]>
Commit | Line | Data |
---|---|---|
2b188cc1 JA |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* | |
3 | * Shared application/kernel submission and completion ring pairs, for | |
4 | * supporting fast/efficient IO. | |
5 | * | |
6 | * A note on the read/write ordering memory barriers that are matched between | |
1e84b97b SB |
7 | * the application and kernel side. |
8 | * | |
9 | * After the application reads the CQ ring tail, it must use an | |
10 | * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses | |
11 | * before writing the tail (using smp_load_acquire to read the tail will | |
12 | * do). It also needs a smp_mb() before updating CQ head (ordering the | |
13 | * entry load(s) with the head store), pairing with an implicit barrier | |
14 | * through a control-dependency in io_get_cqring (smp_store_release to | |
15 | * store head will do). Failure to do so could lead to reading invalid | |
16 | * CQ entries. | |
17 | * | |
18 | * Likewise, the application must use an appropriate smp_wmb() before | |
19 | * writing the SQ tail (ordering SQ entry stores with the tail store), | |
20 | * which pairs with smp_load_acquire in io_get_sqring (smp_store_release | |
21 | * to store the tail will do). And it needs a barrier ordering the SQ | |
22 | * head load before writing new SQ entries (smp_load_acquire to read | |
23 | * head will do). | |
24 | * | |
25 | * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application | |
26 | * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after* | |
27 | * updating the SQ tail; a full memory barrier smp_mb() is needed | |
28 | * between. | |
2b188cc1 JA |
29 | * |
30 | * Also see the examples in the liburing library: | |
31 | * | |
32 | * git://git.kernel.dk/liburing | |
33 | * | |
34 | * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens | |
35 | * from data shared between the kernel and application. This is done both | |
36 | * for ordering purposes, but also to ensure that once a value is loaded from | |
37 | * data that the application could potentially modify, it remains stable. | |
38 | * | |
39 | * Copyright (C) 2018-2019 Jens Axboe | |
c992fe29 | 40 | * Copyright (c) 2018-2019 Christoph Hellwig |
2b188cc1 JA |
41 | */ |
42 | #include <linux/kernel.h> | |
43 | #include <linux/init.h> | |
44 | #include <linux/errno.h> | |
45 | #include <linux/syscalls.h> | |
46 | #include <linux/compat.h> | |
52de1fe1 | 47 | #include <net/compat.h> |
2b188cc1 JA |
48 | #include <linux/refcount.h> |
49 | #include <linux/uio.h> | |
6b47ee6e | 50 | #include <linux/bits.h> |
2b188cc1 JA |
51 | |
52 | #include <linux/sched/signal.h> | |
53 | #include <linux/fs.h> | |
54 | #include <linux/file.h> | |
55 | #include <linux/fdtable.h> | |
56 | #include <linux/mm.h> | |
57 | #include <linux/mman.h> | |
2b188cc1 JA |
58 | #include <linux/percpu.h> |
59 | #include <linux/slab.h> | |
6c271ce2 | 60 | #include <linux/kthread.h> |
2b188cc1 | 61 | #include <linux/blkdev.h> |
edafccee | 62 | #include <linux/bvec.h> |
2b188cc1 JA |
63 | #include <linux/net.h> |
64 | #include <net/sock.h> | |
65 | #include <net/af_unix.h> | |
6b06314c | 66 | #include <net/scm.h> |
2b188cc1 JA |
67 | #include <linux/anon_inodes.h> |
68 | #include <linux/sched/mm.h> | |
69 | #include <linux/uaccess.h> | |
70 | #include <linux/nospec.h> | |
edafccee JA |
71 | #include <linux/sizes.h> |
72 | #include <linux/hugetlb.h> | |
aa4c3967 | 73 | #include <linux/highmem.h> |
15b71abe JA |
74 | #include <linux/namei.h> |
75 | #include <linux/fsnotify.h> | |
4840e418 | 76 | #include <linux/fadvise.h> |
3e4827b0 | 77 | #include <linux/eventpoll.h> |
ff002b30 | 78 | #include <linux/fs_struct.h> |
7d67af2c | 79 | #include <linux/splice.h> |
b41e9852 | 80 | #include <linux/task_work.h> |
bcf5a063 | 81 | #include <linux/pagemap.h> |
0f212204 | 82 | #include <linux/io_uring.h> |
91d8f519 | 83 | #include <linux/blk-cgroup.h> |
4ea33a97 | 84 | #include <linux/audit.h> |
2b188cc1 | 85 | |
c826bd7a DD |
86 | #define CREATE_TRACE_POINTS |
87 | #include <trace/events/io_uring.h> | |
88 | ||
2b188cc1 JA |
89 | #include <uapi/linux/io_uring.h> |
90 | ||
91 | #include "internal.h" | |
561fb04a | 92 | #include "io-wq.h" |
2b188cc1 | 93 | |
5277deaa | 94 | #define IORING_MAX_ENTRIES 32768 |
33a107f0 | 95 | #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES) |
65e19f54 JA |
96 | |
97 | /* | |
98 | * Shift of 9 is 512 entries, or exactly one page on 64-bit archs | |
99 | */ | |
100 | #define IORING_FILE_TABLE_SHIFT 9 | |
101 | #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT) | |
102 | #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1) | |
103 | #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE) | |
21b55dbc SG |
104 | #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \ |
105 | IORING_REGISTER_LAST + IORING_OP_LAST) | |
2b188cc1 JA |
106 | |
107 | struct io_uring { | |
108 | u32 head ____cacheline_aligned_in_smp; | |
109 | u32 tail ____cacheline_aligned_in_smp; | |
110 | }; | |
111 | ||
1e84b97b | 112 | /* |
75b28aff HV |
113 | * This data is shared with the application through the mmap at offsets |
114 | * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING. | |
1e84b97b SB |
115 | * |
116 | * The offsets to the member fields are published through struct | |
117 | * io_sqring_offsets when calling io_uring_setup. | |
118 | */ | |
75b28aff | 119 | struct io_rings { |
1e84b97b SB |
120 | /* |
121 | * Head and tail offsets into the ring; the offsets need to be | |
122 | * masked to get valid indices. | |
123 | * | |
75b28aff HV |
124 | * The kernel controls head of the sq ring and the tail of the cq ring, |
125 | * and the application controls tail of the sq ring and the head of the | |
126 | * cq ring. | |
1e84b97b | 127 | */ |
75b28aff | 128 | struct io_uring sq, cq; |
1e84b97b | 129 | /* |
75b28aff | 130 | * Bitmasks to apply to head and tail offsets (constant, equals |
1e84b97b SB |
131 | * ring_entries - 1) |
132 | */ | |
75b28aff HV |
133 | u32 sq_ring_mask, cq_ring_mask; |
134 | /* Ring sizes (constant, power of 2) */ | |
135 | u32 sq_ring_entries, cq_ring_entries; | |
1e84b97b SB |
136 | /* |
137 | * Number of invalid entries dropped by the kernel due to | |
138 | * invalid index stored in array | |
139 | * | |
140 | * Written by the kernel, shouldn't be modified by the | |
141 | * application (i.e. get number of "new events" by comparing to | |
142 | * cached value). | |
143 | * | |
144 | * After a new SQ head value was read by the application this | |
145 | * counter includes all submissions that were dropped reaching | |
146 | * the new SQ head (and possibly more). | |
147 | */ | |
75b28aff | 148 | u32 sq_dropped; |
1e84b97b | 149 | /* |
0d9b5b3a | 150 | * Runtime SQ flags |
1e84b97b SB |
151 | * |
152 | * Written by the kernel, shouldn't be modified by the | |
153 | * application. | |
154 | * | |
155 | * The application needs a full memory barrier before checking | |
156 | * for IORING_SQ_NEED_WAKEUP after updating the sq tail. | |
157 | */ | |
75b28aff | 158 | u32 sq_flags; |
0d9b5b3a SG |
159 | /* |
160 | * Runtime CQ flags | |
161 | * | |
162 | * Written by the application, shouldn't be modified by the | |
163 | * kernel. | |
164 | */ | |
165 | u32 cq_flags; | |
1e84b97b SB |
166 | /* |
167 | * Number of completion events lost because the queue was full; | |
168 | * this should be avoided by the application by making sure | |
0b4295b5 | 169 | * there are not more requests pending than there is space in |
1e84b97b SB |
170 | * the completion queue. |
171 | * | |
172 | * Written by the kernel, shouldn't be modified by the | |
173 | * application (i.e. get number of "new events" by comparing to | |
174 | * cached value). | |
175 | * | |
176 | * As completion events come in out of order this counter is not | |
177 | * ordered with any other data. | |
178 | */ | |
75b28aff | 179 | u32 cq_overflow; |
1e84b97b SB |
180 | /* |
181 | * Ring buffer of completion events. | |
182 | * | |
183 | * The kernel writes completion events fresh every time they are | |
184 | * produced, so the application is allowed to modify pending | |
185 | * entries. | |
186 | */ | |
75b28aff | 187 | struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp; |
2b188cc1 JA |
188 | }; |
189 | ||
edafccee JA |
190 | struct io_mapped_ubuf { |
191 | u64 ubuf; | |
192 | size_t len; | |
193 | struct bio_vec *bvec; | |
194 | unsigned int nr_bvecs; | |
de293938 | 195 | unsigned long acct_pages; |
edafccee JA |
196 | }; |
197 | ||
65e19f54 JA |
198 | struct fixed_file_table { |
199 | struct file **files; | |
31b51510 JA |
200 | }; |
201 | ||
05589553 XW |
202 | struct fixed_file_ref_node { |
203 | struct percpu_ref refs; | |
204 | struct list_head node; | |
205 | struct list_head file_list; | |
206 | struct fixed_file_data *file_data; | |
4a38aed2 | 207 | struct llist_node llist; |
e297822b | 208 | bool done; |
05589553 XW |
209 | }; |
210 | ||
05f3fb3c JA |
211 | struct fixed_file_data { |
212 | struct fixed_file_table *table; | |
213 | struct io_ring_ctx *ctx; | |
214 | ||
b2e96852 | 215 | struct fixed_file_ref_node *node; |
05f3fb3c | 216 | struct percpu_ref refs; |
05f3fb3c | 217 | struct completion done; |
05589553 XW |
218 | struct list_head ref_list; |
219 | spinlock_t lock; | |
05f3fb3c JA |
220 | }; |
221 | ||
5a2e745d JA |
222 | struct io_buffer { |
223 | struct list_head list; | |
224 | __u64 addr; | |
225 | __s32 len; | |
226 | __u16 bid; | |
227 | }; | |
228 | ||
21b55dbc SG |
229 | struct io_restriction { |
230 | DECLARE_BITMAP(register_op, IORING_REGISTER_LAST); | |
231 | DECLARE_BITMAP(sqe_op, IORING_OP_LAST); | |
232 | u8 sqe_flags_allowed; | |
233 | u8 sqe_flags_required; | |
7e84e1c7 | 234 | bool registered; |
21b55dbc SG |
235 | }; |
236 | ||
534ca6d6 JA |
237 | struct io_sq_data { |
238 | refcount_t refs; | |
69fb2131 JA |
239 | struct mutex lock; |
240 | ||
241 | /* ctx's that are using this sqd */ | |
242 | struct list_head ctx_list; | |
243 | struct list_head ctx_new_list; | |
244 | struct mutex ctx_lock; | |
245 | ||
534ca6d6 JA |
246 | struct task_struct *thread; |
247 | struct wait_queue_head wait; | |
08369246 XW |
248 | |
249 | unsigned sq_thread_idle; | |
534ca6d6 JA |
250 | }; |
251 | ||
2b188cc1 JA |
252 | struct io_ring_ctx { |
253 | struct { | |
254 | struct percpu_ref refs; | |
255 | } ____cacheline_aligned_in_smp; | |
256 | ||
257 | struct { | |
258 | unsigned int flags; | |
e1d85334 | 259 | unsigned int compat: 1; |
aad5d8da | 260 | unsigned int limit_mem: 1; |
e1d85334 RD |
261 | unsigned int cq_overflow_flushed: 1; |
262 | unsigned int drain_next: 1; | |
263 | unsigned int eventfd_async: 1; | |
21b55dbc | 264 | unsigned int restricted: 1; |
d9d05217 | 265 | unsigned int sqo_dead: 1; |
2b188cc1 | 266 | |
75b28aff HV |
267 | /* |
268 | * Ring buffer of indices into array of io_uring_sqe, which is | |
269 | * mmapped by the application using the IORING_OFF_SQES offset. | |
270 | * | |
271 | * This indirection could e.g. be used to assign fixed | |
272 | * io_uring_sqe entries to operations and only submit them to | |
273 | * the queue when needed. | |
274 | * | |
275 | * The kernel modifies neither the indices array nor the entries | |
276 | * array. | |
277 | */ | |
278 | u32 *sq_array; | |
2b188cc1 JA |
279 | unsigned cached_sq_head; |
280 | unsigned sq_entries; | |
281 | unsigned sq_mask; | |
6c271ce2 | 282 | unsigned sq_thread_idle; |
498ccd9e | 283 | unsigned cached_sq_dropped; |
2c3bac6d | 284 | unsigned cached_cq_overflow; |
ad3eb2c8 | 285 | unsigned long sq_check_overflow; |
de0617e4 JA |
286 | |
287 | struct list_head defer_list; | |
5262f567 | 288 | struct list_head timeout_list; |
1d7bb1d5 | 289 | struct list_head cq_overflow_list; |
fcb323cc | 290 | |
ad3eb2c8 | 291 | struct io_uring_sqe *sq_sqes; |
2b188cc1 JA |
292 | } ____cacheline_aligned_in_smp; |
293 | ||
206aefde JA |
294 | struct io_rings *rings; |
295 | ||
2b188cc1 | 296 | /* IO offload */ |
561fb04a | 297 | struct io_wq *io_wq; |
2aede0e4 JA |
298 | |
299 | /* | |
300 | * For SQPOLL usage - we hold a reference to the parent task, so we | |
301 | * have access to the ->files | |
302 | */ | |
303 | struct task_struct *sqo_task; | |
304 | ||
305 | /* Only used for accounting purposes */ | |
306 | struct mm_struct *mm_account; | |
307 | ||
91d8f519 DZ |
308 | #ifdef CONFIG_BLK_CGROUP |
309 | struct cgroup_subsys_state *sqo_blkcg_css; | |
310 | #endif | |
311 | ||
534ca6d6 JA |
312 | struct io_sq_data *sq_data; /* if using sq thread polling */ |
313 | ||
90554200 | 314 | struct wait_queue_head sqo_sq_wait; |
69fb2131 | 315 | struct list_head sqd_list; |
75b28aff | 316 | |
6b06314c JA |
317 | /* |
318 | * If used, fixed file set. Writers must ensure that ->refs is dead, | |
319 | * readers must ensure that ->refs is alive as long as the file* is | |
320 | * used. Only updated through io_uring_register(2). | |
321 | */ | |
05f3fb3c | 322 | struct fixed_file_data *file_data; |
6b06314c JA |
323 | unsigned nr_user_files; |
324 | ||
edafccee JA |
325 | /* if used, fixed mapped user buffers */ |
326 | unsigned nr_user_bufs; | |
327 | struct io_mapped_ubuf *user_bufs; | |
328 | ||
2b188cc1 JA |
329 | struct user_struct *user; |
330 | ||
0b8c0ec7 | 331 | const struct cred *creds; |
181e448d | 332 | |
4ea33a97 JA |
333 | #ifdef CONFIG_AUDIT |
334 | kuid_t loginuid; | |
335 | unsigned int sessionid; | |
336 | #endif | |
337 | ||
0f158b4c JA |
338 | struct completion ref_comp; |
339 | struct completion sq_thread_comp; | |
206aefde | 340 | |
0ddf92e8 JA |
341 | /* if all else fails... */ |
342 | struct io_kiocb *fallback_req; | |
343 | ||
206aefde JA |
344 | #if defined(CONFIG_UNIX) |
345 | struct socket *ring_sock; | |
346 | #endif | |
347 | ||
5a2e745d JA |
348 | struct idr io_buffer_idr; |
349 | ||
071698e1 JA |
350 | struct idr personality_idr; |
351 | ||
206aefde JA |
352 | struct { |
353 | unsigned cached_cq_tail; | |
354 | unsigned cq_entries; | |
355 | unsigned cq_mask; | |
356 | atomic_t cq_timeouts; | |
f010505b | 357 | unsigned cq_last_tm_flush; |
ad3eb2c8 | 358 | unsigned long cq_check_overflow; |
206aefde JA |
359 | struct wait_queue_head cq_wait; |
360 | struct fasync_struct *cq_fasync; | |
361 | struct eventfd_ctx *cq_ev_fd; | |
362 | } ____cacheline_aligned_in_smp; | |
2b188cc1 JA |
363 | |
364 | struct { | |
365 | struct mutex uring_lock; | |
366 | wait_queue_head_t wait; | |
367 | } ____cacheline_aligned_in_smp; | |
368 | ||
369 | struct { | |
370 | spinlock_t completion_lock; | |
e94f141b | 371 | |
def596e9 | 372 | /* |
540e32a0 | 373 | * ->iopoll_list is protected by the ctx->uring_lock for |
def596e9 JA |
374 | * io_uring instances that don't use IORING_SETUP_SQPOLL. |
375 | * For SQPOLL, only the single threaded io_sq_thread() will | |
376 | * manipulate the list, hence no extra locking is needed there. | |
377 | */ | |
540e32a0 | 378 | struct list_head iopoll_list; |
78076bb6 JA |
379 | struct hlist_head *cancel_hash; |
380 | unsigned cancel_hash_bits; | |
e94f141b | 381 | bool poll_multi_file; |
31b51510 | 382 | |
fcb323cc JA |
383 | spinlock_t inflight_lock; |
384 | struct list_head inflight_list; | |
2b188cc1 | 385 | } ____cacheline_aligned_in_smp; |
85faa7b8 | 386 | |
4a38aed2 JA |
387 | struct delayed_work file_put_work; |
388 | struct llist_head file_put_llist; | |
389 | ||
85faa7b8 | 390 | struct work_struct exit_work; |
21b55dbc | 391 | struct io_restriction restrictions; |
2b188cc1 JA |
392 | }; |
393 | ||
09bb8394 JA |
394 | /* |
395 | * First field must be the file pointer in all the | |
396 | * iocb unions! See also 'struct kiocb' in <linux/fs.h> | |
397 | */ | |
221c5eb2 JA |
398 | struct io_poll_iocb { |
399 | struct file *file; | |
018043be | 400 | struct wait_queue_head *head; |
221c5eb2 | 401 | __poll_t events; |
8c838788 | 402 | bool done; |
221c5eb2 | 403 | bool canceled; |
392edb45 | 404 | struct wait_queue_entry wait; |
221c5eb2 JA |
405 | }; |
406 | ||
018043be PB |
407 | struct io_poll_remove { |
408 | struct file *file; | |
409 | u64 addr; | |
410 | }; | |
411 | ||
b5dba59e JA |
412 | struct io_close { |
413 | struct file *file; | |
414 | struct file *put_file; | |
415 | int fd; | |
416 | }; | |
417 | ||
ad8a48ac JA |
418 | struct io_timeout_data { |
419 | struct io_kiocb *req; | |
420 | struct hrtimer timer; | |
421 | struct timespec64 ts; | |
422 | enum hrtimer_mode mode; | |
423 | }; | |
424 | ||
8ed8d3c3 JA |
425 | struct io_accept { |
426 | struct file *file; | |
427 | struct sockaddr __user *addr; | |
428 | int __user *addr_len; | |
429 | int flags; | |
09952e3e | 430 | unsigned long nofile; |
8ed8d3c3 JA |
431 | }; |
432 | ||
433 | struct io_sync { | |
434 | struct file *file; | |
435 | loff_t len; | |
436 | loff_t off; | |
437 | int flags; | |
d63d1b5e | 438 | int mode; |
8ed8d3c3 JA |
439 | }; |
440 | ||
fbf23849 JA |
441 | struct io_cancel { |
442 | struct file *file; | |
443 | u64 addr; | |
444 | }; | |
445 | ||
b29472ee JA |
446 | struct io_timeout { |
447 | struct file *file; | |
bfe68a22 PB |
448 | u32 off; |
449 | u32 target_seq; | |
135fcde8 | 450 | struct list_head list; |
90cd7e42 PB |
451 | /* head of the link, used by linked timeouts only */ |
452 | struct io_kiocb *head; | |
b29472ee JA |
453 | }; |
454 | ||
0bdf7a2d PB |
455 | struct io_timeout_rem { |
456 | struct file *file; | |
457 | u64 addr; | |
9c8e11b3 PB |
458 | |
459 | /* timeout update */ | |
460 | struct timespec64 ts; | |
461 | u32 flags; | |
0bdf7a2d PB |
462 | }; |
463 | ||
9adbd45d JA |
464 | struct io_rw { |
465 | /* NOTE: kiocb has the file as the first member, so don't do it here */ | |
466 | struct kiocb kiocb; | |
467 | u64 addr; | |
468 | u64 len; | |
469 | }; | |
470 | ||
3fbb51c1 JA |
471 | struct io_connect { |
472 | struct file *file; | |
473 | struct sockaddr __user *addr; | |
474 | int addr_len; | |
475 | }; | |
476 | ||
e47293fd JA |
477 | struct io_sr_msg { |
478 | struct file *file; | |
fddaface | 479 | union { |
270a5940 | 480 | struct user_msghdr __user *umsg; |
fddaface JA |
481 | void __user *buf; |
482 | }; | |
e47293fd | 483 | int msg_flags; |
bcda7baa | 484 | int bgid; |
fddaface | 485 | size_t len; |
bcda7baa | 486 | struct io_buffer *kbuf; |
e47293fd JA |
487 | }; |
488 | ||
15b71abe JA |
489 | struct io_open { |
490 | struct file *file; | |
491 | int dfd; | |
944d1444 | 492 | bool ignore_nonblock; |
15b71abe | 493 | struct filename *filename; |
c12cedf2 | 494 | struct open_how how; |
4022e7af | 495 | unsigned long nofile; |
15b71abe JA |
496 | }; |
497 | ||
05f3fb3c JA |
498 | struct io_files_update { |
499 | struct file *file; | |
500 | u64 arg; | |
501 | u32 nr_args; | |
502 | u32 offset; | |
503 | }; | |
504 | ||
4840e418 JA |
505 | struct io_fadvise { |
506 | struct file *file; | |
507 | u64 offset; | |
508 | u32 len; | |
509 | u32 advice; | |
510 | }; | |
511 | ||
c1ca757b JA |
512 | struct io_madvise { |
513 | struct file *file; | |
514 | u64 addr; | |
515 | u32 len; | |
516 | u32 advice; | |
517 | }; | |
518 | ||
3e4827b0 JA |
519 | struct io_epoll { |
520 | struct file *file; | |
521 | int epfd; | |
522 | int op; | |
523 | int fd; | |
524 | struct epoll_event event; | |
e47293fd JA |
525 | }; |
526 | ||
7d67af2c PB |
527 | struct io_splice { |
528 | struct file *file_out; | |
529 | struct file *file_in; | |
530 | loff_t off_out; | |
531 | loff_t off_in; | |
532 | u64 len; | |
533 | unsigned int flags; | |
534 | }; | |
535 | ||
ddf0322d JA |
536 | struct io_provide_buf { |
537 | struct file *file; | |
538 | __u64 addr; | |
539 | __s32 len; | |
540 | __u32 bgid; | |
541 | __u16 nbufs; | |
542 | __u16 bid; | |
543 | }; | |
544 | ||
1d9e1288 BM |
545 | struct io_statx { |
546 | struct file *file; | |
547 | int dfd; | |
548 | unsigned int mask; | |
549 | unsigned int flags; | |
e62753e4 | 550 | const char __user *filename; |
1d9e1288 BM |
551 | struct statx __user *buffer; |
552 | }; | |
553 | ||
36f4fa68 JA |
554 | struct io_shutdown { |
555 | struct file *file; | |
556 | int how; | |
557 | }; | |
558 | ||
80a261fd JA |
559 | struct io_rename { |
560 | struct file *file; | |
561 | int old_dfd; | |
562 | int new_dfd; | |
563 | struct filename *oldpath; | |
564 | struct filename *newpath; | |
565 | int flags; | |
566 | }; | |
567 | ||
14a1143b JA |
568 | struct io_unlink { |
569 | struct file *file; | |
570 | int dfd; | |
571 | int flags; | |
572 | struct filename *filename; | |
573 | }; | |
574 | ||
3ca405eb PB |
575 | struct io_completion { |
576 | struct file *file; | |
577 | struct list_head list; | |
0f7e466b | 578 | int cflags; |
3ca405eb PB |
579 | }; |
580 | ||
f499a021 JA |
581 | struct io_async_connect { |
582 | struct sockaddr_storage address; | |
583 | }; | |
584 | ||
03b1230c JA |
585 | struct io_async_msghdr { |
586 | struct iovec fast_iov[UIO_FASTIOV]; | |
587 | struct iovec *iov; | |
588 | struct sockaddr __user *uaddr; | |
589 | struct msghdr msg; | |
b537916c | 590 | struct sockaddr_storage addr; |
03b1230c JA |
591 | }; |
592 | ||
f67676d1 JA |
593 | struct io_async_rw { |
594 | struct iovec fast_iov[UIO_FASTIOV]; | |
ff6165b2 JA |
595 | const struct iovec *free_iovec; |
596 | struct iov_iter iter; | |
227c0c96 | 597 | size_t bytes_done; |
bcf5a063 | 598 | struct wait_page_queue wpq; |
f67676d1 JA |
599 | }; |
600 | ||
6b47ee6e PB |
601 | enum { |
602 | REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT, | |
603 | REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT, | |
604 | REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT, | |
605 | REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT, | |
606 | REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT, | |
bcda7baa | 607 | REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT, |
6b47ee6e | 608 | |
6b47ee6e PB |
609 | REQ_F_FAIL_LINK_BIT, |
610 | REQ_F_INFLIGHT_BIT, | |
611 | REQ_F_CUR_POS_BIT, | |
612 | REQ_F_NOWAIT_BIT, | |
6b47ee6e | 613 | REQ_F_LINK_TIMEOUT_BIT, |
6b47ee6e | 614 | REQ_F_ISREG_BIT, |
99bc4c38 | 615 | REQ_F_NEED_CLEANUP_BIT, |
d7718a9d | 616 | REQ_F_POLLED_BIT, |
bcda7baa | 617 | REQ_F_BUFFER_SELECTED_BIT, |
5b0bbee4 | 618 | REQ_F_NO_FILE_TABLE_BIT, |
7cdaf587 | 619 | REQ_F_WORK_INITIALIZED_BIT, |
900fad45 | 620 | REQ_F_LTIMEOUT_ACTIVE_BIT, |
84557871 JA |
621 | |
622 | /* not a real bit, just to check we're not overflowing the space */ | |
623 | __REQ_F_LAST_BIT, | |
6b47ee6e PB |
624 | }; |
625 | ||
626 | enum { | |
627 | /* ctx owns file */ | |
628 | REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT), | |
629 | /* drain existing IO first */ | |
630 | REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT), | |
631 | /* linked sqes */ | |
632 | REQ_F_LINK = BIT(REQ_F_LINK_BIT), | |
633 | /* doesn't sever on completion < 0 */ | |
634 | REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT), | |
635 | /* IOSQE_ASYNC */ | |
636 | REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT), | |
bcda7baa JA |
637 | /* IOSQE_BUFFER_SELECT */ |
638 | REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT), | |
6b47ee6e | 639 | |
6b47ee6e PB |
640 | /* fail rest of links */ |
641 | REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT), | |
642 | /* on inflight list */ | |
643 | REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT), | |
644 | /* read/write uses file position */ | |
645 | REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT), | |
646 | /* must not punt to workers */ | |
647 | REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT), | |
900fad45 | 648 | /* has or had linked timeout */ |
6b47ee6e | 649 | REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT), |
6b47ee6e PB |
650 | /* regular file */ |
651 | REQ_F_ISREG = BIT(REQ_F_ISREG_BIT), | |
99bc4c38 PB |
652 | /* needs cleanup */ |
653 | REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT), | |
d7718a9d JA |
654 | /* already went through poll handler */ |
655 | REQ_F_POLLED = BIT(REQ_F_POLLED_BIT), | |
bcda7baa JA |
656 | /* buffer already selected */ |
657 | REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT), | |
5b0bbee4 JA |
658 | /* doesn't need file table for this request */ |
659 | REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT), | |
7cdaf587 XW |
660 | /* io_wq_work is initialized */ |
661 | REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT), | |
900fad45 PB |
662 | /* linked timeout is active, i.e. prepared by link's head */ |
663 | REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT), | |
d7718a9d JA |
664 | }; |
665 | ||
666 | struct async_poll { | |
667 | struct io_poll_iocb poll; | |
807abcb0 | 668 | struct io_poll_iocb *double_poll; |
6b47ee6e PB |
669 | }; |
670 | ||
09bb8394 JA |
671 | /* |
672 | * NOTE! Each of the iocb union members has the file pointer | |
673 | * as the first entry in their struct definition. So you can | |
674 | * access the file pointer through any of the sub-structs, | |
675 | * or directly as just 'ki_filp' in this struct. | |
676 | */ | |
2b188cc1 | 677 | struct io_kiocb { |
221c5eb2 | 678 | union { |
09bb8394 | 679 | struct file *file; |
9adbd45d | 680 | struct io_rw rw; |
221c5eb2 | 681 | struct io_poll_iocb poll; |
018043be | 682 | struct io_poll_remove poll_remove; |
8ed8d3c3 JA |
683 | struct io_accept accept; |
684 | struct io_sync sync; | |
fbf23849 | 685 | struct io_cancel cancel; |
b29472ee | 686 | struct io_timeout timeout; |
0bdf7a2d | 687 | struct io_timeout_rem timeout_rem; |
3fbb51c1 | 688 | struct io_connect connect; |
e47293fd | 689 | struct io_sr_msg sr_msg; |
15b71abe | 690 | struct io_open open; |
b5dba59e | 691 | struct io_close close; |
05f3fb3c | 692 | struct io_files_update files_update; |
4840e418 | 693 | struct io_fadvise fadvise; |
c1ca757b | 694 | struct io_madvise madvise; |
3e4827b0 | 695 | struct io_epoll epoll; |
7d67af2c | 696 | struct io_splice splice; |
ddf0322d | 697 | struct io_provide_buf pbuf; |
1d9e1288 | 698 | struct io_statx statx; |
36f4fa68 | 699 | struct io_shutdown shutdown; |
80a261fd | 700 | struct io_rename rename; |
14a1143b | 701 | struct io_unlink unlink; |
3ca405eb PB |
702 | /* use only after cleaning per-op data, see io_clean_op() */ |
703 | struct io_completion compl; | |
221c5eb2 | 704 | }; |
2b188cc1 | 705 | |
e8c2bc1f JA |
706 | /* opcode allocated if it needs to store data for async defer */ |
707 | void *async_data; | |
d625c6ee | 708 | u8 opcode; |
65a6543d XW |
709 | /* polled IO has completed */ |
710 | u8 iopoll_completed; | |
2b188cc1 | 711 | |
4f4eeba8 | 712 | u16 buf_index; |
9cf7c104 | 713 | u32 result; |
4f4eeba8 | 714 | |
010e8e6b PB |
715 | struct io_ring_ctx *ctx; |
716 | unsigned int flags; | |
717 | refcount_t refs; | |
718 | struct task_struct *task; | |
719 | u64 user_data; | |
d7718a9d | 720 | |
f2f87370 | 721 | struct io_kiocb *link; |
0415767e | 722 | struct percpu_ref *fixed_file_refs; |
fcb323cc | 723 | |
d21ffe7e PB |
724 | /* |
725 | * 1. used with ctx->iopoll_list with reads/writes | |
726 | * 2. to track reqs with ->files (see io_op_def::file_table) | |
727 | */ | |
010e8e6b | 728 | struct list_head inflight_entry; |
010e8e6b PB |
729 | struct callback_head task_work; |
730 | /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */ | |
731 | struct hlist_node hash_node; | |
732 | struct async_poll *apoll; | |
733 | struct io_wq_work work; | |
2b188cc1 | 734 | }; |
05589553 | 735 | |
27dc8338 PB |
736 | struct io_defer_entry { |
737 | struct list_head list; | |
738 | struct io_kiocb *req; | |
9cf7c104 | 739 | u32 seq; |
2b188cc1 JA |
740 | }; |
741 | ||
def596e9 | 742 | #define IO_IOPOLL_BATCH 8 |
2b188cc1 | 743 | |
013538bd JA |
744 | struct io_comp_state { |
745 | unsigned int nr; | |
746 | struct list_head list; | |
747 | struct io_ring_ctx *ctx; | |
748 | }; | |
749 | ||
9a56a232 JA |
750 | struct io_submit_state { |
751 | struct blk_plug plug; | |
752 | ||
2579f913 JA |
753 | /* |
754 | * io_kiocb alloc cache | |
755 | */ | |
756 | void *reqs[IO_IOPOLL_BATCH]; | |
6c8a3134 | 757 | unsigned int free_reqs; |
2579f913 | 758 | |
27926b68 JA |
759 | bool plug_started; |
760 | ||
013538bd JA |
761 | /* |
762 | * Batch completion logic | |
763 | */ | |
764 | struct io_comp_state comp; | |
765 | ||
9a56a232 JA |
766 | /* |
767 | * File reference cache | |
768 | */ | |
769 | struct file *file; | |
770 | unsigned int fd; | |
6e1271e6 | 771 | unsigned int file_refs; |
9a56a232 JA |
772 | unsigned int ios_left; |
773 | }; | |
774 | ||
d3656344 | 775 | struct io_op_def { |
d3656344 JA |
776 | /* needs req->file assigned */ |
777 | unsigned needs_file : 1; | |
fd2206e4 JA |
778 | /* don't fail if file grab fails */ |
779 | unsigned needs_file_no_error : 1; | |
d3656344 JA |
780 | /* hash wq insertion if file is a regular file */ |
781 | unsigned hash_reg_file : 1; | |
782 | /* unbound wq insertion if file is a non-regular file */ | |
783 | unsigned unbound_nonreg_file : 1; | |
66f4af93 JA |
784 | /* opcode is not supported by this kernel */ |
785 | unsigned not_supported : 1; | |
8a72758c JA |
786 | /* set if opcode supports polled "wait" */ |
787 | unsigned pollin : 1; | |
788 | unsigned pollout : 1; | |
bcda7baa JA |
789 | /* op supports buffer selection */ |
790 | unsigned buffer_select : 1; | |
e8c2bc1f JA |
791 | /* must always have async data allocated */ |
792 | unsigned needs_async_data : 1; | |
27926b68 JA |
793 | /* should block plug */ |
794 | unsigned plug : 1; | |
e8c2bc1f JA |
795 | /* size of async data needed, if any */ |
796 | unsigned short async_size; | |
0f203765 | 797 | unsigned work_flags; |
d3656344 JA |
798 | }; |
799 | ||
0918682b | 800 | static const struct io_op_def io_op_defs[] = { |
0463b6c5 PB |
801 | [IORING_OP_NOP] = {}, |
802 | [IORING_OP_READV] = { | |
d3656344 JA |
803 | .needs_file = 1, |
804 | .unbound_nonreg_file = 1, | |
8a72758c | 805 | .pollin = 1, |
4d954c25 | 806 | .buffer_select = 1, |
e8c2bc1f | 807 | .needs_async_data = 1, |
27926b68 | 808 | .plug = 1, |
e8c2bc1f | 809 | .async_size = sizeof(struct io_async_rw), |
0f203765 | 810 | .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG, |
d3656344 | 811 | }, |
0463b6c5 | 812 | [IORING_OP_WRITEV] = { |
d3656344 JA |
813 | .needs_file = 1, |
814 | .hash_reg_file = 1, | |
815 | .unbound_nonreg_file = 1, | |
8a72758c | 816 | .pollout = 1, |
e8c2bc1f | 817 | .needs_async_data = 1, |
27926b68 | 818 | .plug = 1, |
e8c2bc1f | 819 | .async_size = sizeof(struct io_async_rw), |
69228338 JA |
820 | .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG | |
821 | IO_WQ_WORK_FSIZE, | |
d3656344 | 822 | }, |
0463b6c5 | 823 | [IORING_OP_FSYNC] = { |
d3656344 | 824 | .needs_file = 1, |
0f203765 | 825 | .work_flags = IO_WQ_WORK_BLKCG, |
d3656344 | 826 | }, |
0463b6c5 | 827 | [IORING_OP_READ_FIXED] = { |
d3656344 JA |
828 | .needs_file = 1, |
829 | .unbound_nonreg_file = 1, | |
8a72758c | 830 | .pollin = 1, |
27926b68 | 831 | .plug = 1, |
e8c2bc1f | 832 | .async_size = sizeof(struct io_async_rw), |
4017eb91 | 833 | .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM, |
d3656344 | 834 | }, |
0463b6c5 | 835 | [IORING_OP_WRITE_FIXED] = { |
d3656344 JA |
836 | .needs_file = 1, |
837 | .hash_reg_file = 1, | |
838 | .unbound_nonreg_file = 1, | |
8a72758c | 839 | .pollout = 1, |
27926b68 | 840 | .plug = 1, |
e8c2bc1f | 841 | .async_size = sizeof(struct io_async_rw), |
4017eb91 JA |
842 | .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE | |
843 | IO_WQ_WORK_MM, | |
d3656344 | 844 | }, |
0463b6c5 | 845 | [IORING_OP_POLL_ADD] = { |
d3656344 JA |
846 | .needs_file = 1, |
847 | .unbound_nonreg_file = 1, | |
848 | }, | |
0463b6c5 PB |
849 | [IORING_OP_POLL_REMOVE] = {}, |
850 | [IORING_OP_SYNC_FILE_RANGE] = { | |
d3656344 | 851 | .needs_file = 1, |
0f203765 | 852 | .work_flags = IO_WQ_WORK_BLKCG, |
d3656344 | 853 | }, |
0463b6c5 | 854 | [IORING_OP_SENDMSG] = { |
d3656344 JA |
855 | .needs_file = 1, |
856 | .unbound_nonreg_file = 1, | |
8a72758c | 857 | .pollout = 1, |
e8c2bc1f JA |
858 | .needs_async_data = 1, |
859 | .async_size = sizeof(struct io_async_msghdr), | |
10cad2c4 | 860 | .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG, |
d3656344 | 861 | }, |
0463b6c5 | 862 | [IORING_OP_RECVMSG] = { |
d3656344 JA |
863 | .needs_file = 1, |
864 | .unbound_nonreg_file = 1, | |
8a72758c | 865 | .pollin = 1, |
52de1fe1 | 866 | .buffer_select = 1, |
e8c2bc1f JA |
867 | .needs_async_data = 1, |
868 | .async_size = sizeof(struct io_async_msghdr), | |
10cad2c4 | 869 | .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG, |
d3656344 | 870 | }, |
0463b6c5 | 871 | [IORING_OP_TIMEOUT] = { |
e8c2bc1f JA |
872 | .needs_async_data = 1, |
873 | .async_size = sizeof(struct io_timeout_data), | |
0f203765 | 874 | .work_flags = IO_WQ_WORK_MM, |
d3656344 | 875 | }, |
9c8e11b3 PB |
876 | [IORING_OP_TIMEOUT_REMOVE] = { |
877 | /* used by timeout updates' prep() */ | |
878 | .work_flags = IO_WQ_WORK_MM, | |
879 | }, | |
0463b6c5 | 880 | [IORING_OP_ACCEPT] = { |
d3656344 JA |
881 | .needs_file = 1, |
882 | .unbound_nonreg_file = 1, | |
8a72758c | 883 | .pollin = 1, |
0f203765 | 884 | .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES, |
d3656344 | 885 | }, |
0463b6c5 PB |
886 | [IORING_OP_ASYNC_CANCEL] = {}, |
887 | [IORING_OP_LINK_TIMEOUT] = { | |
e8c2bc1f JA |
888 | .needs_async_data = 1, |
889 | .async_size = sizeof(struct io_timeout_data), | |
0f203765 | 890 | .work_flags = IO_WQ_WORK_MM, |
d3656344 | 891 | }, |
0463b6c5 | 892 | [IORING_OP_CONNECT] = { |
d3656344 JA |
893 | .needs_file = 1, |
894 | .unbound_nonreg_file = 1, | |
8a72758c | 895 | .pollout = 1, |
e8c2bc1f JA |
896 | .needs_async_data = 1, |
