each completed | |
1792 | * event using the eventfd_signal() function. | |
1793 | */ | |
74259703 | 1794 | eventfd = eventfd_ctx_fdget(iocb->aio_resfd); |
7316b49c | 1795 | if (IS_ERR(eventfd)) |
18bfb9c6 | 1796 | return PTR_ERR(eventfd); |
7316b49c | 1797 | |
74259703 | 1798 | req->ki_eventfd = eventfd; |
9830f4be GR |
1799 | } |
1800 | ||
7316b49c | 1801 | if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) { |
caf4167a | 1802 | pr_debug("EFAULT: aio_key\n"); |
7316b49c | 1803 | return -EFAULT; |
1da177e4 LT |
1804 | } |
1805 | ||
a9339b78 AV |
1806 | req->ki_res.obj = (u64)(unsigned long)user_iocb; |
1807 | req->ki_res.data = iocb->aio_data; | |
1808 | req->ki_res.res = 0; | |
1809 | req->ki_res.res2 = 0; | |
1da177e4 | 1810 | |
88a6f18b | 1811 | switch (iocb->aio_lio_opcode) { |
89319d31 | 1812 | case IOCB_CMD_PREAD: |
7316b49c | 1813 | return aio_read(&req->rw, iocb, false, compat); |
89319d31 | 1814 | case IOCB_CMD_PWRITE: |
7316b49c | 1815 | return aio_write(&req->rw, iocb, false, compat); |
89319d31 | 1816 | case IOCB_CMD_PREADV: |
7316b49c | 1817 | return aio_read(&req->rw, iocb, true, compat); |
89319d31 | 1818 | case IOCB_CMD_PWRITEV: |
7316b49c | 1819 | return aio_write(&req->rw, iocb, true, compat); |
a3c0d439 | 1820 | case IOCB_CMD_FSYNC: |
7316b49c | 1821 | return aio_fsync(&req->fsync, iocb, false); |
a3c0d439 | 1822 | case IOCB_CMD_FDSYNC: |
7316b49c | 1823 | return aio_fsync(&req->fsync, iocb, true); |
bfe4037e | 1824 | case IOCB_CMD_POLL: |
7316b49c | 1825 | return aio_poll(req, iocb); |
89319d31 | 1826 | default: |
88a6f18b | 1827 | pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode); |
7316b49c | 1828 | return -EINVAL; |
89319d31 | 1829 | } |
1da177e4 LT |
1830 | } |
1831 | ||
88a6f18b JA |
1832 | static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb, |
1833 | bool compat) | |
1834 | { | |
7316b49c | 1835 | struct aio_kiocb *req; |
88a6f18b | 1836 | struct iocb iocb; |
7316b49c | 1837 | int err; |
88a6f18b JA |
1838 | |
1839 | if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb)))) | |
1840 | return -EFAULT; | |
1841 | ||
7316b49c AV |
1842 | /* enforce forwards compatibility on users */ |
1843 | if (unlikely(iocb.aio_reserved2)) { | |
1844 | pr_debug("EINVAL: reserve field set\n"); | |
1845 | return -EINVAL; | |
1846 | } | |
1847 | ||
1848 | /* prevent overflows */ | |
1849 | if (unlikely( | |
1850 | (iocb.aio_buf != (unsigned long)iocb.aio_buf) || | |
1851 | (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) || | |
1852 | ((ssize_t)iocb.aio_nbytes < 0) | |
1853 | )) { | |
1854 | pr_debug("EINVAL: overflow check\n"); | |
1855 | return -EINVAL; | |
1856 | } | |
1857 | ||
1858 | req = aio_get_req(ctx); | |
1859 | if (unlikely(!req)) | |
1860 | return -EAGAIN; | |
1861 | ||
1862 | err = __io_submit_one(ctx, &iocb, user_iocb, req, compat); | |
1863 | ||
1864 | /* Done with the synchronous reference */ | |
1865 | iocb_put(req); | |
1866 | ||
1867 | /* | |
1868 | * If err is 0, we'd either done aio_complete() ourselves or have | |
1869 | * arranged for that to be done asynchronously. Anything non-zero | |
1870 | * means that we need to destroy req ourselves. | |
1871 | */ | |
1872 | if (unlikely(err)) { | |
1873 | iocb_destroy(req); | |
1874 | put_reqs_available(ctx, 1); | |
1875 | } | |
1876 | return err; | |
88a6f18b JA |
1877 | } |
1878 | ||
67ba049f AV |
1879 | /* sys_io_submit: |
1880 | * Queue the nr iocbs pointed to by iocbpp for processing. Returns | |
1881 | * the number of iocbs queued. May return -EINVAL if the aio_context | |
1882 | * specified by ctx_id is invalid, if nr is < 0, if the iocb at | |
1883 | * *iocbpp[0] is not properly initialized, if the operation specified | |
1884 | * is invalid for the file descriptor in the iocb. May fail with | |
1885 | * -EFAULT if any of the data structures point to invalid data. May | |
1886 | * fail with -EBADF if the file descriptor specified in the first | |
1887 | * iocb is invalid. May fail with -EAGAIN if insufficient resources | |
1888 | * are available to queue any iocbs. Will return 0 if nr is 0. Will | |
1889 | * fail with -ENOSYS if not implemented. | |
1890 | */ | |
1891 | SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr, | |
1892 | struct iocb __user * __user *, iocbpp) | |
1da177e4 LT |
1893 | { |
1894 | struct kioctx *ctx; | |
1895 | long ret = 0; | |
080d676d | 1896 | int i = 0; |
9f5b9425 | 1897 | struct blk_plug plug; |
1da177e4 LT |
1898 | |
1899 | if (unlikely(nr < 0)) | |
1900 | return -EINVAL; | |
1901 | ||
1da177e4 LT |
1902 | ctx = lookup_ioctx(ctx_id); |
1903 | if (unlikely(!ctx)) { | |
caf4167a | 1904 | pr_debug("EINVAL: invalid context id\n"); |
1da177e4 LT |
1905 | return -EINVAL; |
1906 | } | |
1907 | ||
1da92779 AV |
1908 | if (nr > ctx->nr_events) |
1909 | nr = ctx->nr_events; | |
1910 | ||
a79d40e9 JA |
1911 | if (nr > AIO_PLUG_THRESHOLD) |
1912 | blk_start_plug(&plug); | |
67ba049f | 1913 | for (i = 0; i < nr; i++) { |
1da177e4 | 1914 | struct iocb __user *user_iocb; |
1da177e4 | 1915 | |
67ba049f | 1916 | if (unlikely(get_user(user_iocb, iocbpp + i))) { |
1da177e4 LT |
1917 | ret = -EFAULT; |
1918 | break; | |
1919 | } | |
1920 | ||
67ba049f | 1921 | ret = io_submit_one(ctx, user_iocb, false); |
1da177e4 LT |
1922 | if (ret) |
1923 | break; | |
1924 | } | |
a79d40e9 JA |
1925 | if (nr > AIO_PLUG_THRESHOLD) |
1926 | blk_finish_plug(&plug); | |
1da177e4 | 1927 | |
723be6e3 | 1928 | percpu_ref_put(&ctx->users); |
1da177e4 LT |
1929 | return i ? i : ret; |
1930 | } | |
1931 | ||
c00d2c7e | 1932 | #ifdef CONFIG_COMPAT |
c00d2c7e | 1933 | COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id, |
67ba049f | 1934 | int, nr, compat_uptr_t __user *, iocbpp) |
c00d2c7e | 1935 | { |
67ba049f AV |
1936 | struct kioctx *ctx; |
1937 | long ret = 0; | |
1938 | int i = 0; | |
1939 | struct blk_plug plug; | |
c00d2c7e AV |
1940 | |
1941 | if (unlikely(nr < 0)) | |
1942 | return -EINVAL; | |
1943 | ||
67ba049f AV |
1944 | ctx = lookup_ioctx(ctx_id); |
1945 | if (unlikely(!ctx)) { | |
1946 | pr_debug("EINVAL: invalid context id\n"); | |
1947 | return -EINVAL; | |
1948 | } | |
1949 | ||
1da92779 AV |
1950 | if (nr > ctx->nr_events) |
1951 | nr = ctx->nr_events; | |
1952 | ||
a79d40e9 JA |
1953 | if (nr > AIO_PLUG_THRESHOLD) |
1954 | blk_start_plug(&plug); | |
67ba049f AV |
1955 | for (i = 0; i < nr; i++) { |
1956 | compat_uptr_t user_iocb; | |
1957 | ||
1958 | if (unlikely(get_user(user_iocb, iocbpp + i))) { | |
1959 | ret = -EFAULT; | |
1960 | break; | |
1961 | } | |
1962 | ||
1963 | ret = io_submit_one(ctx, compat_ptr(user_iocb), true); | |
1964 | if (ret) | |
1965 | break; | |
1966 | } | |
a79d40e9 JA |
1967 | if (nr > AIO_PLUG_THRESHOLD) |
1968 | blk_finish_plug(&plug); | |
67ba049f AV |
1969 | |
1970 | percpu_ref_put(&ctx->users); | |
1971 | return i ? i : ret; | |
c00d2c7e AV |
1972 | } |
1973 | #endif | |
1974 | ||
1da177e4 LT |
1975 | /* sys_io_cancel: |
1976 | * Attempts to cancel an iocb previously passed to io_submit. If | |
1977 | * the operation is successfully cancelled, the resulting event is | |