897 | .async_size = sizeof(struct io_async_connect), | |
0f203765 | 898 | .work_flags = IO_WQ_WORK_MM, |
d3656344 | 899 | }, |
0463b6c5 | 900 | [IORING_OP_FALLOCATE] = { |
d3656344 | 901 | .needs_file = 1, |
69228338 | 902 | .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE, |
d3656344 | 903 | }, |
0463b6c5 | 904 | [IORING_OP_OPENAT] = { |
0f203765 | 905 | .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG | |
14587a46 | 906 | IO_WQ_WORK_FS | IO_WQ_WORK_MM, |
d3656344 | 907 | }, |
0463b6c5 | 908 | [IORING_OP_CLOSE] = { |
fd2206e4 JA |
909 | .needs_file = 1, |
910 | .needs_file_no_error = 1, | |
0f203765 | 911 | .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG, |
d3656344 | 912 | }, |
0463b6c5 | 913 | [IORING_OP_FILES_UPDATE] = { |
0f203765 | 914 | .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM, |
d3656344 | 915 | }, |
0463b6c5 | 916 | [IORING_OP_STATX] = { |
0f203765 JA |
917 | .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM | |
918 | IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG, | |
d3656344 | 919 | }, |
0463b6c5 | 920 | [IORING_OP_READ] = { |
3a6820f2 JA |
921 | .needs_file = 1, |
922 | .unbound_nonreg_file = 1, | |
8a72758c | 923 | .pollin = 1, |
bcda7baa | 924 | .buffer_select = 1, |
27926b68 | 925 | .plug = 1, |
e8c2bc1f | 926 | .async_size = sizeof(struct io_async_rw), |
0f203765 | 927 | .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG, |
3a6820f2 | 928 | }, |
0463b6c5 | 929 | [IORING_OP_WRITE] = { |
3a6820f2 JA |
930 | .needs_file = 1, |
931 | .unbound_nonreg_file = 1, | |
8a72758c | 932 | .pollout = 1, |
27926b68 | 933 | .plug = 1, |
e8c2bc1f | 934 | .async_size = sizeof(struct io_async_rw), |
69228338 JA |
935 | .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG | |
936 | IO_WQ_WORK_FSIZE, | |
3a6820f2 | 937 | }, |
0463b6c5 | 938 | [IORING_OP_FADVISE] = { |
4840e418 | 939 | .needs_file = 1, |
0f203765 | 940 | .work_flags = IO_WQ_WORK_BLKCG, |
4840e418 | 941 | }, |
0463b6c5 | 942 | [IORING_OP_MADVISE] = { |
0f203765 | 943 | .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG, |
c1ca757b | 944 | }, |
0463b6c5 | 945 | [IORING_OP_SEND] = { |
fddaface JA |
946 | .needs_file = 1, |
947 | .unbound_nonreg_file = 1, | |
8a72758c | 948 | .pollout = 1, |
0f203765 | 949 | .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG, |
fddaface | 950 | }, |
0463b6c5 | 951 | [IORING_OP_RECV] = { |
fddaface JA |
952 | .needs_file = 1, |
953 | .unbound_nonreg_file = 1, | |
8a72758c | 954 | .pollin = 1, |
bcda7baa | 955 | .buffer_select = 1, |
0f203765 | 956 | .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG, |
fddaface | 957 | }, |
0463b6c5 | 958 | [IORING_OP_OPENAT2] = { |
0f203765 | 959 | .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_FS | |
14587a46 | 960 | IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM, |
cebdb986 | 961 | }, |
3e4827b0 JA |
962 | [IORING_OP_EPOLL_CTL] = { |
963 | .unbound_nonreg_file = 1, | |
0f203765 | 964 | .work_flags = IO_WQ_WORK_FILES, |
3e4827b0 | 965 | }, |
7d67af2c PB |
966 | [IORING_OP_SPLICE] = { |
967 | .needs_file = 1, | |
968 | .hash_reg_file = 1, | |
969 | .unbound_nonreg_file = 1, | |
0f203765 | 970 | .work_flags = IO_WQ_WORK_BLKCG, |
ddf0322d JA |
971 | }, |
972 | [IORING_OP_PROVIDE_BUFFERS] = {}, | |
067524e9 | 973 | [IORING_OP_REMOVE_BUFFERS] = {}, |
f2a8d5c7 PB |
974 | [IORING_OP_TEE] = { |
975 | .needs_file = 1, | |
976 | .hash_reg_file = 1, | |
977 | .unbound_nonreg_file = 1, | |
978 | }, | |
36f4fa68 JA |
979 | [IORING_OP_SHUTDOWN] = { |
980 | .needs_file = 1, | |
981 | }, | |
80a261fd JA |
982 | [IORING_OP_RENAMEAT] = { |
983 | .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES | | |
984 | IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG, | |
985 | }, | |
14a1143b JA |
986 | [IORING_OP_UNLINKAT] = { |
987 | .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES | | |
988 | IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG, | |
989 | }, | |
d3656344 JA |
990 | }; |
991 | ||
2e0464d4 BM |
992 | enum io_mem_account { |
993 | ACCT_LOCKED, | |
994 | ACCT_PINNED, | |
995 | }; | |
996 | ||
90df0853 PB |
997 | static void __io_uring_cancel_task_requests(struct io_ring_ctx *ctx, |
998 | struct task_struct *task); | |
999 | ||
1ffc5422 PB |
1000 | static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node); |
1001 | static struct fixed_file_ref_node *alloc_fixed_file_ref_node( | |
1002 | struct io_ring_ctx *ctx); | |
1003 | ||
81b68a5c PB |
1004 | static void __io_complete_rw(struct io_kiocb *req, long res, long res2, |
1005 | struct io_comp_state *cs); | |
78e19bbe | 1006 | static void io_cqring_fill_event(struct io_kiocb *req, long res); |
ec9c02ad | 1007 | static void io_put_req(struct io_kiocb *req); |
216578e5 | 1008 | static void io_put_req_deferred(struct io_kiocb *req, int nr); |
c40f6379 | 1009 | static void io_double_put_req(struct io_kiocb *req); |
94ae5e77 | 1010 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req); |
7271ef3a | 1011 | static void __io_queue_linked_timeout(struct io_kiocb *req); |
94ae5e77 | 1012 | static void io_queue_linked_timeout(struct io_kiocb *req); |
05f3fb3c JA |
1013 | static int __io_sqe_files_update(struct io_ring_ctx *ctx, |
1014 | struct io_uring_files_update *ip, | |
1015 | unsigned nr_args); | |
3ca405eb | 1016 | static void __io_clean_op(struct io_kiocb *req); |
8371adf5 PB |
1017 | static struct file *io_file_get(struct io_submit_state *state, |
1018 | struct io_kiocb *req, int fd, bool fixed); | |
c1379e24 | 1019 | static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs); |
4349f30e | 1020 | static void io_file_put_work(struct work_struct *work); |
de0617e4 | 1021 | |
b63534c4 JA |
1022 | static ssize_t io_import_iovec(int rw, struct io_kiocb *req, |
1023 | struct iovec **iovec, struct iov_iter *iter, | |
1024 | bool needs_lock); | |
ff6165b2 JA |
1025 | static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec, |
1026 | const struct iovec *fast_iov, | |
227c0c96 | 1027 | struct iov_iter *iter, bool force); |
9d5c8190 | 1028 | static void io_req_drop_files(struct io_kiocb *req); |
de0617e4 | 1029 | |
2b188cc1 JA |
1030 | static struct kmem_cache *req_cachep; |
1031 | ||
0918682b | 1032 | static const struct file_operations io_uring_fops; |
2b188cc1 JA |
1033 | |
1034 | struct sock *io_uring_get_socket(struct file *file) | |
1035 | { | |
1036 | #if defined(CONFIG_UNIX) | |
1037 | if (file->f_op == &io_uring_fops) { | |
1038 | struct io_ring_ctx *ctx = file->private_data; | |
1039 | ||
1040 | return ctx->ring_sock->sk; | |
1041 | } | |
1042 | #endif | |
1043 | return NULL; | |
1044 | } | |
1045 | EXPORT_SYMBOL(io_uring_get_socket); | |
1046 | ||
f2f87370 PB |
1047 | #define io_for_each_link(pos, head) \ |
1048 | for (pos = (head); pos; pos = pos->link) | |
1049 | ||
3ca405eb PB |
1050 | static inline void io_clean_op(struct io_kiocb *req) |
1051 | { | |
9d5c8190 | 1052 | if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED)) |
3ca405eb PB |
1053 | __io_clean_op(req); |
1054 | } | |
1055 | ||
36f72fe2 PB |
1056 | static inline void io_set_resource_node(struct io_kiocb *req) |
1057 | { | |
1058 | struct io_ring_ctx *ctx = req->ctx; | |
1059 | ||
1060 | if (!req->fixed_file_refs) { | |
1061 | req->fixed_file_refs = &ctx->file_data->node->refs; | |
1062 | percpu_ref_get(req->fixed_file_refs); | |
1063 | } | |
1064 | } | |
1065 | ||
08d23634 PB |
1066 | static bool io_match_task(struct io_kiocb *head, |
1067 | struct task_struct *task, | |
1068 | struct files_struct *files) | |
1069 | { | |
1070 | struct io_kiocb *req; | |
1071 | ||
1072 | if (task && head->task != task) | |
1073 | return false; | |
1074 | if (!files) | |
1075 | return true; | |
1076 | ||
1077 | io_for_each_link(req, head) { | |
02a13674 JA |
1078 | if (!(req->flags & REQ_F_WORK_INITIALIZED)) |
1079 | continue; | |
1080 | if (req->file && req->file->f_op == &io_uring_fops) | |
1081 | return true; | |
1082 | if ((req->work.flags & IO_WQ_WORK_FILES) && | |
08d23634 PB |
1083 | req->work.identity->files == files) |
1084 | return true; | |
1085 | } | |
1086 | return false; | |
1087 | } | |
1088 | ||
28cea78a | 1089 | static void io_sq_thread_drop_mm_files(void) |
c40f6379 | 1090 | { |
28cea78a | 1091 | struct files_struct *files = current->files; |
c40f6379 JA |
1092 | struct mm_struct *mm = current->mm; |
1093 | ||
1094 | if (mm) { | |
1095 | kthread_unuse_mm(mm); | |
1096 | mmput(mm); | |
4b70cf9d | 1097 | current->mm = NULL; |
c40f6379 | 1098 | } |
28cea78a JA |
1099 | if (files) { |
1100 | struct nsproxy *nsproxy = current->nsproxy; | |
1101 | ||
1102 | task_lock(current); | |
1103 | current->files = NULL; | |
1104 | current->nsproxy = NULL; | |
1105 | task_unlock(current); | |
1106 | put_files_struct(files); | |
1107 | put_nsproxy(nsproxy); | |
1108 | } | |
1109 | } | |
1110 | ||
1a38ffc9 | 1111 | static int __io_sq_thread_acquire_files(struct io_ring_ctx *ctx) |
28cea78a | 1112 | { |
621fadc2 PB |
1113 | if (current->flags & PF_EXITING) |
1114 | return -EFAULT; | |
1115 | ||
28cea78a JA |
1116 | if (!current->files) { |
1117 | struct files_struct *files; | |
1118 | struct nsproxy *nsproxy; | |
1119 | ||
1120 | task_lock(ctx->sqo_task); | |
1121 | files = ctx->sqo_task->files; | |
1122 | if (!files) { | |
1123 | task_unlock(ctx->sqo_task); | |
1a38ffc9 | 1124 | return -EOWNERDEAD; |
28cea78a JA |
1125 | } |
1126 | atomic_inc(&files->count); | |
1127 | get_nsproxy(ctx->sqo_task->nsproxy); | |
1128 | nsproxy = ctx->sqo_task->nsproxy; | |
1129 | task_unlock(ctx->sqo_task); | |
1130 | ||
1131 | task_lock(current); | |
1132 | current->files = files; | |
1133 | current->nsproxy = nsproxy; | |
1134 | task_unlock(current); | |
1135 | } | |
1a38ffc9 | 1136 | return 0; |
c40f6379 JA |
1137 | } |
1138 | ||
1139 | static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx) | |
1140 | { | |
4b70cf9d JA |
1141 | struct mm_struct *mm; |
1142 | ||
621fadc2 PB |
1143 | if (current->flags & PF_EXITING) |
1144 | return -EFAULT; | |
4b70cf9d JA |
1145 | if (current->mm) |
1146 | return 0; | |
1147 | ||
1148 | /* Should never happen */ | |
1149 | if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL))) | |
1150 | return -EFAULT; | |
1151 | ||
1152 | task_lock(ctx->sqo_task); | |
1153 | mm = ctx->sqo_task->mm; | |
1154 | if (unlikely(!mm || !mmget_not_zero(mm))) | |
1155 | mm = NULL; | |
1156 | task_unlock(ctx->sqo_task); | |
1157 | ||
1158 | if (mm) { | |
1159 | kthread_use_mm(mm); | |
1160 | return 0; | |
c40f6379 JA |
1161 | } |
1162 | ||
4b70cf9d | 1163 | return -EFAULT; |
c40f6379 JA |
1164 | } |
1165 | ||
28cea78a JA |
1166 | static int io_sq_thread_acquire_mm_files(struct io_ring_ctx *ctx, |
1167 | struct io_kiocb *req) | |
c40f6379 | 1168 | { |
28cea78a | 1169 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
1a38ffc9 | 1170 | int ret; |
28cea78a JA |
1171 | |
1172 | if (def->work_flags & IO_WQ_WORK_MM) { | |
1a38ffc9 | 1173 | ret = __io_sq_thread_acquire_mm(ctx); |
28cea78a JA |
1174 | if (unlikely(ret)) |
1175 | return ret; | |
1176 | } | |
1177 | ||
1a38ffc9 PB |
1178 | if (def->needs_file || (def->work_flags & IO_WQ_WORK_FILES)) { |
1179 | ret = __io_sq_thread_acquire_files(ctx); | |
1180 | if (unlikely(ret)) | |
1181 | return ret; | |
1182 | } | |
28cea78a JA |
1183 | |
1184 | return 0; | |
c40f6379 JA |
1185 | } |
1186 | ||
91d8f519 DZ |
1187 | static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx, |
1188 | struct cgroup_subsys_state **cur_css) | |
1189 | ||
1190 | { | |
1191 | #ifdef CONFIG_BLK_CGROUP | |
1192 | /* puts the old one when swapping */ | |
1193 | if (*cur_css != ctx->sqo_blkcg_css) { | |
1194 | kthread_associate_blkcg(ctx->sqo_blkcg_css); | |
1195 | *cur_css = ctx->sqo_blkcg_css; | |
1196 | } | |
1197 | #endif | |
1198 | } | |
1199 | ||
1200 | static void io_sq_thread_unassociate_blkcg(void) | |
1201 | { | |
1202 | #ifdef CONFIG_BLK_CGROUP | |
1203 | kthread_associate_blkcg(NULL); | |
1204 | #endif | |
1205 | } | |
1206 | ||
c40f6379 JA |
1207 | static inline void req_set_fail_links(struct io_kiocb *req) |
1208 | { | |
1209 | if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK) | |
1210 | req->flags |= REQ_F_FAIL_LINK; | |
1211 | } | |
4a38aed2 | 1212 | |
1e6fa521 JA |
1213 | /* |
1214 | * None of these are dereferenced, they are simply used to check if any of | |
1215 | * them have changed. If we're under current and check they are still the | |
1216 | * same, we're fine to grab references to them for actual out-of-line use. | |
1217 | */ | |
1218 | static void io_init_identity(struct io_identity *id) | |
1219 | { | |
1220 | id->files = current->files; | |
1221 | id->mm = current->mm; | |
1222 | #ifdef CONFIG_BLK_CGROUP | |
1223 | rcu_read_lock(); | |
1224 | id->blkcg_css = blkcg_css(); | |
1225 | rcu_read_unlock(); | |
1226 | #endif | |
1227 | id->creds = current_cred(); | |
1228 | id->nsproxy = current->nsproxy; | |
1229 | id->fs = current->fs; | |
1230 | id->fsize = rlimit(RLIMIT_FSIZE); | |
4ea33a97 JA |
1231 | #ifdef CONFIG_AUDIT |
1232 | id->loginuid = current->loginuid; | |
1233 | id->sessionid = current->sessionid; | |
1234 | #endif | |
1e6fa521 JA |
1235 | refcount_set(&id->count, 1); |
1236 | } | |
1237 | ||
ec99ca6c PB |
1238 | static inline void __io_req_init_async(struct io_kiocb *req) |
1239 | { | |
1240 | memset(&req->work, 0, sizeof(req->work)); | |
1241 | req->flags |= REQ_F_WORK_INITIALIZED; | |
1242 | } | |
1243 | ||
7cdaf587 XW |
1244 | /* |
1245 | * Note: must call io_req_init_async() for the first time you | |
1246 | * touch any members of io_wq_work. | |
1247 | */ | |
1248 | static inline void io_req_init_async(struct io_kiocb *req) | |
1249 | { | |
500a373d JA |
1250 | struct io_uring_task *tctx = current->io_uring; |
1251 | ||
7cdaf587 XW |
1252 | if (req->flags & REQ_F_WORK_INITIALIZED) |
1253 | return; | |
1254 | ||
ec99ca6c | 1255 | __io_req_init_async(req); |
500a373d JA |
1256 | |
1257 | /* Grab a ref if this isn't our static identity */ | |
1258 | req->work.identity = tctx->identity; | |
1259 | if (tctx->identity != &tctx->__identity) | |
1260 | refcount_inc(&req->work.identity->count); | |
7cdaf587 XW |
1261 | } |
1262 | ||
0cdaf760 PB |
1263 | static inline bool io_async_submit(struct io_ring_ctx *ctx) |
1264 | { | |
1265 | return ctx->flags & IORING_SETUP_SQPOLL; | |
1266 | } | |
1267 | ||
2b188cc1 JA |
1268 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) |
1269 | { | |
1270 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
1271 | ||
0f158b4c | 1272 | complete(&ctx->ref_comp); |
2b188cc1 JA |
1273 | } |
1274 | ||
8eb7e2d0 PB |
1275 | static inline bool io_is_timeout_noseq(struct io_kiocb *req) |
1276 | { | |
1277 | return !req->timeout.off; | |
1278 | } | |
1279 | ||
2b188cc1 JA |
1280 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) |
1281 | { | |
1282 | struct io_ring_ctx *ctx; | |
78076bb6 | 1283 | int hash_bits; |
2b188cc1 JA |
1284 | |
1285 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
1286 | if (!ctx) | |
1287 | return NULL; | |
1288 | ||
0ddf92e8 JA |
1289 | ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL); |
1290 | if (!ctx->fallback_req) | |
1291 | goto err; | |
1292 | ||
78076bb6 JA |
1293 | /* |
1294 | * Use 5 bits less than the max cq entries, that should give us around | |
1295 | * 32 entries per hash list if totally full and uniformly spread. | |
1296 | */ | |
1297 | hash_bits = ilog2(p->cq_entries); | |
1298 | hash_bits -= 5; | |
1299 | if (hash_bits <= 0) | |
1300 | hash_bits = 1; | |
1301 | ctx->cancel_hash_bits = hash_bits; | |
1302 | ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head), | |
1303 | GFP_KERNEL); | |
1304 | if (!ctx->cancel_hash) | |
1305 | goto err; | |
1306 | __hash_init(ctx->cancel_hash, 1U << hash_bits); | |
1307 | ||
21482896 | 1308 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
206aefde JA |
1309 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) |
1310 | goto err; | |
2b188cc1 JA |
1311 | |
1312 | ctx->flags = p->flags; | |
90554200 | 1313 | init_waitqueue_head(&ctx->sqo_sq_wait); |
69fb2131 | 1314 | INIT_LIST_HEAD(&ctx->sqd_list); |
2b188cc1 | 1315 | init_waitqueue_head(&ctx->cq_wait); |
1d7bb1d5 | 1316 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
0f158b4c JA |
1317 | init_completion(&ctx->ref_comp); |
1318 | init_completion(&ctx->sq_thread_comp); | |
5a2e745d | 1319 | idr_init(&ctx->io_buffer_idr); |
071698e1 | 1320 | idr_init(&ctx->personality_idr); |
2b188cc1 JA |
1321 | mutex_init(&ctx->uring_lock); |
1322 | init_waitqueue_head(&ctx->wait); | |
1323 | spin_lock_init(&ctx->completion_lock); | |
540e32a0 | 1324 | INIT_LIST_HEAD(&ctx->iopoll_list); |
de0617e4 | 1325 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 1326 | INIT_LIST_HEAD(&ctx->timeout_list); |
fcb323cc JA |
1327 | spin_lock_init(&ctx->inflight_lock); |
1328 | INIT_LIST_HEAD(&ctx->inflight_list); | |
4a38aed2 JA |
1329 | INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work); |
1330 | init_llist_head(&ctx->file_put_llist); | |
2b188cc1 | 1331 | return ctx; |
206aefde | 1332 | err: |
0ddf92e8 JA |
1333 | if (ctx->fallback_req) |
1334 | kmem_cache_free(req_cachep, ctx->fallback_req); | |
78076bb6 | 1335 | kfree(ctx->cancel_hash); |
206aefde JA |
1336 | kfree(ctx); |
1337 | return NULL; | |
2b188cc1 JA |
1338 | } |
1339 | ||
9cf7c104 | 1340 | static bool req_need_defer(struct io_kiocb *req, u32 seq) |
7adf4eaf | 1341 | { |
2bc9930e JA |
1342 | if (unlikely(req->flags & REQ_F_IO_DRAIN)) { |
1343 | struct io_ring_ctx *ctx = req->ctx; | |
a197f664 | 1344 | |
9cf7c104 | 1345 | return seq != ctx->cached_cq_tail |
2c3bac6d | 1346 | + READ_ONCE(ctx->cached_cq_overflow); |
2bc9930e | 1347 | } |
de0617e4 | 1348 | |
9d858b21 | 1349 | return false; |
de0617e4 JA |
1350 | } |
1351 | ||
de0617e4 | 1352 | static void __io_commit_cqring(struct io_ring_ctx *ctx) |
2b188cc1 | 1353 | { |
75b28aff | 1354 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 1355 | |
07910158 PB |
1356 | /* order cqe stores with ring update */ |
1357 | smp_store_release(&rings->cq.tail, ctx->cached_cq_tail); | |
2b188cc1 JA |
1358 | } |
1359 | ||
5c3462cf | 1360 | static void io_put_identity(struct io_uring_task *tctx, struct io_kiocb *req) |
1e6fa521 | 1361 | { |
500a373d | 1362 | if (req->work.identity == &tctx->__identity) |
1e6fa521 JA |
1363 | return; |
1364 | if (refcount_dec_and_test(&req->work.identity->count)) | |
1365 | kfree(req->work.identity); | |
1366 | } | |
1367 | ||
4edf20f9 | 1368 | static void io_req_clean_work(struct io_kiocb *req) |
18d9be1a | 1369 | { |
7cdaf587 | 1370 | if (!(req->flags & REQ_F_WORK_INITIALIZED)) |
4edf20f9 | 1371 | return; |
51a4cc11 JA |
1372 | |
1373 | req->flags &= ~REQ_F_WORK_INITIALIZED; | |
7cdaf587 | 1374 | |
dfead8a8 | 1375 | if (req->work.flags & IO_WQ_WORK_MM) { |
98447d65 | 1376 | mmdrop(req->work.identity->mm); |
dfead8a8 | 1377 | req->work.flags &= ~IO_WQ_WORK_MM; |
cccf0ee8 | 1378 | } |
91d8f519 | 1379 | #ifdef CONFIG_BLK_CGROUP |
dfead8a8 | 1380 | if (req->work.flags & IO_WQ_WORK_BLKCG) { |
98447d65 | 1381 | css_put(req->work.identity->blkcg_css); |
dfead8a8 JA |
1382 | req->work.flags &= ~IO_WQ_WORK_BLKCG; |
1383 | } | |
91d8f519 | 1384 | #endif |
dfead8a8 | 1385 | if (req->work.flags & IO_WQ_WORK_CREDS) { |
98447d65 | 1386 | put_cred(req->work.identity->creds); |
dfead8a8 | 1387 | req->work.flags &= ~IO_WQ_WORK_CREDS; |
cccf0ee8 | 1388 | } |
dfead8a8 | 1389 | if (req->work.flags & IO_WQ_WORK_FS) { |
98447d65 | 1390 | struct fs_struct *fs = req->work.identity->fs; |
51a4cc11 | 1391 | |
98447d65 | 1392 | spin_lock(&req->work.identity->fs->lock); |
ff002b30 JA |
1393 | if (--fs->users) |
1394 | fs = NULL; | |
98447d65 | 1395 | spin_unlock(&req->work.identity->fs->lock); |
ff002b30 JA |
1396 | if (fs) |
1397 | free_fs_struct(fs); | |
dfead8a8 | 1398 | req->work.flags &= ~IO_WQ_WORK_FS; |
ff002b30 | 1399 | } |
9d5c8190 PB |
1400 | if (req->flags & REQ_F_INFLIGHT) |
1401 | io_req_drop_files(req); | |
51a4cc11 | 1402 | |
5c3462cf | 1403 | io_put_identity(req->task->io_uring, req); |
561fb04a JA |
1404 | } |
1405 | ||
1e6fa521 JA |
1406 | /* |
1407 | * Create a private copy of io_identity, since some fields don't match | |
1408 | * the current context. | |
1409 | */ | |
1410 | static bool io_identity_cow(struct io_kiocb *req) | |
1411 | { | |
5c3462cf | 1412 | struct io_uring_task *tctx = current->io_uring; |
1e6fa521 JA |
1413 | const struct cred *creds = NULL; |
1414 | struct io_identity *id; | |
1415 | ||
1416 | if (req->work.flags & IO_WQ_WORK_CREDS) | |
1417 | creds = req->work.identity->creds; | |
1418 | ||
1419 | id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL); | |
1420 | if (unlikely(!id)) { | |
1421 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
1422 | return false; | |
1423 | } | |
1424 | ||
1425 | /* | |
1426 | * We can safely just re-init the creds we copied Either the field | |
1427 | * matches the current one, or we haven't grabbed it yet. The only | |
1428 | * exception is ->creds, through registered personalities, so handle | |
1429 | * that one separately. | |
1430 | */ | |
1431 | io_init_identity(id); | |
1432 | if (creds) | |
e8c954df | 1433 | id->creds = creds; |
1e6fa521 JA |
1434 | |
1435 | /* add one for this request */ | |
1436 | refcount_inc(&id->count); | |
1437 | ||
cb8a8ae3 JA |
1438 | /* drop tctx and req identity references, if needed */ |
1439 | if (tctx->identity != &tctx->__identity && | |
1440 | refcount_dec_and_test(&tctx->identity->count)) | |
1441 | kfree(tctx->identity); | |
1442 | if (req->work.identity != &tctx->__identity && | |
1443 | refcount_dec_and_test(&req->work.identity->count)) | |
1e6fa521 JA |
1444 | kfree(req->work.identity); |
1445 | ||
1446 | req->work.identity = id; | |
500a373d | 1447 | tctx->identity = id; |
1e6fa521 JA |
1448 | return true; |
1449 | } | |
1450 | ||
1451 | static bool io_grab_identity(struct io_kiocb *req) | |
18d9be1a | 1452 | { |
d3656344 | 1453 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
5c3462cf | 1454 | struct io_identity *id = req->work.identity; |
23329513 | 1455 | struct io_ring_ctx *ctx = req->ctx; |
54a91f3b | 1456 | |
69228338 JA |
1457 | if (def->work_flags & IO_WQ_WORK_FSIZE) { |
1458 | if (id->fsize != rlimit(RLIMIT_FSIZE)) | |
1459 | return false; | |
1460 | req->work.flags |= IO_WQ_WORK_FSIZE; | |
1461 | } | |
91d8f519 | 1462 | #ifdef CONFIG_BLK_CGROUP |
dfead8a8 JA |
1463 | if (!(req->work.flags & IO_WQ_WORK_BLKCG) && |
1464 | (def->work_flags & IO_WQ_WORK_BLKCG)) { | |
91d8f519 | 1465 | rcu_read_lock(); |
1e6fa521 JA |
1466 | if (id->blkcg_css != blkcg_css()) { |
1467 | rcu_read_unlock(); | |
1468 | return false; | |
1469 | } | |
91d8f519 DZ |
1470 | /* |
1471 | * This should be rare, either the cgroup is dying or the task | |
1472 | * is moving cgroups. Just punt to root for the handful of ios. | |
1473 | */ | |
1e6fa521 | 1474 | if (css_tryget_online(id->blkcg_css)) |
dfead8a8 | 1475 | req->work.flags |= IO_WQ_WORK_BLKCG; |
91d8f519 DZ |
1476 | rcu_read_unlock(); |
1477 | } | |
1478 | #endif | |
dfead8a8 | 1479 | if (!(req->work.flags & IO_WQ_WORK_CREDS)) { |
1e6fa521 JA |
1480 | if (id->creds != current_cred()) |
1481 | return false; | |
1482 | get_cred(id->creds); | |
dfead8a8 JA |
1483 | req->work.flags |= IO_WQ_WORK_CREDS; |
1484 | } | |
4ea33a97 JA |
1485 | #ifdef CONFIG_AUDIT |
1486 | if (!uid_eq(current->loginuid, id->loginuid) || | |
1487 | current->sessionid != id->sessionid) | |
1488 | return false; | |
1489 | #endif | |
dfead8a8 JA |
1490 | if (!(req->work.flags & IO_WQ_WORK_FS) && |
1491 | (def->work_flags & IO_WQ_WORK_FS)) { | |
1e6fa521 JA |
1492 | if (current->fs != id->fs) |
1493 | return false; | |
1494 | spin_lock(&id->fs->lock); | |
1495 | if (!id->fs->in_exec) { | |
1496 | id->fs->users++; | |
dfead8a8 | 1497 | req->work.flags |= IO_WQ_WORK_FS; |
dca9cf8b PB |
1498 | } else { |
1499 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
1500 | } | |
1501 | spin_unlock(¤t->fs->lock); | |
1502 | } | |
af604703 PB |
1503 | if (!(req->work.flags & IO_WQ_WORK_FILES) && |
1504 | (def->work_flags & IO_WQ_WORK_FILES) && | |
1505 | !(req->flags & REQ_F_NO_FILE_TABLE)) { | |
1506 | if (id->files != current->files || | |
1507 | id->nsproxy != current->nsproxy) | |
1508 | return false; | |
1509 | atomic_inc(&id->files->count); | |
1510 | get_nsproxy(id->nsproxy); | |
af604703 | 1511 | |
02a13674 JA |
1512 | if (!(req->flags & REQ_F_INFLIGHT)) { |
1513 | req->flags |= REQ_F_INFLIGHT; | |
1514 | ||
1515 | spin_lock_irq(&ctx->inflight_lock); | |
1516 | list_add(&req->inflight_entry, &ctx->inflight_list); | |
1517 | spin_unlock_irq(&ctx->inflight_lock); | |
1518 | } | |
af604703 PB |
1519 | req->work.flags |= IO_WQ_WORK_FILES; |
1520 | } | |
77788775 JA |
1521 | if (!(req->work.flags & IO_WQ_WORK_MM) && |
1522 | (def->work_flags & IO_WQ_WORK_MM)) { | |
1523 | if (id->mm != current->mm) | |
1524 | return false; | |
1525 | mmgrab(id->mm); | |
1526 | req->work.flags |= IO_WQ_WORK_MM; | |
1527 | } | |
1e6fa521 JA |
1528 | |
1529 | return true; | |
1530 | } | |
1531 | ||
1532 | static void io_prep_async_work(struct io_kiocb *req) | |
1533 | { | |
1534 | const struct io_op_def *def = &io_op_defs[req->opcode]; | |
1e6fa521 JA |
1535 | struct io_ring_ctx *ctx = req->ctx; |
1536 | ||
1537 | io_req_init_async(req); | |
1538 | ||
feaadc4f PB |
1539 | if (req->flags & REQ_F_FORCE_ASYNC) |
1540 | req->work.flags |= IO_WQ_WORK_CONCURRENT; | |
1541 | ||
1e6fa521 JA |
1542 | if (req->flags & REQ_F_ISREG) { |
1543 | if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL)) | |
1544 | io_wq_hash_work(&req->work, file_inode(req->file)); | |
1545 | } else { | |
1546 | if (def->unbound_nonreg_file) | |
1547 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
1548 | } | |
1549 | ||
1e6fa521 JA |
1550 | /* if we fail grabbing identity, we must COW, regrab, and retry */ |
1551 | if (io_grab_identity(req)) | |
1552 | return; | |
1553 | ||
1554 | if (!io_identity_cow(req)) | |
1555 | return; | |
1556 | ||
1557 | /* can't fail at this point */ | |
1558 | if (!io_grab_identity(req)) | |
1559 | WARN_ON(1); | |
561fb04a | 1560 | } |
cccf0ee8 | 1561 | |
cbdcb435 | 1562 | static void io_prep_async_link(struct io_kiocb *req) |
561fb04a | 1563 | { |
cbdcb435 | 1564 | struct io_kiocb *cur; |
54a91f3b | 1565 | |
f2f87370 PB |
1566 | io_for_each_link(cur, req) |
1567 | io_prep_async_work(cur); | |
561fb04a JA |
1568 | } |
1569 | ||
7271ef3a | 1570 | static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req) |
561fb04a | 1571 | { |
a197f664 | 1572 | struct io_ring_ctx *ctx = req->ctx; |
cbdcb435 | 1573 | struct io_kiocb *link = io_prep_linked_timeout(req); |
561fb04a | 1574 | |
8766dd51 PB |
1575 | trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req, |
1576 | &req->work, req->flags); | |
1577 | io_wq_enqueue(ctx->io_wq, &req->work); | |
7271ef3a | 1578 | return link; |
18d9be1a JA |
1579 | } |
1580 | ||
cbdcb435 PB |
1581 | static void io_queue_async_work(struct io_kiocb *req) |
1582 | { | |
7271ef3a JA |
1583 | struct io_kiocb *link; |
1584 | ||
cbdcb435 PB |
1585 | /* init ->work of the whole link before punting */ |
1586 | io_prep_async_link(req); | |
7271ef3a JA |
1587 | link = __io_queue_async_work(req); |
1588 | ||
1589 | if (link) | |
1590 | io_queue_linked_timeout(link); | |
cbdcb435 PB |
1591 | } |
1592 | ||
5262f567 JA |
1593 | static void io_kill_timeout(struct io_kiocb *req) |
1594 | { | |
e8c2bc1f | 1595 | struct io_timeout_data *io = req->async_data; |
5262f567 JA |
1596 | int ret; |
1597 | ||
e8c2bc1f | 1598 | ret = hrtimer_try_to_cancel(&io->timer); |
5262f567 | 1599 | if (ret != -1) { |
01cec8c1 PB |
1600 | atomic_set(&req->ctx->cq_timeouts, |
1601 | atomic_read(&req->ctx->cq_timeouts) + 1); | |
135fcde8 | 1602 | list_del_init(&req->timeout.list); |
78e19bbe | 1603 | io_cqring_fill_event(req, 0); |
216578e5 | 1604 | io_put_req_deferred(req, 1); |
5262f567 JA |
1605 | } |
1606 | } | |
1607 | ||
76e1b642 JA |
1608 | /* |
1609 | * Returns true if we found and killed one or more timeouts | |
1610 | */ | |
6b81928d PB |
1611 | static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk, |
1612 | struct files_struct *files) | |
5262f567 JA |
1613 | { |
1614 | struct io_kiocb *req, *tmp; | |
76e1b642 | 1615 | int canceled = 0; |
5262f567 JA |
1616 | |
1617 | spin_lock_irq(&ctx->completion_lock); | |
f3606e3a | 1618 | list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) { |
6b81928d | 1619 | if (io_match_task(req, tsk, files)) { |
f3606e3a | 1620 | io_kill_timeout(req); |
76e1b642 JA |
1621 | canceled++; |
1622 | } | |
f3606e3a | 1623 | } |
5262f567 | 1624 | spin_unlock_irq(&ctx->completion_lock); |
76e1b642 | 1625 | return canceled != 0; |
5262f567 JA |
1626 | } |
1627 | ||
04518945 | 1628 | static void __io_queue_deferred(struct io_ring_ctx *ctx) |
de0617e4 | 1629 | { |
04518945 | 1630 | do { |
27dc8338 PB |
1631 | struct io_defer_entry *de = list_first_entry(&ctx->defer_list, |
1632 | struct io_defer_entry, list); | |
7271ef3a | 1633 | struct io_kiocb *link; |
de0617e4 | 1634 | |
9cf7c104 | 1635 | if (req_need_defer(de->req, de->seq)) |
04518945 | 1636 | break; |
27dc8338 | 1637 | list_del_init(&de->list); |
cbdcb435 | 1638 | /* punt-init is done before queueing for defer */ |
7271ef3a JA |
1639 | link = __io_queue_async_work(de->req); |
1640 | if (link) { | |
1641 | __io_queue_linked_timeout(link); | |
1642 | /* drop submission reference */ | |
216578e5 | 1643 | io_put_req_deferred(link, 1); |
7271ef3a | 1644 | } |
27dc8338 | 1645 | kfree(de); |
04518945 PB |
1646 | } while (!list_empty(&ctx->defer_list)); |
1647 | } | |
1648 | ||
360428f8 | 1649 | static void io_flush_timeouts(struct io_ring_ctx *ctx) |
de0617e4 | 1650 | { |
f010505b MDG |
1651 | u32 seq; |
1652 | ||
1653 | if (list_empty(&ctx->timeout_list)) | |
1654 | return; | |
1655 | ||
1656 | seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts); | |
1657 | ||
1658 | do { | |
1659 | u32 events_needed, events_got; | |
360428f8 | 1660 | struct io_kiocb *req = list_first_entry(&ctx->timeout_list, |
135fcde8 | 1661 | struct io_kiocb, timeout.list); |
de0617e4 | 1662 | |
8eb7e2d0 | 1663 | if (io_is_timeout_noseq(req)) |
360428f8 | 1664 | break; |
f010505b MDG |
1665 | |
1666 | /* | |
1667 | * Since seq can easily wrap around over time, subtract | |
1668 | * the last seq at which timeouts were flushed before comparing. | |
1669 | * Assuming not more than 2^31-1 events have happened since, | |
1670 | * these subtractions won't have wrapped, so we can check if | |
1671 | * target is in [last_seq, current_seq] by comparing the two. | |
1672 | */ | |
1673 | events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush; | |
1674 | events_got = seq - ctx->cq_last_tm_flush; | |
1675 | if (events_got < events_needed) | |
360428f8 | 1676 | break; |
bfe68a22 | 1677 | |
135fcde8 | 1678 | list_del_init(&req->timeout.list); |
5262f567 | 1679 | io_kill_timeout(req); |
f010505b MDG |
1680 | } while (!list_empty(&ctx->timeout_list)); |
1681 | ||
1682 | ctx->cq_last_tm_flush = seq; | |
360428f8 | 1683 | } |
5262f567 | 1684 | |
360428f8 PB |
1685 | static void io_commit_cqring(struct io_ring_ctx *ctx) |
1686 | { | |
1687 | io_flush_timeouts(ctx); | |
de0617e4 JA |
1688 | __io_commit_cqring(ctx); |
1689 | ||
04518945 PB |
1690 | if (unlikely(!list_empty(&ctx->defer_list))) |
1691 | __io_queue_deferred(ctx); | |
de0617e4 JA |
1692 | } |
1693 | ||
90554200 JA |
1694 | static inline bool io_sqring_full(struct io_ring_ctx *ctx) |
1695 | { | |
1696 | struct io_rings *r = ctx->rings; | |
1697 | ||
1698 | return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries; | |
1699 | } | |
1700 | ||
2b188cc1 JA |
1701 | static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx) |
1702 | { | |
75b28aff | 1703 | struct io_rings *rings = ctx->rings; |
2b188cc1 JA |
1704 | unsigned tail; |
1705 | ||
1706 | tail = ctx->cached_cq_tail; | |
115e12e5 SB |
1707 | /* |
1708 | * writes to the cq entry need to come after reading head; the | |
1709 | * control dependency is enough as we're using WRITE_ONCE to | |
1710 | * fill the cq entry | |
1711 | */ | |
75b28aff | 1712 | if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries) |
2b188cc1 JA |
1713 | return NULL; |
1714 | ||
1715 | ctx->cached_cq_tail++; | |
75b28aff | 1716 | return &rings->cqes[tail & ctx->cq_mask]; |
2b188cc1 JA |
1717 | } |
1718 | ||
f2842ab5 JA |
1719 | static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx) |
1720 | { | |
f0b493e6 JA |
1721 | if (!ctx->cq_ev_fd) |
1722 | return false; | |
7e55a19c SG |
1723 | if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) |
1724 | return false; | |
f2842ab5 JA |
1725 | if (!ctx->eventfd_async) |
1726 | return true; | |
b41e9852 | 1727 | return io_wq_current_is_worker(); |
f2842ab5 JA |
1728 | } |
1729 | ||
e23de15f PB |
1730 | static inline unsigned __io_cqring_events(struct io_ring_ctx *ctx) |
1731 | { | |
1732 | return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head); | |
1733 | } | |
1734 | ||
b41e9852 | 1735 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
1d7bb1d5 | 1736 | { |
b1445e59 PB |
1737 | /* see waitqueue_active() comment */ |
1738 | smp_mb(); | |
1739 | ||
1d7bb1d5 JA |
1740 | if (waitqueue_active(&ctx->wait)) |
1741 | wake_up(&ctx->wait); | |
534ca6d6 JA |
1742 | if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait)) |
1743 | wake_up(&ctx->sq_data->wait); | |
b41e9852 | 1744 | if (io_should_trigger_evfd(ctx)) |
1d7bb1d5 | 1745 | eventfd_signal(ctx->cq_ev_fd, 1); |
b1445e59 | 1746 | if (waitqueue_active(&ctx->cq_wait)) { |
4aa84f2f PB |
1747 | wake_up_interruptible(&ctx->cq_wait); |
1748 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
1749 | } | |
1d7bb1d5 JA |
1750 | } |
1751 | ||
80c18e4a PB |
1752 | static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx) |
1753 | { | |
b1445e59 PB |
1754 | /* see waitqueue_active() comment */ |
1755 | smp_mb(); | |
1756 | ||
80c18e4a PB |
1757 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
1758 | if (waitqueue_active(&ctx->wait)) | |
1759 | wake_up(&ctx->wait); | |
1760 | } | |
1761 | if (io_should_trigger_evfd(ctx)) | |
1762 | eventfd_signal(ctx->cq_ev_fd, 1); | |
b1445e59 | 1763 | if (waitqueue_active(&ctx->cq_wait)) { |
4aa84f2f PB |
1764 | wake_up_interruptible(&ctx->cq_wait); |
1765 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
1766 | } | |
80c18e4a PB |
1767 | } |
1768 | ||
c4a2ed72 | 1769 | /* Returns true if there are no backlogged entries after the flush */ |
6c503150 PB |
1770 | static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force, |
1771 | struct task_struct *tsk, | |
1772 | struct files_struct *files) | |
1d7bb1d5 JA |
1773 | { |
1774 | struct io_rings *rings = ctx->rings; | |
e6c8aa9a | 1775 | struct io_kiocb *req, *tmp; |
1d7bb1d5 | 1776 | struct io_uring_cqe *cqe; |
1d7bb1d5 | 1777 | unsigned long flags; |
09e88404 | 1778 | bool all_flushed; |
1d7bb1d5 JA |
1779 | LIST_HEAD(list); |
1780 | ||
e23de15f PB |
1781 | if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries) |
1782 | return false; | |
1d7bb1d5 JA |
1783 | |
1784 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
e6c8aa9a | 1785 | list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) { |
08d23634 | 1786 | if (!io_match_task(req, tsk, files)) |
e6c8aa9a JA |
1787 | continue; |
1788 | ||
1d7bb1d5 JA |
1789 | cqe = io_get_cqring(ctx); |
1790 | if (!cqe && !force) | |
1791 | break; | |
1792 | ||
40d8ddd4 | 1793 | list_move(&req->compl.list, &list); |
1d7bb1d5 JA |
1794 | if (cqe) { |
1795 | WRITE_ONCE(cqe->user_data, req->user_data); | |
1796 | WRITE_ONCE(cqe->res, req->result); | |
0f7e466b | 1797 | WRITE_ONCE(cqe->flags, req->compl.cflags); |
1d7bb1d5 | 1798 | } else { |
2c3bac6d | 1799 | ctx->cached_cq_overflow++; |
1d7bb1d5 | 1800 | WRITE_ONCE(ctx->rings->cq_overflow, |
2c3bac6d | 1801 | ctx->cached_cq_overflow); |
1d7bb1d5 JA |
1802 | } |
1803 | } | |
1804 | ||
09e88404 PB |
1805 | all_flushed = list_empty(&ctx->cq_overflow_list); |
1806 | if (all_flushed) { | |
1807 | clear_bit(0, &ctx->sq_check_overflow); | |
1808 | clear_bit(0, &ctx->cq_check_overflow); | |
1809 | ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW; | |
1810 | } | |
46930143 | 1811 | |
09e88404 | 1812 | io_commit_cqring(ctx); |
1d7bb1d5 JA |
1813 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
1814 | io_cqring_ev_posted(ctx); | |
1815 | ||
1816 | while (!list_empty(&list)) { | |
40d8ddd4 PB |
1817 | req = list_first_entry(&list, struct io_kiocb, compl.list); |
1818 | list_del(&req->compl.list); | |
ec9c02ad | 1819 | io_put_req(req); |
1d7bb1d5 | 1820 | } |
c4a2ed72 | 1821 | |
09e88404 | 1822 | return all_flushed; |
1d7bb1d5 JA |
1823 | } |
1824 | ||
6c503150 PB |
1825 | static void io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force, |
1826 | struct task_struct *tsk, | |
1827 | struct files_struct *files) | |
1828 | { | |
1829 | if (test_bit(0, &ctx->cq_check_overflow)) { | |
1830 | /* iopoll syncs against uring_lock, not completion_lock */ | |
1831 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
1832 | mutex_lock(&ctx->uring_lock); | |
1833 | __io_cqring_overflow_flush(ctx, force, tsk, files); | |
1834 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
1835 | mutex_unlock(&ctx->uring_lock); | |
1836 | } | |
1837 | } | |
1838 | ||
bcda7baa | 1839 | static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags) |
2b188cc1 | 1840 | { |
78e19bbe | 1841 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
1842 | struct io_uring_cqe *cqe; |
1843 | ||
78e19bbe | 1844 | trace_io_uring_complete(ctx, req->user_data, res); |
51c3ff62 | 1845 | |
2b188cc1 JA |
1846 | /* |
1847 | * If we can't get a cq entry, userspace overflowed the | |
1848 | * submission (by quite a lot). Increment the overflow count in | |
1849 | * the ring. | |
1850 | */ | |
1851 | cqe = io_get_cqring(ctx); | |
1d7bb1d5 | 1852 | if (likely(cqe)) { |
78e19bbe | 1853 | WRITE_ONCE(cqe->user_data, req->user_data); |
2b188cc1 | 1854 | WRITE_ONCE(cqe->res, res); |
bcda7baa | 1855 | WRITE_ONCE(cqe->flags, cflags); |
fdaf083c JA |
1856 | } else if (ctx->cq_overflow_flushed || |
1857 | atomic_read(&req->task->io_uring->in_idle)) { | |
0f212204 JA |
1858 | /* |
1859 | * If we're in ring overflow flush mode, or in task cancel mode, | |
1860 | * then we cannot store the request for later flushing, we need | |
1861 | * to drop it on the floor. | |
1862 | */ | |
2c3bac6d PB |
1863 | ctx->cached_cq_overflow++; |
1864 | WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow); | |
1d7bb1d5 | 1865 | } else { |
ad3eb2c8 JA |
1866 | if (list_empty(&ctx->cq_overflow_list)) { |
1867 | set_bit(0, &ctx->sq_check_overflow); | |
1868 | set_bit(0, &ctx->cq_check_overflow); | |
6d5f9049 | 1869 | ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW; |
ad3eb2c8 | 1870 | } |
40d8ddd4 | 1871 | io_clean_op(req); |
1d7bb1d5 | 1872 | req->result = res; |
0f7e466b | 1873 | req->compl.cflags = cflags; |
40d8ddd4 PB |
1874 | refcount_inc(&req->refs); |
1875 | list_add_tail(&req->compl.list, &ctx->cq_overflow_list); | |
2b188cc1 JA |
1876 | } |
1877 | } | |
1878 | ||
bcda7baa JA |
1879 | static void io_cqring_fill_event(struct io_kiocb *req, long res) |
1880 | { | |
1881 | __io_cqring_fill_event(req, res, 0); | |
1882 | } | |
1883 | ||
e1e16097 | 1884 | static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags) |
2b188cc1 | 1885 | { |
78e19bbe | 1886 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
1887 | unsigned long flags; |
1888 | ||
1889 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
bcda7baa | 1890 | __io_cqring_fill_event(req, res, cflags); |
2b188cc1 JA |
1891 | io_commit_cqring(ctx); |
1892 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1893 | ||
8c838788 | 1894 | io_cqring_ev_posted(ctx); |
2b188cc1 JA |
1895 | } |
1896 | ||
229a7b63 | 1897 | static void io_submit_flush_completions(struct io_comp_state *cs) |
bcda7baa | 1898 | { |
229a7b63 JA |
1899 | struct io_ring_ctx *ctx = cs->ctx; |
1900 | ||
1901 | spin_lock_irq(&ctx->completion_lock); | |
1902 | while (!list_empty(&cs->list)) { | |
1903 | struct io_kiocb *req; | |
1904 | ||
3ca405eb PB |
1905 | req = list_first_entry(&cs->list, struct io_kiocb, compl.list); |
1906 | list_del(&req->compl.list); | |
0f7e466b | 1907 | __io_cqring_fill_event(req, req->result, req->compl.cflags); |
216578e5 PB |
1908 | |
1909 | /* | |
1910 | * io_free_req() doesn't care about completion_lock unless one | |
1911 | * of these flags is set. REQ_F_WORK_INITIALIZED is in the list | |
1912 | * because of a potential deadlock with req->work.fs->lock | |
1913 | */ | |
1914 | if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT | |
1915 | |REQ_F_WORK_INITIALIZED)) { | |
229a7b63 JA |
1916 | spin_unlock_irq(&ctx->completion_lock); |
1917 | io_put_req(req); | |
1918 | spin_lock_irq(&ctx->completion_lock); | |
216578e5 PB |
1919 | } else { |
1920 | io_put_req(req); | |
229a7b63 JA |
1921 | } |
1922 | } | |
1923 | io_commit_cqring(ctx); | |
1924 | spin_unlock_irq(&ctx->completion_lock); | |
1925 | ||
1926 | io_cqring_ev_posted(ctx); | |
1927 | cs->nr = 0; | |
1928 | } | |
1929 | ||
1930 | static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags, | |
1931 | struct io_comp_state *cs) | |
1932 | { | |
1933 | if (!cs) { | |
1934 | io_cqring_add_event(req, res, cflags); | |
1935 | io_put_req(req); | |
1936 | } else { | |
3ca405eb | 1937 | io_clean_op(req); |
229a7b63 | 1938 | req->result = res; |
0f7e466b | 1939 | req->compl.cflags = cflags; |
3ca405eb | 1940 | list_add_tail(&req->compl.list, &cs->list); |
229a7b63 JA |
1941 | if (++cs->nr >= 32) |
1942 | io_submit_flush_completions(cs); | |
1943 | } | |
e1e16097 JA |
1944 | } |
1945 | ||
1946 | static void io_req_complete(struct io_kiocb *req, long res) | |
bcda7baa | 1947 | { |
229a7b63 | 1948 | __io_req_complete(req, res, 0, NULL); |
bcda7baa JA |
1949 | } |
1950 | ||
0ddf92e8 JA |
1951 | static inline bool io_is_fallback_req(struct io_kiocb *req) |
1952 | { | |
1953 | return req == (struct io_kiocb *) | |
1954 | ((unsigned long) req->ctx->fallback_req & ~1UL); | |
1955 | } | |
1956 | ||
1957 | static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx) | |
1958 | { | |
1959 | struct io_kiocb *req; | |
1960 | ||
1961 | req = ctx->fallback_req; | |
dd461af6 | 1962 | if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req)) |
0ddf92e8 JA |
1963 | return req; |
1964 | ||
1965 | return NULL; | |
1966 | } | |
1967 | ||
0553b8bd PB |
1968 | static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx, |
1969 | struct io_submit_state *state) | |
2b188cc1 | 1970 | { |
f6b6c7d6 | 1971 | if (!state->free_reqs) { |
291b2821 | 1972 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
2579f913 JA |
1973 | size_t sz; |
1974 | int ret; | |
1975 | ||
1976 | sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs)); | |
fd6fab2c JA |
1977 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs); |
1978 | ||
1979 | /* | |
1980 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
1981 | * retry single alloc to be on the safe side. | |
1982 | */ | |
1983 | if (unlikely(ret <= 0)) { | |
1984 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
1985 | if (!state->reqs[0]) | |
0ddf92e8 | 1986 | goto fallback; |
fd6fab2c JA |
1987 | ret = 1; |
1988 | } | |
291b2821 | 1989 | state->free_reqs = ret; |
2b188cc1 JA |
1990 | } |
1991 | ||
291b2821 PB |
1992 | state->free_reqs--; |
1993 | return state->reqs[state->free_reqs]; | |
0ddf92e8 | 1994 | fallback: |
0553b8bd | 1995 | return io_get_fallback_req(ctx); |
2b188cc1 JA |
1996 | } |
1997 | ||
8da11c19 PB |
1998 | static inline void io_put_file(struct io_kiocb *req, struct file *file, |
1999 | bool fixed) | |
2000 | { | |
36f72fe2 | 2001 | if (!fixed) |
8da11c19 PB |
2002 | fput(file); |
2003 | } | |
2004 | ||
4edf20f9 | 2005 | static void io_dismantle_req(struct io_kiocb *req) |
2b188cc1 | 2006 | { |
3ca405eb | 2007 | io_clean_op(req); |
929a3af9 | 2008 | |
e8c2bc1f JA |
2009 | if (req->async_data) |
2010 | kfree(req->async_data); | |
8da11c19 PB |
2011 | if (req->file) |
2012 | io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE)); | |
36f72fe2 PB |
2013 | if (req->fixed_file_refs) |
2014 | percpu_ref_put(req->fixed_file_refs); | |
4edf20f9 | 2015 | io_req_clean_work(req); |
e65ef56d JA |
2016 | } |
2017 | ||
216578e5 | 2018 | static void __io_free_req(struct io_kiocb *req) |
c6ca97b3 | 2019 | { |
0f212204 | 2020 | struct io_uring_task *tctx = req->task->io_uring; |
51a4cc11 | 2021 | struct io_ring_ctx *ctx = req->ctx; |
c6ca97b3 | 2022 | |
216578e5 | 2023 | io_dismantle_req(req); |
c6ca97b3 | 2024 | |
d8a6df10 | 2025 | percpu_counter_dec(&tctx->inflight); |
fdaf083c | 2026 | if (atomic_read(&tctx->in_idle)) |
0f212204 | 2027 | wake_up(&tctx->wait); |
e3bc8e9d JA |
2028 | put_task_struct(req->task); |
2029 | ||
b1e50e54 PB |
2030 | if (likely(!io_is_fallback_req(req))) |
2031 | kmem_cache_free(req_cachep, req); | |
2032 | else | |
ecfc5177 PB |
2033 | clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req); |
2034 | percpu_ref_put(&ctx->refs); | |
e65ef56d JA |
2035 | } |
2036 | ||
f2f87370 PB |
2037 | static inline void io_remove_next_linked(struct io_kiocb *req) |
2038 | { | |
2039 | struct io_kiocb *nxt = req->link; | |
2040 | ||
2041 | req->link = nxt->link; | |
2042 | nxt->link = NULL; | |
2043 | } | |
2044 | ||
c9abd7ad | 2045 | static void io_kill_linked_timeout(struct io_kiocb *req) |
2665abfd | 2046 | { |
a197f664 | 2047 | struct io_ring_ctx *ctx = req->ctx; |
7c86ffee | 2048 | struct io_kiocb *link; |
c9abd7ad PB |
2049 | bool cancelled = false; |
2050 | unsigned long flags; | |
7c86ffee | 2051 | |
c9abd7ad | 2052 | spin_lock_irqsave(&ctx->completion_lock, flags); |
f2f87370 PB |
2053 | link = req->link; |
2054 | ||
900fad45 PB |
2055 | /* |
2056 | * Can happen if a linked timeout fired and link had been like | |
2057 | * req -> link t-out -> link t-out [-> ...] | |
2058 | */ | |
c9abd7ad PB |
2059 | if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) { |
2060 | struct io_timeout_data *io = link->async_data; | |
2061 | int ret; | |
7c86ffee | 2062 | |
f2f87370 | 2063 | io_remove_next_linked(req); |
90cd7e42 | 2064 | link->timeout.head = NULL; |
c9abd7ad PB |
2065 | ret = hrtimer_try_to_cancel(&io->timer); |
2066 | if (ret != -1) { | |
2067 | io_cqring_fill_event(link, -ECANCELED); | |
2068 | io_commit_cqring(ctx); | |
2069 | cancelled = true; | |
2070 | } | |
2071 | } | |
7c86ffee | 2072 | req->flags &= ~REQ_F_LINK_TIMEOUT; |
216578e5 | 2073 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
ab0b6451 | 2074 | |
c9abd7ad | 2075 | if (cancelled) { |
7c86ffee | 2076 | io_cqring_ev_posted(ctx); |
c9abd7ad PB |
2077 | io_put_req(link); |
2078 | } | |
7c86ffee PB |
2079 | } |
2080 | ||
9e645e11 | 2081 | |
d148ca4b | 2082 | static void io_fail_links(struct io_kiocb *req) |
9e645e11 | 2083 | { |
f2f87370 | 2084 | struct io_kiocb *link, *nxt; |
2665abfd | 2085 | struct io_ring_ctx *ctx = req->ctx; |
d148ca4b | 2086 | unsigned long flags; |
9e645e11 | 2087 | |
d148ca4b | 2088 | spin_lock_irqsave(&ctx->completion_lock, flags); |
f2f87370 PB |
2089 | link = req->link; |
2090 | req->link = NULL; | |
9e645e11 | 2091 | |
f2f87370 PB |
2092 | while (link) { |
2093 | nxt = link->link; | |
2094 | link->link = NULL; | |
2665abfd | 2095 | |
f2f87370 | 2096 | trace_io_uring_fail_link(req, link); |
7c86ffee | 2097 | io_cqring_fill_event(link, -ECANCELED); |
216578e5 PB |
2098 | |
2099 | /* | |
2100 | * It's ok to free under spinlock as they're not linked anymore, | |
2101 | * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on | |
2102 | * work.fs->lock. | |
2103 | */ | |
2104 | if (link->flags & REQ_F_WORK_INITIALIZED) | |
2105 | io_put_req_deferred(link, 2); | |
2106 | else | |
2107 | io_double_put_req(link); | |
f2f87370 | 2108 | link = nxt; |
9e645e11 | 2109 | } |
2665abfd | 2110 | io_commit_cqring(ctx); |
216578e5 | 2111 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
9e645e11 | 2112 | |
2665abfd | 2113 | io_cqring_ev_posted(ctx); |
9e645e11 JA |
2114 | } |
2115 | ||
3fa5e0f3 | 2116 | static struct io_kiocb *__io_req_find_next(struct io_kiocb *req) |
c69f8dbe | 2117 | { |
7c86ffee PB |
2118 | if (req->flags & REQ_F_LINK_TIMEOUT) |
2119 | io_kill_linked_timeout(req); | |
944e58bf | 2120 | |
9e645e11 JA |
2121 | /* |
2122 | * If LINK is set, we have dependent requests in this chain. If we | |
2123 | * didn't fail this request, queue the first one up, moving any other | |
2124 | * dependencies to the next request. In case of failure, fail the rest | |
2125 | * of the chain. | |
2126 | */ | |
f2f87370 PB |
2127 | if (likely(!(req->flags & REQ_F_FAIL_LINK))) { |
2128 | struct io_kiocb *nxt = req->link; | |
2129 | ||
2130 | req->link = NULL; | |
2131 | return nxt; | |
2132 | } | |
9b5f7bd9 PB |
2133 | io_fail_links(req); |
2134 | return NULL; | |
4d7dd462 | 2135 | } |
9e645e11 | 2136 | |
f2f87370 | 2137 | static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req) |
3fa5e0f3 | 2138 | { |
f2f87370 | 2139 | if (likely(!(req->link) && !(req->flags & REQ_F_LINK_TIMEOUT))) |
3fa5e0f3 PB |
2140 | return NULL; |
2141 | return __io_req_find_next(req); | |
2142 | } | |
2143 | ||
355fb9e2 | 2144 | static int io_req_task_work_add(struct io_kiocb *req) |
c2c4c83c JA |
2145 | { |
2146 | struct task_struct *tsk = req->task; | |
2147 | struct io_ring_ctx *ctx = req->ctx; | |
91989c70 JA |
2148 | enum task_work_notify_mode notify; |
2149 | int ret; | |
c2c4c83c | 2150 | |
6200b0ae JA |
2151 | if (tsk->flags & PF_EXITING) |
2152 | return -ESRCH; | |
2153 | ||
c2c4c83c | 2154 | /* |
0ba9c9ed JA |
2155 | * SQPOLL kernel thread doesn't need notification, just a wakeup. For |
2156 | * all other cases, use TWA_SIGNAL unconditionally to ensure we're | |
2157 | * processing task_work. There's no reliable way to tell if TWA_RESUME | |
2158 | * will do the job. | |
c2c4c83c | 2159 | */ |
91989c70 | 2160 | notify = TWA_NONE; |
355fb9e2 | 2161 | if (!(ctx->flags & IORING_SETUP_SQPOLL)) |
c2c4c83c JA |
2162 | notify = TWA_SIGNAL; |
2163 | ||
87c4311f | 2164 | ret = task_work_add(tsk, &req->task_work, notify); |
c2c4c83c JA |
2165 | if (!ret) |
2166 | wake_up_process(tsk); | |
0ba9c9ed | 2167 | |
c2c4c83c JA |
2168 | return ret; |
2169 | } | |
2170 | ||
c40f6379 JA |
2171 | static void __io_req_task_cancel(struct io_kiocb *req, int error) |
2172 | { | |
2173 | struct io_ring_ctx *ctx = req->ctx; | |
2174 | ||
2175 | spin_lock_irq(&ctx->completion_lock); | |
2176 | io_cqring_fill_event(req, error); | |
2177 | io_commit_cqring(ctx); | |
2178 | spin_unlock_irq(&ctx->completion_lock); | |
2179 | ||
2180 | io_cqring_ev_posted(ctx); | |
2181 | req_set_fail_links(req); | |
2182 | io_double_put_req(req); | |
2183 | } | |
2184 | ||
2185 | static void io_req_task_cancel(struct callback_head *cb) | |
2186 | { | |
2187 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); | |
87ceb6a6 | 2188 | struct io_ring_ctx *ctx = req->ctx; |
c40f6379 JA |
2189 | |
2190 | __io_req_task_cancel(req, -ECANCELED); | |
87ceb6a6 | 2191 | percpu_ref_put(&ctx->refs); |
c40f6379 JA |
2192 | } |
2193 | ||
2194 | static void __io_req_task_submit(struct io_kiocb *req) | |
2195 | { | |
2196 | struct io_ring_ctx *ctx = req->ctx; | |
2197 | ||
81b6d05c | 2198 | mutex_lock(&ctx->uring_lock); |
d9d05217 PB |
2199 | if (!ctx->sqo_dead && |
2200 | !__io_sq_thread_acquire_mm(ctx) && | |
2201 | !__io_sq_thread_acquire_files(ctx)) | |
c1379e24 | 2202 | __io_queue_sqe(req, NULL); |
81b6d05c | 2203 | else |
c40f6379 | 2204 | __io_req_task_cancel(req, -EFAULT); |
81b6d05c | 2205 | mutex_unlock(&ctx->uring_lock); |
c40f6379 JA |
2206 | } |
2207 | ||
2208 | static void io_req_task_submit(struct callback_head *cb) | |
2209 | { | |
2210 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); | |
6d816e08 | 2211 | struct io_ring_ctx *ctx = req->ctx; |
c40f6379 JA |
2212 | |
2213 | __io_req_task_submit(req); | |
6d816e08 | 2214 | percpu_ref_put(&ctx->refs); |
c40f6379 JA |
2215 | } |
2216 | ||
2217 | static void io_req_task_queue(struct io_kiocb *req) | |
2218 | { | |
c40f6379 JA |
2219 | int ret; |
2220 | ||
2221 | init_task_work(&req->task_work, io_req_task_submit); | |
6d816e08 | 2222 | percpu_ref_get(&req->ctx->refs); |
c40f6379 | 2223 | |
355fb9e2 | 2224 | ret = io_req_task_work_add(req); |
c40f6379 | 2225 | if (unlikely(ret)) { |
c2c4c83c JA |
2226 | struct task_struct *tsk; |
2227 | ||
c40f6379 JA |
2228 | init_task_work(&req->task_work, io_req_task_cancel); |
2229 | tsk = io_wq_get_task(req->ctx->io_wq); | |
91989c70 | 2230 | task_work_add(tsk, &req->task_work, TWA_NONE); |
c2c4c83c | 2231 | wake_up_process(tsk); |
c40f6379 | 2232 | } |
c40f6379 JA |
2233 | } |
2234 | ||
f2f87370 | 2235 | static inline void io_queue_next(struct io_kiocb *req) |
c69f8dbe | 2236 | { |
9b5f7bd9 | 2237 | struct io_kiocb *nxt = io_req_find_next(req); |
944e58bf PB |
2238 | |
2239 | if (nxt) | |
906a8c3f | 2240 | io_req_task_queue(nxt); |
c69f8dbe JL |
2241 | } |
2242 | ||
c3524383 | 2243 | static void io_free_req(struct io_kiocb *req) |
7a743e22 | 2244 | { |
c3524383 PB |
2245 | io_queue_next(req); |
2246 | __io_free_req(req); | |
2247 | } | |
8766dd51 | 2248 | |
2d6500d4 PB |
2249 | struct req_batch { |
2250 | void *reqs[IO_IOPOLL_BATCH]; | |
2251 | int to_free; | |
7a743e22 | 2252 | |
5af1d13e PB |
2253 | struct task_struct *task; |
2254 | int task_refs; | |
2d6500d4 PB |
2255 | }; |
2256 | ||
5af1d13e PB |
2257 | static inline void io_init_req_batch(struct req_batch *rb) |
2258 | { | |
2259 | rb->to_free = 0; | |
2260 | rb->task_refs = 0; | |
2261 | rb->task = NULL; | |
2262 | } | |
2263 | ||
2d6500d4 PB |
2264 | static void __io_req_free_batch_flush(struct io_ring_ctx *ctx, |
2265 | struct req_batch *rb) | |
2266 | { | |
2267 | kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs); | |
2268 | percpu_ref_put_many(&ctx->refs, rb->to_free); | |
2269 | rb->to_free = 0; | |
2270 | } | |
2271 | ||
2272 | static void io_req_free_batch_finish(struct io_ring_ctx *ctx, | |
2273 | struct req_batch *rb) | |
2274 | { | |
2275 | if (rb->to_free) | |
2276 | __io_req_free_batch_flush(ctx, rb); | |
5af1d13e | 2277 | if (rb->task) { |
d8a6df10 JA |
2278 | struct io_uring_task *tctx = rb->task->io_uring; |
2279 | ||
2280 | percpu_counter_sub(&tctx->inflight, rb->task_refs); | |
c93cc9e1 JA |
2281 | if (atomic_read(&tctx->in_idle)) |
2282 | wake_up(&tctx->wait); | |
5af1d13e PB |
2283 | put_task_struct_many(rb->task, rb->task_refs); |
2284 | rb->task = NULL; | |
2285 | } | |
2d6500d4 PB |
2286 | } |
2287 | ||
2288 | static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req) | |
2289 | { | |
2290 | if (unlikely(io_is_fallback_req(req))) { | |
2291 | io_free_req(req); | |
2292 | return; | |
2293 | } | |
f2f87370 | 2294 | io_queue_next(req); |
2d6500d4 | 2295 | |
e3bc8e9d | 2296 | if (req->task != rb->task) { |
0f212204 | 2297 | if (rb->task) { |
d8a6df10 JA |
2298 | struct io_uring_task *tctx = rb->task->io_uring; |
2299 | ||
2300 | percpu_counter_sub(&tctx->inflight, rb->task_refs); | |
c93cc9e1 JA |
2301 | if (atomic_read(&tctx->in_idle)) |
2302 | wake_up(&tctx->wait); | |
e3bc8e9d | 2303 | put_task_struct_many(rb->task, rb->task_refs); |
5af1d13e | 2304 | } |
e3bc8e9d JA |
2305 | rb->task = req->task; |
2306 | rb->task_refs = 0; | |
5af1d13e | 2307 | } |
e3bc8e9d | 2308 | rb->task_refs++; |
5af1d13e | 2309 | |
4edf20f9 | 2310 | io_dismantle_req(req); |
2d6500d4 PB |
2311 | rb->reqs[rb->to_free++] = req; |
2312 | if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs))) | |
2313 | __io_req_free_batch_flush(req->ctx, rb); | |
7a743e22 PB |
2314 | } |
2315 | ||
ba816ad6 JA |
2316 | /* |
2317 | * Drop reference to request, return next in chain (if there is one) if this | |
2318 | * was the last reference to this request. | |
2319 | */ | |
9b5f7bd9 | 2320 | static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req) |
e65ef56d | 2321 | { |
9b5f7bd9 PB |
2322 | struct io_kiocb *nxt = NULL; |
2323 | ||
2a44f467 | 2324 | if (refcount_dec_and_test(&req->refs)) { |
9b5f7bd9 | 2325 | nxt = io_req_find_next(req); |
4d7dd462 | 2326 | __io_free_req(req); |
2a44f467 | 2327 | } |
9b5f7bd9 | 2328 | return nxt; |
2b188cc1 JA |
2329 | } |
2330 | ||
e65ef56d JA |
2331 | static void io_put_req(struct io_kiocb *req) |
2332 | { | |
2333 | if (refcount_dec_and_test(&req->refs)) | |
2334 | io_free_req(req); | |
2b188cc1 JA |
2335 | } |
2336 | ||
216578e5 PB |
2337 | static void io_put_req_deferred_cb(struct callback_head *cb) |
2338 | { | |
2339 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); | |
2340 | ||
2341 | io_free_req(req); | |
2342 | } | |
2343 | ||
2344 | static void io_free_req_deferred(struct io_kiocb *req) | |
2345 | { | |
2346 | int ret; | |
2347 | ||
2348 | init_task_work(&req->task_work, io_put_req_deferred_cb); | |
355fb9e2 | 2349 | ret = io_req_task_work_add(req); |
216578e5 PB |
2350 | if (unlikely(ret)) { |
2351 | struct task_struct *tsk; | |
2352 | ||
2353 | tsk = io_wq_get_task(req->ctx->io_wq); | |
91989c70 | 2354 | task_work_add(tsk, &req->task_work, TWA_NONE); |
216578e5 PB |
2355 | wake_up_process(tsk); |
2356 | } | |
2357 | } | |
2358 | ||
2359 | static inline void io_put_req_deferred(struct io_kiocb *req, int refs) | |
2360 | { | |
2361 | if (refcount_sub_and_test(refs, &req->refs)) | |
2362 | io_free_req_deferred(req); | |
2363 | } | |
2364 | ||
f4db7182 | 2365 | static struct io_wq_work *io_steal_work(struct io_kiocb *req) |
7a743e22 | 2366 | { |
6df1db6b | 2367 | struct io_kiocb *nxt; |
f4db7182 | 2368 | |
7a743e22 | 2369 | /* |
f4db7182 PB |
2370 | * A ref is owned by io-wq in which context we're. So, if that's the |
2371 | * last one, it's safe to steal next work. False negatives are Ok, | |
2372 | * it just will be re-punted async in io_put_work() | |
7a743e22 | 2373 | */ |
f4db7182 PB |
2374 | if (refcount_read(&req->refs) != 1) |
2375 | return NULL; | |
7a743e22 | 2376 | |
9b5f7bd9 | 2377 | nxt = io_req_find_next(req); |
6df1db6b | 2378 | return nxt ? &nxt->work : NULL; |
7a743e22 PB |
2379 | } |
2380 | ||
978db57e JA |
2381 | static void io_double_put_req(struct io_kiocb *req) |
2382 | { | |
2383 | /* drop both submit and complete references */ | |
2384 | if (refcount_sub_and_test(2, &req->refs)) | |
2385 | io_free_req(req); | |
2386 | } | |
2387 | ||
6c503150 | 2388 | static unsigned io_cqring_events(struct io_ring_ctx *ctx) |
a3a0e43f JA |
2389 | { |
2390 | /* See comment at the top of this file */ | |
2391 | smp_rmb(); | |
e23de15f | 2392 | return __io_cqring_events(ctx); |
a3a0e43f JA |
2393 | } |
2394 | ||
fb5ccc98 PB |
2395 | static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) |
2396 | { | |
2397 | struct io_rings *rings = ctx->rings; | |
2398 | ||
2399 | /* make sure SQ entry isn't read before tail */ | |
2400 | return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; | |
2401 | } | |
2402 | ||
8ff069bf | 2403 | static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf) |
e94f141b | 2404 | { |
8ff069bf | 2405 | unsigned int cflags; |
e94f141b | 2406 | |
bcda7baa JA |
2407 | cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT; |
2408 | cflags |= IORING_CQE_F_BUFFER; | |
0e1b6fe3 | 2409 | req->flags &= ~REQ_F_BUFFER_SELECTED; |
bcda7baa JA |
2410 | kfree(kbuf); |
2411 | return cflags; | |
e94f141b JA |
2412 | } |
2413 | ||
8ff069bf | 2414 | static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req) |
bcda7baa | 2415 | { |
4d954c25 | 2416 | struct io_buffer *kbuf; |
bcda7baa | 2417 | |
4d954c25 | 2418 | kbuf = (struct io_buffer *) (unsigned long) req->rw.addr; |
8ff069bf PB |
2419 | return io_put_kbuf(req, kbuf); |
2420 | } | |
2421 | ||
4c6e277c JA |
2422 | static inline bool io_run_task_work(void) |
2423 | { | |
6200b0ae JA |
2424 | /* |
2425 | * Not safe to run on exiting task, and the task_work handling will | |
2426 | * not add work to such a task. | |
2427 | */ | |
2428 | if (unlikely(current->flags & PF_EXITING)) | |
2429 | return false; | |
4c6e277c JA |
2430 | if (current->task_works) { |
2431 | __set_current_state(TASK_RUNNING); | |
2432 | task_work_run(); | |
2433 | return true; | |
2434 | } | |
2435 | ||
2436 | return false; | |
bcda7baa JA |
2437 | } |
2438 | ||
bbde017a XW |
2439 | static void io_iopoll_queue(struct list_head *again) |
2440 | { | |
2441 | struct io_kiocb *req; | |
2442 | ||
2443 | do { | |
d21ffe7e PB |
2444 | req = list_first_entry(again, struct io_kiocb, inflight_entry); |
2445 | list_del(&req->inflight_entry); | |
81b68a5c | 2446 | __io_complete_rw(req, -EAGAIN, 0, NULL); |
bbde017a XW |
2447 | } while (!list_empty(again)); |
2448 | } | |
2449 | ||
def596e9 JA |
2450 | /* |
2451 | * Find and free completed poll iocbs | |
2452 | */ | |
2453 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
2454 | struct list_head *done) | |
2455 | { | |
8237e045 | 2456 | struct req_batch rb; |
def596e9 | 2457 | struct io_kiocb *req; |
bbde017a XW |
2458 | LIST_HEAD(again); |
2459 | ||
2460 | /* order with ->result store in io_complete_rw_iopoll() */ | |
2461 | smp_rmb(); | |
def596e9 | 2462 | |
5af1d13e | 2463 | io_init_req_batch(&rb); |
def596e9 | 2464 | while (!list_empty(done)) { |
bcda7baa JA |
2465 | int cflags = 0; |
2466 | ||
d21ffe7e | 2467 | req = list_first_entry(done, struct io_kiocb, inflight_entry); |
bbde017a | 2468 | if (READ_ONCE(req->result) == -EAGAIN) { |
56450c20 | 2469 | req->result = 0; |
bbde017a | 2470 | req->iopoll_completed = 0; |
d21ffe7e | 2471 | list_move_tail(&req->inflight_entry, &again); |
bbde017a XW |
2472 | continue; |
2473 | } | |
d21ffe7e | 2474 | list_del(&req->inflight_entry); |
def596e9 | 2475 | |
bcda7baa | 2476 | if (req->flags & REQ_F_BUFFER_SELECTED) |
8ff069bf | 2477 | cflags = io_put_rw_kbuf(req); |
bcda7baa JA |
2478 | |
2479 | __io_cqring_fill_event(req, req->result, cflags); | |
def596e9 JA |
2480 | (*nr_events)++; |
2481 | ||
c3524383 | 2482 | if (refcount_dec_and_test(&req->refs)) |
2d6500d4 | 2483 | io_req_free_batch(&rb, req); |
def596e9 | 2484 | } |
def596e9 | 2485 | |
09bb8394 | 2486 | io_commit_cqring(ctx); |
80c18e4a | 2487 | io_cqring_ev_posted_iopoll(ctx); |
2d6500d4 | 2488 | io_req_free_batch_finish(ctx, &rb); |
581f9810 | 2489 | |
bbde017a XW |
2490 | if (!list_empty(&again)) |
2491 | io_iopoll_queue(&again); | |
581f9810 BM |
2492 | } |
2493 | ||
def596e9 JA |
2494 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, |
2495 | long min) | |
2496 | { | |
2497 | struct io_kiocb *req, *tmp; | |
2498 | LIST_HEAD(done); | |
2499 | bool spin; | |
2500 | int ret; | |
2501 | ||
2502 | /* | |
2503 | * Only spin for completions if we don't have multiple devices hanging | |
2504 | * off our complete list, and we're under the requested amount. | |
2505 | */ | |
2506 | spin = !ctx->poll_multi_file && *nr_events < min; | |
2507 | ||
2508 | ret = 0; | |
d21ffe7e | 2509 | list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) { |
9adbd45d | 2510 | struct kiocb *kiocb = &req->rw.kiocb; |
def596e9 JA |
2511 | |
2512 | /* | |
581f9810 BM |
2513 | * Move completed and retryable entries to our local lists. |
2514 | * If we find a request that requires polling, break out | |
2515 | * and complete those lists first, if we have entries there. | |
def596e9 | 2516 | */ |
65a6543d | 2517 | if (READ_ONCE(req->iopoll_completed)) { |
d21ffe7e | 2518 | list_move_tail(&req->inflight_entry, &done); |
def596e9 JA |
2519 | continue; |
2520 | } | |
2521 | if (!list_empty(&done)) | |
2522 | break; | |
2523 | ||
2524 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
2525 | if (ret < 0) | |
2526 | break; | |
2527 | ||
3aadc23e PB |
2528 | /* iopoll may have completed current req */ |
2529 | if (READ_ONCE(req->iopoll_completed)) | |
d21ffe7e | 2530 | list_move_tail(&req->inflight_entry, &done); |
3aadc23e | 2531 | |
def596e9 JA |
2532 | if (ret && spin) |
2533 | spin = false; | |
2534 | ret = 0; | |
2535 | } | |
2536 | ||
2537 | if (!list_empty(&done)) | |
2538 | io_iopoll_complete(ctx, nr_events, &done); | |
2539 | ||
2540 | return ret; | |
2541 | } | |
2542 | ||
2543 | /* | |
d195a66e | 2544 | * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a |
def596e9 JA |
2545 | * non-spinning poll check - we'll still enter the driver poll loop, but only |
2546 | * as a non-spinning completion check. | |
2547 | */ | |
2548 | static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
2549 | long min) | |
2550 | { | |
540e32a0 | 2551 | while (!list_empty(&ctx->iopoll_list) && !need_resched()) { |
def596e9 JA |
2552 | int ret; |
2553 | ||
2554 | ret = io_do_iopoll(ctx, nr_events, min); | |
2555 | if (ret < 0) | |
2556 | return ret; | |
eba0a4dd | 2557 | if (*nr_events >= min) |
def596e9 JA |
2558 | return 0; |
2559 | } | |
2560 | ||
2561 | return 1; | |
2562 | } | |
2563 | ||
2564 | /* | |
2565 | * We can't just wait for polled events to come to us, we have to actively | |
2566 | * find and complete them. | |
2567 | */ | |
b2edc0a7 | 2568 | static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) |
def596e9 JA |
2569 | { |
2570 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
2571 | return; | |
2572 | ||
2573 | mutex_lock(&ctx->uring_lock); | |
540e32a0 | 2574 | while (!list_empty(&ctx->iopoll_list)) { |
def596e9 JA |
2575 | unsigned int nr_events = 0; |
2576 | ||
b2edc0a7 | 2577 | io_do_iopoll(ctx, &nr_events, 0); |
08f5439f | 2578 | |
b2edc0a7 PB |
2579 | /* let it sleep and repeat later if can't complete a request */ |
2580 | if (nr_events == 0) | |
2581 | break; | |
08f5439f JA |
2582 | /* |
2583 | * Ensure we allow local-to-the-cpu processing to take place, | |
2584 | * in this case we need to ensure that we reap all events. | |
3fcee5a6 | 2585 | * Also let task_work, etc. to progress by releasing the mutex |
08f5439f | 2586 | */ |
3fcee5a6 PB |
2587 | if (need_resched()) { |
2588 | mutex_unlock(&ctx->uring_lock); | |
2589 | cond_resched(); | |
2590 | mutex_lock(&ctx->uring_lock); | |
2591 | } | |
def596e9 JA |
2592 | } |
2593 | mutex_unlock(&ctx->uring_lock); | |
2594 | } | |
2595 | ||
7668b92a | 2596 | static int io_iopoll_check(struct io_ring_ctx *ctx, long min) |
def596e9 | 2597 | { |
7668b92a | 2598 | unsigned int nr_events = 0; |
2b2ed975 | 2599 | int iters = 0, ret = 0; |
500f9fba | 2600 | |
c7849be9 XW |
2601 | /* |
2602 | * We disallow the app entering submit/complete with polling, but we | |
2603 | * still need to lock the ring to prevent racing with polled issue | |
2604 | * that got punted to a workqueue. | |
2605 | */ | |
2606 | mutex_lock(&ctx->uring_lock); | |
def596e9 | 2607 | do { |
a3a0e43f JA |
2608 | /* |
2609 | * Don't enter poll loop if we already have events pending. | |
2610 | * If we do, we can potentially be spinning for commands that | |
2611 | * already triggered a CQE (eg in error). | |
2612 | */ | |
6c503150 PB |
2613 | if (test_bit(0, &ctx->cq_check_overflow)) |
2614 | __io_cqring_overflow_flush(ctx, false, NULL, NULL); | |
2615 | if (io_cqring_events(ctx)) | |
a3a0e43f JA |
2616 | break; |
2617 | ||
500f9fba JA |
2618 | /* |