1978 | * copied into the memory pointed to by result without being placed | |
1979 | * into the completion queue and 0 is returned. May fail with | |
1980 | * -EFAULT if any of the data structures pointed to are invalid. | |
1981 | * May fail with -EINVAL if aio_context specified by ctx_id is | |
1982 | * invalid. May fail with -EAGAIN if the iocb specified was not | |
1983 | * cancelled. Will fail with -ENOSYS if not implemented. | |
1984 | */ | |
002c8976 HC |
1985 | SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb, |
1986 | struct io_event __user *, result) | |
1da177e4 | 1987 | { |
1da177e4 | 1988 | struct kioctx *ctx; |
04b2fa9f | 1989 | struct aio_kiocb *kiocb; |
888933f8 | 1990 | int ret = -EINVAL; |
1da177e4 | 1991 | u32 key; |
a9339b78 | 1992 | u64 obj = (u64)(unsigned long)iocb; |
1da177e4 | 1993 | |
f3a2752a | 1994 | if (unlikely(get_user(key, &iocb->aio_key))) |
1da177e4 | 1995 | return -EFAULT; |
f3a2752a CH |
1996 | if (unlikely(key != KIOCB_KEY)) |
1997 | return -EINVAL; | |
1da177e4 LT |
1998 | |
1999 | ctx = lookup_ioctx(ctx_id); | |
2000 | if (unlikely(!ctx)) | |
2001 | return -EINVAL; | |
2002 | ||
2003 | spin_lock_irq(&ctx->ctx_lock); | |
833f4154 AV |
2004 | /* TODO: use a hash or array, this sucks. */ |
2005 | list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) { | |
a9339b78 | 2006 | if (kiocb->ki_res.obj == obj) { |
833f4154 AV |
2007 | ret = kiocb->ki_cancel(&kiocb->rw); |
2008 | list_del_init(&kiocb->ki_list); | |
2009 | break; | |
2010 | } | |
888933f8 | 2011 | } |
1da177e4 LT |
2012 | spin_unlock_irq(&ctx->ctx_lock); |
2013 | ||
906b973c | 2014 | if (!ret) { |
bec68faa KO |
2015 | /* |
2016 | * The result argument is no longer used - the io_event is | |
2017 | * always delivered via the ring buffer. -EINPROGRESS indicates | |
2018 | * cancellation is progress: | |
906b973c | 2019 | */ |
bec68faa | 2020 | ret = -EINPROGRESS; |
906b973c | 2021 | } |
1da177e4 | 2022 | |
723be6e3 | 2023 | percpu_ref_put(&ctx->users); |
1da177e4 LT |
2024 | |
2025 | return ret; | |
2026 | } | |
2027 | ||
fa2e62a5 DD |
2028 | static long do_io_getevents(aio_context_t ctx_id, |
2029 | long min_nr, | |
2030 | long nr, | |
2031 | struct io_event __user *events, | |
2032 | struct timespec64 *ts) | |
2033 | { | |
2034 | ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX; | |
2035 | struct kioctx *ioctx = lookup_ioctx(ctx_id); | |
2036 | long ret = -EINVAL; | |
2037 | ||
2038 | if (likely(ioctx)) { | |
2039 | if (likely(min_nr <= nr && min_nr >= 0)) | |
2040 | ret = read_events(ioctx, min_nr, nr, events, until); | |
2041 | percpu_ref_put(&ioctx->users); | |
2042 | } | |
2043 | ||
2044 | return ret; | |
2045 | } | |
2046 | ||
1da177e4 LT |
2047 | /* io_getevents: |
2048 | * Attempts to read at least min_nr events and up to nr events from | |
642b5123 ST |
2049 | * the completion queue for the aio_context specified by ctx_id. If |
2050 | * it succeeds, the number of read events is returned. May fail with | |
2051 | * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is | |
2052 | * out of range, if timeout is out of range. May fail with -EFAULT | |
2053 | * if any of the memory specified is invalid. May return 0 or | |
2054 | * < min_nr if the timeout specified by timeout has elapsed | |
2055 | * before sufficient events are available, where timeout == NULL | |
2056 | * specifies an infinite timeout. Note that the timeout pointed to by | |
6900807c | 2057 | * timeout is relative. Will fail with -ENOSYS if not implemented. |
1da177e4 | 2058 | */ |
7a35397f DD |