2619 | * If a submit got punted to a workqueue, we can have the | |
2620 | * application entering polling for a command before it gets | |
2621 | * issued. That app will hold the uring_lock for the duration | |
2622 | * of the poll right here, so we need to take a breather every | |
2623 | * now and then to ensure that the issue has a chance to add | |
2624 | * the poll to the issued list. Otherwise we can spin here | |
2625 | * forever, while the workqueue is stuck trying to acquire the | |
2626 | * very same mutex. | |
2627 | */ | |
2628 | if (!(++iters & 7)) { | |
2629 | mutex_unlock(&ctx->uring_lock); | |
4c6e277c | 2630 | io_run_task_work(); |
500f9fba JA |
2631 | mutex_lock(&ctx->uring_lock); |
2632 | } | |
2633 | ||
7668b92a | 2634 | ret = io_iopoll_getevents(ctx, &nr_events, min); |
def596e9 JA |
2635 | if (ret <= 0) |
2636 | break; | |
2637 | ret = 0; | |
7668b92a | 2638 | } while (min && !nr_events && !need_resched()); |
def596e9 | 2639 | |
500f9fba | 2640 | mutex_unlock(&ctx->uring_lock); |
def596e9 JA |
2641 | return ret; |
2642 | } | |
2643 | ||
491381ce | 2644 | static void kiocb_end_write(struct io_kiocb *req) |
2b188cc1 | 2645 | { |
491381ce JA |
2646 | /* |
2647 | * Tell lockdep we inherited freeze protection from submission | |
2648 | * thread. | |
2649 | */ | |
2650 | if (req->flags & REQ_F_ISREG) { | |
2651 | struct inode *inode = file_inode(req->file); | |
2b188cc1 | 2652 | |
491381ce | 2653 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); |
2b188cc1 | 2654 | } |
491381ce | 2655 | file_end_write(req->file); |
2b188cc1 JA |
2656 | } |
2657 | ||
a1d7c393 JA |
2658 | static void io_complete_rw_common(struct kiocb *kiocb, long res, |
2659 | struct io_comp_state *cs) | |
2b188cc1 | 2660 | { |
9adbd45d | 2661 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
bcda7baa | 2662 | int cflags = 0; |
2b188cc1 | 2663 | |
491381ce JA |
2664 | if (kiocb->ki_flags & IOCB_WRITE) |
2665 | kiocb_end_write(req); | |
2b188cc1 | 2666 | |
4e88d6e7 JA |
2667 | if (res != req->result) |
2668 | req_set_fail_links(req); | |
bcda7baa | 2669 | if (req->flags & REQ_F_BUFFER_SELECTED) |
8ff069bf | 2670 | cflags = io_put_rw_kbuf(req); |
a1d7c393 | 2671 | __io_req_complete(req, res, cflags, cs); |
ba816ad6 JA |
2672 | } |
2673 | ||
b63534c4 JA |
2674 | #ifdef CONFIG_BLOCK |
2675 | static bool io_resubmit_prep(struct io_kiocb *req, int error) | |
2676 | { | |
2677 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
2678 | ssize_t ret = -ECANCELED; | |
2679 | struct iov_iter iter; | |
2680 | int rw; | |
2681 | ||
2682 | if (error) { | |
2683 | ret = error; | |
2684 | goto end_req; | |
2685 | } | |
2686 | ||
2687 | switch (req->opcode) { | |
2688 | case IORING_OP_READV: | |
2689 | case IORING_OP_READ_FIXED: | |
2690 | case IORING_OP_READ: | |
2691 | rw = READ; | |
2692 | break; | |
2693 | case IORING_OP_WRITEV: | |
2694 | case IORING_OP_WRITE_FIXED: | |
2695 | case IORING_OP_WRITE: | |
2696 | rw = WRITE; | |
2697 | break; | |
2698 | default: | |
2699 | printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n", | |
2700 | req->opcode); | |
2701 | goto end_req; | |
2702 | } | |
2703 | ||
e8c2bc1f | 2704 | if (!req->async_data) { |
8f3d7496 JA |
2705 | ret = io_import_iovec(rw, req, &iovec, &iter, false); |
2706 | if (ret < 0) | |
2707 | goto end_req; | |
2708 | ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false); | |
2709 | if (!ret) | |
2710 | return true; | |
2711 | kfree(iovec); | |
2712 | } else { | |
b63534c4 | 2713 | return true; |
8f3d7496 | 2714 | } |
b63534c4 | 2715 | end_req: |
b63534c4 | 2716 | req_set_fail_links(req); |
b63534c4 JA |
2717 | return false; |
2718 | } | |
b63534c4 JA |
2719 | #endif |
2720 | ||
2721 | static bool io_rw_reissue(struct io_kiocb *req, long res) | |
2722 | { | |
2723 | #ifdef CONFIG_BLOCK | |
355afaeb | 2724 | umode_t mode = file_inode(req->file)->i_mode; |
b63534c4 JA |
2725 | int ret; |
2726 | ||
355afaeb JA |
2727 | if (!S_ISBLK(mode) && !S_ISREG(mode)) |
2728 | return false; | |
b63534c4 JA |
2729 | if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker()) |
2730 | return false; | |
2731 | ||
55e6ac1e PB |
2732 | lockdep_assert_held(&req->ctx->uring_lock); |
2733 | ||
28cea78a | 2734 | ret = io_sq_thread_acquire_mm_files(req->ctx, req); |
6d816e08 | 2735 | |
fdee946d JA |
2736 | if (io_resubmit_prep(req, ret)) { |
2737 | refcount_inc(&req->refs); | |
2738 | io_queue_async_work(req); | |
b63534c4 | 2739 | return true; |
fdee946d JA |
2740 | } |
2741 | ||
b63534c4 JA |
2742 | #endif |
2743 | return false; | |
2744 | } | |
2745 | ||
a1d7c393 JA |
2746 | static void __io_complete_rw(struct io_kiocb *req, long res, long res2, |
2747 | struct io_comp_state *cs) | |
2748 | { | |
2749 | if (!io_rw_reissue(req, res)) | |
2750 | io_complete_rw_common(&req->rw.kiocb, res, cs); | |
ba816ad6 JA |
2751 | } |
2752 | ||
2753 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
2754 | { | |
9adbd45d | 2755 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
ba816ad6 | 2756 | |
a1d7c393 | 2757 | __io_complete_rw(req, res, res2, NULL); |
2b188cc1 JA |
2758 | } |
2759 | ||
def596e9 JA |
2760 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
2761 | { | |
9adbd45d | 2762 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
def596e9 | 2763 | |
491381ce JA |
2764 | if (kiocb->ki_flags & IOCB_WRITE) |
2765 | kiocb_end_write(req); | |
def596e9 | 2766 | |
2d7d6792 | 2767 | if (res != -EAGAIN && res != req->result) |
4e88d6e7 | 2768 | req_set_fail_links(req); |
bbde017a XW |
2769 | |
2770 | WRITE_ONCE(req->result, res); | |
2771 | /* order with io_poll_complete() checking ->result */ | |
cd664b0e PB |
2772 | smp_wmb(); |
2773 | WRITE_ONCE(req->iopoll_completed, 1); | |
def596e9 JA |
2774 | } |
2775 | ||
2776 | /* | |
2777 | * After the iocb has been issued, it's safe to be found on the poll list. | |
2778 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
2779 | * find it from a io_iopoll_getevents() thread before the issuer is done | |
2780 | * accessing the kiocb cookie. | |
2781 | */ | |
2e9dbe90 | 2782 | static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async) |
def596e9 JA |
2783 | { |
2784 | struct io_ring_ctx *ctx = req->ctx; | |
2785 | ||
2786 | /* | |
2787 | * Track whether we have multiple files in our lists. This will impact | |
2788 | * how we do polling eventually, not spinning if we're on potentially | |
2789 | * different devices. | |
2790 | */ | |
540e32a0 | 2791 | if (list_empty(&ctx->iopoll_list)) { |
def596e9 JA |
2792 | ctx->poll_multi_file = false; |
2793 | } else if (!ctx->poll_multi_file) { | |
2794 | struct io_kiocb *list_req; | |
2795 | ||
540e32a0 | 2796 | list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb, |
d21ffe7e | 2797 | inflight_entry); |
9adbd45d | 2798 | if (list_req->file != req->file) |
def596e9 JA |
2799 | ctx->poll_multi_file = true; |
2800 | } | |
2801 | ||
2802 | /* | |
2803 | * For fast devices, IO may have already completed. If it has, add | |
2804 | * it to the front so we find it first. | |
2805 | */ | |
65a6543d | 2806 | if (READ_ONCE(req->iopoll_completed)) |
d21ffe7e | 2807 | list_add(&req->inflight_entry, &ctx->iopoll_list); |
def596e9 | 2808 | else |
d21ffe7e | 2809 | list_add_tail(&req->inflight_entry, &ctx->iopoll_list); |
bdcd3eab | 2810 | |
2e9dbe90 XW |
2811 | /* |
2812 | * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread | |
2813 | * task context or in io worker task context. If current task context is | |
2814 | * sq thread, we don't need to check whether should wake up sq thread. | |
2815 | */ | |
2816 | if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) && | |
534ca6d6 JA |
2817 | wq_has_sleeper(&ctx->sq_data->wait)) |
2818 | wake_up(&ctx->sq_data->wait); | |
def596e9 JA |
2819 | } |
2820 | ||
6e1271e6 | 2821 | static inline void __io_state_file_put(struct io_submit_state *state) |
9a56a232 | 2822 | { |
6e1271e6 PB |
2823 | fput_many(state->file, state->file_refs); |
2824 | state->file_refs = 0; | |
9f13c35b PB |
2825 | } |
2826 | ||
2827 | static inline void io_state_file_put(struct io_submit_state *state) | |
2828 | { | |
6e1271e6 | 2829 | if (state->file_refs) |
9f13c35b | 2830 | __io_state_file_put(state); |
9a56a232 JA |
2831 | } |
2832 | ||
2833 | /* | |
2834 | * Get as many references to a file as we have IOs left in this submission, | |
2835 | * assuming most submissions are for one file, or at least that each file | |
2836 | * has more than one submission. | |
2837 | */ | |
8da11c19 | 2838 | static struct file *__io_file_get(struct io_submit_state *state, int fd) |
9a56a232 JA |
2839 | { |
2840 | if (!state) | |
2841 | return fget(fd); | |
2842 | ||
6e1271e6 | 2843 | if (state->file_refs) { |
9a56a232 | 2844 | if (state->fd == fd) { |
6e1271e6 | 2845 | state->file_refs--; |
9a56a232 JA |
2846 | return state->file; |
2847 | } | |
9f13c35b | 2848 | __io_state_file_put(state); |
9a56a232 JA |
2849 | } |
2850 | state->file = fget_many(fd, state->ios_left); | |
6e1271e6 | 2851 | if (unlikely(!state->file)) |
9a56a232 JA |
2852 | return NULL; |
2853 | ||
2854 | state->fd = fd; | |
6e1271e6 | 2855 | state->file_refs = state->ios_left - 1; |
9a56a232 JA |
2856 | return state->file; |
2857 | } | |
2858 | ||
4503b767 JA |
2859 | static bool io_bdev_nowait(struct block_device *bdev) |
2860 | { | |
9ba0d0c8 | 2861 | return !bdev || blk_queue_nowait(bdev_get_queue(bdev)); |
4503b767 JA |
2862 | } |
2863 | ||
2b188cc1 JA |
2864 | /* |
2865 | * If we tracked the file through the SCM inflight mechanism, we could support | |
2866 | * any file. For now, just ensure that anything potentially problematic is done | |
2867 | * inline. | |
2868 | */ | |
af197f50 | 2869 | static bool io_file_supports_async(struct file *file, int rw) |
2b188cc1 JA |
2870 | { |
2871 | umode_t mode = file_inode(file)->i_mode; | |
2872 | ||
4503b767 | 2873 | if (S_ISBLK(mode)) { |
4e7b5671 CH |
2874 | if (IS_ENABLED(CONFIG_BLOCK) && |
2875 | io_bdev_nowait(I_BDEV(file->f_mapping->host))) | |
4503b767 JA |
2876 | return true; |
2877 | return false; | |
2878 | } | |
2879 | if (S_ISCHR(mode) || S_ISSOCK(mode)) | |
2b188cc1 | 2880 | return true; |
4503b767 | 2881 | if (S_ISREG(mode)) { |
4e7b5671 CH |
2882 | if (IS_ENABLED(CONFIG_BLOCK) && |
2883 | io_bdev_nowait(file->f_inode->i_sb->s_bdev) && | |
4503b767 JA |
2884 | file->f_op != &io_uring_fops) |
2885 | return true; | |
2886 | return false; | |
2887 | } | |
2b188cc1 | 2888 | |
c5b85625 JA |
2889 | /* any ->read/write should understand O_NONBLOCK */ |
2890 | if (file->f_flags & O_NONBLOCK) | |
2891 | return true; | |
2892 | ||
af197f50 JA |
2893 | if (!(file->f_mode & FMODE_NOWAIT)) |
2894 | return false; | |
2895 | ||
2896 | if (rw == READ) | |
2897 | return file->f_op->read_iter != NULL; | |
2898 | ||
2899 | return file->f_op->write_iter != NULL; | |
2b188cc1 JA |
2900 | } |
2901 | ||
a88fc400 | 2902 | static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2b188cc1 | 2903 | { |
def596e9 | 2904 | struct io_ring_ctx *ctx = req->ctx; |
9adbd45d | 2905 | struct kiocb *kiocb = &req->rw.kiocb; |
09bb8394 JA |
2906 | unsigned ioprio; |
2907 | int ret; | |
2b188cc1 | 2908 | |
491381ce JA |
2909 | if (S_ISREG(file_inode(req->file)->i_mode)) |
2910 | req->flags |= REQ_F_ISREG; | |
2911 | ||
2b188cc1 | 2912 | kiocb->ki_pos = READ_ONCE(sqe->off); |
ba04291e JA |
2913 | if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) { |
2914 | req->flags |= REQ_F_CUR_POS; | |
2915 | kiocb->ki_pos = req->file->f_pos; | |
2916 | } | |
2b188cc1 | 2917 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); |
3e577dcd PB |
2918 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); |
2919 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
2920 | if (unlikely(ret)) | |
2921 | return ret; | |
2b188cc1 JA |
2922 | |
2923 | ioprio = READ_ONCE(sqe->ioprio); | |
2924 | if (ioprio) { | |
2925 | ret = ioprio_check_cap(ioprio); | |
2926 | if (ret) | |
09bb8394 | 2927 | return ret; |
2b188cc1 JA |
2928 | |
2929 | kiocb->ki_ioprio = ioprio; | |
2930 | } else | |
2931 | kiocb->ki_ioprio = get_current_ioprio(); | |
2932 | ||
8449eeda | 2933 | /* don't allow async punt if RWF_NOWAIT was requested */ |
c5b85625 | 2934 | if (kiocb->ki_flags & IOCB_NOWAIT) |
8449eeda SB |
2935 | req->flags |= REQ_F_NOWAIT; |
2936 | ||
def596e9 | 2937 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
2938 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
2939 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 2940 | return -EOPNOTSUPP; |
2b188cc1 | 2941 | |
def596e9 JA |
2942 | kiocb->ki_flags |= IOCB_HIPRI; |
2943 | kiocb->ki_complete = io_complete_rw_iopoll; | |
65a6543d | 2944 | req->iopoll_completed = 0; |
def596e9 | 2945 | } else { |
09bb8394 JA |
2946 | if (kiocb->ki_flags & IOCB_HIPRI) |
2947 | return -EINVAL; | |
def596e9 JA |
2948 | kiocb->ki_complete = io_complete_rw; |
2949 | } | |
9adbd45d | 2950 | |
3529d8c2 JA |
2951 | req->rw.addr = READ_ONCE(sqe->addr); |
2952 | req->rw.len = READ_ONCE(sqe->len); | |
4f4eeba8 | 2953 | req->buf_index = READ_ONCE(sqe->buf_index); |
2b188cc1 | 2954 | return 0; |
2b188cc1 JA |
2955 | } |
2956 | ||
2957 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
2958 | { | |
2959 | switch (ret) { | |
2960 | case -EIOCBQUEUED: | |
2961 | break; | |
2962 | case -ERESTARTSYS: | |
2963 | case -ERESTARTNOINTR: | |
2964 | case -ERESTARTNOHAND: | |
2965 | case -ERESTART_RESTARTBLOCK: | |
2966 | /* | |
2967 | * We can't just restart the syscall, since previously | |
2968 | * submitted sqes may already be in progress. Just fail this | |
2969 | * IO with EINTR. | |
2970 | */ | |
2971 | ret = -EINTR; | |
df561f66 | 2972 | fallthrough; |
2b188cc1 JA |
2973 | default: |
2974 | kiocb->ki_complete(kiocb, ret, 0); | |
2975 | } | |
2976 | } | |
2977 | ||
a1d7c393 JA |
2978 | static void kiocb_done(struct kiocb *kiocb, ssize_t ret, |
2979 | struct io_comp_state *cs) | |
ba816ad6 | 2980 | { |
ba04291e | 2981 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
e8c2bc1f | 2982 | struct io_async_rw *io = req->async_data; |
ba04291e | 2983 | |
227c0c96 | 2984 | /* add previously done IO, if any */ |
e8c2bc1f | 2985 | if (io && io->bytes_done > 0) { |
227c0c96 | 2986 | if (ret < 0) |
e8c2bc1f | 2987 | ret = io->bytes_done; |
227c0c96 | 2988 | else |
e8c2bc1f | 2989 | ret += io->bytes_done; |
227c0c96 JA |
2990 | } |
2991 | ||
ba04291e JA |
2992 | if (req->flags & REQ_F_CUR_POS) |
2993 | req->file->f_pos = kiocb->ki_pos; | |
bcaec089 | 2994 | if (ret >= 0 && kiocb->ki_complete == io_complete_rw) |
a1d7c393 | 2995 | __io_complete_rw(req, ret, 0, cs); |
ba816ad6 JA |
2996 | else |
2997 | io_rw_done(kiocb, ret); | |
2998 | } | |
2999 | ||
9adbd45d | 3000 | static ssize_t io_import_fixed(struct io_kiocb *req, int rw, |
7d009165 | 3001 | struct iov_iter *iter) |
edafccee | 3002 | { |
9adbd45d JA |
3003 | struct io_ring_ctx *ctx = req->ctx; |
3004 | size_t len = req->rw.len; | |
edafccee | 3005 | struct io_mapped_ubuf *imu; |
4be1c615 | 3006 | u16 index, buf_index = req->buf_index; |
edafccee JA |
3007 | size_t offset; |
3008 | u64 buf_addr; | |
3009 | ||
edafccee JA |
3010 | if (unlikely(buf_index >= ctx->nr_user_bufs)) |
3011 | return -EFAULT; | |
edafccee JA |
3012 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); |
3013 | imu = &ctx->user_bufs[index]; | |
9adbd45d | 3014 | buf_addr = req->rw.addr; |
edafccee JA |
3015 | |
3016 | /* overflow */ | |
3017 | if (buf_addr + len < buf_addr) | |
3018 | return -EFAULT; | |
3019 | /* not inside the mapped region */ | |
3020 | if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len) | |
3021 | return -EFAULT; | |
3022 | ||
3023 | /* | |
3024 | * May not be a start of buffer, set size appropriately | |
3025 | * and advance us to the beginning. | |
3026 | */ | |
3027 | offset = buf_addr - imu->ubuf; | |
3028 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
bd11b3a3 JA |
3029 | |
3030 | if (offset) { | |
3031 | /* | |
3032 | * Don't use iov_iter_advance() here, as it's really slow for | |
3033 | * using the latter parts of a big fixed buffer - it iterates | |
3034 | * over each segment manually. We can cheat a bit here, because | |
3035 | * we know that: | |
3036 | * | |
3037 | * 1) it's a BVEC iter, we set it up | |
3038 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
3039 | * first and last bvec | |
3040 | * | |
3041 | * So just find our index, and adjust the iterator afterwards. | |
3042 | * If the offset is within the first bvec (or the whole first | |
3043 | * bvec, just use iov_iter_advance(). This makes it easier | |
3044 | * since we can just skip the first segment, which may not | |
3045 | * be PAGE_SIZE aligned. | |
3046 | */ | |
3047 | const struct bio_vec *bvec = imu->bvec; | |
3048 | ||
3049 | if (offset <= bvec->bv_len) { | |
3050 | iov_iter_advance(iter, offset); | |
3051 | } else { | |
3052 | unsigned long seg_skip; | |
3053 | ||
3054 | /* skip first vec */ | |
3055 | offset -= bvec->bv_len; | |
3056 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
3057 | ||
3058 | iter->bvec = bvec + seg_skip; | |
3059 | iter->nr_segs -= seg_skip; | |
99c79f66 | 3060 | iter->count -= bvec->bv_len + offset; |
bd11b3a3 | 3061 | iter->iov_offset = offset & ~PAGE_MASK; |
bd11b3a3 JA |
3062 | } |
3063 | } | |
3064 | ||
5e559561 | 3065 | return len; |
edafccee JA |
3066 | } |
3067 | ||
bcda7baa JA |
3068 | static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock) |
3069 | { | |
3070 | if (needs_lock) | |
3071 | mutex_unlock(&ctx->uring_lock); | |
3072 | } | |
3073 | ||
3074 | static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock) | |
3075 | { | |
3076 | /* | |
3077 | * "Normal" inline submissions always hold the uring_lock, since we | |
3078 | * grab it from the system call. Same is true for the SQPOLL offload. | |
3079 | * The only exception is when we've detached the request and issue it | |
3080 | * from an async worker thread, grab the lock for that case. | |
3081 | */ | |
3082 | if (needs_lock) | |
3083 | mutex_lock(&ctx->uring_lock); | |
3084 | } | |
3085 | ||
3086 | static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len, | |
3087 | int bgid, struct io_buffer *kbuf, | |
3088 | bool needs_lock) | |
3089 | { | |
3090 | struct io_buffer *head; | |
3091 | ||
3092 | if (req->flags & REQ_F_BUFFER_SELECTED) | |
3093 | return kbuf; | |
3094 | ||
3095 | io_ring_submit_lock(req->ctx, needs_lock); | |
3096 | ||
3097 | lockdep_assert_held(&req->ctx->uring_lock); | |
3098 | ||
3099 | head = idr_find(&req->ctx->io_buffer_idr, bgid); | |
3100 | if (head) { | |
3101 | if (!list_empty(&head->list)) { | |
3102 | kbuf = list_last_entry(&head->list, struct io_buffer, | |
3103 | list); | |
3104 | list_del(&kbuf->list); | |
3105 | } else { | |
3106 | kbuf = head; | |
3107 | idr_remove(&req->ctx->io_buffer_idr, bgid); | |
3108 | } | |
3109 | if (*len > kbuf->len) | |
3110 | *len = kbuf->len; | |
3111 | } else { | |
3112 | kbuf = ERR_PTR(-ENOBUFS); | |
3113 | } | |
3114 | ||
3115 | io_ring_submit_unlock(req->ctx, needs_lock); | |
3116 | ||
3117 | return kbuf; | |
3118 | } | |
3119 | ||
4d954c25 JA |
3120 | static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len, |
3121 | bool needs_lock) | |
3122 | { | |
3123 | struct io_buffer *kbuf; | |
4f4eeba8 | 3124 | u16 bgid; |
4d954c25 JA |
3125 | |
3126 | kbuf = (struct io_buffer *) (unsigned long) req->rw.addr; | |
4f4eeba8 | 3127 | bgid = req->buf_index; |
4d954c25 JA |
3128 | kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock); |
3129 | if (IS_ERR(kbuf)) | |
3130 | return kbuf; | |
3131 | req->rw.addr = (u64) (unsigned long) kbuf; | |
3132 | req->flags |= REQ_F_BUFFER_SELECTED; | |
3133 | return u64_to_user_ptr(kbuf->addr); | |
3134 | } | |
3135 | ||
3136 | #ifdef CONFIG_COMPAT | |
3137 | static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov, | |
3138 | bool needs_lock) | |
3139 | { | |
3140 | struct compat_iovec __user *uiov; | |
3141 | compat_ssize_t clen; | |
3142 | void __user *buf; | |
3143 | ssize_t len; | |
3144 | ||
3145 | uiov = u64_to_user_ptr(req->rw.addr); | |
3146 | if (!access_ok(uiov, sizeof(*uiov))) | |
3147 | return -EFAULT; | |
3148 | if (__get_user(clen, &uiov->iov_len)) | |
3149 | return -EFAULT; | |
3150 | if (clen < 0) | |
3151 | return -EINVAL; | |
3152 | ||
3153 | len = clen; | |
3154 | buf = io_rw_buffer_select(req, &len, needs_lock); | |
3155 | if (IS_ERR(buf)) | |
3156 | return PTR_ERR(buf); | |
3157 | iov[0].iov_base = buf; | |
3158 | iov[0].iov_len = (compat_size_t) len; | |
3159 | return 0; | |
3160 | } | |
3161 | #endif | |
3162 | ||
3163 | static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov, | |
3164 | bool needs_lock) | |
3165 | { | |
3166 | struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr); | |
3167 | void __user *buf; | |
3168 | ssize_t len; | |
3169 | ||
3170 | if (copy_from_user(iov, uiov, sizeof(*uiov))) | |
3171 | return -EFAULT; | |
3172 | ||
3173 | len = iov[0].iov_len; | |
3174 | if (len < 0) | |
3175 | return -EINVAL; | |
3176 | buf = io_rw_buffer_select(req, &len, needs_lock); | |
3177 | if (IS_ERR(buf)) | |
3178 | return PTR_ERR(buf); | |
3179 | iov[0].iov_base = buf; | |
3180 | iov[0].iov_len = len; | |
3181 | return 0; | |
3182 | } | |
3183 | ||
3184 | static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov, | |
3185 | bool needs_lock) | |
3186 | { | |
dddb3e26 JA |
3187 | if (req->flags & REQ_F_BUFFER_SELECTED) { |
3188 | struct io_buffer *kbuf; | |
3189 | ||
3190 | kbuf = (struct io_buffer *) (unsigned long) req->rw.addr; | |
3191 | iov[0].iov_base = u64_to_user_ptr(kbuf->addr); | |
3192 | iov[0].iov_len = kbuf->len; | |
4d954c25 | 3193 | return 0; |
dddb3e26 | 3194 | } |
dd201662 | 3195 | if (req->rw.len != 1) |
4d954c25 JA |
3196 | return -EINVAL; |
3197 | ||
3198 | #ifdef CONFIG_COMPAT | |
3199 | if (req->ctx->compat) | |
3200 | return io_compat_import(req, iov, needs_lock); | |
3201 | #endif | |
3202 | ||
3203 | return __io_iov_buffer_select(req, iov, needs_lock); | |
3204 | } | |
3205 | ||
2846c481 | 3206 | static ssize_t io_import_iovec(int rw, struct io_kiocb *req, |
8452fd0c JA |
3207 | struct iovec **iovec, struct iov_iter *iter, |
3208 | bool needs_lock) | |
2b188cc1 | 3209 | { |
9adbd45d JA |
3210 | void __user *buf = u64_to_user_ptr(req->rw.addr); |
3211 | size_t sqe_len = req->rw.len; | |
4d954c25 | 3212 | ssize_t ret; |
edafccee JA |
3213 | u8 opcode; |
3214 | ||
d625c6ee | 3215 | opcode = req->opcode; |
7d009165 | 3216 | if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) { |
edafccee | 3217 | *iovec = NULL; |
9adbd45d | 3218 | return io_import_fixed(req, rw, iter); |
edafccee | 3219 | } |
2b188cc1 | 3220 | |
bcda7baa | 3221 | /* buffer index only valid with fixed read/write, or buffer select */ |
4f4eeba8 | 3222 | if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT)) |
9adbd45d JA |
3223 | return -EINVAL; |
3224 | ||
3a6820f2 | 3225 | if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) { |
bcda7baa | 3226 | if (req->flags & REQ_F_BUFFER_SELECT) { |
4d954c25 | 3227 | buf = io_rw_buffer_select(req, &sqe_len, needs_lock); |
867a23ea | 3228 | if (IS_ERR(buf)) |
4d954c25 | 3229 | return PTR_ERR(buf); |
3f9d6441 | 3230 | req->rw.len = sqe_len; |
bcda7baa JA |
3231 | } |
3232 | ||
3a6820f2 JA |
3233 | ret = import_single_range(rw, buf, sqe_len, *iovec, iter); |
3234 | *iovec = NULL; | |
10fc72e4 | 3235 | return ret; |
3a6820f2 JA |
3236 | } |
3237 | ||
4d954c25 JA |
3238 | if (req->flags & REQ_F_BUFFER_SELECT) { |
3239 | ret = io_iov_buffer_select(req, *iovec, needs_lock); | |
3f9d6441 JA |
3240 | if (!ret) { |
3241 | ret = (*iovec)->iov_len; | |
3242 | iov_iter_init(iter, rw, *iovec, 1, ret); | |
3243 | } | |
4d954c25 JA |
3244 | *iovec = NULL; |
3245 | return ret; | |
3246 | } | |
3247 | ||
89cd35c5 CH |
3248 | return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter, |
3249 | req->ctx->compat); | |
2b188cc1 JA |
3250 | } |
3251 | ||
0fef9483 JA |
3252 | static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb) |
3253 | { | |
5b09e37e | 3254 | return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos; |
0fef9483 JA |
3255 | } |
3256 | ||
31b51510 | 3257 | /* |
32960613 JA |
3258 | * For files that don't have ->read_iter() and ->write_iter(), handle them |
3259 | * by looping over ->read() or ->write() manually. | |
31b51510 | 3260 | */ |
4017eb91 | 3261 | static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter) |
32960613 | 3262 | { |
4017eb91 JA |
3263 | struct kiocb *kiocb = &req->rw.kiocb; |
3264 | struct file *file = req->file; | |
32960613 JA |
3265 | ssize_t ret = 0; |
3266 | ||
3267 | /* | |
3268 | * Don't support polled IO through this interface, and we can't | |
3269 | * support non-blocking either. For the latter, this just causes | |
3270 | * the kiocb to be handled from an async context. | |
3271 | */ | |
3272 | if (kiocb->ki_flags & IOCB_HIPRI) | |
3273 | return -EOPNOTSUPP; | |
3274 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
3275 | return -EAGAIN; | |
3276 | ||
3277 | while (iov_iter_count(iter)) { | |
311ae9e1 | 3278 | struct iovec iovec; |
32960613 JA |
3279 | ssize_t nr; |
3280 | ||
311ae9e1 PB |
3281 | if (!iov_iter_is_bvec(iter)) { |
3282 | iovec = iov_iter_iovec(iter); | |
3283 | } else { | |
4017eb91 JA |
3284 | iovec.iov_base = u64_to_user_ptr(req->rw.addr); |
3285 | iovec.iov_len = req->rw.len; | |
311ae9e1 PB |
3286 | } |
3287 | ||
32960613 JA |
3288 | if (rw == READ) { |
3289 | nr = file->f_op->read(file, iovec.iov_base, | |
0fef9483 | 3290 | iovec.iov_len, io_kiocb_ppos(kiocb)); |
32960613 JA |
3291 | } else { |
3292 | nr = file->f_op->write(file, iovec.iov_base, | |
0fef9483 | 3293 | iovec.iov_len, io_kiocb_ppos(kiocb)); |
32960613 JA |
3294 | } |
3295 | ||
3296 | if (nr < 0) { | |
3297 | if (!ret) | |
3298 | ret = nr; | |
3299 | break; | |
3300 | } | |
3301 | ret += nr; | |
3302 | if (nr != iovec.iov_len) | |
3303 | break; | |
4017eb91 JA |
3304 | req->rw.len -= nr; |
3305 | req->rw.addr += nr; | |
32960613 JA |
3306 | iov_iter_advance(iter, nr); |
3307 | } | |
3308 | ||
3309 | return ret; | |
3310 | } | |
3311 | ||
ff6165b2 JA |
3312 | static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec, |
3313 | const struct iovec *fast_iov, struct iov_iter *iter) | |
f67676d1 | 3314 | { |
e8c2bc1f | 3315 | struct io_async_rw *rw = req->async_data; |
b64e3444 | 3316 | |
ff6165b2 | 3317 | memcpy(&rw->iter, iter, sizeof(*iter)); |
afb87658 | 3318 | rw->free_iovec = iovec; |
227c0c96 | 3319 | rw->bytes_done = 0; |
ff6165b2 | 3320 | /* can only be fixed buffers, no need to do anything */ |
9c3a205c | 3321 | if (iov_iter_is_bvec(iter)) |
ff6165b2 | 3322 | return; |
b64e3444 | 3323 | if (!iovec) { |
ff6165b2 JA |
3324 | unsigned iov_off = 0; |
3325 | ||
3326 | rw->iter.iov = rw->fast_iov; | |
3327 | if (iter->iov != fast_iov) { | |
3328 | iov_off = iter->iov - fast_iov; | |
3329 | rw->iter.iov += iov_off; | |
3330 | } | |
3331 | if (rw->fast_iov != fast_iov) | |
3332 | memcpy(rw->fast_iov + iov_off, fast_iov + iov_off, | |
45097dae | 3333 | sizeof(struct iovec) * iter->nr_segs); |
99bc4c38 PB |
3334 | } else { |
3335 | req->flags |= REQ_F_NEED_CLEANUP; | |
f67676d1 JA |
3336 | } |
3337 | } | |
3338 | ||
e8c2bc1f | 3339 | static inline int __io_alloc_async_data(struct io_kiocb *req) |
3d9932a8 | 3340 | { |
e8c2bc1f JA |
3341 | WARN_ON_ONCE(!io_op_defs[req->opcode].async_size); |
3342 | req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL); | |
3343 | return req->async_data == NULL; | |
3d9932a8 XW |
3344 | } |
3345 | ||
e8c2bc1f | 3346 | static int io_alloc_async_data(struct io_kiocb *req) |
f67676d1 | 3347 | { |
e8c2bc1f | 3348 | if (!io_op_defs[req->opcode].needs_async_data) |
d3656344 | 3349 | return 0; |
3d9932a8 | 3350 | |
e8c2bc1f | 3351 | return __io_alloc_async_data(req); |
b7bb4f7d JA |
3352 | } |
3353 | ||
ff6165b2 JA |
3354 | static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec, |
3355 | const struct iovec *fast_iov, | |
227c0c96 | 3356 | struct iov_iter *iter, bool force) |
b7bb4f7d | 3357 | { |
e8c2bc1f | 3358 | if (!force && !io_op_defs[req->opcode].needs_async_data) |
74566df3 | 3359 | return 0; |
e8c2bc1f JA |
3360 | if (!req->async_data) { |
3361 | if (__io_alloc_async_data(req)) | |
5d204bcf | 3362 | return -ENOMEM; |
b7bb4f7d | 3363 | |
ff6165b2 | 3364 | io_req_map_rw(req, iovec, fast_iov, iter); |
5d204bcf | 3365 | } |
b7bb4f7d | 3366 | return 0; |
f67676d1 JA |
3367 | } |
3368 | ||
73debe68 | 3369 | static inline int io_rw_prep_async(struct io_kiocb *req, int rw) |
c3e330a4 | 3370 | { |
e8c2bc1f | 3371 | struct io_async_rw *iorw = req->async_data; |
f4bff104 | 3372 | struct iovec *iov = iorw->fast_iov; |
c3e330a4 PB |
3373 | ssize_t ret; |
3374 | ||
2846c481 | 3375 | ret = io_import_iovec(rw, req, &iov, &iorw->iter, false); |
c3e330a4 PB |
3376 | if (unlikely(ret < 0)) |
3377 | return ret; | |
3378 | ||
ab0b196c PB |
3379 | iorw->bytes_done = 0; |
3380 | iorw->free_iovec = iov; | |
3381 | if (iov) | |
3382 | req->flags |= REQ_F_NEED_CLEANUP; | |
c3e330a4 PB |
3383 | return 0; |
3384 | } | |
3385 | ||
73debe68 | 3386 | static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f67676d1 JA |
3387 | { |
3388 | ssize_t ret; | |
3389 | ||
a88fc400 | 3390 | ret = io_prep_rw(req, sqe); |
3529d8c2 JA |
3391 | if (ret) |
3392 | return ret; | |
f67676d1 | 3393 | |
3529d8c2 JA |
3394 | if (unlikely(!(req->file->f_mode & FMODE_READ))) |
3395 | return -EBADF; | |
f67676d1 | 3396 | |
5f798bea | 3397 | /* either don't need iovec imported or already have it */ |
2d199895 | 3398 | if (!req->async_data) |
3529d8c2 | 3399 | return 0; |
73debe68 | 3400 | return io_rw_prep_async(req, READ); |
f67676d1 JA |
3401 | } |
3402 | ||
c1dd91d1 JA |
3403 | /* |
3404 | * This is our waitqueue callback handler, registered through lock_page_async() | |
3405 | * when we initially tried to do the IO with the iocb armed our waitqueue. | |
3406 | * This gets called when the page is unlocked, and we generally expect that to | |
3407 | * happen when the page IO is completed and the page is now uptodate. This will | |
3408 | * queue a task_work based retry of the operation, attempting to copy the data | |
3409 | * again. If the latter fails because the page was NOT uptodate, then we will | |
3410 | * do a thread based blocking retry of the operation. That's the unexpected | |
3411 | * slow path. | |
3412 | */ | |
bcf5a063 JA |
3413 | static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode, |
3414 | int sync, void *arg) | |
3415 | { | |
3416 | struct wait_page_queue *wpq; | |
3417 | struct io_kiocb *req = wait->private; | |
bcf5a063 | 3418 | struct wait_page_key *key = arg; |
bcf5a063 JA |
3419 | int ret; |
3420 | ||
3421 | wpq = container_of(wait, struct wait_page_queue, wait); | |
3422 | ||
cdc8fcb4 LT |
3423 | if (!wake_page_match(wpq, key)) |
3424 | return 0; | |
3425 | ||
c8d317aa | 3426 | req->rw.kiocb.ki_flags &= ~IOCB_WAITQ; |
bcf5a063 JA |
3427 | list_del_init(&wait->entry); |
3428 | ||
e7375122 | 3429 | init_task_work(&req->task_work, io_req_task_submit); |
6d816e08 JA |
3430 | percpu_ref_get(&req->ctx->refs); |
3431 | ||
bcf5a063 JA |
3432 | /* submit ref gets dropped, acquire a new one */ |
3433 | refcount_inc(&req->refs); | |
355fb9e2 | 3434 | ret = io_req_task_work_add(req); |
bcf5a063 | 3435 | if (unlikely(ret)) { |
c2c4c83c JA |
3436 | struct task_struct *tsk; |
3437 | ||
bcf5a063 | 3438 | /* queue just for cancelation */ |
e7375122 | 3439 | init_task_work(&req->task_work, io_req_task_cancel); |
bcf5a063 | 3440 | tsk = io_wq_get_task(req->ctx->io_wq); |
91989c70 | 3441 | task_work_add(tsk, &req->task_work, TWA_NONE); |
c2c4c83c | 3442 | wake_up_process(tsk); |
bcf5a063 | 3443 | } |
bcf5a063 JA |
3444 | return 1; |
3445 | } | |
3446 | ||
c1dd91d1 JA |
3447 | /* |
3448 | * This controls whether a given IO request should be armed for async page | |
3449 | * based retry. If we return false here, the request is handed to the async | |
3450 | * worker threads for retry. If we're doing buffered reads on a regular file, | |
3451 | * we prepare a private wait_page_queue entry and retry the operation. This | |
3452 | * will either succeed because the page is now uptodate and unlocked, or it | |
3453 | * will register a callback when the page is unlocked at IO completion. Through | |
3454 | * that callback, io_uring uses task_work to setup a retry of the operation. | |
3455 | * That retry will attempt the buffered read again. The retry will generally | |
3456 | * succeed, or in rare cases where it fails, we then fall back to using the | |
3457 | * async worker threads for a blocking retry. | |
3458 | */ | |
227c0c96 | 3459 | static bool io_rw_should_retry(struct io_kiocb *req) |
f67676d1 | 3460 | { |
e8c2bc1f JA |
3461 | struct io_async_rw *rw = req->async_data; |
3462 | struct wait_page_queue *wait = &rw->wpq; | |
bcf5a063 | 3463 | struct kiocb *kiocb = &req->rw.kiocb; |
f67676d1 | 3464 | |
bcf5a063 JA |
3465 | /* never retry for NOWAIT, we just complete with -EAGAIN */ |
3466 | if (req->flags & REQ_F_NOWAIT) | |
3467 | return false; | |
f67676d1 | 3468 | |
227c0c96 | 3469 | /* Only for buffered IO */ |
3b2a4439 | 3470 | if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI)) |
bcf5a063 | 3471 | return false; |
3b2a4439 | 3472 | |
bcf5a063 JA |
3473 | /* |
3474 | * just use poll if we can, and don't attempt if the fs doesn't | |
3475 | * support callback based unlocks | |
3476 | */ | |
3477 | if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC)) | |
3478 | return false; | |