2059 | #if !defined(CONFIG_64BIT_TIME) || defined(CONFIG_64BIT) |
2060 | ||
002c8976 HC |
2061 | SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id, |
2062 | long, min_nr, | |
2063 | long, nr, | |
2064 | struct io_event __user *, events, | |
7a35397f | 2065 | struct __kernel_timespec __user *, timeout) |
1da177e4 | 2066 | { |
fa2e62a5 | 2067 | struct timespec64 ts; |
7a074e96 CH |
2068 | int ret; |
2069 | ||
2070 | if (timeout && unlikely(get_timespec64(&ts, timeout))) | |
2071 | return -EFAULT; | |
2072 | ||
2073 | ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL); | |
2074 | if (!ret && signal_pending(current)) | |
2075 | ret = -EINTR; | |
2076 | return ret; | |
2077 | } | |
1da177e4 | 2078 | |
7a35397f DD |
2079 | #endif |
2080 | ||
9ba546c0 CH |
2081 | struct __aio_sigset { |
2082 | const sigset_t __user *sigmask; | |
2083 | size_t sigsetsize; | |
2084 | }; | |
2085 | ||
7a074e96 CH |
2086 | SYSCALL_DEFINE6(io_pgetevents, |
2087 | aio_context_t, ctx_id, | |
2088 | long, min_nr, | |
2089 | long, nr, | |
2090 | struct io_event __user *, events, | |
7a35397f | 2091 | struct __kernel_timespec __user *, timeout, |
7a074e96 CH |
2092 | const struct __aio_sigset __user *, usig) |
2093 | { | |
2094 | struct __aio_sigset ksig = { NULL, }; | |
7a074e96 | 2095 | struct timespec64 ts; |
97abc889 | 2096 | bool interrupted; |
7a074e96 CH |
2097 | int ret; |
2098 | ||
2099 | if (timeout && unlikely(get_timespec64(&ts, timeout))) | |
2100 | return -EFAULT; | |
2101 | ||
2102 | if (usig && copy_from_user(&ksig, usig, sizeof(ksig))) | |
2103 | return -EFAULT; | |
2104 | ||
b772434b | 2105 | ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize); |
7a35397f DD |
2106 | if (ret) |
2107 | return ret; | |
7a074e96 CH |
2108 | |
2109 | ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL); | |
97abc889 ON |
2110 | |
2111 | interrupted = signal_pending(current); | |
b772434b | 2112 | restore_saved_sigmask_unless(interrupted); |
97abc889 | 2113 | if (interrupted && !ret) |
7a35397f | 2114 | ret = -ERESTARTNOHAND; |
7a074e96 | 2115 | |
7a35397f DD |
2116 | return ret; |
2117 | } | |
2118 | ||
2119 | #if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT) | |
2120 | ||
2121 | SYSCALL_DEFINE6(io_pgetevents_time32, | |
2122 | aio_context_t, ctx_id, | |
2123 | long, min_nr, | |
2124 | long, nr, | |
2125 | struct io_event __user *, events, | |
2126 | struct old_timespec32 __user *, timeout, | |
2127 | const struct __aio_sigset __user *, usig) | |
2128 | { | |
2129 | struct __aio_sigset ksig = { NULL, }; | |
7a35397f | 2130 | struct timespec64 ts; |
97abc889 | 2131 | bool interrupted; |
7a35397f DD |
2132 | int ret; |
2133 | ||
2134 | if (timeout && unlikely(get_old_timespec32(&ts, timeout))) | |
2135 | return -EFAULT; | |
2136 | ||
2137 | if (usig && copy_from_user(&ksig, usig, sizeof(ksig))) | |
2138 | return -EFAULT; | |
2139 | ||
ded653cc | 2140 | |
b772434b | 2141 | ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize); |
ded653cc DD |
2142 | if (ret) |
2143 | return ret; | |
7a074e96 CH |
2144 | |
2145 | ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL); | |
97abc889 ON |
2146 | |
2147 | interrupted = signal_pending(current); | |
b772434b | 2148 | restore_saved_sigmask_unless(interrupted); |
97abc889 | 2149 | if (interrupted && !ret) |
854a6ed5 | 2150 | ret = -ERESTARTNOHAND; |
fa2e62a5 | 2151 | |
7a074e96 | 2152 | return ret; |
1da177e4 | 2153 | } |
c00d2c7e | 2154 | |
7a35397f DD |
2155 | #endif |
2156 | ||
2157 | #if defined(CONFIG_COMPAT_32BIT_TIME) | |
2158 | ||
8dabe724 AB |
2159 | SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id, |
2160 | __s32, min_nr, | |
2161 | __s32, nr, | |
2162 | struct io_event __user *, events, | |
2163 | struct old_timespec32 __user *, timeout) | |
c00d2c7e | 2164 | { |
fa2e62a5 | 2165 | struct timespec64 t; |
7a074e96 CH |
2166 | int ret; |
2167 | ||
9afc5eee | 2168 | if (timeout && get_old_timespec32(&t, timeout)) |
7a074e96 CH |
2169 | return -EFAULT; |
2170 | ||
2171 | ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL); | |
2172 | if (!ret && signal_pending(current)) | |
2173 | ret = -EINTR; | |
2174 | return ret; | |
2175 | } | |
2176 | ||
7a35397f DD |
2177 | #endif |
2178 | ||
2179 | #ifdef CONFIG_COMPAT | |
c00d2c7e | 2180 | |
7a074e96 | 2181 | struct __compat_aio_sigset { |
97eba80f | 2182 | compat_uptr_t sigmask; |
7a074e96 CH |
2183 | compat_size_t sigsetsize; |
2184 | }; | |
2185 | ||
7a35397f DD |
2186 | #if defined(CONFIG_COMPAT_32BIT_TIME) |
2187 | ||
7a074e96 CH |
2188 | COMPAT_SYSCALL_DEFINE6(io_pgetevents, |
2189 | compat_aio_context_t, ctx_id, | |
2190 | compat_long_t, min_nr, | |
2191 | compat_long_t, nr, | |
2192 | struct io_event __user *, events, | |
9afc5eee | 2193 | struct old_timespec32 __user *, timeout, |
7a074e96 CH |
2194 | const struct __compat_aio_sigset __user *, usig) |
2195 | { | |
97eba80f | 2196 | struct __compat_aio_sigset ksig = { 0, }; |
7a074e96 | 2197 | struct timespec64 t; |
97abc889 | 2198 | bool interrupted; |
7a074e96 CH |
2199 | int ret; |
2200 | ||
9afc5eee | 2201 | if (timeout && get_old_timespec32(&t, timeout)) |
7a074e96 CH |
2202 | return -EFAULT; |
2203 | ||
2204 | if (usig && copy_from_user(&ksig, usig, sizeof(ksig))) | |
2205 | return -EFAULT; | |
2206 | ||
97eba80f | 2207 | ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize); |
ded653cc DD |
2208 | if (ret) |
2209 | return ret; | |
c00d2c7e | 2210 | |
7a074e96 | 2211 | ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL); |
97abc889 ON |
2212 | |
2213 | interrupted = signal_pending(current); | |
b772434b | 2214 | restore_saved_sigmask_unless(interrupted); |
97abc889 | 2215 | if (interrupted && !ret) |
854a6ed5 | 2216 | ret = -ERESTARTNOHAND; |
fa2e62a5 | 2217 | |
7a074e96 | 2218 | return ret; |
c00d2c7e | 2219 | } |
7a35397f DD |
2220 | |
2221 | #endif | |
2222 | ||
2223 | COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64, | |
2224 | compat_aio_context_t, ctx_id, | |
2225 | compat_long_t, min_nr, | |
2226 | compat_long_t, nr, | |
2227 | struct io_event __user *, events, | |
2228 | struct __kernel_timespec __user *, timeout, | |
2229 | const struct __compat_aio_sigset __user *, usig) | |
2230 | { | |
97eba80f | 2231 | struct __compat_aio_sigset ksig = { 0, }; |
7a35397f | 2232 | struct timespec64 t; |
97abc889 | 2233 | bool interrupted; |
7a35397f DD |
2234 | int ret; |
2235 | ||
2236 | if (timeout && get_timespec64(&t, timeout)) | |
2237 | return -EFAULT; | |
2238 | ||
2239 | if (usig && copy_from_user(&ksig, usig, sizeof(ksig))) | |
2240 | return -EFAULT; | |
2241 | ||
97eba80f | 2242 | ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize); |
7a35397f DD |
2243 | if (ret) |
2244 | return ret; | |
2245 | ||
2246 | ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL); | |
97abc889 ON |
2247 | |
2248 | interrupted = signal_pending(current); | |
b772434b | 2249 | restore_saved_sigmask_unless(interrupted); |
97abc889 | 2250 | if (interrupted && !ret) |
7a35397f | 2251 | ret = -ERESTARTNOHAND; |
fa2e62a5 | 2252 | |
7a074e96 | 2253 | return ret; |
c00d2c7e AV |
2254 | } |
2255 | #endif |