f67676d1 | 3479 | |
3b2a4439 JA |
3480 | wait->wait.func = io_async_buf_func; |
3481 | wait->wait.private = req; | |
3482 | wait->wait.flags = 0; | |
3483 | INIT_LIST_HEAD(&wait->wait.entry); | |
3484 | kiocb->ki_flags |= IOCB_WAITQ; | |
c8d317aa | 3485 | kiocb->ki_flags &= ~IOCB_NOWAIT; |
3b2a4439 | 3486 | kiocb->ki_waitq = wait; |
3b2a4439 | 3487 | return true; |
bcf5a063 JA |
3488 | } |
3489 | ||
3490 | static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter) | |
3491 | { | |
3492 | if (req->file->f_op->read_iter) | |
3493 | return call_read_iter(req->file, &req->rw.kiocb, iter); | |
2dd2111d | 3494 | else if (req->file->f_op->read) |
4017eb91 | 3495 | return loop_rw_iter(READ, req, iter); |
2dd2111d GH |
3496 | else |
3497 | return -EINVAL; | |
f67676d1 JA |
3498 | } |
3499 | ||
a1d7c393 JA |
3500 | static int io_read(struct io_kiocb *req, bool force_nonblock, |
3501 | struct io_comp_state *cs) | |
2b188cc1 JA |
3502 | { |
3503 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
9adbd45d | 3504 | struct kiocb *kiocb = &req->rw.kiocb; |
ff6165b2 | 3505 | struct iov_iter __iter, *iter = &__iter; |
e8c2bc1f | 3506 | struct io_async_rw *rw = req->async_data; |
227c0c96 | 3507 | ssize_t io_size, ret, ret2; |
f5cac8b1 | 3508 | bool no_async; |
ff6165b2 | 3509 | |
2846c481 | 3510 | if (rw) { |
e8c2bc1f | 3511 | iter = &rw->iter; |
2846c481 PB |
3512 | iovec = NULL; |
3513 | } else { | |
3514 | ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock); | |
3515 | if (ret < 0) | |
3516 | return ret; | |
3517 | } | |
632546c4 | 3518 | io_size = iov_iter_count(iter); |
fa15bafb | 3519 | req->result = io_size; |
227c0c96 | 3520 | ret = 0; |
2b188cc1 | 3521 | |
fd6c2e4c JA |
3522 | /* Ensure we clear previously set non-block flag */ |
3523 | if (!force_nonblock) | |
29de5f6a | 3524 | kiocb->ki_flags &= ~IOCB_NOWAIT; |
a88fc400 PB |
3525 | else |
3526 | kiocb->ki_flags |= IOCB_NOWAIT; | |
3527 | ||
fd6c2e4c | 3528 | |
24c74678 | 3529 | /* If the file doesn't support async, just async punt */ |
f5cac8b1 JA |
3530 | no_async = force_nonblock && !io_file_supports_async(req->file, READ); |
3531 | if (no_async) | |
f67676d1 | 3532 | goto copy_iov; |
9e645e11 | 3533 | |
632546c4 | 3534 | ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size); |
fa15bafb PB |
3535 | if (unlikely(ret)) |
3536 | goto out_free; | |
2b188cc1 | 3537 | |
227c0c96 | 3538 | ret = io_iter_do_read(req, iter); |
32960613 | 3539 | |
227c0c96 JA |
3540 | if (!ret) { |
3541 | goto done; | |
3542 | } else if (ret == -EIOCBQUEUED) { | |
3543 | ret = 0; | |
3544 | goto out_free; | |
3545 | } else if (ret == -EAGAIN) { | |
eefdf30f JA |
3546 | /* IOPOLL retry should happen for io-wq threads */ |
3547 | if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
f91daf56 | 3548 | goto done; |
355afaeb JA |
3549 | /* no retry on NONBLOCK marked file */ |
3550 | if (req->file->f_flags & O_NONBLOCK) | |
3551 | goto done; | |
84216315 | 3552 | /* some cases will consume bytes even on error returns */ |
632546c4 | 3553 | iov_iter_revert(iter, io_size - iov_iter_count(iter)); |
f38c7e3a JA |
3554 | ret = 0; |
3555 | goto copy_iov; | |
227c0c96 | 3556 | } else if (ret < 0) { |
00d23d51 JA |
3557 | /* make sure -ERESTARTSYS -> -EINTR is done */ |
3558 | goto done; | |
227c0c96 JA |
3559 | } |
3560 | ||
3561 | /* read it all, or we did blocking attempt. no retry. */ | |
f91daf56 | 3562 | if (!iov_iter_count(iter) || !force_nonblock || |
9a173346 | 3563 | (req->file->f_flags & O_NONBLOCK) || !(req->flags & REQ_F_ISREG)) |
227c0c96 JA |
3564 | goto done; |
3565 | ||
3566 | io_size -= ret; | |
3567 | copy_iov: | |
3568 | ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true); | |
3569 | if (ret2) { | |
3570 | ret = ret2; | |
3571 | goto out_free; | |
3572 | } | |
f5cac8b1 JA |
3573 | if (no_async) |
3574 | return -EAGAIN; | |
e8c2bc1f | 3575 | rw = req->async_data; |
227c0c96 JA |
3576 | /* it's copied and will be cleaned with ->io */ |
3577 | iovec = NULL; | |
3578 | /* now use our persistent iterator, if we aren't already */ | |
e8c2bc1f | 3579 | iter = &rw->iter; |
227c0c96 | 3580 | retry: |
e8c2bc1f | 3581 | rw->bytes_done += ret; |
227c0c96 JA |
3582 | /* if we can retry, do so with the callbacks armed */ |
3583 | if (!io_rw_should_retry(req)) { | |
fa15bafb PB |
3584 | kiocb->ki_flags &= ~IOCB_WAITQ; |
3585 | return -EAGAIN; | |
2b188cc1 | 3586 | } |
227c0c96 JA |
3587 | |
3588 | /* | |
3589 | * Now retry read with the IOCB_WAITQ parts set in the iocb. If we | |
3590 | * get -EIOCBQUEUED, then we'll get a notification when the desired | |
3591 | * page gets unlocked. We can also get a partial read here, and if we | |
3592 | * do, then just retry at the new offset. | |
3593 | */ | |
3594 | ret = io_iter_do_read(req, iter); | |
3595 | if (ret == -EIOCBQUEUED) { | |
3596 | ret = 0; | |
3597 | goto out_free; | |
3598 | } else if (ret > 0 && ret < io_size) { | |
3599 | /* we got some bytes, but not all. retry. */ | |
3600 | goto retry; | |
3601 | } | |
3602 | done: | |
3603 | kiocb_done(kiocb, ret, cs); | |
3604 | ret = 0; | |
f67676d1 | 3605 | out_free: |
f261c168 | 3606 | /* it's reportedly faster than delegating the null check to kfree() */ |
252917c3 | 3607 | if (iovec) |
6f2cc166 | 3608 | kfree(iovec); |
2b188cc1 JA |
3609 | return ret; |
3610 | } | |
3611 | ||
73debe68 | 3612 | static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f67676d1 JA |
3613 | { |
3614 | ssize_t ret; | |
3615 | ||
a88fc400 | 3616 | ret = io_prep_rw(req, sqe); |
3529d8c2 JA |
3617 | if (ret) |
3618 | return ret; | |
f67676d1 | 3619 | |
3529d8c2 JA |
3620 | if (unlikely(!(req->file->f_mode & FMODE_WRITE))) |
3621 | return -EBADF; | |
f67676d1 | 3622 | |
5f798bea | 3623 | /* either don't need iovec imported or already have it */ |
2d199895 | 3624 | if (!req->async_data) |
3529d8c2 | 3625 | return 0; |
73debe68 | 3626 | return io_rw_prep_async(req, WRITE); |
f67676d1 JA |
3627 | } |
3628 | ||
a1d7c393 JA |
3629 | static int io_write(struct io_kiocb *req, bool force_nonblock, |
3630 | struct io_comp_state *cs) | |
2b188cc1 JA |
3631 | { |
3632 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
9adbd45d | 3633 | struct kiocb *kiocb = &req->rw.kiocb; |
ff6165b2 | 3634 | struct iov_iter __iter, *iter = &__iter; |
e8c2bc1f | 3635 | struct io_async_rw *rw = req->async_data; |
fa15bafb | 3636 | ssize_t ret, ret2, io_size; |
2b188cc1 | 3637 | |
2846c481 | 3638 | if (rw) { |
e8c2bc1f | 3639 | iter = &rw->iter; |
2846c481 PB |
3640 | iovec = NULL; |
3641 | } else { | |
3642 | ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock); | |
3643 | if (ret < 0) | |
3644 | return ret; | |
3645 | } | |
632546c4 | 3646 | io_size = iov_iter_count(iter); |
fa15bafb | 3647 | req->result = io_size; |
2b188cc1 | 3648 | |
fd6c2e4c JA |
3649 | /* Ensure we clear previously set non-block flag */ |
3650 | if (!force_nonblock) | |
a88fc400 PB |
3651 | kiocb->ki_flags &= ~IOCB_NOWAIT; |
3652 | else | |
3653 | kiocb->ki_flags |= IOCB_NOWAIT; | |
fd6c2e4c | 3654 | |
24c74678 | 3655 | /* If the file doesn't support async, just async punt */ |
af197f50 | 3656 | if (force_nonblock && !io_file_supports_async(req->file, WRITE)) |
f67676d1 | 3657 | goto copy_iov; |
31b51510 | 3658 | |
10d59345 JA |
3659 | /* file path doesn't support NOWAIT for non-direct_IO */ |
3660 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) && | |
3661 | (req->flags & REQ_F_ISREG)) | |
f67676d1 | 3662 | goto copy_iov; |
31b51510 | 3663 | |
632546c4 | 3664 | ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size); |
fa15bafb PB |
3665 | if (unlikely(ret)) |
3666 | goto out_free; | |
4ed734b0 | 3667 | |
fa15bafb PB |
3668 | /* |
3669 | * Open-code file_start_write here to grab freeze protection, | |
3670 | * which will be released by another thread in | |
3671 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
3672 | * released so that it doesn't complain about the held lock when | |
3673 | * we return to userspace. | |
3674 | */ | |
3675 | if (req->flags & REQ_F_ISREG) { | |
8a3c84b6 | 3676 | sb_start_write(file_inode(req->file)->i_sb); |
fa15bafb PB |
3677 | __sb_writers_release(file_inode(req->file)->i_sb, |
3678 | SB_FREEZE_WRITE); | |
3679 | } | |
3680 | kiocb->ki_flags |= IOCB_WRITE; | |
4ed734b0 | 3681 | |
fa15bafb | 3682 | if (req->file->f_op->write_iter) |
ff6165b2 | 3683 | ret2 = call_write_iter(req->file, kiocb, iter); |
2dd2111d | 3684 | else if (req->file->f_op->write) |
4017eb91 | 3685 | ret2 = loop_rw_iter(WRITE, req, iter); |
2dd2111d GH |
3686 | else |
3687 | ret2 = -EINVAL; | |
4ed734b0 | 3688 | |
fa15bafb PB |
3689 | /* |
3690 | * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just | |
3691 | * retry them without IOCB_NOWAIT. | |
3692 | */ | |
3693 | if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT)) | |
3694 | ret2 = -EAGAIN; | |
355afaeb JA |
3695 | /* no retry on NONBLOCK marked file */ |
3696 | if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK)) | |
3697 | goto done; | |
fa15bafb | 3698 | if (!force_nonblock || ret2 != -EAGAIN) { |
eefdf30f JA |
3699 | /* IOPOLL retry should happen for io-wq threads */ |
3700 | if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN) | |
3701 | goto copy_iov; | |
355afaeb | 3702 | done: |
fa15bafb PB |
3703 | kiocb_done(kiocb, ret2, cs); |
3704 | } else { | |
f67676d1 | 3705 | copy_iov: |
84216315 | 3706 | /* some cases will consume bytes even on error returns */ |
632546c4 | 3707 | iov_iter_revert(iter, io_size - iov_iter_count(iter)); |
227c0c96 | 3708 | ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false); |
ff6165b2 JA |
3709 | if (!ret) |
3710 | return -EAGAIN; | |
2b188cc1 | 3711 | } |
31b51510 | 3712 | out_free: |
f261c168 | 3713 | /* it's reportedly faster than delegating the null check to kfree() */ |
252917c3 | 3714 | if (iovec) |
6f2cc166 | 3715 | kfree(iovec); |
2b188cc1 JA |
3716 | return ret; |
3717 | } | |
3718 | ||
80a261fd JA |
3719 | static int io_renameat_prep(struct io_kiocb *req, |
3720 | const struct io_uring_sqe *sqe) | |
3721 | { | |
3722 | struct io_rename *ren = &req->rename; | |
3723 | const char __user *oldf, *newf; | |
3724 | ||
3725 | if (unlikely(req->flags & REQ_F_FIXED_FILE)) | |
3726 | return -EBADF; | |
3727 | ||
3728 | ren->old_dfd = READ_ONCE(sqe->fd); | |
3729 | oldf = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
3730 | newf = u64_to_user_ptr(READ_ONCE(sqe->addr2)); | |
3731 | ren->new_dfd = READ_ONCE(sqe->len); | |
3732 | ren->flags = READ_ONCE(sqe->rename_flags); | |
3733 | ||
3734 | ren->oldpath = getname(oldf); | |
3735 | if (IS_ERR(ren->oldpath)) | |
3736 | return PTR_ERR(ren->oldpath); | |
3737 | ||
3738 | ren->newpath = getname(newf); | |
3739 | if (IS_ERR(ren->newpath)) { | |
3740 | putname(ren->oldpath); | |
3741 | return PTR_ERR(ren->newpath); | |
3742 | } | |
3743 | ||
3744 | req->flags |= REQ_F_NEED_CLEANUP; | |
3745 | return 0; | |
3746 | } | |
3747 | ||
3748 | static int io_renameat(struct io_kiocb *req, bool force_nonblock) | |
3749 | { | |
3750 | struct io_rename *ren = &req->rename; | |
3751 | int ret; | |
3752 | ||
3753 | if (force_nonblock) | |
3754 | return -EAGAIN; | |
3755 | ||
3756 | ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd, | |
3757 | ren->newpath, ren->flags); | |
3758 | ||
3759 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
3760 | if (ret < 0) | |
3761 | req_set_fail_links(req); | |
3762 | io_req_complete(req, ret); | |
3763 | return 0; | |
3764 | } | |
3765 | ||
14a1143b JA |
3766 | static int io_unlinkat_prep(struct io_kiocb *req, |
3767 | const struct io_uring_sqe *sqe) | |
3768 | { | |
3769 | struct io_unlink *un = &req->unlink; | |
3770 | const char __user *fname; | |
3771 | ||
3772 | if (unlikely(req->flags & REQ_F_FIXED_FILE)) | |
3773 | return -EBADF; | |
3774 | ||
3775 | un->dfd = READ_ONCE(sqe->fd); | |
3776 | ||
3777 | un->flags = READ_ONCE(sqe->unlink_flags); | |
3778 | if (un->flags & ~AT_REMOVEDIR) | |
3779 | return -EINVAL; | |
3780 | ||
3781 | fname = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
3782 | un->filename = getname(fname); | |
3783 | if (IS_ERR(un->filename)) | |
3784 | return PTR_ERR(un->filename); | |
3785 | ||
3786 | req->flags |= REQ_F_NEED_CLEANUP; | |
3787 | return 0; | |
3788 | } | |
3789 | ||
3790 | static int io_unlinkat(struct io_kiocb *req, bool force_nonblock) | |
3791 | { | |
3792 | struct io_unlink *un = &req->unlink; | |
3793 | int ret; | |
3794 | ||
3795 | if (force_nonblock) | |
3796 | return -EAGAIN; | |
3797 | ||
3798 | if (un->flags & AT_REMOVEDIR) | |
3799 | ret = do_rmdir(un->dfd, un->filename); | |
3800 | else | |
3801 | ret = do_unlinkat(un->dfd, un->filename); | |
3802 | ||
3803 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
3804 | if (ret < 0) | |
3805 | req_set_fail_links(req); | |
3806 | io_req_complete(req, ret); | |
3807 | return 0; | |
3808 | } | |
3809 | ||
36f4fa68 JA |
3810 | static int io_shutdown_prep(struct io_kiocb *req, |
3811 | const struct io_uring_sqe *sqe) | |
3812 | { | |
3813 | #if defined(CONFIG_NET) | |
3814 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
3815 | return -EINVAL; | |
3816 | if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags || | |
3817 | sqe->buf_index) | |
3818 | return -EINVAL; | |
3819 | ||
3820 | req->shutdown.how = READ_ONCE(sqe->len); | |
3821 | return 0; | |
3822 | #else | |
3823 | return -EOPNOTSUPP; | |
3824 | #endif | |
3825 | } | |
3826 | ||
3827 | static int io_shutdown(struct io_kiocb *req, bool force_nonblock) | |
3828 | { | |
3829 | #if defined(CONFIG_NET) | |
3830 | struct socket *sock; | |
3831 | int ret; | |
3832 | ||
3833 | if (force_nonblock) | |
3834 | return -EAGAIN; | |
3835 | ||
48aba79b | 3836 | sock = sock_from_file(req->file); |
36f4fa68 | 3837 | if (unlikely(!sock)) |
48aba79b | 3838 | return -ENOTSOCK; |
36f4fa68 JA |
3839 | |
3840 | ret = __sys_shutdown_sock(sock, req->shutdown.how); | |
a146468d JA |
3841 | if (ret < 0) |
3842 | req_set_fail_links(req); | |
36f4fa68 JA |
3843 | io_req_complete(req, ret); |
3844 | return 0; | |
3845 | #else | |
3846 | return -EOPNOTSUPP; | |
3847 | #endif | |
3848 | } | |
3849 | ||
f2a8d5c7 PB |
3850 | static int __io_splice_prep(struct io_kiocb *req, |
3851 | const struct io_uring_sqe *sqe) | |
7d67af2c PB |
3852 | { |
3853 | struct io_splice* sp = &req->splice; | |
3854 | unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL; | |
7d67af2c | 3855 | |
3232dd02 PB |
3856 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3857 | return -EINVAL; | |
7d67af2c PB |
3858 | |
3859 | sp->file_in = NULL; | |
7d67af2c PB |
3860 | sp->len = READ_ONCE(sqe->len); |
3861 | sp->flags = READ_ONCE(sqe->splice_flags); | |
3862 | ||
3863 | if (unlikely(sp->flags & ~valid_flags)) | |
3864 | return -EINVAL; | |
3865 | ||
8371adf5 PB |
3866 | sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), |
3867 | (sp->flags & SPLICE_F_FD_IN_FIXED)); | |
3868 | if (!sp->file_in) | |
3869 | return -EBADF; | |
7d67af2c PB |
3870 | req->flags |= REQ_F_NEED_CLEANUP; |
3871 | ||
7cdaf587 XW |
3872 | if (!S_ISREG(file_inode(sp->file_in)->i_mode)) { |
3873 | /* | |
3874 | * Splice operation will be punted aync, and here need to | |
3875 | * modify io_wq_work.flags, so initialize io_wq_work firstly. | |
3876 | */ | |
3877 | io_req_init_async(req); | |
7d67af2c | 3878 | req->work.flags |= IO_WQ_WORK_UNBOUND; |
7cdaf587 | 3879 | } |
7d67af2c PB |
3880 | |
3881 | return 0; | |
3882 | } | |
3883 | ||
f2a8d5c7 PB |
3884 | static int io_tee_prep(struct io_kiocb *req, |
3885 | const struct io_uring_sqe *sqe) | |
3886 | { | |
3887 | if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off)) | |
3888 | return -EINVAL; | |
3889 | return __io_splice_prep(req, sqe); | |
3890 | } | |
3891 | ||
3892 | static int io_tee(struct io_kiocb *req, bool force_nonblock) | |
3893 | { | |
3894 | struct io_splice *sp = &req->splice; | |
3895 | struct file *in = sp->file_in; | |
3896 | struct file *out = sp->file_out; | |
3897 | unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED; | |
3898 | long ret = 0; | |
3899 | ||
3900 | if (force_nonblock) | |
3901 | return -EAGAIN; | |
3902 | if (sp->len) | |
3903 | ret = do_tee(in, out, sp->len, flags); | |
3904 | ||
3905 | io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED)); | |
3906 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
3907 | ||
f2a8d5c7 PB |
3908 | if (ret != sp->len) |
3909 | req_set_fail_links(req); | |
e1e16097 | 3910 | io_req_complete(req, ret); |
f2a8d5c7 PB |
3911 | return 0; |
3912 | } | |
3913 | ||
3914 | static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
3915 | { | |
3916 | struct io_splice* sp = &req->splice; | |
3917 | ||
3918 | sp->off_in = READ_ONCE(sqe->splice_off_in); | |
3919 | sp->off_out = READ_ONCE(sqe->off); | |
3920 | return __io_splice_prep(req, sqe); | |
3921 | } | |
3922 | ||
014db007 | 3923 | static int io_splice(struct io_kiocb *req, bool force_nonblock) |
7d67af2c PB |
3924 | { |
3925 | struct io_splice *sp = &req->splice; | |
3926 | struct file *in = sp->file_in; | |
3927 | struct file *out = sp->file_out; | |
3928 | unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED; | |
3929 | loff_t *poff_in, *poff_out; | |
c9687426 | 3930 | long ret = 0; |
7d67af2c | 3931 | |
2fb3e822 PB |
3932 | if (force_nonblock) |
3933 | return -EAGAIN; | |
7d67af2c PB |
3934 | |
3935 | poff_in = (sp->off_in == -1) ? NULL : &sp->off_in; | |
3936 | poff_out = (sp->off_out == -1) ? NULL : &sp->off_out; | |
c9687426 | 3937 | |
948a7749 | 3938 | if (sp->len) |
c9687426 | 3939 | ret = do_splice(in, poff_in, out, poff_out, sp->len, flags); |
7d67af2c PB |
3940 | |
3941 | io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED)); | |
3942 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
3943 | ||
7d67af2c PB |
3944 | if (ret != sp->len) |
3945 | req_set_fail_links(req); | |
e1e16097 | 3946 | io_req_complete(req, ret); |
7d67af2c PB |
3947 | return 0; |
3948 | } | |
3949 | ||
2b188cc1 JA |
3950 | /* |
3951 | * IORING_OP_NOP just posts a completion event, nothing else. | |
3952 | */ | |
229a7b63 | 3953 | static int io_nop(struct io_kiocb *req, struct io_comp_state *cs) |
2b188cc1 JA |
3954 | { |
3955 | struct io_ring_ctx *ctx = req->ctx; | |
2b188cc1 | 3956 | |
def596e9 JA |
3957 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
3958 | return -EINVAL; | |
3959 | ||
229a7b63 | 3960 | __io_req_complete(req, 0, 0, cs); |
2b188cc1 JA |
3961 | return 0; |
3962 | } | |
3963 | ||
3529d8c2 | 3964 | static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
c992fe29 | 3965 | { |
6b06314c | 3966 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 3967 | |
09bb8394 JA |
3968 | if (!req->file) |
3969 | return -EBADF; | |
c992fe29 | 3970 | |
6b06314c | 3971 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 3972 | return -EINVAL; |
edafccee | 3973 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
3974 | return -EINVAL; |
3975 | ||
8ed8d3c3 JA |
3976 | req->sync.flags = READ_ONCE(sqe->fsync_flags); |
3977 | if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC)) | |
3978 | return -EINVAL; | |
3979 | ||
3980 | req->sync.off = READ_ONCE(sqe->off); | |
3981 | req->sync.len = READ_ONCE(sqe->len); | |
c992fe29 CH |
3982 | return 0; |
3983 | } | |
3984 | ||
ac45abc0 | 3985 | static int io_fsync(struct io_kiocb *req, bool force_nonblock) |
8ed8d3c3 | 3986 | { |
8ed8d3c3 | 3987 | loff_t end = req->sync.off + req->sync.len; |
8ed8d3c3 JA |
3988 | int ret; |
3989 | ||
ac45abc0 PB |
3990 | /* fsync always requires a blocking context */ |
3991 | if (force_nonblock) | |
3992 | return -EAGAIN; | |
3993 | ||
9adbd45d | 3994 | ret = vfs_fsync_range(req->file, req->sync.off, |
8ed8d3c3 JA |
3995 | end > 0 ? end : LLONG_MAX, |
3996 | req->sync.flags & IORING_FSYNC_DATASYNC); | |
3997 | if (ret < 0) | |
3998 | req_set_fail_links(req); | |
e1e16097 | 3999 | io_req_complete(req, ret); |
c992fe29 CH |
4000 | return 0; |
4001 | } | |
4002 | ||
d63d1b5e JA |
4003 | static int io_fallocate_prep(struct io_kiocb *req, |
4004 | const struct io_uring_sqe *sqe) | |
4005 | { | |
4006 | if (sqe->ioprio || sqe->buf_index || sqe->rw_flags) | |
4007 | return -EINVAL; | |
3232dd02 PB |
4008 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4009 | return -EINVAL; | |
d63d1b5e JA |
4010 | |
4011 | req->sync.off = READ_ONCE(sqe->off); | |
4012 | req->sync.len = READ_ONCE(sqe->addr); | |
4013 | req->sync.mode = READ_ONCE(sqe->len); | |
4014 | return 0; | |
4015 | } | |
4016 | ||
014db007 | 4017 | static int io_fallocate(struct io_kiocb *req, bool force_nonblock) |
5d17b4a4 | 4018 | { |
ac45abc0 PB |
4019 | int ret; |
4020 | ||
d63d1b5e | 4021 | /* fallocate always requiring blocking context */ |
ac45abc0 | 4022 | if (force_nonblock) |
5d17b4a4 | 4023 | return -EAGAIN; |
ac45abc0 PB |
4024 | ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off, |
4025 | req->sync.len); | |
ac45abc0 PB |
4026 | if (ret < 0) |
4027 | req_set_fail_links(req); | |
e1e16097 | 4028 | io_req_complete(req, ret); |
5d17b4a4 JA |
4029 | return 0; |
4030 | } | |
4031 | ||
ec65fea5 | 4032 | static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
b7bb4f7d | 4033 | { |
f8748881 | 4034 | const char __user *fname; |
15b71abe | 4035 | int ret; |
b7bb4f7d | 4036 | |
ec65fea5 | 4037 | if (unlikely(sqe->ioprio || sqe->buf_index)) |
15b71abe | 4038 | return -EINVAL; |
ec65fea5 | 4039 | if (unlikely(req->flags & REQ_F_FIXED_FILE)) |
cf3040ca | 4040 | return -EBADF; |
03b1230c | 4041 | |
ec65fea5 PB |
4042 | /* open.how should be already initialised */ |
4043 | if (!(req->open.how.flags & O_PATH) && force_o_largefile()) | |
08a1d26e | 4044 | req->open.how.flags |= O_LARGEFILE; |
3529d8c2 | 4045 | |
25e72d10 PB |
4046 | req->open.dfd = READ_ONCE(sqe->fd); |
4047 | fname = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
f8748881 | 4048 | req->open.filename = getname(fname); |
15b71abe JA |
4049 | if (IS_ERR(req->open.filename)) { |
4050 | ret = PTR_ERR(req->open.filename); | |
4051 | req->open.filename = NULL; | |
4052 | return ret; | |
4053 | } | |
4022e7af | 4054 | req->open.nofile = rlimit(RLIMIT_NOFILE); |
944d1444 | 4055 | req->open.ignore_nonblock = false; |
8fef80bf | 4056 | req->flags |= REQ_F_NEED_CLEANUP; |
15b71abe | 4057 | return 0; |
03b1230c JA |
4058 | } |
4059 | ||
ec65fea5 PB |
4060 | static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4061 | { | |
4062 | u64 flags, mode; | |
4063 | ||
14587a46 | 4064 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4eb8dded | 4065 | return -EINVAL; |
ec65fea5 PB |
4066 | mode = READ_ONCE(sqe->len); |
4067 | flags = READ_ONCE(sqe->open_flags); | |
4068 | req->open.how = build_open_how(flags, mode); | |
4069 | return __io_openat_prep(req, sqe); | |
4070 | } | |
4071 | ||
cebdb986 | 4072 | static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
aa1fa28f | 4073 | { |
cebdb986 | 4074 | struct open_how __user *how; |
cebdb986 | 4075 | size_t len; |
0fa03c62 JA |
4076 | int ret; |
4077 | ||
14587a46 | 4078 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4eb8dded | 4079 | return -EINVAL; |
cebdb986 JA |
4080 | how = u64_to_user_ptr(READ_ONCE(sqe->addr2)); |
4081 | len = READ_ONCE(sqe->len); | |
cebdb986 JA |
4082 | if (len < OPEN_HOW_SIZE_VER0) |
4083 | return -EINVAL; | |
3529d8c2 | 4084 | |
cebdb986 JA |
4085 | ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how, |
4086 | len); | |
4087 | if (ret) | |
4088 | return ret; | |
3529d8c2 | 4089 | |
ec65fea5 | 4090 | return __io_openat_prep(req, sqe); |
cebdb986 JA |
4091 | } |
4092 | ||
014db007 | 4093 | static int io_openat2(struct io_kiocb *req, bool force_nonblock) |
15b71abe JA |
4094 | { |
4095 | struct open_flags op; | |
15b71abe JA |
4096 | struct file *file; |
4097 | int ret; | |
4098 | ||
944d1444 | 4099 | if (force_nonblock && !req->open.ignore_nonblock) |
15b71abe | 4100 | return -EAGAIN; |
15b71abe | 4101 | |
cebdb986 | 4102 | ret = build_open_flags(&req->open.how, &op); |
15b71abe JA |
4103 | if (ret) |
4104 | goto err; | |
4105 | ||
4022e7af | 4106 | ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile); |
15b71abe JA |
4107 | if (ret < 0) |
4108 | goto err; | |
4109 | ||
4110 | file = do_filp_open(req->open.dfd, req->open.filename, &op); | |
4111 | if (IS_ERR(file)) { | |
4112 | put_unused_fd(ret); | |
4113 | ret = PTR_ERR(file); | |
944d1444 JA |
4114 | /* |
4115 | * A work-around to ensure that /proc/self works that way | |
4116 | * that it should - if we get -EOPNOTSUPP back, then assume | |
4117 | * that proc_self_get_link() failed us because we're in async | |
4118 | * context. We should be safe to retry this from the task | |
4119 | * itself with force_nonblock == false set, as it should not | |
4120 | * block on lookup. Would be nice to know this upfront and | |
4121 | * avoid the async dance, but doesn't seem feasible. | |
4122 | */ | |
4123 | if (ret == -EOPNOTSUPP && io_wq_current_is_worker()) { | |
4124 | req->open.ignore_nonblock = true; | |
4125 | refcount_inc(&req->refs); | |
4126 | io_req_task_queue(req); | |
4127 | return 0; | |
4128 | } | |
15b71abe JA |
4129 | } else { |
4130 | fsnotify_open(file); | |
4131 | fd_install(ret, file); | |
4132 | } | |
4133 | err: | |
4134 | putname(req->open.filename); | |
8fef80bf | 4135 | req->flags &= ~REQ_F_NEED_CLEANUP; |
15b71abe JA |
4136 | if (ret < 0) |
4137 | req_set_fail_links(req); | |
e1e16097 | 4138 | io_req_complete(req, ret); |
15b71abe JA |
4139 | return 0; |
4140 | } | |
4141 | ||
014db007 | 4142 | static int io_openat(struct io_kiocb *req, bool force_nonblock) |
cebdb986 | 4143 | { |
014db007 | 4144 | return io_openat2(req, force_nonblock); |
cebdb986 JA |
4145 | } |
4146 | ||
067524e9 JA |
4147 | static int io_remove_buffers_prep(struct io_kiocb *req, |
4148 | const struct io_uring_sqe *sqe) | |
4149 | { | |
4150 | struct io_provide_buf *p = &req->pbuf; | |
4151 | u64 tmp; | |
4152 | ||
4153 | if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off) | |
4154 | return -EINVAL; | |
4155 | ||
4156 | tmp = READ_ONCE(sqe->fd); | |
4157 | if (!tmp || tmp > USHRT_MAX) | |
4158 | return -EINVAL; | |
4159 | ||
4160 | memset(p, 0, sizeof(*p)); | |
4161 | p->nbufs = tmp; | |
4162 | p->bgid = READ_ONCE(sqe->buf_group); | |
4163 | return 0; | |
4164 | } | |
4165 | ||
4166 | static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf, | |
4167 | int bgid, unsigned nbufs) | |
4168 | { | |
4169 | unsigned i = 0; | |
4170 | ||
4171 | /* shouldn't happen */ | |
4172 | if (!nbufs) | |
4173 | return 0; | |
4174 | ||
4175 | /* the head kbuf is the list itself */ | |
4176 | while (!list_empty(&buf->list)) { | |
4177 | struct io_buffer *nxt; | |
4178 | ||
4179 | nxt = list_first_entry(&buf->list, struct io_buffer, list); | |
4180 | list_del(&nxt->list); | |
4181 | kfree(nxt); | |
4182 | if (++i == nbufs) | |
4183 | return i; | |
4184 | } | |
4185 | i++; | |
4186 | kfree(buf); | |
4187 | idr_remove(&ctx->io_buffer_idr, bgid); | |
4188 | ||
4189 | return i; | |
4190 | } | |
4191 | ||
229a7b63 JA |
4192 | static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock, |
4193 | struct io_comp_state *cs) | |
067524e9 JA |
4194 | { |
4195 | struct io_provide_buf *p = &req->pbuf; | |
4196 | struct io_ring_ctx *ctx = req->ctx; | |
4197 | struct io_buffer *head; | |
4198 | int ret = 0; | |
4199 | ||
4200 | io_ring_submit_lock(ctx, !force_nonblock); | |
4201 | ||
4202 | lockdep_assert_held(&ctx->uring_lock); | |
4203 | ||
4204 | ret = -ENOENT; | |
4205 | head = idr_find(&ctx->io_buffer_idr, p->bgid); | |
4206 | if (head) | |
4207 | ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs); | |
067524e9 JA |
4208 | if (ret < 0) |
4209 | req_set_fail_links(req); | |
067524e9 | 4210 | |
31bff9a5 PB |
4211 | /* need to hold the lock to complete IOPOLL requests */ |
4212 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
4213 | __io_req_complete(req, ret, 0, cs); | |
4214 | io_ring_submit_unlock(ctx, !force_nonblock); | |
4215 | } else { | |
4216 | io_ring_submit_unlock(ctx, !force_nonblock); | |
4217 | __io_req_complete(req, ret, 0, cs); | |
4218 | } | |
067524e9 JA |
4219 | return 0; |
4220 | } | |
4221 | ||
ddf0322d JA |
4222 | static int io_provide_buffers_prep(struct io_kiocb *req, |
4223 | const struct io_uring_sqe *sqe) | |
4224 | { | |
4225 | struct io_provide_buf *p = &req->pbuf; | |
4226 | u64 tmp; | |
4227 | ||
4228 | if (sqe->ioprio || sqe->rw_flags) | |
4229 | return -EINVAL; | |
4230 | ||
4231 | tmp = READ_ONCE(sqe->fd); | |
4232 | if (!tmp || tmp > USHRT_MAX) | |
4233 | return -E2BIG; | |
4234 | p->nbufs = tmp; | |
4235 | p->addr = READ_ONCE(sqe->addr); | |
4236 | p->len = READ_ONCE(sqe->len); | |
4237 | ||
efe68c1c | 4238 | if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs))) |
ddf0322d JA |
4239 | return -EFAULT; |
4240 | ||
4241 | p->bgid = READ_ONCE(sqe->buf_group); | |
4242 | tmp = READ_ONCE(sqe->off); | |
4243 | if (tmp > USHRT_MAX) | |
4244 | return -E2BIG; | |
4245 | p->bid = tmp; | |
4246 | return 0; | |
4247 | } | |
4248 | ||
4249 | static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head) | |
4250 | { | |
4251 | struct io_buffer *buf; | |
4252 | u64 addr = pbuf->addr; | |
4253 | int i, bid = pbuf->bid; | |
4254 | ||
4255 | for (i = 0; i < pbuf->nbufs; i++) { | |
4256 | buf = kmalloc(sizeof(*buf), GFP_KERNEL); | |
4257 | if (!buf) | |
4258 | break; | |
4259 | ||
4260 | buf->addr = addr; | |
4261 | buf->len = pbuf->len; | |
4262 | buf->bid = bid; | |
4263 | addr += pbuf->len; | |
4264 | bid++; | |
4265 | if (!*head) { | |
4266 | INIT_LIST_HEAD(&buf->list); | |
4267 | *head = buf; | |
4268 | } else { | |
4269 | list_add_tail(&buf->list, &(*head)->list); | |
4270 | } | |
4271 | } | |
4272 | ||
4273 | return i ? i : -ENOMEM; | |
4274 | } | |
4275 | ||
229a7b63 JA |
4276 | static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock, |
4277 | struct io_comp_state *cs) | |
ddf0322d JA |
4278 | { |
4279 | struct io_provide_buf *p = &req->pbuf; | |
4280 | struct io_ring_ctx *ctx = req->ctx; | |
4281 | struct io_buffer *head, *list; | |
4282 | int ret = 0; | |
4283 | ||
4284 | io_ring_submit_lock(ctx, !force_nonblock); | |
4285 | ||
4286 | lockdep_assert_held(&ctx->uring_lock); | |
4287 | ||
4288 | list = head = idr_find(&ctx->io_buffer_idr, p->bgid); | |
4289 | ||
4290 | ret = io_add_buffers(p, &head); | |
4291 | if (ret < 0) | |
4292 | goto out; | |
4293 | ||
4294 | if (!list) { | |
4295 | ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1, | |
4296 | GFP_KERNEL); | |
4297 | if (ret < 0) { | |
067524e9 | 4298 | __io_remove_buffers(ctx, head, p->bgid, -1U); |
ddf0322d JA |
4299 | goto out; |
4300 | } | |
4301 | } | |
4302 | out: | |
ddf0322d JA |
4303 | if (ret < 0) |
4304 | req_set_fail_links(req); | |
31bff9a5 PB |
4305 | |
4306 | /* need to hold the lock to complete IOPOLL requests */ | |
4307 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
4308 | __io_req_complete(req, ret, 0, cs); | |
4309 | io_ring_submit_unlock(ctx, !force_nonblock); | |
4310 | } else { | |
4311 | io_ring_submit_unlock(ctx, !force_nonblock); | |
4312 | __io_req_complete(req, ret, 0, cs); | |
4313 | } | |
ddf0322d | 4314 | return 0; |
cebdb986 JA |
4315 | } |
4316 | ||
3e4827b0 JA |
4317 | static int io_epoll_ctl_prep(struct io_kiocb *req, |
4318 | const struct io_uring_sqe *sqe) | |
4319 | { | |
4320 | #if defined(CONFIG_EPOLL) | |
4321 | if (sqe->ioprio || sqe->buf_index) | |
4322 | return -EINVAL; | |
6ca56f84 | 4323 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL))) |
3232dd02 | 4324 | return -EINVAL; |
3e4827b0 JA |
4325 | |
4326 | req->epoll.epfd = READ_ONCE(sqe->fd); | |
4327 | req->epoll.op = READ_ONCE(sqe->len); | |
4328 | req->epoll.fd = READ_ONCE(sqe->off); | |
4329 | ||
4330 | if (ep_op_has_event(req->epoll.op)) { | |
4331 | struct epoll_event __user *ev; | |
4332 | ||
4333 | ev = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
4334 | if (copy_from_user(&req->epoll.event, ev, sizeof(*ev))) | |
4335 | return -EFAULT; | |
4336 | } | |
4337 | ||
4338 | return 0; | |
4339 | #else | |
4340 | return -EOPNOTSUPP; | |
4341 | #endif | |
4342 | } | |
4343 | ||
229a7b63 JA |
4344 | static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock, |
4345 | struct io_comp_state *cs) | |
3e4827b0 JA |
4346 | { |
4347 | #if defined(CONFIG_EPOLL) | |
4348 | struct io_epoll *ie = &req->epoll; | |
4349 | int ret; | |
4350 | ||
4351 | ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock); | |
4352 | if (force_nonblock && ret == -EAGAIN) | |
4353 | return -EAGAIN; | |
4354 | ||
4355 | if (ret < 0) | |
4356 | req_set_fail_links(req); | |
229a7b63 | 4357 | __io_req_complete(req, ret, 0, cs); |
3e4827b0 JA |
4358 | return 0; |
4359 | #else | |
4360 | return -EOPNOTSUPP; | |
4361 | #endif | |
4362 | } | |
4363 | ||
c1ca757b JA |
4364 | static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4365 | { | |
4366 | #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU) | |
4367 | if (sqe->ioprio || sqe->buf_index || sqe->off) | |
4368 | return -EINVAL; | |
3232dd02 PB |
4369 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4370 | return -EINVAL; | |
c1ca757b JA |
4371 | |
4372 | req->madvise.addr = READ_ONCE(sqe->addr); | |
4373 | req->madvise.len = READ_ONCE(sqe->len); | |
4374 | req->madvise.advice = READ_ONCE(sqe->fadvise_advice); | |
4375 | return 0; | |
4376 | #else | |
4377 | return -EOPNOTSUPP; | |
4378 | #endif | |
4379 | } | |
4380 | ||
014db007 | 4381 | static int io_madvise(struct io_kiocb *req, bool force_nonblock) |
c1ca757b JA |
4382 | { |
4383 | #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU) | |
4384 | struct io_madvise *ma = &req->madvise; | |
4385 | int ret; | |
4386 | ||
4387 | if (force_nonblock) | |
4388 | return -EAGAIN; | |
4389 | ||
0726b01e | 4390 | ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice); |
c1ca757b JA |
4391 | if (ret < 0) |
4392 | req_set_fail_links(req); | |
e1e16097 | 4393 | io_req_complete(req, ret); |
c1ca757b JA |
4394 | return 0; |
4395 | #else | |
4396 | return -EOPNOTSUPP; | |
4397 | #endif | |
4398 | } | |
4399 | ||
4840e418 JA |
4400 | static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4401 | { | |
4402 | if (sqe->ioprio || sqe->buf_index || sqe->addr) | |
4403 | return -EINVAL; | |
3232dd02 PB |
4404 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4405 | return -EINVAL; | |
4840e418 JA |
4406 | |
4407 | req->fadvise.offset = READ_ONCE(sqe->off); | |
4408 | req->fadvise.len = READ_ONCE(sqe->len); | |
4409 | req->fadvise.advice = READ_ONCE(sqe->fadvise_advice); | |
4410 | return 0; | |
4411 | } | |
4412 | ||
014db007 | 4413 | static int io_fadvise(struct io_kiocb *req, bool force_nonblock) |
4840e418 JA |
4414 | { |
4415 | struct io_fadvise *fa = &req->fadvise; | |
4416 | int ret; | |
4417 | ||
3e69426d JA |
4418 | if (force_nonblock) { |
4419 | switch (fa->advice) { | |
4420 | case POSIX_FADV_NORMAL: | |
4421 | case POSIX_FADV_RANDOM: | |
4422 | case POSIX_FADV_SEQUENTIAL: | |
4423 | break; | |
4424 | default: | |
4425 | return -EAGAIN; | |
4426 | } | |
4427 | } | |
4840e418 JA |
4428 | |
4429 | ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice); | |
4430 | if (ret < 0) | |
4431 | req_set_fail_links(req); | |
e1e16097 | 4432 | io_req_complete(req, ret); |
4840e418 JA |
4433 | return 0; |
4434 | } | |
4435 | ||
eddc7ef5 JA |
4436 | static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4437 | { | |
6ca56f84 | 4438 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL))) |
3232dd02 | 4439 | return -EINVAL; |
eddc7ef5 JA |
4440 | if (sqe->ioprio || sqe->buf_index) |
4441 | return -EINVAL; | |
9c280f90 | 4442 | if (req->flags & REQ_F_FIXED_FILE) |
cf3040ca | 4443 | return -EBADF; |
eddc7ef5 | 4444 | |
1d9e1288 BM |
4445 | req->statx.dfd = READ_ONCE(sqe->fd); |
4446 | req->statx.mask = READ_ONCE(sqe->len); | |
e62753e4 | 4447 | req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
1d9e1288 BM |
4448 | req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2)); |
4449 | req->statx.flags = READ_ONCE(sqe->statx_flags); | |
eddc7ef5 JA |
4450 | |
4451 | return 0; | |
4452 | } | |
4453 | ||
014db007 | 4454 | static int io_statx(struct io_kiocb *req, bool force_nonblock) |
eddc7ef5 | 4455 | { |
1d9e1288 | 4456 | struct io_statx *ctx = &req->statx; |
eddc7ef5 JA |
4457 | int ret; |
4458 | ||
5b0bbee4 JA |
4459 | if (force_nonblock) { |
4460 | /* only need file table for an actual valid fd */ | |
4461 | if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD) | |
4462 | req->flags |= REQ_F_NO_FILE_TABLE; | |
eddc7ef5 | 4463 | return -EAGAIN; |
5b0bbee4 | 4464 | } |
eddc7ef5 | 4465 | |
e62753e4 BM |
4466 | ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask, |
4467 | ctx->buffer); | |
eddc7ef5 | 4468 | |
eddc7ef5 JA |
4469 | if (ret < 0) |
4470 | req_set_fail_links(req); | |
e1e16097 | 4471 | io_req_complete(req, ret); |
eddc7ef5 JA |
4472 | return 0; |
4473 | } | |
4474 | ||
b5dba59e JA |
4475 | static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4476 | { | |
4477 | /* | |
4478 | * If we queue this for async, it must not be cancellable. That would | |
7cdaf587 XW |
4479 | * leave the 'file' in an undeterminate state, and here need to modify |
4480 | * io_wq_work.flags, so initialize io_wq_work firstly. | |
b5dba59e | 4481 | */ |
7cdaf587 | 4482 | io_req_init_async(req); |
b5dba59e | 4483 | |
14587a46 | 4484 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3232dd02 | 4485 | return -EINVAL; |
b5dba59e JA |
4486 | if (sqe->ioprio || sqe->off || sqe->addr || sqe->len || |
4487 | sqe->rw_flags || sqe->buf_index) | |
4488 | return -EINVAL; | |
9c280f90 | 4489 | if (req->flags & REQ_F_FIXED_FILE) |
cf3040ca | 4490 | return -EBADF; |
b5dba59e JA |
4491 | |
4492 | req->close.fd = READ_ONCE(sqe->fd); | |
0f212204 | 4493 | if ((req->file && req->file->f_op == &io_uring_fops)) |
fd2206e4 | 4494 | return -EBADF; |
b5dba59e | 4495 | |
3af73b28 | 4496 | req->close.put_file = NULL; |
b5dba59e | 4497 | return 0; |
b5dba59e JA |
4498 | } |
4499 | ||
229a7b63 JA |
4500 | static int io_close(struct io_kiocb *req, bool force_nonblock, |
4501 | struct io_comp_state *cs) | |
b5dba59e | 4502 | { |
3af73b28 | 4503 | struct io_close *close = &req->close; |
b5dba59e JA |
4504 | int ret; |
4505 | ||
3af73b28 PB |
4506 | /* might be already done during nonblock submission */ |
4507 | if (!close->put_file) { | |
9fe83c43 | 4508 | ret = close_fd_get_file(close->fd, &close->put_file); |
3af73b28 PB |
4509 | if (ret < 0) |
4510 | return (ret == -ENOENT) ? -EBADF : ret; | |
4511 | } | |
b5dba59e JA |
4512 | |
4513 | /* if the file has a flush method, be safe and punt to async */ | |
3af73b28 | 4514 | if (close->put_file->f_op->flush && force_nonblock) { |
607ec89e JA |
4515 | /* not safe to cancel at this point */ |
4516 | req->work.flags |= IO_WQ_WORK_NO_CANCEL; | |
24c74678 PB |
4517 | /* was never set, but play safe */ |
4518 | req->flags &= ~REQ_F_NOWAIT; | |
0bf0eefd | 4519 | /* avoid grabbing files - we don't need the files */ |
24c74678 | 4520 | req->flags |= REQ_F_NO_FILE_TABLE; |
0bf0eefd | 4521 | return -EAGAIN; |
a2100672 | 4522 | } |
b5dba59e | 4523 | |
3af73b28 | 4524 | /* No ->flush() or already async, safely close from here */ |
98447d65 | 4525 | ret = filp_close(close->put_file, req->work.identity->files); |
3af73b28 PB |
4526 | if (ret < 0) |
4527 | req_set_fail_links(req); | |
3af73b28 PB |
4528 | fput(close->put_file); |
4529 | close->put_file = NULL; | |
229a7b63 | 4530 | __io_req_complete(req, ret, 0, cs); |
1a417f4e | 4531 | return 0; |
b5dba59e JA |
4532 | } |
4533 | ||
3529d8c2 | 4534 | static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
5d17b4a4 JA |
4535 | { |
4536 | struct io_ring_ctx *ctx = req->ctx; | |
5d17b4a4 JA |
4537 | |
4538 | if (!req->file) | |
4539 | return -EBADF; | |
5d17b4a4 JA |
4540 | |
4541 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
4542 | return -EINVAL; | |
4543 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) | |
4544 | return -EINVAL; | |
4545 | ||
8ed8d3c3 JA |
4546 | req->sync.off = READ_ONCE(sqe->off); |
4547 | req->sync.len = READ_ONCE(sqe->len); | |
4548 | req->sync.flags = READ_ONCE(sqe->sync_range_flags); | |
8ed8d3c3 JA |
4549 | return 0; |
4550 | } | |
4551 | ||
ac45abc0 | 4552 | static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock) |
8ed8d3c3 | 4553 | { |
8ed8d3c3 JA |
4554 | int ret; |
4555 | ||
ac45abc0 PB |
4556 | /* sync_file_range always requires a blocking context */ |
4557 | if (force_nonblock) | |
4558 | return -EAGAIN; | |
4559 | ||
9adbd45d | 4560 | ret = sync_file_range(req->file, req->sync.off, req->sync.len, |
8ed8d3c3 JA |
4561 | req->sync.flags); |
4562 | if (ret < 0) | |
4563 | req_set_fail_links(req); | |
e1e16097 | 4564 | io_req_complete(req, ret); |
5d17b4a4 JA |
4565 | return 0; |
4566 | } | |
4567 | ||
469956e8 | 4568 | #if defined(CONFIG_NET) |
02d27d89 PB |
4569 | static int io_setup_async_msg(struct io_kiocb *req, |
4570 | struct io_async_msghdr *kmsg) | |
4571 | { | |
e8c2bc1f JA |
4572 | struct io_async_msghdr *async_msg = req->async_data; |
4573 | ||
4574 | if (async_msg) | |
02d27d89 | 4575 | return -EAGAIN; |
e8c2bc1f | 4576 | if (io_alloc_async_data(req)) { |
02d27d89 PB |
4577 | if (kmsg->iov != kmsg->fast_iov) |
4578 | kfree(kmsg->iov); | |
4579 | return -ENOMEM; | |
4580 | } | |
e8c2bc1f | 4581 | async_msg = req->async_data; |
02d27d89 | 4582 | req->flags |= REQ_F_NEED_CLEANUP; |
e8c2bc1f | 4583 | memcpy(async_msg, kmsg, sizeof(*kmsg)); |
02d27d89 PB |
4584 | return -EAGAIN; |
4585 | } | |
4586 | ||
2ae523ed PB |
4587 | static int io_sendmsg_copy_hdr(struct io_kiocb *req, |
4588 | struct io_async_msghdr *iomsg) | |
4589 | { | |
4590 | iomsg->iov = iomsg->fast_iov; | |
4591 | iomsg->msg.msg_name = &iomsg->addr; | |
4592 | return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg, | |
4593 | req->sr_msg.msg_flags, &iomsg->iov); | |
4594 | } | |
4595 | ||
3529d8c2 | 4596 | static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
03b1230c | 4597 | { |
e8c2bc1f | 4598 | struct io_async_msghdr *async_msg = req->async_data; |
e47293fd | 4599 | struct io_sr_msg *sr = &req->sr_msg; |
99bc4c38 | 4600 | int ret; |
03b1230c | 4601 | |
d2b6f48b PB |
4602 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4603 | return -EINVAL; | |
4604 | ||
e47293fd | 4605 | sr->msg_flags = READ_ONCE(sqe->msg_flags); |
270a5940 | 4606 | sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
fddaface | 4607 | sr->len = READ_ONCE(sqe->len); |
3529d8c2 | 4608 | |
d8768362 JA |
4609 | #ifdef CONFIG_COMPAT |
4610 | if (req->ctx->compat) | |
4611 | sr->msg_flags |= MSG_CMSG_COMPAT; | |
4612 | #endif | |
4613 | ||
e8c2bc1f | 4614 | if (!async_msg || !io_op_defs[req->opcode].needs_async_data) |
3529d8c2 | 4615 | return 0; |
e8c2bc1f | 4616 | ret = io_sendmsg_copy_hdr(req, async_msg); |
99bc4c38 PB |
4617 | if (!ret) |
4618 | req->flags |= REQ_F_NEED_CLEANUP; | |
4619 | return ret; | |
03b1230c JA |
4620 | } |
4621 | ||
229a7b63 JA |
4622 | static int io_sendmsg(struct io_kiocb *req, bool force_nonblock, |
4623 | struct io_comp_state *cs) | |
aa1fa28f | 4624 | { |
6b754c8b | 4625 | struct io_async_msghdr iomsg, *kmsg; |
0fa03c62 | 4626 | struct socket *sock; |
7a7cacba | 4627 | unsigned flags; |
0fa03c62 JA |
4628 | int ret; |
4629 | ||
dba4a925 | 4630 | sock = sock_from_file(req->file); |
7a7cacba | 4631 | if (unlikely(!sock)) |
dba4a925 | 4632 | return -ENOTSOCK; |
3529d8c2 | 4633 | |
e8c2bc1f JA |
4634 | if (req->async_data) { |
4635 | kmsg = req->async_data; | |
4636 | kmsg->msg.msg_name = &kmsg->addr; | |
7a7cacba PB |
4637 | /* if iov is set, it's allocated already */ |
4638 | if (!kmsg->iov) | |
4639 | kmsg->iov = kmsg->fast_iov; | |
4640 | kmsg->msg.msg_iter.iov = kmsg->iov; | |
4641 | } else { | |
4642 | ret = io_sendmsg_copy_hdr(req, &iomsg); | |
4643 | if (ret) | |
4644 | return ret; | |
4645 | kmsg = &iomsg; | |
0fa03c62 | 4646 | } |
0fa03c62 | 4647 | |
7a7cacba PB |
4648 | flags = req->sr_msg.msg_flags; |
4649 | if (flags & MSG_DONTWAIT) | |
4650 | req->flags |= REQ_F_NOWAIT; | |
4651 | else if (force_nonblock) | |
4652 | flags |= MSG_DONTWAIT; | |
e47293fd | 4653 | |
7a7cacba PB |
4654 | ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags); |
4655 | if (force_nonblock && ret == -EAGAIN) | |
4656 | return io_setup_async_msg(req, kmsg); | |
4657 | if (ret == -ERESTARTSYS) | |
4658 | ret = -EINTR; | |
0fa03c62 | 4659 | |
6b754c8b | 4660 | if (kmsg->iov != kmsg->fast_iov) |
0b416c3e | 4661 | kfree(kmsg->iov); |
99bc4c38 | 4662 | req->flags &= ~REQ_F_NEED_CLEANUP; |
4e88d6e7 JA |
4663 | if (ret < 0) |
4664 | req_set_fail_links(req); | |
229a7b63 | 4665 | __io_req_complete(req, ret, 0, cs); |
5d17b4a4 | 4666 | return 0; |
03b1230c | 4667 | } |
aa1fa28f | 4668 | |
229a7b63 JA |
4669 | static int io_send(struct io_kiocb *req, bool force_nonblock, |
4670 | struct io_comp_state *cs) | |
fddaface | 4671 | { |
7a7cacba PB |
4672 | struct io_sr_msg *sr = &req->sr_msg; |
4673 | struct msghdr msg; | |
4674 | struct iovec iov; | |
fddaface | 4675 | struct socket *sock; |
7a7cacba | 4676 | unsigned flags; |
fddaface JA |
4677 | int ret; |
4678 | ||
dba4a925 | 4679 | sock = sock_from_file(req->file); |
7a7cacba | 4680 | if (unlikely(!sock)) |
dba4a925 | 4681 | return -ENOTSOCK; |
fddaface | 4682 | |
7a7cacba PB |
4683 | ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter); |
4684 | if (unlikely(ret)) | |
14db8411 | 4685 | return ret; |
fddaface | 4686 | |
7a7cacba PB |
4687 | msg.msg_name = NULL; |
4688 | msg.msg_control = NULL; | |
4689 | msg.msg_controllen = 0; | |
4690 | msg.msg_namelen = 0; | |
fddaface | 4691 | |
7a7cacba PB |
4692 | flags = req->sr_msg.msg_flags; |
4693 | if (flags & MSG_DONTWAIT) | |
4694 | req->flags |= REQ_F_NOWAIT; | |
4695 | else if (force_nonblock) | |
4696 | flags |= MSG_DONTWAIT; | |
fddaface | 4697 | |
7a7cacba PB |
4698 | msg.msg_flags = flags; |
4699 | ret = sock_sendmsg(sock, &msg); | |
4700 | if (force_nonblock && ret == -EAGAIN) | |
4701 | return -EAGAIN; | |
4702 | if (ret == -ERESTARTSYS) | |
4703 | ret = -EINTR; | |
fddaface | 4704 | |
fddaface JA |
4705 | if (ret < 0) |
4706 | req_set_fail_links(req); | |
229a7b63 | 4707 | __io_req_complete(req, ret, 0, cs); |
fddaface | 4708 | return 0; |
fddaface JA |
4709 | } |
4710 | ||
1400e697 PB |
4711 | static int __io_recvmsg_copy_hdr(struct io_kiocb *req, |
4712 | struct io_async_msghdr *iomsg) | |
52de1fe1 JA |
4713 | { |
4714 | struct io_sr_msg *sr = &req->sr_msg; | |
4715 | struct iovec __user *uiov; | |
4716 | size_t iov_len; | |
4717 | int ret; | |
4718 | ||
1400e697 PB |
4719 | ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg, |
4720 | &iomsg->uaddr, &uiov, &iov_len); | |
52de1fe1 JA |
4721 | if (ret) |
4722 | return ret; | |
4723 | ||
4724 | if (req->flags & REQ_F_BUFFER_SELECT) { | |
4725 | if (iov_len > 1) | |
4726 | return -EINVAL; | |
1400e697 | 4727 | if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov))) |
52de1fe1 | 4728 | return -EFAULT; |
1400e697 PB |
4729 | sr->len = iomsg->iov[0].iov_len; |
4730 | iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1, | |
52de1fe1 | 4731 | sr->len); |
1400e697 | 4732 | iomsg->iov = NULL; |
52de1fe1 | 4733 | } else { |
89cd35c5 CH |
4734 | ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV, |
4735 | &iomsg->iov, &iomsg->msg.msg_iter, | |
4736 | false); | |
52de1fe1 JA |
4737 | if (ret > 0) |
4738 | ret = 0; | |
4739 | } | |
4740 | ||
4741 | return ret; | |
4742 | } | |
4743 | ||
4744 | #ifdef CONFIG_COMPAT | |
4745 | static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req, | |
1400e697 | 4746 | struct io_async_msghdr *iomsg) |
52de1fe1 JA |
4747 | { |
4748 | struct compat_msghdr __user *msg_compat; | |
4749 | struct io_sr_msg *sr = &req->sr_msg; | |
4750 | struct compat_iovec __user *uiov; | |
4751 | compat_uptr_t ptr; | |
4752 | compat_size_t len; | |
4753 | int ret; | |
4754 | ||
270a5940 | 4755 | msg_compat = (struct compat_msghdr __user *) sr->umsg; |
1400e697 | 4756 | ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr, |
52de1fe1 JA |
4757 | &ptr, &len); |
4758 | if (ret) | |
4759 | return ret; | |
4760 | ||
4761 | uiov = compat_ptr(ptr); | |
4762 | if (req->flags & REQ_F_BUFFER_SELECT) { | |
4763 | compat_ssize_t clen; | |
4764 | ||
4765 | if (len > 1) | |
4766 | return -EINVAL; | |
4767 | if (!access_ok(uiov, sizeof(*uiov))) | |
4768 | return -EFAULT; | |
4769 | if (__get_user(clen, &uiov->iov_len)) | |
4770 | return -EFAULT; | |
4771 | if (clen < 0) | |
4772 | return -EINVAL; | |
2d280bc8 PB |
4773 | sr->len = clen; |
4774 | iomsg->iov[0].iov_len = clen; | |
1400e697 | 4775 | iomsg->iov = NULL; |
52de1fe1 | 4776 | } else { |
89cd35c5 CH |
4777 | ret = __import_iovec(READ, (struct iovec __user *)uiov, len, |
4778 | UIO_FASTIOV, &iomsg->iov, | |
4779 | &iomsg->msg.msg_iter, true); | |
52de1fe1 JA |
4780 | if (ret < 0) |
4781 | return ret; | |
4782 | } | |
4783 | ||
4784 | return 0; | |
4785 | } | |
4786 | #endif | |
4787 | ||
1400e697 PB |
4788 | static int io_recvmsg_copy_hdr(struct io_kiocb *req, |
4789 | struct io_async_msghdr *iomsg) | |
52de1fe1 | 4790 | { |
1400e697 PB |
4791 | iomsg->msg.msg_name = &iomsg->addr; |
4792 | iomsg->iov = iomsg->fast_iov; | |
52de1fe1 JA |
4793 | |
4794 | #ifdef CONFIG_COMPAT | |
4795 | if (req->ctx->compat) | |
1400e697 | 4796 | return __io_compat_recvmsg_copy_hdr(req, iomsg); |
fddaface | 4797 | #endif |
52de1fe1 | 4798 | |
1400e697 | 4799 | return __io_recvmsg_copy_hdr(req, iomsg); |
52de1fe1 JA |
4800 | } |
4801 | ||
bcda7baa | 4802 | static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req, |
7fbb1b54 | 4803 | bool needs_lock) |
bcda7baa JA |
4804 | { |
4805 | struct io_sr_msg *sr = &req->sr_msg; | |
4806 | struct io_buffer *kbuf; | |
4807 | ||
bcda7baa JA |
4808 | kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock); |
4809 | if (IS_ERR(kbuf)) | |
4810 | return kbuf; | |
4811 | ||
4812 | sr->kbuf = kbuf; | |
4813 | req->flags |= REQ_F_BUFFER_SELECTED; | |
bcda7baa | 4814 | return kbuf; |
fddaface JA |
4815 | } |
4816 | ||
7fbb1b54 PB |
4817 | static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req) |
4818 | { | |
4819 | return io_put_kbuf(req, req->sr_msg.kbuf); | |
4820 | } | |
4821 | ||
3529d8c2 JA |
4822 | static int io_recvmsg_prep(struct io_kiocb *req, |
4823 | const struct io_uring_sqe *sqe) | |
aa1fa28f | 4824 | { |
e8c2bc1f | 4825 | struct io_async_msghdr *async_msg = req->async_data; |
e47293fd | 4826 | struct io_sr_msg *sr = &req->sr_msg; |
99bc4c38 | 4827 | int ret; |
3529d8c2 | 4828 | |
d2b6f48b PB |
4829 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4830 | return -EINVAL; | |
4831 | ||
3529d8c2 | 4832 | sr->msg_flags = READ_ONCE(sqe->msg_flags); |
270a5940 | 4833 | sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
0b7b21e4 | 4834 | sr->len = READ_ONCE(sqe->len); |
bcda7baa | 4835 | sr->bgid = READ_ONCE(sqe->buf_group); |
06b76d44 | 4836 | |
d8768362 JA |
4837 | #ifdef CONFIG_COMPAT |
4838 | if (req->ctx->compat) | |
4839 | sr->msg_flags |= MSG_CMSG_COMPAT; | |
4840 | #endif | |
4841 | ||
e8c2bc1f | 4842 | if (!async_msg || !io_op_defs[req->opcode].needs_async_data) |
5f798bea | 4843 | return 0; |
e8c2bc1f | 4844 | ret = io_recvmsg_copy_hdr(req, async_msg); |
99bc4c38 PB |
4845 | if (!ret) |
4846 | req->flags |= REQ_F_NEED_CLEANUP; | |
4847 | return ret; | |
aa1fa28f JA |
4848 | } |
4849 | ||
229a7b63 JA |
4850 | static int io_recvmsg(struct io_kiocb *req, bool force_nonblock, |
4851 | struct io_comp_state *cs) | |
aa1fa28f | 4852 | { |
6b754c8b | 4853 | struct io_async_msghdr iomsg, *kmsg; |
03b1230c | 4854 | struct socket *sock; |
7fbb1b54 | 4855 | struct io_buffer *kbuf; |
7a7cacba | 4856 | unsigned flags; |
52de1fe1 | 4857 | int ret, cflags = 0; |
03b1230c | 4858 | |
dba4a925 | 4859 | sock = sock_from_file(req->file); |
7a7cacba | 4860 | if (unlikely(!sock)) |
dba4a925 | 4861 | return -ENOTSOCK; |
3529d8c2 | 4862 | |
e8c2bc1f JA |
4863 | if (req->async_data) { |
4864 | kmsg = req->async_data; | |
4865 | kmsg->msg.msg_name = &kmsg->addr; | |
7a7cacba PB |
4866 | /* if iov is set, it's allocated already */ |
4867 | if (!kmsg->iov) | |
4868 | kmsg->iov = kmsg->fast_iov; | |
4869 | kmsg->msg.msg_iter.iov = kmsg->iov; | |
4870 | } else { | |
4871 | ret = io_recvmsg_copy_hdr(req, &iomsg); | |
4872 | if (ret) | |
681fda8d | 4873 | return ret; |
7a7cacba PB |
4874 | kmsg = &iomsg; |
4875 | } | |
03b1230c | 4876 | |
bc02ef33 | 4877 | if (req->flags & REQ_F_BUFFER_SELECT) { |
7fbb1b54 | 4878 | kbuf = io_recv_buffer_select(req, !force_nonblock); |
bc02ef33 | 4879 | if (IS_ERR(kbuf)) |
52de1fe1 | 4880 | return PTR_ERR(kbuf); |
7a7cacba PB |
4881 | kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr); |
4882 | iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov, | |
4883 | 1, req->sr_msg.len); | |
4884 | } | |
52de1fe1 | 4885 | |
7a7cacba PB |
4886 | flags = req->sr_msg.msg_flags; |
4887 | if (flags & MSG_DONTWAIT) | |
4888 | req->flags |= REQ_F_NOWAIT; | |
4889 | else if (force_nonblock) | |
4890 | flags |= MSG_DONTWAIT; | |
e47293fd | 4891 | |
7a7cacba PB |
4892 | ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg, |
4893 | kmsg->uaddr, flags); | |
0e1b6fe3 PB |
4894 | if (force_nonblock && ret == -EAGAIN) |
4895 | return io_setup_async_msg(req, kmsg); | |
7a7cacba PB |
4896 | if (ret == -ERESTARTSYS) |
4897 | ret = -EINTR; | |
03b1230c | 4898 | |
7fbb1b54 PB |
4899 | if (req->flags & REQ_F_BUFFER_SELECTED) |
4900 | cflags = io_put_recv_kbuf(req); | |
6b754c8b | 4901 | if (kmsg->iov != kmsg->fast_iov) |
0b416c3e | 4902 | kfree(kmsg->iov); |
99bc4c38 | 4903 | req->flags &= ~REQ_F_NEED_CLEANUP; |
4e88d6e7 JA |
4904 | if (ret < 0) |
4905 | req_set_fail_links(req); | |
229a7b63 | 4906 | __io_req_complete(req, ret, cflags, cs); |
03b1230c | 4907 | return 0; |
0fa03c62 | 4908 | } |
5d17b4a4 | 4909 | |
229a7b63 JA |
4910 | static int io_recv(struct io_kiocb *req, bool force_nonblock, |
4911 | struct io_comp_state *cs) | |
fddaface | 4912 | { |
6b754c8b | 4913 | struct io_buffer *kbuf; |
7a7cacba PB |
4914 | struct io_sr_msg *sr = &req->sr_msg; |
4915 | struct msghdr msg; | |
4916 | void __user *buf = sr->buf; | |
fddaface | 4917 | struct socket *sock; |
7a7cacba PB |
4918 | struct iovec iov; |
4919 | unsigned flags; | |
bcda7baa | 4920 | int ret, cflags = 0; |
fddaface | 4921 | |
dba4a925 | 4922 | sock = sock_from_file(req->file); |
7a7cacba | 4923 | if (unlikely(!sock)) |
dba4a925 | 4924 | return -ENOTSOCK; |
fddaface | 4925 | |
bc02ef33 | 4926 | if (req->flags & REQ_F_BUFFER_SELECT) { |
7fbb1b54 | 4927 | kbuf = io_recv_buffer_select(req, !force_nonblock); |
bcda7baa JA |
4928 | if (IS_ERR(kbuf)) |
4929 | return PTR_ERR(kbuf); | |
7a7cacba | 4930 | buf = u64_to_user_ptr(kbuf->addr); |
bc02ef33 | 4931 | } |
bcda7baa | 4932 | |
7a7cacba | 4933 | ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter); |
14c32eee PB |
4934 | if (unlikely(ret)) |
4935 | goto out_free; | |
fddaface | 4936 | |
7a7cacba PB |
4937 | msg.msg_name = NULL; |
4938 | msg.msg_control = NULL; | |
4939 | msg.msg_controllen = 0; | |
4940 | msg.msg_namelen = 0; | |
4941 | msg.msg_iocb = NULL; | |
4942 | msg.msg_flags = 0; | |
fddaface | 4943 | |
7a7cacba PB |
4944 | flags = req->sr_msg.msg_flags; |
4945 | if (flags & MSG_DONTWAIT) | |
4946 | req->flags |= REQ_F_NOWAIT; | |
4947 | else if (force_nonblock) | |
4948 | flags |= MSG_DONTWAIT; | |
4949 | ||
4950 | ret = sock_recvmsg(sock, &msg, flags); | |
4951 | if (force_nonblock && ret == -EAGAIN) | |
4952 | return -EAGAIN; | |
4953 | if (ret == -ERESTARTSYS) | |
4954 | ret = -EINTR; | |
14c32eee | 4955 | out_free: |
7fbb1b54 PB |
4956 | if (req->flags & REQ_F_BUFFER_SELECTED) |
4957 | cflags = io_put_recv_kbuf(req); | |
fddaface JA |
4958 | if (ret < 0) |
4959 | req_set_fail_links(req); | |
229a7b63 | 4960 | __io_req_complete(req, ret, cflags, cs); |
fddaface | 4961 | return 0; |
fddaface JA |
4962 | } |
4963 | ||
3529d8c2 | 4964 | static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
17f2fe35 | 4965 | { |
8ed8d3c3 JA |
4966 | struct io_accept *accept = &req->accept; |
4967 | ||
14587a46 | 4968 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
17f2fe35 | 4969 | return -EINVAL; |
8042d6ce | 4970 | if (sqe->ioprio || sqe->len || sqe->buf_index) |
17f2fe35 JA |
4971 | return -EINVAL; |
4972 | ||
d55e5f5b JA |
4973 | accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
4974 | accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2)); | |
8ed8d3c3 | 4975 | accept->flags = READ_ONCE(sqe->accept_flags); |
09952e3e | 4976 | accept->nofile = rlimit(RLIMIT_NOFILE); |
8ed8d3c3 | 4977 | return 0; |
8ed8d3c3 | 4978 | } |
17f2fe35 | 4979 | |
229a7b63 JA |
4980 | static int io_accept(struct io_kiocb *req, bool force_nonblock, |
4981 | struct io_comp_state *cs) | |
8ed8d3c3 JA |
4982 | { |
4983 | struct io_accept *accept = &req->accept; | |
ac45abc0 | 4984 | unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0; |
8ed8d3c3 JA |
4985 | int ret; |
4986 | ||
e697deed JX |
4987 | if (req->file->f_flags & O_NONBLOCK) |
4988 | req->flags |= REQ_F_NOWAIT; | |
4989 | ||
8ed8d3c3 | 4990 | ret = __sys_accept4_file(req->file, file_flags, accept->addr, |
09952e3e JA |
4991 | accept->addr_len, accept->flags, |
4992 | accept->nofile); | |
8ed8d3c3 | 4993 | if (ret == -EAGAIN && force_nonblock) |
17f2fe35 | 4994 | return -EAGAIN; |
ac45abc0 PB |
4995 | if (ret < 0) { |
4996 | if (ret == -ERESTARTSYS) | |
4997 | ret = -EINTR; | |
4e88d6e7 | 4998 | req_set_fail_links(req); |
ac45abc0 | 4999 | } |
229a7b63 | 5000 | __io_req_complete(req, ret, 0, cs); |
17f2fe35 | 5001 | return 0; |
8ed8d3c3 JA |
5002 | } |
5003 | ||
3529d8c2 | 5004 | static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f499a021 | 5005 | { |
3529d8c2 | 5006 | struct io_connect *conn = &req->connect; |
e8c2bc1f | 5007 | struct io_async_connect *io = req->async_data; |
f499a021 | 5008 | |
14587a46 | 5009 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3fbb51c1 JA |
5010 | return -EINVAL; |
5011 | if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags) | |
5012 | return -EINVAL; | |
5013 | ||
3529d8c2 JA |
5014 | conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
5015 | conn->addr_len = READ_ONCE(sqe->addr2); | |
5016 | ||
5017 | if (!io) | |
5018 | return 0; | |
5019 | ||
5020 | return move_addr_to_kernel(conn->addr, conn->addr_len, | |
e8c2bc1f | 5021 | &io->address); |
f499a021 JA |
5022 | } |
5023 | ||
229a7b63 JA |
5024 | static int io_connect(struct io_kiocb *req, bool force_nonblock, |
5025 | struct io_comp_state *cs) | |
f8e85cf2 | 5026 | { |
e8c2bc1f | 5027 | struct io_async_connect __io, *io; |
f8e85cf2 | 5028 | unsigned file_flags; |
3fbb51c1 | 5029 | int ret; |
f8e85cf2 | 5030 | |
e8c2bc1f JA |
5031 | if (req->async_data) { |
5032 | io = req->async_data; | |
f499a021 | 5033 | } else { |
3529d8c2 JA |
5034 | ret = move_addr_to_kernel(req->connect.addr, |
5035 | req->connect.addr_len, | |
e8c2bc1f | 5036 | &__io.address); |
f499a021 JA |
5037 | if (ret) |
5038 | goto out; | |
5039 | io = &__io; | |
5040 | } | |
5041 | ||
3fbb51c1 JA |
5042 | file_flags = force_nonblock ? O_NONBLOCK : 0; |
5043 | ||
e8c2bc1f | 5044 | ret = __sys_connect_file(req->file, &io->address, |
3fbb51c1 | 5045 | req->connect.addr_len, file_flags); |
87f80d62 | 5046 | if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) { |
e8c2bc1f | 5047 | if (req->async_data) |
b7bb4f7d | 5048 | return -EAGAIN; |
e8c2bc1f | 5049 | if (io_alloc_async_data(req)) { |
f499a021 JA |
5050 | ret = -ENOMEM; |
5051 | goto out; | |
5052 | } | |
e8c2bc1f JA |
5053 | io = req->async_data; |
5054 | memcpy(req->async_data, &__io, sizeof(__io)); | |
f8e85cf2 | 5055 | return -EAGAIN; |
f499a021 | 5056 | } |
f8e85cf2 JA |
5057 | if (ret == -ERESTARTSYS) |
5058 | ret = -EINTR; | |
f499a021 | 5059 | out: |
4e88d6e7 JA |
5060 | if (ret < 0) |
5061 | req_set_fail_links(req); | |
229a7b63 | 5062 | __io_req_complete(req, ret, 0, cs); |
f8e85cf2 | 5063 | return 0; |
469956e8 Y |
5064 | } |
5065 | #else /* !CONFIG_NET */ | |
5066 | static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
5067 | { | |
f8e85cf2 | 5068 | return -EOPNOTSUPP; |
f8e85cf2 JA |
5069 | } |
5070 | ||
1e16c2f9 RD |
5071 | static int io_sendmsg(struct io_kiocb *req, bool force_nonblock, |
5072 | struct io_comp_state *cs) | |
469956e8 Y |
5073 | { |
5074 | return -EOPNOTSUPP; | |
5075 | } | |
5076 | ||
1e16c2f9 RD |
5077 | static int io_send(struct io_kiocb *req, bool force_nonblock, |
5078 | struct io_comp_state *cs) | |
469956e8 Y |
5079 | { |
5080 | return -EOPNOTSUPP; | |
5081 | } | |
5082 | ||
5083 | static int io_recvmsg_prep(struct io_kiocb *req, | |
5084 | const struct io_uring_sqe *sqe) | |
5085 | { | |
5086 | return -EOPNOTSUPP; | |
5087 | } | |
5088 | ||
1e16c2f9 RD |
5089 | static int io_recvmsg(struct io_kiocb *req, bool force_nonblock, |
5090 | struct io_comp_state *cs) | |
469956e8 Y |
5091 | { |
5092 | return -EOPNOTSUPP; | |
5093 | } | |
5094 | ||
1e16c2f9 RD |
5095 | static int io_recv(struct io_kiocb *req, bool force_nonblock, |
5096 | struct io_comp_state *cs) | |
469956e8 Y |
5097 | { |
5098 | return -EOPNOTSUPP; | |
5099 | } | |
5100 | ||
5101 | static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
5102 | { | |
5103 | return -EOPNOTSUPP; | |
5104 | } | |
5105 | ||
1e16c2f9 RD |
5106 | static int io_accept(struct io_kiocb *req, bool force_nonblock, |
5107 | struct io_comp_state *cs) | |
469956e8 Y |
5108 | { |
5109 | return -EOPNOTSUPP; | |
5110 | } | |
ce593a6c | 5111 | |
469956e8 Y |
5112 | static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
5113 | { | |
5114 | return -EOPNOTSUPP; | |
5115 | } | |
5116 | ||
1e16c2f9 RD |
5117 | static int io_connect(struct io_kiocb *req, bool force_nonblock, |
5118 | struct io_comp_state *cs) | |
469956e8 | 5119 | { |
f8e85cf2 | 5120 | return -EOPNOTSUPP; |
ce593a6c | 5121 | } |
469956e8 | 5122 | #endif /* CONFIG_NET */ |
f8e85cf2 | 5123 | |
d7718a9d JA |
5124 | struct io_poll_table { |
5125 | struct poll_table_struct pt; | |
5126 | struct io_kiocb *req; | |
5127 | int error; | |
5128 | }; | |
ce593a6c | 5129 | |
d7718a9d JA |
5130 | static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll, |
5131 | __poll_t mask, task_work_func_t func) | |
5132 | { | |
aa96bf8a | 5133 | int ret; |
d7718a9d JA |
5134 | |
5135 | /* for instances that support it check for an event match first: */ | |
5136 | if (mask && !(mask & poll->events)) | |
5137 | return 0; | |
5138 | ||
5139 | trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask); | |
5140 | ||
5141 | list_del_init(&poll->wait.entry); | |
5142 | ||
d7718a9d JA |
5143 | req->result = mask; |
5144 | init_task_work(&req->task_work, func); | |
6d816e08 JA |
5145 | percpu_ref_get(&req->ctx->refs); |
5146 | ||
d7718a9d | 5147 | /* |
e3aabf95 JA |
5148 | * If this fails, then the task is exiting. When a task exits, the |
5149 | * work gets canceled, so just cancel this request as well instead | |
5150 | * of executing it. We can't safely execute it anyway, as we may not | |
5151 | * have the needed state needed for it anyway. | |
d7718a9d | 5152 | */ |
355fb9e2 | 5153 | ret = io_req_task_work_add(req); |
aa96bf8a | 5154 | if (unlikely(ret)) { |
c2c4c83c JA |
5155 | struct task_struct *tsk; |
5156 | ||
e3aabf95 | 5157 | WRITE_ONCE(poll->canceled, true); |
aa96bf8a | 5158 | tsk = io_wq_get_task(req->ctx->io_wq); |
91989c70 | 5159 | task_work_add(tsk, &req->task_work, TWA_NONE); |
ce593a6c | 5160 | wake_up_process(tsk); |
aa96bf8a | 5161 | } |
d7718a9d JA |
5162 | return 1; |
5163 | } | |
5164 | ||
74ce6ce4 JA |
5165 | static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll) |
5166 | __acquires(&req->ctx->completion_lock) | |
5167 | { | |
5168 | struct io_ring_ctx *ctx = req->ctx; | |
5169 | ||
5170 | if (!req->result && !READ_ONCE(poll->canceled)) { | |
5171 | struct poll_table_struct pt = { ._key = poll->events }; | |
5172 | ||
5173 | req->result = vfs_poll(req->file, &pt) & poll->events; | |
5174 | } | |
5175 | ||
5176 | spin_lock_irq(&ctx->completion_lock); | |
5177 | if (!req->result && !READ_ONCE(poll->canceled)) { | |
5178 | add_wait_queue(poll->head, &poll->wait); | |
5179 | return true; | |
5180 | } | |
5181 | ||
5182 | return false; | |
5183 | } | |
5184 | ||
d4e7cd36 | 5185 | static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req) |
18bceab1 | 5186 | { |
e8c2bc1f | 5187 | /* pure poll stashes this in ->async_data, poll driven retry elsewhere */ |
d4e7cd36 | 5188 | if (req->opcode == IORING_OP_POLL_ADD) |
e8c2bc1f | 5189 | return req->async_data; |
d4e7cd36 JA |
5190 | return req->apoll->double_poll; |
5191 | } | |
5192 | ||
5193 | static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req) | |
5194 | { | |
5195 | if (req->opcode == IORING_OP_POLL_ADD) | |
5196 | return &req->poll; | |
5197 | return &req->apoll->poll; | |
5198 | } | |
5199 | ||
5200 | static void io_poll_remove_double(struct io_kiocb *req) | |
5201 | { | |
5202 | struct io_poll_iocb *poll = io_poll_get_double(req); | |
18bceab1 JA |
5203 | |
5204 | lockdep_assert_held(&req->ctx->completion_lock); | |
5205 | ||
5206 | if (poll && poll->head) { | |
5207 | struct wait_queue_head *head = poll->head; | |
5208 | ||
5209 | spin_lock(&head->lock); | |
5210 | list_del_init(&poll->wait.entry); | |
5211 | if (poll->wait.private) | |
5212 | refcount_dec(&req->refs); | |
5213 | poll->head = NULL; | |
5214 | spin_unlock(&head->lock); | |
5215 | } | |
5216 | } | |
5217 | ||
5218 | static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error) | |
5219 | { | |
5220 | struct io_ring_ctx *ctx = req->ctx; | |
5221 | ||
d4e7cd36 | 5222 | io_poll_remove_double(req); |
18bceab1 JA |
5223 | req->poll.done = true; |
5224 | io_cqring_fill_event(req, error ? error : mangle_poll(mask)); | |
5225 | io_commit_cqring(ctx); | |
5226 | } | |
5227 | ||
dd221f46 | 5228 | static void io_poll_task_func(struct callback_head *cb) |
18bceab1 | 5229 | { |
dd221f46 | 5230 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); |
18bceab1 | 5231 | struct io_ring_ctx *ctx = req->ctx; |
dd221f46 | 5232 | struct io_kiocb *nxt; |
18bceab1 JA |
5233 | |
5234 | if (io_poll_rewait(req, &req->poll)) { | |
5235 | spin_unlock_irq(&ctx->completion_lock); | |
dd221f46 PB |
5236 | } else { |
5237 | hash_del(&req->hash_node); | |
5238 | io_poll_complete(req, req->result, 0); | |
5239 | spin_unlock_irq(&ctx->completion_lock); | |
18bceab1 | 5240 | |
dd221f46 PB |
5241 | nxt = io_put_req_find_next(req); |
5242 | io_cqring_ev_posted(ctx); | |
5243 | if (nxt) | |
5244 | __io_req_task_submit(nxt); | |
5245 | } | |
18bceab1 | 5246 | |
6d816e08 | 5247 | percpu_ref_put(&ctx->refs); |
18bceab1 JA |
5248 | } |
5249 | ||
5250 | static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode, | |
5251 | int sync, void *key) | |
5252 | { | |
5253 | struct io_kiocb *req = wait->private; | |
d4e7cd36 | 5254 | struct io_poll_iocb *poll = io_poll_get_single(req); |
18bceab1 JA |
5255 | __poll_t mask = key_to_poll(key); |
5256 | ||
5257 | /* for instances that support it check for an event match first: */ | |
5258 | if (mask && !(mask & poll->events)) | |
5259 | return 0; | |
5260 | ||
8706e04e JA |
5261 | list_del_init(&wait->entry); |
5262 | ||
807abcb0 | 5263 | if (poll && poll->head) { |
18bceab1 JA |
5264 | bool done; |
5265 | ||
807abcb0 JA |
5266 | spin_lock(&poll->head->lock); |
5267 | done = list_empty(&poll->wait.entry); | |
18bceab1 | 5268 | if (!done) |
807abcb0 | 5269 | list_del_init(&poll->wait.entry); |
d4e7cd36 JA |
5270 | /* make sure double remove sees this as being gone */ |
5271 | wait->private = NULL; | |
807abcb0 | 5272 | spin_unlock(&poll->head->lock); |
c8b5e260 JA |
5273 | if (!done) { |
5274 | /* use wait func handler, so it matches the rq type */ | |
5275 | poll->wait.func(&poll->wait, mode, sync, key); | |
5276 | } | |
18bceab1 JA |
5277 | } |
5278 | refcount_dec(&req->refs); | |
5279 | return 1; | |
5280 | } | |
5281 | ||
5282 | static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events, | |
5283 | wait_queue_func_t wake_func) | |
5284 | { | |
5285 | poll->head = NULL; | |
5286 | poll->done = false; | |
5287 | poll->canceled = false; | |
5288 | poll->events = events; | |
5289 | INIT_LIST_HEAD(&poll->wait.entry); | |
5290 | init_waitqueue_func_entry(&poll->wait, wake_func); | |
5291 | } | |
5292 | ||
5293 | static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt, | |
807abcb0 JA |
5294 | struct wait_queue_head *head, |
5295 | struct io_poll_iocb **poll_ptr) | |
18bceab1 JA |
5296 | { |
5297 | struct io_kiocb *req = pt->req; | |
5298 | ||
5299 | /* | |
5300 | * If poll->head is already set, it's because the file being polled | |
5301 | * uses multiple waitqueues for poll handling (eg one for read, one | |
5302 | * for write). Setup a separate io_poll_iocb if this happens. | |
5303 | */ | |
5304 | if (unlikely(poll->head)) { | |
58852d4d PB |
5305 | struct io_poll_iocb *poll_one = poll; |
5306 | ||
18bceab1 | 5307 | /* already have a 2nd entry, fail a third attempt */ |
807abcb0 | 5308 | if (*poll_ptr) { |
18bceab1 JA |
5309 | pt->error = -EINVAL; |
5310 | return; | |
5311 | } | |
5312 | poll = kmalloc(sizeof(*poll), GFP_ATOMIC); | |
5313 | if (!poll) { | |
5314 | pt->error = -ENOMEM; | |
5315 | return; | |
5316 | } | |
58852d4d | 5317 | io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake); |
18bceab1 JA |
5318 | refcount_inc(&req->refs); |
5319 | poll->wait.private = req; | |
807abcb0 | 5320 | *poll_ptr = poll; |
18bceab1 JA |
5321 | } |
5322 | ||
5323 | pt->error = 0; | |
5324 | poll->head = head; | |
a31eb4a2 JX |
5325 | |
5326 | if (poll->events & EPOLLEXCLUSIVE) | |
5327 | add_wait_queue_exclusive(head, &poll->wait); | |
5328 | else | |
5329 | add_wait_queue(head, &poll->wait); | |
18bceab1 JA |
5330 | } |
5331 | ||
5332 | static void io_async_queue_proc(struct file *file, struct wait_queue_head *head, | |
5333 | struct poll_table_struct *p) | |
5334 | { | |
5335 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
807abcb0 | 5336 | struct async_poll *apoll = pt->req->apoll; |
18bceab1 | 5337 | |
807abcb0 | 5338 | __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll); |
18bceab1 JA |
5339 | } |
5340 | ||
d7718a9d JA |
5341 | static void io_async_task_func(struct callback_head *cb) |
5342 | { | |
5343 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); | |
5344 | struct async_poll *apoll = req->apoll; | |
5345 | struct io_ring_ctx *ctx = req->ctx; | |
5346 | ||
5347 | trace_io_uring_task_run(req->ctx, req->opcode, req->user_data); | |
5348 | ||
74ce6ce4 | 5349 | if (io_poll_rewait(req, &apoll->poll)) { |
d7718a9d | 5350 | spin_unlock_irq(&ctx->completion_lock); |
6d816e08 | 5351 | percpu_ref_put(&ctx->refs); |
74ce6ce4 | 5352 | return; |
d7718a9d JA |
5353 | } |
5354 | ||
31067255 | 5355 | /* If req is still hashed, it cannot have been canceled. Don't check. */ |
0be0b0e3 | 5356 | if (hash_hashed(&req->hash_node)) |
74ce6ce4 | 5357 | hash_del(&req->hash_node); |
2bae047e | 5358 | |
d4e7cd36 | 5359 | io_poll_remove_double(req); |
74ce6ce4 JA |
5360 | spin_unlock_irq(&ctx->completion_lock); |
5361 | ||
0be0b0e3 PB |
5362 | if (!READ_ONCE(apoll->poll.canceled)) |
5363 | __io_req_task_submit(req); | |
5364 | else | |
5365 | __io_req_task_cancel(req, -ECANCELED); | |
aa340845 | 5366 | |
6d816e08 | 5367 | percpu_ref_put(&ctx->refs); |
807abcb0 | 5368 | kfree(apoll->double_poll); |
31067255 | 5369 | kfree(apoll); |
d7718a9d JA |
5370 | } |
5371 | ||
5372 | static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
5373 | void *key) | |
5374 | { | |
5375 | struct io_kiocb *req = wait->private; | |
5376 | struct io_poll_iocb *poll = &req->apoll->poll; | |
5377 | ||
5378 | trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data, | |
5379 | key_to_poll(key)); | |
5380 | ||
5381 | return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func); | |
5382 | } | |
5383 | ||
5384 | static void io_poll_req_insert(struct io_kiocb *req) | |
5385 | { | |
5386 | struct io_ring_ctx *ctx = req->ctx; | |
5387 | struct hlist_head *list; | |
5388 | ||
5389 | list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)]; | |
5390 | hlist_add_head(&req->hash_node, list); | |
5391 | } | |
5392 | ||
5393 | static __poll_t __io_arm_poll_handler(struct io_kiocb *req, | |
5394 | struct io_poll_iocb *poll, | |
5395 | struct io_poll_table *ipt, __poll_t mask, | |
5396 | wait_queue_func_t wake_func) | |
5397 | __acquires(&ctx->completion_lock) | |
5398 | { | |
5399 | struct io_ring_ctx *ctx = req->ctx; | |
5400 | bool cancel = false; | |
5401 | ||
4d52f338 | 5402 | INIT_HLIST_NODE(&req->hash_node); |
18bceab1 | 5403 | io_init_poll_iocb(poll, mask, wake_func); |
b90cd197 | 5404 | poll->file = req->file; |
18bceab1 | 5405 | poll->wait.private = req; |
d7718a9d JA |
5406 | |
5407 | ipt->pt._key = mask; | |
5408 | ipt->req = req; | |
5409 | ipt->error = -EINVAL; | |
5410 | ||
d7718a9d JA |
5411 | mask = vfs_poll(req->file, &ipt->pt) & poll->events; |
5412 | ||
5413 | spin_lock_irq(&ctx->completion_lock); | |
5414 | if (likely(poll->head)) { | |
5415 | spin_lock(&poll->head->lock); | |
5416 | if (unlikely(list_empty(&poll->wait.entry))) { | |
5417 | if (ipt->error) | |
5418 | cancel = true; | |
5419 | ipt->error = 0; | |
5420 | mask = 0; | |
5421 | } | |
5422 | if (mask || ipt->error) | |
5423 | list_del_init(&poll->wait.entry); | |
5424 | else if (cancel) | |
5425 | WRITE_ONCE(poll->canceled, true); | |
5426 | else if (!poll->done) /* actually waiting for an event */ | |
5427 | io_poll_req_insert(req); | |
5428 | spin_unlock(&poll->head->lock); | |
5429 | } | |
5430 | ||
5431 | return mask; | |
5432 | } | |
5433 | ||
5434 | static bool io_arm_poll_handler(struct io_kiocb *req) | |
5435 | { | |
5436 | const struct io_op_def *def = &io_op_defs[req->opcode]; | |
5437 | struct io_ring_ctx *ctx = req->ctx; | |
5438 | struct async_poll *apoll; | |
5439 | struct io_poll_table ipt; | |
5440 | __poll_t mask, ret; | |
9dab14b8 | 5441 | int rw; |
d7718a9d JA |
5442 | |
5443 | if (!req->file || !file_can_poll(req->file)) | |
5444 | return false; | |
24c74678 | 5445 | if (req->flags & REQ_F_POLLED) |
d7718a9d | 5446 | return false; |
9dab14b8 JA |
5447 | if (def->pollin) |
5448 | rw = READ; | |
5449 | else if (def->pollout) | |
5450 | rw = WRITE; | |
5451 | else | |
5452 | return false; | |
5453 | /* if we can't nonblock try, then no point in arming a poll handler */ | |
5454 | if (!io_file_supports_async(req->file, rw)) | |
d7718a9d JA |
5455 | return false; |
5456 | ||
5457 | apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC); | |
5458 | if (unlikely(!apoll)) | |
5459 | return false; | |
807abcb0 | 5460 | apoll->double_poll = NULL; |
d7718a9d JA |
5461 | |
5462 | req->flags |= REQ_F_POLLED; | |
d7718a9d | 5463 | req->apoll = apoll; |
d7718a9d | 5464 | |
8755d97a | 5465 | mask = 0; |
d7718a9d | 5466 | if (def->pollin) |
8755d97a | 5467 | mask |= POLLIN | POLLRDNORM; |
d7718a9d JA |
5468 | if (def->pollout) |
5469 | mask |= POLLOUT | POLLWRNORM; | |
901341bb LH |
5470 | |
5471 | /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */ | |
5472 | if ((req->opcode == IORING_OP_RECVMSG) && | |
5473 | (req->sr_msg.msg_flags & MSG_ERRQUEUE)) | |
5474 | mask &= ~POLLIN; | |
5475 | ||
d7718a9d JA |
5476 | mask |= POLLERR | POLLPRI; |
5477 | ||
5478 | ipt.pt._qproc = io_async_queue_proc; | |
5479 | ||
5480 | ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, | |
5481 | io_async_wake); | |
a36da65c | 5482 | if (ret || ipt.error) { |
d4e7cd36 | 5483 | io_poll_remove_double(req); |
d7718a9d | 5484 | spin_unlock_irq(&ctx->completion_lock); |
807abcb0 | 5485 | kfree(apoll->double_poll); |
d7718a9d JA |
5486 | kfree(apoll); |
5487 | return false; | |
5488 | } | |
5489 | spin_unlock_irq(&ctx->completion_lock); | |
5490 | trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask, | |
5491 | apoll->poll.events); | |
5492 | return true; | |
5493 | } | |
5494 | ||
5495 | static bool __io_poll_remove_one(struct io_kiocb *req, | |
5496 | struct io_poll_iocb *poll) | |
221c5eb2 | 5497 | { |
b41e9852 | 5498 | bool do_complete = false; |
221c5eb2 JA |
5499 | |
5500 | spin_lock(&poll->head->lock); | |
5501 | WRITE_ONCE(poll->canceled, true); | |
392edb45 JA |
5502 | if (!list_empty(&poll->wait.entry)) { |
5503 | list_del_init(&poll->wait.entry); | |
b41e9852 | 5504 | do_complete = true; |
221c5eb2 JA |
5505 | } |
5506 | spin_unlock(&poll->head->lock); | |
3bfa5bcb | 5507 | hash_del(&req->hash_node); |
d7718a9d JA |
5508 | return do_complete; |
5509 | } | |
5510 | ||
5511 | static bool io_poll_remove_one(struct io_kiocb *req) | |
5512 | { | |
5513 | bool do_complete; | |
5514 | ||
d4e7cd36 JA |
5515 | io_poll_remove_double(req); |
5516 | ||
d7718a9d JA |
5517 | if (req->opcode == IORING_OP_POLL_ADD) { |
5518 | do_complete = __io_poll_remove_one(req, &req->poll); | |
5519 | } else { | |
3bfa5bcb JA |
5520 | struct async_poll *apoll = req->apoll; |
5521 | ||
d7718a9d | 5522 | /* non-poll requests have submit ref still */ |
3bfa5bcb JA |
5523 | do_complete = __io_poll_remove_one(req, &apoll->poll); |
5524 | if (do_complete) { | |
d7718a9d | 5525 | io_put_req(req); |
807abcb0 | 5526 | kfree(apoll->double_poll); |
3bfa5bcb JA |
5527 | kfree(apoll); |
5528 | } | |
b1f573bd XW |
5529 | } |
5530 | ||
b41e9852 JA |
5531 | if (do_complete) { |
5532 | io_cqring_fill_event(req, -ECANCELED); | |
5533 | io_commit_cqring(req->ctx); | |
f254ac04 | 5534 | req_set_fail_links(req); |
216578e5 | 5535 | io_put_req_deferred(req, 1); |
b41e9852 JA |
5536 | } |
5537 | ||
5538 | return do_complete; | |
221c5eb2 JA |
5539 | } |
5540 | ||
76e1b642 JA |
5541 | /* |
5542 | * Returns true if we found and killed one or more poll requests | |
5543 | */ | |
6b81928d PB |
5544 | static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk, |
5545 | struct files_struct *files) | |
221c5eb2 | 5546 | { |
78076bb6 | 5547 | struct hlist_node *tmp; |
221c5eb2 | 5548 | struct io_kiocb *req; |
8e2e1faf | 5549 | int posted = 0, i; |
221c5eb2 JA |
5550 | |
5551 | spin_lock_irq(&ctx->completion_lock); | |
78076bb6 JA |
5552 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { |
5553 | struct hlist_head *list; | |
5554 | ||
5555 | list = &ctx->cancel_hash[i]; | |
f3606e3a | 5556 | hlist_for_each_entry_safe(req, tmp, list, hash_node) { |
6b81928d | 5557 | if (io_match_task(req, tsk, files)) |
f3606e3a JA |
5558 | posted += io_poll_remove_one(req); |
5559 | } | |
221c5eb2 JA |
5560 | } |
5561 | spin_unlock_irq(&ctx->completion_lock); | |
b41e9852 | 5562 | |
8e2e1faf JA |
5563 | if (posted) |
5564 | io_cqring_ev_posted(ctx); | |
76e1b642 JA |
5565 | |
5566 | return posted != 0; | |
221c5eb2 JA |
5567 | } |
5568 | ||
47f46768 JA |
5569 | static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr) |
5570 | { | |
78076bb6 | 5571 | struct hlist_head *list; |
47f46768 JA |
5572 | struct io_kiocb *req; |
5573 | ||
78076bb6 JA |
5574 | list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)]; |
5575 | hlist_for_each_entry(req, list, hash_node) { | |
b41e9852 JA |
5576 | if (sqe_addr != req->user_data) |
5577 | continue; | |
5578 | if (io_poll_remove_one(req)) | |
eac406c6 | 5579 | return 0; |
b41e9852 | 5580 | return -EALREADY; |
47f46768 JA |
5581 | } |
5582 | ||
5583 | return -ENOENT; | |
5584 | } | |
5585 | ||
3529d8c2 JA |
5586 | static int io_poll_remove_prep(struct io_kiocb *req, |
5587 | const struct io_uring_sqe *sqe) | |
0969e783 | 5588 | { |
0969e783 JA |
5589 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5590 | return -EINVAL; | |
5591 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index || | |
5592 | sqe->poll_events) | |
5593 | return -EINVAL; | |
5594 | ||
018043be | 5595 | req->poll_remove.addr = READ_ONCE(sqe->addr); |
0969e783 JA |
5596 | return 0; |
5597 | } | |
5598 | ||
221c5eb2 JA |
5599 | /* |
5600 | * Find a running poll command that matches one specified in sqe->addr, | |
5601 | * and remove it if found. | |
5602 | */ | |
fc4df999 | 5603 | static int io_poll_remove(struct io_kiocb *req) |
221c5eb2 JA |
5604 | { |
5605 | struct io_ring_ctx *ctx = req->ctx; | |
47f46768 | 5606 | int ret; |
221c5eb2 | 5607 | |
221c5eb2 | 5608 | spin_lock_irq(&ctx->completion_lock); |
018043be | 5609 | ret = io_poll_cancel(ctx, req->poll_remove.addr); |
221c5eb2 JA |
5610 | spin_unlock_irq(&ctx->completion_lock); |
5611 | ||
4e88d6e7 JA |
5612 | if (ret < 0) |
5613 | req_set_fail_links(req); | |
e1e16097 | 5614 | io_req_complete(req, ret); |
221c5eb2 JA |
5615 | return 0; |
5616 | } | |
5617 | ||
221c5eb2 JA |
5618 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, |
5619 | void *key) | |
5620 | { | |
c2f2eb7d JA |
5621 | struct io_kiocb *req = wait->private; |
5622 | struct io_poll_iocb *poll = &req->poll; | |
221c5eb2 | 5623 | |
d7718a9d | 5624 | return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func); |
221c5eb2 JA |
5625 | } |
5626 | ||
221c5eb2 JA |
5627 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, |
5628 | struct poll_table_struct *p) | |
5629 | { | |
5630 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
5631 | ||
e8c2bc1f | 5632 | __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data); |
eac406c6 JA |
5633 | } |
5634 | ||
3529d8c2 | 5635 | static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
221c5eb2 JA |
5636 | { |
5637 | struct io_poll_iocb *poll = &req->poll; | |
5769a351 | 5638 | u32 events; |
221c5eb2 JA |
5639 | |
5640 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
5641 | return -EINVAL; | |
5642 | if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
5643 | return -EINVAL; | |
5644 | ||
5769a351 JX |
5645 | events = READ_ONCE(sqe->poll32_events); |
5646 | #ifdef __BIG_ENDIAN | |
5647 | events = swahw32(events); | |
5648 | #endif | |
a31eb4a2 JX |
5649 | poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP | |
5650 | (events & EPOLLEXCLUSIVE); | |
0969e783 JA |
5651 | return 0; |
5652 | } | |
5653 | ||
014db007 | 5654 | static int io_poll_add(struct io_kiocb *req) |
0969e783 JA |
5655 | { |
5656 | struct io_poll_iocb *poll = &req->poll; | |
5657 | struct io_ring_ctx *ctx = req->ctx; | |
5658 | struct io_poll_table ipt; | |
0969e783 | 5659 | __poll_t mask; |
0969e783 | 5660 | |
d7718a9d | 5661 | ipt.pt._qproc = io_poll_queue_proc; |
36703247 | 5662 | |
d7718a9d JA |
5663 | mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events, |
5664 | io_poll_wake); | |
221c5eb2 | 5665 | |
8c838788 | 5666 | if (mask) { /* no async, we'd stolen it */ |
221c5eb2 | 5667 | ipt.error = 0; |
b0dd8a41 | 5668 | io_poll_complete(req, mask, 0); |
221c5eb2 | 5669 | } |
221c5eb2 JA |
5670 | spin_unlock_irq(&ctx->completion_lock); |
5671 | ||
8c838788 JA |
5672 | if (mask) { |
5673 | io_cqring_ev_posted(ctx); | |
014db007 | 5674 | io_put_req(req); |
221c5eb2 | 5675 | } |
8c838788 | 5676 | return ipt.error; |
221c5eb2 JA |
5677 | } |
5678 | ||
5262f567 JA |
5679 | static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer) |
5680 | { | |
ad8a48ac JA |
5681 | struct io_timeout_data *data = container_of(timer, |
5682 | struct io_timeout_data, timer); | |
5683 | struct io_kiocb *req = data->req; | |
5684 | struct io_ring_ctx *ctx = req->ctx; | |
5262f567 JA |
5685 | unsigned long flags; |
5686 | ||
5262f567 | 5687 | spin_lock_irqsave(&ctx->completion_lock, flags); |
a71976f3 | 5688 | list_del_init(&req->timeout.list); |
01cec8c1 PB |
5689 | atomic_set(&req->ctx->cq_timeouts, |
5690 | atomic_read(&req->ctx->cq_timeouts) + 1); | |
5691 | ||
78e19bbe | 5692 | io_cqring_fill_event(req, -ETIME); |
5262f567 JA |
5693 | io_commit_cqring(ctx); |
5694 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
5695 | ||
5696 | io_cqring_ev_posted(ctx); | |
4e88d6e7 | 5697 | req_set_fail_links(req); |
5262f567 JA |
5698 | io_put_req(req); |
5699 | return HRTIMER_NORESTART; | |
5700 | } | |
5701 | ||
fbd15848 PB |
5702 | static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx, |
5703 | __u64 user_data) | |
f254ac04 | 5704 | { |
fbd15848 | 5705 | struct io_timeout_data *io; |
47f46768 JA |
5706 | struct io_kiocb *req; |
5707 | int ret = -ENOENT; | |
f254ac04 | 5708 | |
135fcde8 | 5709 | list_for_each_entry(req, &ctx->timeout_list, timeout.list) { |
47f46768 | 5710 | if (user_data == req->user_data) { |
47f46768 JA |
5711 | ret = 0; |
5712 | break; | |
5713 | } | |
5714 | } | |
5715 | ||
5716 | if (ret == -ENOENT) | |
fbd15848 PB |
5717 | return ERR_PTR(ret); |
5718 | ||
5719 | io = req->async_data; | |
e8c2bc1f | 5720 | ret = hrtimer_try_to_cancel(&io->timer); |
f254ac04 | 5721 | if (ret == -1) |
fbd15848 | 5722 | return ERR_PTR(-EALREADY); |
a71976f3 | 5723 | list_del_init(&req->timeout.list); |
fbd15848 PB |
5724 | return req; |
5725 | } | |
47f46768 | 5726 | |
fbd15848 PB |
5727 | static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data) |
5728 | { | |
5729 | struct io_kiocb *req = io_timeout_extract(ctx, user_data); | |
5730 | ||
5731 | if (IS_ERR(req)) | |
5732 | return PTR_ERR(req); | |
f254ac04 JA |
5733 | |
5734 | req_set_fail_links(req); | |
f254ac04 | 5735 | io_cqring_fill_event(req, -ECANCELED); |
216578e5 | 5736 | io_put_req_deferred(req, 1); |
f254ac04 JA |
5737 | return 0; |
5738 | } | |
5739 | ||
9c8e11b3 PB |
5740 | static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data, |
5741 | struct timespec64 *ts, enum hrtimer_mode mode) | |
47f46768 | 5742 | { |
9c8e11b3 PB |
5743 | struct io_kiocb *req = io_timeout_extract(ctx, user_data); |
5744 | struct io_timeout_data *data; | |
47f46768 | 5745 | |
9c8e11b3 PB |
5746 | if (IS_ERR(req)) |
5747 | return PTR_ERR(req); | |
47f46768 | 5748 | |
9c8e11b3 PB |
5749 | req->timeout.off = 0; /* noseq */ |
5750 | data = req->async_data; | |
5751 | list_add_tail(&req->timeout.list, &ctx->timeout_list); | |
5752 | hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode); | |
5753 | data->timer.function = io_timeout_fn; | |
5754 | hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode); | |
5755 | return 0; | |
47f46768 JA |
5756 | } |
5757 | ||
3529d8c2 JA |
5758 | static int io_timeout_remove_prep(struct io_kiocb *req, |
5759 | const struct io_uring_sqe *sqe) | |
b29472ee | 5760 | { |
9c8e11b3 PB |
5761 | struct io_timeout_rem *tr = &req->timeout_rem; |
5762 | ||
b29472ee JA |
5763 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5764 | return -EINVAL; | |
61710e43 DA |
5765 | if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) |
5766 | return -EINVAL; | |
9c8e11b3 | 5767 | if (sqe->ioprio || sqe->buf_index || sqe->len) |
b29472ee JA |
5768 | return -EINVAL; |
5769 | ||
9c8e11b3 PB |
5770 | tr->addr = READ_ONCE(sqe->addr); |
5771 | tr->flags = READ_ONCE(sqe->timeout_flags); | |
5772 | if (tr->flags & IORING_TIMEOUT_UPDATE) { | |
5773 | if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS)) | |
5774 | return -EINVAL; | |
5775 | if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2))) | |
5776 | return -EFAULT; | |
5777 | } else if (tr->flags) { | |
5778 | /* timeout removal doesn't support flags */ | |
b29472ee | 5779 | return -EINVAL; |
9c8e11b3 | 5780 | } |
b29472ee | 5781 | |
b29472ee JA |
5782 | return 0; |
5783 | } | |
5784 | ||
11365043 JA |
5785 | /* |
5786 | * Remove or update an existing timeout command | |
5787 | */ | |
fc4df999 | 5788 | static int io_timeout_remove(struct io_kiocb *req) |
11365043 | 5789 | { |
9c8e11b3 | 5790 | struct io_timeout_rem *tr = &req->timeout_rem; |
11365043 | 5791 | struct io_ring_ctx *ctx = req->ctx; |
47f46768 | 5792 | int ret; |
11365043 | 5793 | |
11365043 | 5794 | spin_lock_irq(&ctx->completion_lock); |
9c8e11b3 PB |
5795 | if (req->timeout_rem.flags & IORING_TIMEOUT_UPDATE) { |
5796 | enum hrtimer_mode mode = (tr->flags & IORING_TIMEOUT_ABS) | |
5797 | ? HRTIMER_MODE_ABS : HRTIMER_MODE_REL; | |
5798 | ||
5799 | ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode); | |
5800 | } else { | |
5801 | ret = io_timeout_cancel(ctx, tr->addr); | |
5802 | } | |
11365043 | 5803 | |
47f46768 | 5804 | io_cqring_fill_event(req, ret); |
11365043 JA |
5805 | io_commit_cqring(ctx); |
5806 | spin_unlock_irq(&ctx->completion_lock); | |
5262f567 | 5807 | io_cqring_ev_posted(ctx); |
4e88d6e7 JA |
5808 | if (ret < 0) |
5809 | req_set_fail_links(req); | |
ec9c02ad | 5810 | io_put_req(req); |
11365043 | 5811 | return 0; |
5262f567 JA |
5812 | } |
5813 | ||
3529d8c2 | 5814 | static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2d28390a | 5815 | bool is_timeout_link) |
5262f567 | 5816 | { |
ad8a48ac | 5817 | struct io_timeout_data *data; |
a41525ab | 5818 | unsigned flags; |
56080b02 | 5819 | u32 off = READ_ONCE(sqe->off); |
5262f567 | 5820 | |
ad8a48ac | 5821 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5262f567 | 5822 | return -EINVAL; |
ad8a48ac | 5823 | if (sqe->ioprio || sqe->buf_index || sqe->len != 1) |
a41525ab | 5824 | return -EINVAL; |
56080b02 | 5825 | if (off && is_timeout_link) |
2d28390a | 5826 | return -EINVAL; |
a41525ab JA |
5827 | flags = READ_ONCE(sqe->timeout_flags); |
5828 | if (flags & ~IORING_TIMEOUT_ABS) | |
5262f567 | 5829 | return -EINVAL; |
bdf20073 | 5830 | |
bfe68a22 | 5831 | req->timeout.off = off; |
26a61679 | 5832 | |
e8c2bc1f | 5833 | if (!req->async_data && io_alloc_async_data(req)) |
26a61679 JA |
5834 | return -ENOMEM; |
5835 | ||
e8c2bc1f | 5836 | data = req->async_data; |
ad8a48ac | 5837 | data->req = req; |
ad8a48ac JA |
5838 | |
5839 | if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr))) | |
5262f567 JA |
5840 | return -EFAULT; |
5841 | ||
11365043 | 5842 | if (flags & IORING_TIMEOUT_ABS) |
ad8a48ac | 5843 | data->mode = HRTIMER_MODE_ABS; |
11365043 | 5844 | else |
ad8a48ac | 5845 | data->mode = HRTIMER_MODE_REL; |
11365043 | 5846 | |
ad8a48ac JA |
5847 | hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode); |
5848 | return 0; | |
5849 | } | |
5850 | ||
fc4df999 | 5851 | static int io_timeout(struct io_kiocb *req) |
ad8a48ac | 5852 | { |
ad8a48ac | 5853 | struct io_ring_ctx *ctx = req->ctx; |
e8c2bc1f | 5854 | struct io_timeout_data *data = req->async_data; |
ad8a48ac | 5855 | struct list_head *entry; |
bfe68a22 | 5856 | u32 tail, off = req->timeout.off; |
ad8a48ac | 5857 | |
733f5c95 | 5858 | spin_lock_irq(&ctx->completion_lock); |
93bd25bb | 5859 | |
5262f567 JA |
5860 | /* |
5861 | * sqe->off holds how many events that need to occur for this | |
93bd25bb JA |
5862 | * timeout event to be satisfied. If it isn't set, then this is |
5863 | * a pure timeout request, sequence isn't used. | |
5262f567 | 5864 | */ |
8eb7e2d0 | 5865 | if (io_is_timeout_noseq(req)) { |
93bd25bb JA |
5866 | entry = ctx->timeout_list.prev; |
5867 | goto add; | |
5868 | } | |
5262f567 | 5869 | |
bfe68a22 PB |
5870 | tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts); |
5871 | req->timeout.target_seq = tail + off; | |
5262f567 | 5872 | |
f010505b MDG |
5873 | /* Update the last seq here in case io_flush_timeouts() hasn't. |
5874 | * This is safe because ->completion_lock is held, and submissions | |
5875 | * and completions are never mixed in the same ->completion_lock section. | |
5876 | */ | |
5877 | ctx->cq_last_tm_flush = tail; | |
5878 | ||
5262f567 JA |
5879 | /* |
5880 | * Insertion sort, ensuring the first entry in the list is always | |
5881 | * the one we need first. | |
5882 | */ | |
5262f567 | 5883 | list_for_each_prev(entry, &ctx->timeout_list) { |
135fcde8 PB |
5884 | struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, |
5885 | timeout.list); | |
5262f567 | 5886 | |
8eb7e2d0 | 5887 | if (io_is_timeout_noseq(nxt)) |
93bd25bb | 5888 | continue; |
bfe68a22 PB |
5889 | /* nxt.seq is behind @tail, otherwise would've been completed */ |
5890 | if (off >= nxt->timeout.target_seq - tail) | |
5262f567 JA |
5891 | break; |
5892 | } | |
93bd25bb | 5893 | add: |
135fcde8 | 5894 | list_add(&req->timeout.list, entry); |
ad8a48ac JA |
5895 | data->timer.function = io_timeout_fn; |
5896 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode); | |
5262f567 | 5897 | spin_unlock_irq(&ctx->completion_lock); |
5262f567 JA |
5898 | return 0; |
5899 | } | |
5262f567 | 5900 | |
62755e35 JA |
5901 | static bool io_cancel_cb(struct io_wq_work *work, void *data) |
5902 | { | |
5903 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
5904 | ||
5905 | return req->user_data == (unsigned long) data; | |
5906 | } | |
5907 | ||
e977d6d3 | 5908 | static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr) |
62755e35 | 5909 | { |
62755e35 | 5910 | enum io_wq_cancel cancel_ret; |
62755e35 JA |
5911 | int ret = 0; |
5912 | ||
4f26bda1 | 5913 | cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false); |
62755e35 JA |
5914 | switch (cancel_ret) { |
5915 | case IO_WQ_CANCEL_OK: | |
5916 | ret = 0; | |
5917 | break; | |
5918 | case IO_WQ_CANCEL_RUNNING: | |
5919 | ret = -EALREADY; | |
5920 | break; | |
5921 | case IO_WQ_CANCEL_NOTFOUND: | |
5922 | ret = -ENOENT; | |
5923 | break; | |
5924 | } | |
5925 | ||
e977d6d3 JA |
5926 | return ret; |
5927 | } | |
5928 | ||
47f46768 JA |
5929 | static void io_async_find_and_cancel(struct io_ring_ctx *ctx, |
5930 | struct io_kiocb *req, __u64 sqe_addr, | |
014db007 | 5931 | int success_ret) |
47f46768 JA |
5932 | { |
5933 | unsigned long flags; | |
5934 | int ret; | |
5935 | ||
5936 | ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr); | |
5937 | if (ret != -ENOENT) { | |
5938 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
5939 | goto done; | |
5940 | } | |
5941 | ||
5942 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
5943 | ret = io_timeout_cancel(ctx, sqe_addr); | |
5944 | if (ret != -ENOENT) | |
5945 | goto done; | |
5946 | ret = io_poll_cancel(ctx, sqe_addr); | |
5947 | done: | |
b0dd8a41 JA |
5948 | if (!ret) |
5949 | ret = success_ret; | |
47f46768 JA |
5950 | io_cqring_fill_event(req, ret); |
5951 | io_commit_cqring(ctx); | |
5952 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
5953 | io_cqring_ev_posted(ctx); | |
5954 | ||
4e88d6e7 JA |
5955 | if (ret < 0) |
5956 | req_set_fail_links(req); | |
014db007 | 5957 | io_put_req(req); |
47f46768 JA |
5958 | } |
5959 | ||
3529d8c2 JA |
5960 | static int io_async_cancel_prep(struct io_kiocb *req, |
5961 | const struct io_uring_sqe *sqe) | |
e977d6d3 | 5962 | { |
fbf23849 | 5963 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
e977d6d3 | 5964 | return -EINVAL; |
61710e43 DA |
5965 | if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) |
5966 | return -EINVAL; | |
5967 | if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags) | |
e977d6d3 JA |
5968 | return -EINVAL; |
5969 | ||
fbf23849 JA |
5970 | req->cancel.addr = READ_ONCE(sqe->addr); |
5971 | return 0; | |
5972 | } | |
5973 | ||
014db007 | 5974 | static int io_async_cancel(struct io_kiocb *req) |
fbf23849 JA |
5975 | { |
5976 | struct io_ring_ctx *ctx = req->ctx; | |
fbf23849 | 5977 | |
014db007 | 5978 | io_async_find_and_cancel(ctx, req, req->cancel.addr, 0); |
5262f567 JA |
5979 | return 0; |
5980 | } | |
5981 | ||
05f3fb3c JA |
5982 | static int io_files_update_prep(struct io_kiocb *req, |
5983 | const struct io_uring_sqe *sqe) | |
5984 | { | |
6ca56f84 JA |
5985 | if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL)) |
5986 | return -EINVAL; | |
61710e43 DA |
5987 | if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) |
5988 | return -EINVAL; | |
5989 | if (sqe->ioprio || sqe->rw_flags) | |
05f3fb3c JA |
5990 | return -EINVAL; |
5991 | ||
5992 | req->files_update.offset = READ_ONCE(sqe->off); | |
5993 | req->files_update.nr_args = READ_ONCE(sqe->len); | |
5994 | if (!req->files_update.nr_args) | |
5995 | return -EINVAL; | |
5996 | req->files_update.arg = READ_ONCE(sqe->addr); | |
5997 | return 0; | |
5998 | } | |
5999 | ||
229a7b63 JA |
6000 | static int io_files_update(struct io_kiocb *req, bool force_nonblock, |
6001 | struct io_comp_state *cs) | |
fbf23849 JA |
6002 | { |
6003 | struct io_ring_ctx *ctx = req->ctx; | |
05f3fb3c JA |
6004 | struct io_uring_files_update up; |
6005 | int ret; | |
fbf23849 | 6006 | |
f86cd20c | 6007 | if (force_nonblock) |
05f3fb3c | 6008 | return -EAGAIN; |
05f3fb3c JA |
6009 | |
6010 | up.offset = req->files_update.offset; | |
6011 | up.fds = req->files_update.arg; | |
6012 | ||
6013 | mutex_lock(&ctx->uring_lock); | |
6014 | ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args); | |
6015 | mutex_unlock(&ctx->uring_lock); | |
6016 | ||
6017 | if (ret < 0) | |
6018 | req_set_fail_links(req); | |
229a7b63 | 6019 | __io_req_complete(req, ret, 0, cs); |
5262f567 JA |
6020 | return 0; |
6021 | } | |
6022 | ||
bfe76559 | 6023 | static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f67676d1 | 6024 | { |
d625c6ee | 6025 | switch (req->opcode) { |
e781573e | 6026 | case IORING_OP_NOP: |
bfe76559 | 6027 | return 0; |
f67676d1 JA |
6028 | case IORING_OP_READV: |
6029 | case IORING_OP_READ_FIXED: | |
3a6820f2 | 6030 | case IORING_OP_READ: |
bfe76559 | 6031 | return io_read_prep(req, sqe); |
f67676d1 JA |
6032 | case IORING_OP_WRITEV: |
6033 | case IORING_OP_WRITE_FIXED: | |
3a6820f2 | 6034 | case IORING_OP_WRITE: |
bfe76559 | 6035 | return io_write_prep(req, sqe); |
0969e783 | 6036 | case IORING_OP_POLL_ADD: |
bfe76559 | 6037 | return io_poll_add_prep(req, sqe); |
0969e783 | 6038 | case IORING_OP_POLL_REMOVE: |
bfe76559 | 6039 | return io_poll_remove_prep(req, sqe); |
8ed8d3c3 | 6040 | case IORING_OP_FSYNC: |
bfe76559 | 6041 | return io_prep_fsync(req, sqe); |
8ed8d3c3 | 6042 | case IORING_OP_SYNC_FILE_RANGE: |
bfe76559 | 6043 | return io_prep_sfr(req, sqe); |
03b1230c | 6044 | case IORING_OP_SENDMSG: |
fddaface | 6045 | case IORING_OP_SEND: |
bfe76559 | 6046 | return io_sendmsg_prep(req, sqe); |
03b1230c | 6047 | case IORING_OP_RECVMSG: |
fddaface | 6048 | case IORING_OP_RECV: |
bfe76559 | 6049 | return io_recvmsg_prep(req, sqe); |
f499a021 | 6050 | case IORING_OP_CONNECT: |
bfe76559 | 6051 | return io_connect_prep(req, sqe); |
2d28390a | 6052 | case IORING_OP_TIMEOUT: |
bfe76559 | 6053 | return io_timeout_prep(req, sqe, false); |
b29472ee | 6054 | case IORING_OP_TIMEOUT_REMOVE: |
bfe76559 | 6055 | return io_timeout_remove_prep(req, sqe); |
fbf23849 | 6056 | case IORING_OP_ASYNC_CANCEL: |
bfe76559 | 6057 | return io_async_cancel_prep(req, sqe); |
2d28390a | 6058 | case IORING_OP_LINK_TIMEOUT: |
bfe76559 | 6059 | return io_timeout_prep(req, sqe, true); |
8ed8d3c3 | 6060 | case IORING_OP_ACCEPT: |
bfe76559 | 6061 | return io_accept_prep(req, sqe); |
d63d1b5e | 6062 | case IORING_OP_FALLOCATE: |
bfe76559 | 6063 | return io_fallocate_prep(req, sqe); |
15b71abe | 6064 | case IORING_OP_OPENAT: |
bfe76559 | 6065 | return io_openat_prep(req, sqe); |
b5dba59e | 6066 | case IORING_OP_CLOSE: |
bfe76559 | 6067 | return io_close_prep(req, sqe); |
05f3fb3c | 6068 | case IORING_OP_FILES_UPDATE: |
bfe76559 | 6069 | return io_files_update_prep(req, sqe); |
eddc7ef5 | 6070 | case IORING_OP_STATX: |
bfe76559 | 6071 | return io_statx_prep(req, sqe); |
4840e418 | 6072 | case IORING_OP_FADVISE: |
bfe76559 | 6073 | return io_fadvise_prep(req, sqe); |
c1ca757b | 6074 | case IORING_OP_MADVISE: |
bfe76559 | 6075 | return io_madvise_prep(req, sqe); |
cebdb986 | 6076 | case IORING_OP_OPENAT2: |
bfe76559 | 6077 | return io_openat2_prep(req, sqe); |
3e4827b0 | 6078 | case IORING_OP_EPOLL_CTL: |
bfe76559 | 6079 | return io_epoll_ctl_prep(req, sqe); |
7d67af2c | 6080 | case IORING_OP_SPLICE: |
bfe76559 | 6081 | return io_splice_prep(req, sqe); |
ddf0322d | 6082 | case IORING_OP_PROVIDE_BUFFERS: |
bfe76559 | 6083 | return io_provide_buffers_prep(req, sqe); |
067524e9 | 6084 | case IORING_OP_REMOVE_BUFFERS: |
bfe76559 | 6085 | return io_remove_buffers_prep(req, sqe); |
f2a8d5c7 | 6086 | case IORING_OP_TEE: |
bfe76559 | 6087 | return io_tee_prep(req, sqe); |
36f4fa68 JA |
6088 | case IORING_OP_SHUTDOWN: |
6089 | return io_shutdown_prep(req, sqe); | |
80a261fd JA |
6090 | case IORING_OP_RENAMEAT: |
6091 | return io_renameat_prep(req, sqe); | |
14a1143b JA |
6092 | case IORING_OP_UNLINKAT: |
6093 | return io_unlinkat_prep(req, sqe); | |
f67676d1 JA |
6094 | } |
6095 | ||
bfe76559 PB |
6096 | printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n", |
6097 | req->opcode); | |
6098 | return-EINVAL; | |
6099 | } | |
6100 | ||
6101 | static int io_req_defer_prep(struct io_kiocb *req, | |
6102 | const struct io_uring_sqe *sqe) | |
6103 | { | |
bfe76559 PB |
6104 | if (!sqe) |
6105 | return 0; | |
6106 | if (io_alloc_async_data(req)) | |
6107 | return -EAGAIN; | |
bfe76559 | 6108 | return io_req_prep(req, sqe); |
f67676d1 JA |
6109 | } |
6110 | ||
9cf7c104 PB |
6111 | static u32 io_get_sequence(struct io_kiocb *req) |
6112 | { | |
6113 | struct io_kiocb *pos; | |
6114 | struct io_ring_ctx *ctx = req->ctx; | |
f2f87370 | 6115 | u32 total_submitted, nr_reqs = 0; |
9cf7c104 | 6116 | |
f2f87370 PB |
6117 | io_for_each_link(pos, req) |
6118 | nr_reqs++; | |
9cf7c104 PB |
6119 | |
6120 | total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped; | |
6121 | return total_submitted - nr_reqs; | |
6122 | } | |
6123 | ||
3529d8c2 | 6124 | static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
de0617e4 | 6125 | { |
a197f664 | 6126 | struct io_ring_ctx *ctx = req->ctx; |
27dc8338 | 6127 | struct io_defer_entry *de; |
f67676d1 | 6128 | int ret; |
9cf7c104 | 6129 | u32 seq; |
de0617e4 | 6130 | |
9d858b21 | 6131 | /* Still need defer if there is pending req in defer list. */ |
9cf7c104 PB |
6132 | if (likely(list_empty_careful(&ctx->defer_list) && |
6133 | !(req->flags & REQ_F_IO_DRAIN))) | |
6134 | return 0; | |
6135 | ||
6136 | seq = io_get_sequence(req); | |
6137 | /* Still a chance to pass the sequence check */ | |
6138 | if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) | |
de0617e4 JA |
6139 | return 0; |
6140 | ||
e8c2bc1f | 6141 | if (!req->async_data) { |
650b5481 | 6142 | ret = io_req_defer_prep(req, sqe); |
327d6d96 | 6143 | if (ret) |
650b5481 PB |
6144 | return ret; |
6145 | } | |
cbdcb435 | 6146 | io_prep_async_link(req); |
27dc8338 PB |
6147 | de = kmalloc(sizeof(*de), GFP_KERNEL); |
6148 | if (!de) | |
6149 | return -ENOMEM; | |
2d28390a | 6150 | |
de0617e4 | 6151 | spin_lock_irq(&ctx->completion_lock); |
9cf7c104 | 6152 | if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { |
de0617e4 | 6153 | spin_unlock_irq(&ctx->completion_lock); |
27dc8338 | 6154 | kfree(de); |
ae34817b PB |
6155 | io_queue_async_work(req); |
6156 | return -EIOCBQUEUED; | |
de0617e4 JA |
6157 | } |
6158 | ||
915967f6 | 6159 | trace_io_uring_defer(ctx, req, req->user_data); |
27dc8338 | 6160 | de->req = req; |
9cf7c104 | 6161 | de->seq = seq; |
27dc8338 | 6162 | list_add_tail(&de->list, &ctx->defer_list); |
de0617e4 JA |
6163 | spin_unlock_irq(&ctx->completion_lock); |
6164 | return -EIOCBQUEUED; | |
6165 | } | |
6166 | ||
f573d384 | 6167 | static void io_req_drop_files(struct io_kiocb *req) |
99bc4c38 | 6168 | { |
f573d384 | 6169 | struct io_ring_ctx *ctx = req->ctx; |
c98de08c | 6170 | struct io_uring_task *tctx = req->task->io_uring; |
f573d384 JA |
6171 | unsigned long flags; |
6172 | ||
02a13674 JA |
6173 | if (req->work.flags & IO_WQ_WORK_FILES) { |
6174 | put_files_struct(req->work.identity->files); | |
6175 | put_nsproxy(req->work.identity->nsproxy); | |
6176 | } | |
f573d384 JA |
6177 | spin_lock_irqsave(&ctx->inflight_lock, flags); |
6178 | list_del(&req->inflight_entry); | |
f573d384 JA |
6179 | spin_unlock_irqrestore(&ctx->inflight_lock, flags); |
6180 | req->flags &= ~REQ_F_INFLIGHT; | |
dfead8a8 | 6181 | req->work.flags &= ~IO_WQ_WORK_FILES; |
dfea9fce PB |
6182 | if (atomic_read(&tctx->in_idle)) |
6183 | wake_up(&tctx->wait); | |
f573d384 | 6184 | } |
99bc4c38 | 6185 | |
3ca405eb | 6186 | static void __io_clean_op(struct io_kiocb *req) |
99bc4c38 | 6187 | { |
0e1b6fe3 PB |
6188 | if (req->flags & REQ_F_BUFFER_SELECTED) { |
6189 | switch (req->opcode) { | |
6190 | case IORING_OP_READV: | |
6191 | case IORING_OP_READ_FIXED: | |
6192 | case IORING_OP_READ: | |
bcda7baa | 6193 | kfree((void *)(unsigned long)req->rw.addr); |
0e1b6fe3 PB |
6194 | break; |
6195 | case IORING_OP_RECVMSG: | |
6196 | case IORING_OP_RECV: | |
bcda7baa | 6197 | kfree(req->sr_msg.kbuf); |
0e1b6fe3 PB |
6198 | break; |
6199 | } | |
6200 | req->flags &= ~REQ_F_BUFFER_SELECTED; | |
99bc4c38 PB |
6201 | } |
6202 | ||
0e1b6fe3 PB |
6203 | if (req->flags & REQ_F_NEED_CLEANUP) { |
6204 | switch (req->opcode) { | |
6205 | case IORING_OP_READV: | |
6206 | case IORING_OP_READ_FIXED: | |
6207 | case IORING_OP_READ: | |
6208 | case IORING_OP_WRITEV: | |
6209 | case IORING_OP_WRITE_FIXED: | |
e8c2bc1f JA |
6210 | case IORING_OP_WRITE: { |
6211 | struct io_async_rw *io = req->async_data; | |
6212 | if (io->free_iovec) | |
6213 | kfree(io->free_iovec); | |
0e1b6fe3 | 6214 | break; |
e8c2bc1f | 6215 | } |
0e1b6fe3 | 6216 | case IORING_OP_RECVMSG: |
e8c2bc1f JA |
6217 | case IORING_OP_SENDMSG: { |
6218 | struct io_async_msghdr *io = req->async_data; | |
6219 | if (io->iov != io->fast_iov) | |
6220 | kfree(io->iov); | |
0e1b6fe3 | 6221 | break; |
e8c2bc1f | 6222 | } |
0e1b6fe3 PB |
6223 | case IORING_OP_SPLICE: |
6224 | case IORING_OP_TEE: | |
6225 | io_put_file(req, req->splice.file_in, | |
6226 | (req->splice.flags & SPLICE_F_FD_IN_FIXED)); | |
6227 | break; | |
f3cd4850 JA |
6228 | case IORING_OP_OPENAT: |
6229 | case IORING_OP_OPENAT2: | |
6230 | if (req->open.filename) | |
6231 | putname(req->open.filename); | |
6232 | break; | |
80a261fd JA |
6233 | case IORING_OP_RENAMEAT: |
6234 | putname(req->rename.oldpath); | |
6235 | putname(req->rename.newpath); | |
6236 | break; | |
14a1143b JA |
6237 | case IORING_OP_UNLINKAT: |
6238 | putname(req->unlink.filename); | |
6239 | break; | |
0e1b6fe3 PB |
6240 | } |
6241 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
99bc4c38 | 6242 | } |
99bc4c38 PB |
6243 | } |
6244 | ||
c1379e24 PB |
6245 | static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock, |
6246 | struct io_comp_state *cs) | |
2b188cc1 | 6247 | { |
a197f664 | 6248 | struct io_ring_ctx *ctx = req->ctx; |
d625c6ee | 6249 | int ret; |
2b188cc1 | 6250 | |
d625c6ee | 6251 | switch (req->opcode) { |
2b188cc1 | 6252 | case IORING_OP_NOP: |
229a7b63 | 6253 | ret = io_nop(req, cs); |
2b188cc1 JA |
6254 | break; |
6255 | case IORING_OP_READV: | |
edafccee | 6256 | case IORING_OP_READ_FIXED: |
3a6820f2 | 6257 | case IORING_OP_READ: |
a1d7c393 | 6258 | ret = io_read(req, force_nonblock, cs); |
edafccee | 6259 | break; |
3529d8c2 | 6260 | case IORING_OP_WRITEV: |
edafccee | 6261 | case IORING_OP_WRITE_FIXED: |
3a6820f2 | 6262 | case IORING_OP_WRITE: |
a1d7c393 | 6263 | ret = io_write(req, force_nonblock, cs); |
2b188cc1 | 6264 | break; |
c992fe29 | 6265 | case IORING_OP_FSYNC: |
014db007 | 6266 | ret = io_fsync(req, force_nonblock); |
c992fe29 | 6267 | break; |
221c5eb2 | 6268 | case IORING_OP_POLL_ADD: |
014db007 | 6269 | ret = io_poll_add(req); |
221c5eb2 JA |
6270 | break; |
6271 | case IORING_OP_POLL_REMOVE: | |
fc4df999 | 6272 | ret = io_poll_remove(req); |
221c5eb2 | 6273 | break; |
5d17b4a4 | 6274 | case IORING_OP_SYNC_FILE_RANGE: |
014db007 | 6275 | ret = io_sync_file_range(req, force_nonblock); |
5d17b4a4 | 6276 | break; |
0fa03c62 | 6277 | case IORING_OP_SENDMSG: |
062d04d7 PB |
6278 | ret = io_sendmsg(req, force_nonblock, cs); |
6279 | break; | |
fddaface | 6280 | case IORING_OP_SEND: |
062d04d7 | 6281 | ret = io_send(req, force_nonblock, cs); |
0fa03c62 | 6282 | break; |
aa1fa28f | 6283 | case IORING_OP_RECVMSG: |
062d04d7 PB |
6284 | ret = io_recvmsg(req, force_nonblock, cs); |
6285 | break; | |
fddaface | 6286 | case IORING_OP_RECV: |
062d04d7 | 6287 | ret = io_recv(req, force_nonblock, cs); |
aa1fa28f | 6288 | break; |
5262f567 | 6289 | case IORING_OP_TIMEOUT: |
fc4df999 | 6290 | ret = io_timeout(req); |
5262f567 | 6291 | break; |
11365043 | 6292 | case IORING_OP_TIMEOUT_REMOVE: |
fc4df999 | 6293 | ret = io_timeout_remove(req); |
11365043 | 6294 | break; |
17f2fe35 | 6295 | case IORING_OP_ACCEPT: |
229a7b63 | 6296 | ret = io_accept(req, force_nonblock, cs); |
17f2fe35 | 6297 | break; |
f8e85cf2 | 6298 | case IORING_OP_CONNECT: |
229a7b63 | 6299 | ret = io_connect(req, force_nonblock, cs); |
f8e85cf2 | 6300 | break; |
62755e35 | 6301 | case IORING_OP_ASYNC_CANCEL: |
014db007 | 6302 | ret = io_async_cancel(req); |
62755e35 | 6303 | break; |
d63d1b5e | 6304 | case IORING_OP_FALLOCATE: |
014db007 | 6305 | ret = io_fallocate(req, force_nonblock); |
d63d1b5e | 6306 | break; |
15b71abe | 6307 | case IORING_OP_OPENAT: |
014db007 | 6308 | ret = io_openat(req, force_nonblock); |
15b71abe | 6309 | break; |
b5dba59e | 6310 | case IORING_OP_CLOSE: |
229a7b63 | 6311 | ret = io_close(req, force_nonblock, cs); |
b5dba59e | 6312 | break; |
05f3fb3c | 6313 | case IORING_OP_FILES_UPDATE: |
229a7b63 | 6314 | ret = io_files_update(req, force_nonblock, cs); |
05f3fb3c | 6315 | break; |
eddc7ef5 | 6316 | case IORING_OP_STATX: |
014db007 | 6317 | ret = io_statx(req, force_nonblock); |
eddc7ef5 | 6318 | break; |
4840e418 | 6319 | case IORING_OP_FADVISE: |
014db007 | 6320 | ret = io_fadvise(req, force_nonblock); |
4840e418 | 6321 | break; |
c1ca757b | 6322 | case IORING_OP_MADVISE: |
014db007 | 6323 | ret = io_madvise(req, force_nonblock); |
c1ca757b | 6324 | break; |
cebdb986 | 6325 | case IORING_OP_OPENAT2: |
014db007 | 6326 | ret = io_openat2(req, force_nonblock); |
cebdb986 | 6327 | break; |
3e4827b0 | 6328 | case IORING_OP_EPOLL_CTL: |
229a7b63 | 6329 | ret = io_epoll_ctl(req, force_nonblock, cs); |
3e4827b0 | 6330 | break; |
7d67af2c | 6331 | case IORING_OP_SPLICE: |
014db007 | 6332 | ret = io_splice(req, force_nonblock); |
7d67af2c | 6333 | break; |
ddf0322d | 6334 | case IORING_OP_PROVIDE_BUFFERS: |
229a7b63 | 6335 | ret = io_provide_buffers(req, force_nonblock, cs); |
ddf0322d | 6336 | break; |
067524e9 | 6337 | case IORING_OP_REMOVE_BUFFERS: |
229a7b63 | 6338 | ret = io_remove_buffers(req, force_nonblock, cs); |
3e4827b0 | 6339 | break; |
f2a8d5c7 | 6340 | case IORING_OP_TEE: |
f2a8d5c7 PB |
6341 | ret = io_tee(req, force_nonblock); |
6342 | break; | |
36f4fa68 JA |
6343 | case IORING_OP_SHUTDOWN: |
6344 | ret = io_shutdown(req, force_nonblock); | |
6345 | break; | |
80a261fd JA |
6346 | case IORING_OP_RENAMEAT: |
6347 | ret = io_renameat(req, force_nonblock); | |
6348 | break; | |
14a1143b JA |
6349 | case IORING_OP_UNLINKAT: |
6350 | ret = io_unlinkat(req, force_nonblock); | |
6351 | break; | |
2b188cc1 JA |
6352 | default: |
6353 | ret = -EINVAL; | |
6354 | break; | |
6355 | } | |
6356 | ||
def596e9 JA |
6357 | if (ret) |
6358 | return ret; | |
6359 | ||
b532576e JA |
6360 | /* If the op doesn't have a file, we're not polling for it */ |
6361 | if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) { | |
11ba820b JA |
6362 | const bool in_async = io_wq_current_is_worker(); |
6363 | ||
11ba820b JA |
6364 | /* workqueue context doesn't hold uring_lock, grab it now */ |
6365 | if (in_async) | |
6366 | mutex_lock(&ctx->uring_lock); | |
6367 | ||
2e9dbe90 | 6368 | io_iopoll_req_issued(req, in_async); |
11ba820b JA |
6369 | |
6370 | if (in_async) | |
6371 | mutex_unlock(&ctx->uring_lock); | |
def596e9 JA |
6372 | } |
6373 | ||
6374 | return 0; | |
2b188cc1 JA |
6375 | } |
6376 | ||
f4db7182 | 6377 | static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work) |
2b188cc1 JA |
6378 | { |
6379 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
6df1db6b | 6380 | struct io_kiocb *timeout; |
561fb04a | 6381 | int ret = 0; |
2b188cc1 | 6382 | |
6df1db6b PB |
6383 | timeout = io_prep_linked_timeout(req); |
6384 | if (timeout) | |
6385 | io_queue_linked_timeout(timeout); | |
d4c81f38 | 6386 | |
0c9d5ccd JA |
6387 | /* if NO_CANCEL is set, we must still run the work */ |
6388 | if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) == | |
6389 | IO_WQ_WORK_CANCEL) { | |
561fb04a | 6390 | ret = -ECANCELED; |
0c9d5ccd | 6391 | } |
31b51510 | 6392 | |
561fb04a | 6393 | if (!ret) { |
561fb04a | 6394 | do { |
c1379e24 | 6395 | ret = io_issue_sqe(req, false, NULL); |
561fb04a JA |
6396 | /* |
6397 | * We can get EAGAIN for polled IO even though we're | |
6398 | * forcing a sync submission from here, since we can't | |
6399 | * wait for request slots on the block side. | |
6400 | */ | |
6401 | if (ret != -EAGAIN) | |
6402 | break; | |
6403 | cond_resched(); | |
6404 | } while (1); | |
6405 | } | |
31b51510 | 6406 | |
561fb04a | 6407 | if (ret) { |
c07e6719 XW |
6408 | struct io_ring_ctx *lock_ctx = NULL; |
6409 | ||
6410 | if (req->ctx->flags & IORING_SETUP_IOPOLL) | |
6411 | lock_ctx = req->ctx; | |
6412 | ||
dad1b124 | 6413 | /* |
c07e6719 XW |
6414 | * io_iopoll_complete() does not hold completion_lock to |
6415 | * complete polled io, so here for polled io, we can not call | |
6416 | * io_req_complete() directly, otherwise there maybe concurrent | |
6417 | * access to cqring, defer_list, etc, which is not safe. Given | |
6418 | * that io_iopoll_complete() is always called under uring_lock, | |
6419 | * so here for polled io, we also get uring_lock to complete | |
6420 | * it. | |
dad1b124 | 6421 | */ |
c07e6719 XW |
6422 | if (lock_ctx) |
6423 | mutex_lock(&lock_ctx->uring_lock); | |
dad1b124 | 6424 | |
c07e6719 XW |
6425 | req_set_fail_links(req); |
6426 | io_req_complete(req, ret); | |
6427 | ||
6428 | if (lock_ctx) | |
6429 | mutex_unlock(&lock_ctx->uring_lock); | |
edafccee | 6430 | } |
2b188cc1 | 6431 | |
f4db7182 | 6432 | return io_steal_work(req); |
2b188cc1 JA |
6433 | } |
6434 | ||
65e19f54 JA |
6435 | static inline struct file *io_file_from_index(struct io_ring_ctx *ctx, |
6436 | int index) | |
6437 | { | |
6438 | struct fixed_file_table *table; | |
6439 | ||
05f3fb3c | 6440 | table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT]; |
84695089 | 6441 | return table->files[index & IORING_FILE_TABLE_MASK]; |
65e19f54 JA |
6442 | } |
6443 | ||
8371adf5 PB |
6444 | static struct file *io_file_get(struct io_submit_state *state, |
6445 | struct io_kiocb *req, int fd, bool fixed) | |
09bb8394 | 6446 | { |
a197f664 | 6447 | struct io_ring_ctx *ctx = req->ctx; |
8da11c19 | 6448 | struct file *file; |
09bb8394 | 6449 | |
8da11c19 | 6450 | if (fixed) { |
479f517b | 6451 | if (unlikely((unsigned int)fd >= ctx->nr_user_files)) |
8371adf5 | 6452 | return NULL; |
b7620121 | 6453 | fd = array_index_nospec(fd, ctx->nr_user_files); |
8da11c19 | 6454 | file = io_file_from_index(ctx, fd); |
36f72fe2 | 6455 | io_set_resource_node(req); |
09bb8394 | 6456 | } else { |
c826bd7a | 6457 | trace_io_uring_file_get(ctx, fd); |
8da11c19 | 6458 | file = __io_file_get(state, fd); |
09bb8394 JA |
6459 | } |
6460 | ||
02a13674 JA |
6461 | if (file && file->f_op == &io_uring_fops) { |
6462 | io_req_init_async(req); | |
6463 | req->flags |= REQ_F_INFLIGHT; | |
6464 | ||
6465 | spin_lock_irq(&ctx->inflight_lock); | |
6466 | list_add(&req->inflight_entry, &ctx->inflight_list); | |
6467 | spin_unlock_irq(&ctx->inflight_lock); | |
6468 | } | |
6469 | ||
8371adf5 | 6470 | return file; |
09bb8394 JA |
6471 | } |
6472 | ||
2665abfd | 6473 | static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer) |
2b188cc1 | 6474 | { |
ad8a48ac JA |
6475 | struct io_timeout_data *data = container_of(timer, |
6476 | struct io_timeout_data, timer); | |
90cd7e42 | 6477 | struct io_kiocb *prev, *req = data->req; |
2665abfd | 6478 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd | 6479 | unsigned long flags; |
2665abfd JA |
6480 | |
6481 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
90cd7e42 PB |
6482 | prev = req->timeout.head; |
6483 | req->timeout.head = NULL; | |
2665abfd JA |
6484 | |
6485 | /* | |
6486 | * We don't expect the list to be empty, that will only happen if we | |
6487 | * race with the completion of the linked work. | |
6488 | */ | |
90cd7e42 | 6489 | if (prev && refcount_inc_not_zero(&prev->refs)) |
f2f87370 | 6490 | io_remove_next_linked(prev); |
90cd7e42 PB |
6491 | else |
6492 | prev = NULL; | |
2665abfd JA |
6493 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
6494 | ||
6495 | if (prev) { | |
4e88d6e7 | 6496 | req_set_fail_links(prev); |
014db007 | 6497 | io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME); |
76a46e06 | 6498 | io_put_req(prev); |
47f46768 | 6499 | } else { |
e1e16097 | 6500 | io_req_complete(req, -ETIME); |
2665abfd | 6501 | } |
2665abfd JA |
6502 | return HRTIMER_NORESTART; |
6503 | } | |
6504 | ||
7271ef3a | 6505 | static void __io_queue_linked_timeout(struct io_kiocb *req) |
2665abfd | 6506 | { |
76a46e06 | 6507 | /* |
f2f87370 PB |
6508 | * If the back reference is NULL, then our linked request finished |
6509 | * before we got a chance to setup the timer | |
76a46e06 | 6510 | */ |
90cd7e42 | 6511 | if (req->timeout.head) { |
e8c2bc1f | 6512 | struct io_timeout_data *data = req->async_data; |
94ae5e77 | 6513 | |
ad8a48ac JA |
6514 | data->timer.function = io_link_timeout_fn; |
6515 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), | |
6516 | data->mode); | |
2665abfd | 6517 | } |
7271ef3a JA |
6518 | } |
6519 | ||
6520 | static void io_queue_linked_timeout(struct io_kiocb *req) | |
6521 | { | |
6522 | struct io_ring_ctx *ctx = req->ctx; | |
6523 | ||
6524 | spin_lock_irq(&ctx->completion_lock); | |
6525 | __io_queue_linked_timeout(req); | |
76a46e06 | 6526 | spin_unlock_irq(&ctx->completion_lock); |
2665abfd | 6527 | |
2665abfd | 6528 | /* drop submission reference */ |
76a46e06 JA |
6529 | io_put_req(req); |
6530 | } | |
2665abfd | 6531 | |
ad8a48ac | 6532 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) |
2665abfd | 6533 | { |
f2f87370 | 6534 | struct io_kiocb *nxt = req->link; |
2665abfd | 6535 | |
f2f87370 PB |
6536 | if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) || |
6537 | nxt->opcode != IORING_OP_LINK_TIMEOUT) | |
76a46e06 | 6538 | return NULL; |
2665abfd | 6539 | |
90cd7e42 | 6540 | nxt->timeout.head = req; |
900fad45 | 6541 | nxt->flags |= REQ_F_LTIMEOUT_ACTIVE; |
76a46e06 | 6542 | req->flags |= REQ_F_LINK_TIMEOUT; |
76a46e06 | 6543 | return nxt; |
2665abfd JA |
6544 | } |
6545 | ||
c1379e24 | 6546 | static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs) |
2b188cc1 | 6547 | { |
4a0a7a18 | 6548 | struct io_kiocb *linked_timeout; |
193155c8 | 6549 | const struct cred *old_creds = NULL; |
e0c5c576 | 6550 | int ret; |
2b188cc1 | 6551 | |
4a0a7a18 JA |
6552 | again: |
6553 | linked_timeout = io_prep_linked_timeout(req); | |
6554 | ||
2e5aa6cb PB |
6555 | if ((req->flags & REQ_F_WORK_INITIALIZED) && |
6556 | (req->work.flags & IO_WQ_WORK_CREDS) && | |
98447d65 | 6557 | req->work.identity->creds != current_cred()) { |
193155c8 JA |
6558 | if (old_creds) |
6559 | revert_creds(old_creds); | |
98447d65 | 6560 | if (old_creds == req->work.identity->creds) |
193155c8 JA |
6561 | old_creds = NULL; /* restored original creds */ |
6562 | else | |
98447d65 | 6563 | old_creds = override_creds(req->work.identity->creds); |
193155c8 JA |
6564 | } |
6565 | ||
c1379e24 | 6566 | ret = io_issue_sqe(req, true, cs); |
491381ce JA |
6567 | |
6568 | /* | |
6569 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
6570 | * doesn't support non-blocking read/write attempts | |
6571 | */ | |
24c74678 | 6572 | if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) { |
f063c547 | 6573 | if (!io_arm_poll_handler(req)) { |
f063c547 PB |
6574 | /* |
6575 | * Queued up for async execution, worker will release | |
6576 | * submit reference when the iocb is actually submitted. | |
6577 | */ | |
6578 | io_queue_async_work(req); | |
2b188cc1 | 6579 | } |
bbad27b2 | 6580 | |
f063c547 PB |
6581 | if (linked_timeout) |
6582 | io_queue_linked_timeout(linked_timeout); | |
0d63c148 PB |
6583 | } else if (likely(!ret)) { |
6584 | /* drop submission reference */ | |
6585 | req = io_put_req_find_next(req); | |
6586 | if (linked_timeout) | |
6587 | io_queue_linked_timeout(linked_timeout); | |
e65ef56d | 6588 | |
0d63c148 PB |
6589 | if (req) { |
6590 | if (!(req->flags & REQ_F_FORCE_ASYNC)) | |
6591 | goto again; | |
6592 | io_queue_async_work(req); | |
6593 | } | |
6594 | } else { | |
652532ad PB |
6595 | /* un-prep timeout, so it'll be killed as any other linked */ |
6596 | req->flags &= ~REQ_F_LINK_TIMEOUT; | |
4e88d6e7 | 6597 | req_set_fail_links(req); |
e65ef56d | 6598 | io_put_req(req); |
e1e16097 | 6599 | io_req_complete(req, ret); |
9e645e11 | 6600 | } |
652532ad | 6601 | |
193155c8 JA |
6602 | if (old_creds) |
6603 | revert_creds(old_creds); | |
2b188cc1 JA |
6604 | } |
6605 | ||
f13fad7b JA |
6606 | static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
6607 | struct io_comp_state *cs) | |
4fe2c963 JL |
6608 | { |
6609 | int ret; | |
6610 | ||
3529d8c2 | 6611 | ret = io_req_defer(req, sqe); |
4fe2c963 JL |
6612 | if (ret) { |
6613 | if (ret != -EIOCBQUEUED) { | |
1118591a | 6614 | fail_req: |
4e88d6e7 | 6615 | req_set_fail_links(req); |
e1e16097 JA |
6616 | io_put_req(req); |
6617 | io_req_complete(req, ret); | |
4fe2c963 | 6618 | } |
2550878f | 6619 | } else if (req->flags & REQ_F_FORCE_ASYNC) { |
e8c2bc1f | 6620 | if (!req->async_data) { |
bd2ab18a | 6621 | ret = io_req_defer_prep(req, sqe); |
327d6d96 | 6622 | if (unlikely(ret)) |
bd2ab18a PB |
6623 | goto fail_req; |
6624 | } | |
ce35a47a JA |
6625 | io_queue_async_work(req); |
6626 | } else { | |
c1379e24 PB |
6627 | if (sqe) { |
6628 | ret = io_req_prep(req, sqe); | |
6629 | if (unlikely(ret)) | |
6630 | goto fail_req; | |
6631 | } | |
6632 | __io_queue_sqe(req, cs); | |
ce35a47a | 6633 | } |
4fe2c963 JL |
6634 | } |
6635 | ||
f13fad7b JA |
6636 | static inline void io_queue_link_head(struct io_kiocb *req, |
6637 | struct io_comp_state *cs) | |
4fe2c963 | 6638 | { |
94ae5e77 | 6639 | if (unlikely(req->flags & REQ_F_FAIL_LINK)) { |
e1e16097 JA |
6640 | io_put_req(req); |
6641 | io_req_complete(req, -ECANCELED); | |
1b4a51b6 | 6642 | } else |
f13fad7b | 6643 | io_queue_sqe(req, NULL, cs); |
4fe2c963 JL |
6644 | } |
6645 | ||
863e0560 PB |
6646 | struct io_submit_link { |
6647 | struct io_kiocb *head; | |
6648 | struct io_kiocb *last; | |
6649 | }; | |
6650 | ||
1d4240cc | 6651 | static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
863e0560 | 6652 | struct io_submit_link *link, struct io_comp_state *cs) |
9e645e11 | 6653 | { |
a197f664 | 6654 | struct io_ring_ctx *ctx = req->ctx; |
ef4ff581 | 6655 | int ret; |
9e645e11 | 6656 | |
9e645e11 JA |
6657 | /* |
6658 | * If we already have a head request, queue this one for async | |
6659 | * submittal once the head completes. If we don't have a head but | |
6660 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
6661 | * submitted sync once the chain is complete. If none of those | |
6662 | * conditions are true (normal request), then just queue it. | |
6663 | */ | |
863e0560 PB |
6664 | if (link->head) { |
6665 | struct io_kiocb *head = link->head; | |
4e88d6e7 | 6666 | |
8cdf2193 PB |
6667 | /* |
6668 | * Taking sequential execution of a link, draining both sides | |
6669 | * of the link also fullfils IOSQE_IO_DRAIN semantics for all | |
6670 | * requests in the link. So, it drains the head and the | |
6671 | * next after the link request. The last one is done via | |
6672 | * drain_next flag to persist the effect across calls. | |
6673 | */ | |
ef4ff581 | 6674 | if (req->flags & REQ_F_IO_DRAIN) { |
711be031 PB |
6675 | head->flags |= REQ_F_IO_DRAIN; |
6676 | ctx->drain_next = 1; | |
6677 | } | |
3529d8c2 | 6678 | ret = io_req_defer_prep(req, sqe); |
327d6d96 | 6679 | if (unlikely(ret)) { |
4e88d6e7 | 6680 | /* fail even hard links since we don't submit */ |
9d76377f | 6681 | head->flags |= REQ_F_FAIL_LINK; |
1d4240cc | 6682 | return ret; |
2d28390a | 6683 | } |
9d76377f | 6684 | trace_io_uring_link(ctx, req, head); |
f2f87370 | 6685 | link->last->link = req; |
863e0560 | 6686 | link->last = req; |
32fe525b PB |
6687 | |
6688 | /* last request of a link, enqueue the link */ | |
ef4ff581 | 6689 | if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) { |
f13fad7b | 6690 | io_queue_link_head(head, cs); |
863e0560 | 6691 | link->head = NULL; |
32fe525b | 6692 | } |
9e645e11 | 6693 | } else { |
711be031 PB |
6694 | if (unlikely(ctx->drain_next)) { |
6695 | req->flags |= REQ_F_IO_DRAIN; | |
ef4ff581 | 6696 | ctx->drain_next = 0; |
711be031 | 6697 | } |
ef4ff581 | 6698 | if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) { |
711be031 | 6699 | ret = io_req_defer_prep(req, sqe); |
327d6d96 | 6700 | if (unlikely(ret)) |
711be031 | 6701 | req->flags |= REQ_F_FAIL_LINK; |
863e0560 PB |
6702 | link->head = req; |
6703 | link->last = req; | |
711be031 | 6704 | } else { |
f13fad7b | 6705 | io_queue_sqe(req, sqe, cs); |
711be031 | 6706 | } |
9e645e11 | 6707 | } |
2e6e1fde | 6708 | |
1d4240cc | 6709 | return 0; |
9e645e11 JA |
6710 | } |
6711 | ||
9a56a232 JA |
6712 | /* |
6713 | * Batched submission is done, ensure local IO is flushed out. | |
6714 | */ | |
6715 | static void io_submit_state_end(struct io_submit_state *state) | |
6716 | { | |
f13fad7b JA |
6717 | if (!list_empty(&state->comp.list)) |
6718 | io_submit_flush_completions(&state->comp); | |
27926b68 JA |
6719 | if (state->plug_started) |
6720 | blk_finish_plug(&state->plug); | |
9f13c35b | 6721 | io_state_file_put(state); |
2579f913 | 6722 | if (state->free_reqs) |
6c8a3134 | 6723 | kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs); |
9a56a232 JA |
6724 | } |
6725 | ||
6726 | /* | |
6727 | * Start submission side cache. | |
6728 | */ | |
6729 | static void io_submit_state_start(struct io_submit_state *state, | |
013538bd | 6730 | struct io_ring_ctx *ctx, unsigned int max_ios) |
9a56a232 | 6731 | { |
27926b68 | 6732 | state->plug_started = false; |
013538bd JA |
6733 | state->comp.nr = 0; |
6734 | INIT_LIST_HEAD(&state->comp.list); | |
6735 | state->comp.ctx = ctx; | |
2579f913 | 6736 | state->free_reqs = 0; |
6e1271e6 | 6737 | state->file_refs = 0; |
9a56a232 JA |
6738 | state->ios_left = max_ios; |
6739 | } | |
6740 | ||
2b188cc1 JA |
6741 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
6742 | { | |
75b28aff | 6743 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 6744 | |
caf582c6 PB |
6745 | /* |
6746 | * Ensure any loads from the SQEs are done at this point, | |
6747 | * since once we write the new head, the application could | |
6748 | * write new data to them. | |
6749 | */ | |
6750 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
6751 | } |
6752 | ||
2b188cc1 | 6753 | /* |
3529d8c2 | 6754 | * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory |
2b188cc1 JA |
6755 | * that is mapped by userspace. This means that care needs to be taken to |
6756 | * ensure that reads are stable, as we cannot rely on userspace always | |
6757 | * being a good citizen. If members of the sqe are validated and then later | |
6758 | * used, it's important that those reads are done through READ_ONCE() to | |
6759 | * prevent a re-load down the line. | |
6760 | */ | |
709b302f | 6761 | static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx) |
2b188cc1 | 6762 | { |
75b28aff | 6763 | u32 *sq_array = ctx->sq_array; |
2b188cc1 JA |
6764 | unsigned head; |
6765 | ||
6766 | /* | |
6767 | * The cached sq head (or cq tail) serves two purposes: | |
6768 | * | |
6769 | * 1) allows us to batch the cost of updating the user visible | |
6770 | * head updates. | |
6771 | * 2) allows the kernel side to track the head on its own, even | |
6772 | * though the application is the one updating it. | |
6773 | */ | |
ee7d46d9 | 6774 | head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]); |
709b302f PB |
6775 | if (likely(head < ctx->sq_entries)) |
6776 | return &ctx->sq_sqes[head]; | |
2b188cc1 JA |
6777 | |
6778 | /* drop invalid entries */ | |
498ccd9e | 6779 | ctx->cached_sq_dropped++; |
ee7d46d9 | 6780 | WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped); |
709b302f PB |
6781 | return NULL; |
6782 | } | |
6783 | ||
6784 | static inline void io_consume_sqe(struct io_ring_ctx *ctx) | |
6785 | { | |
6786 | ctx->cached_sq_head++; | |
2b188cc1 JA |
6787 | } |
6788 | ||
21b55dbc SG |
6789 | /* |
6790 | * Check SQE restrictions (opcode and flags). | |
6791 | * | |
6792 | * Returns 'true' if SQE is allowed, 'false' otherwise. | |
6793 | */ | |
6794 | static inline bool io_check_restriction(struct io_ring_ctx *ctx, | |
6795 | struct io_kiocb *req, | |
6796 | unsigned int sqe_flags) | |
6797 | { | |
6798 | if (!ctx->restricted) | |
6799 | return true; | |
6800 | ||
6801 | if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) | |
6802 | return false; | |
6803 | ||
6804 | if ((sqe_flags & ctx->restrictions.sqe_flags_required) != | |
6805 | ctx->restrictions.sqe_flags_required) | |
6806 | return false; | |
6807 | ||
6808 | if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | | |
6809 | ctx->restrictions.sqe_flags_required)) | |
6810 | return false; | |
6811 | ||
6812 | return true; | |
6813 | } | |
6814 | ||
ef4ff581 PB |
6815 | #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \ |
6816 | IOSQE_IO_HARDLINK | IOSQE_ASYNC | \ | |
6817 | IOSQE_BUFFER_SELECT) | |
6818 | ||
6819 | static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, | |
6820 | const struct io_uring_sqe *sqe, | |
0cdaf760 | 6821 | struct io_submit_state *state) |
0553b8bd | 6822 | { |
ef4ff581 | 6823 | unsigned int sqe_flags; |
71b547c0 | 6824 | int id, ret; |
ef4ff581 | 6825 | |
0553b8bd PB |
6826 | req->opcode = READ_ONCE(sqe->opcode); |
6827 | req->user_data = READ_ONCE(sqe->user_data); | |
e8c2bc1f | 6828 | req->async_data = NULL; |
0553b8bd PB |
6829 | req->file = NULL; |
6830 | req->ctx = ctx; | |
6831 | req->flags = 0; | |
f2f87370 | 6832 | req->link = NULL; |
36f72fe2 | 6833 | req->fixed_file_refs = NULL; |
0553b8bd PB |
6834 | /* one is dropped after submission, the other at completion */ |
6835 | refcount_set(&req->refs, 2); | |
4dd2824d | 6836 | req->task = current; |
0553b8bd | 6837 | req->result = 0; |
ef4ff581 PB |
6838 | |
6839 | if (unlikely(req->opcode >= IORING_OP_LAST)) | |
6840 | return -EINVAL; | |
6841 | ||
28cea78a | 6842 | if (unlikely(io_sq_thread_acquire_mm_files(ctx, req))) |
9d8426a0 | 6843 | return -EFAULT; |
ef4ff581 PB |
6844 | |
6845 | sqe_flags = READ_ONCE(sqe->flags); | |
6846 | /* enforce forwards compatibility on users */ | |
6847 | if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) | |
6848 | return -EINVAL; | |
6849 | ||
21b55dbc SG |
6850 | if (unlikely(!io_check_restriction(ctx, req, sqe_flags))) |
6851 | return -EACCES; | |
6852 | ||
ef4ff581 PB |
6853 | if ((sqe_flags & IOSQE_BUFFER_SELECT) && |
6854 | !io_op_defs[req->opcode].buffer_select) | |
6855 | return -EOPNOTSUPP; | |
6856 | ||
6857 | id = READ_ONCE(sqe->personality); | |
6858 | if (id) { | |
1e6fa521 JA |
6859 | struct io_identity *iod; |
6860 | ||
1e6fa521 JA |
6861 | iod = idr_find(&ctx->personality_idr, id); |
6862 | if (unlikely(!iod)) | |
ef4ff581 | 6863 | return -EINVAL; |
1e6fa521 | 6864 | refcount_inc(&iod->count); |
ec99ca6c PB |
6865 | |
6866 | __io_req_init_async(req); | |
1e6fa521 JA |
6867 | get_cred(iod->creds); |
6868 | req->work.identity = iod; | |
dfead8a8 | 6869 | req->work.flags |= IO_WQ_WORK_CREDS; |
ef4ff581 PB |
6870 | } |
6871 | ||
6872 | /* same numerical values with corresponding REQ_F_*, safe to copy */ | |
c11368a5 | 6873 | req->flags |= sqe_flags; |
ef4ff581 | 6874 | |
27926b68 JA |
6875 | /* |
6876 | * Plug now if we have more than 1 IO left after this, and the target | |
6877 | * is potentially a read/write to block based storage. | |
6878 | */ | |
6879 | if (!state->plug_started && state->ios_left > 1 && | |
6880 | io_op_defs[req->opcode].plug) { | |
6881 | blk_start_plug(&state->plug); | |
6882 | state->plug_started = true; | |
6883 | } | |
6884 | ||
bd5bbda7 PB |
6885 | ret = 0; |
6886 | if (io_op_defs[req->opcode].needs_file) { | |
6887 | bool fixed = req->flags & REQ_F_FIXED_FILE; | |
6888 | ||
6889 | req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed); | |
6890 | if (unlikely(!req->file && | |
6891 | !io_op_defs[req->opcode].needs_file_no_error)) | |
6892 | ret = -EBADF; | |
6893 | } | |
63ff8223 | 6894 | |
71b547c0 PB |
6895 | state->ios_left--; |
6896 | return ret; | |
0553b8bd PB |
6897 | } |
6898 | ||
0f212204 | 6899 | static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) |
6c271ce2 | 6900 | { |
ac8691c4 | 6901 | struct io_submit_state state; |
863e0560 | 6902 | struct io_submit_link link; |
9e645e11 | 6903 | int i, submitted = 0; |
6c271ce2 | 6904 | |
c4a2ed72 | 6905 | /* if we have a backlog and couldn't flush it all, return BUSY */ |
ad3eb2c8 | 6906 | if (test_bit(0, &ctx->sq_check_overflow)) { |
6c503150 | 6907 | if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL)) |
ad3eb2c8 JA |
6908 | return -EBUSY; |
6909 | } | |
6c271ce2 | 6910 | |
ee7d46d9 PB |
6911 | /* make sure SQ entry isn't read before tail */ |
6912 | nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx)); | |
9ef4f124 | 6913 | |
2b85edfc PB |
6914 | if (!percpu_ref_tryget_many(&ctx->refs, nr)) |
6915 | return -EAGAIN; | |
6c271ce2 | 6916 | |
d8a6df10 | 6917 | percpu_counter_add(¤t->io_uring->inflight, nr); |
faf7b51c | 6918 | refcount_add(nr, ¤t->usage); |
6c271ce2 | 6919 | |
013538bd | 6920 | io_submit_state_start(&state, ctx, nr); |
863e0560 | 6921 | link.head = NULL; |
b14cca0c | 6922 | |
6c271ce2 | 6923 | for (i = 0; i < nr; i++) { |
3529d8c2 | 6924 | const struct io_uring_sqe *sqe; |
196be95c | 6925 | struct io_kiocb *req; |
1cb1edb2 | 6926 | int err; |
fb5ccc98 | 6927 | |
b1e50e54 PB |
6928 | sqe = io_get_sqe(ctx); |
6929 | if (unlikely(!sqe)) { | |
6930 | io_consume_sqe(ctx); | |
6931 | break; | |
6932 | } | |
ac8691c4 | 6933 | req = io_alloc_req(ctx, &state); |
196be95c PB |
6934 | if (unlikely(!req)) { |
6935 | if (!submitted) | |
6936 | submitted = -EAGAIN; | |
fb5ccc98 | 6937 | break; |
196be95c | 6938 | } |
709b302f | 6939 | io_consume_sqe(ctx); |
d3656344 JA |
6940 | /* will complete beyond this point, count as submitted */ |
6941 | submitted++; | |
6942 | ||
692d8363 | 6943 | err = io_init_req(ctx, req, sqe, &state); |
ef4ff581 | 6944 | if (unlikely(err)) { |
1cb1edb2 | 6945 | fail_req: |
e1e16097 JA |
6946 | io_put_req(req); |
6947 | io_req_complete(req, err); | |
196be95c PB |
6948 | break; |
6949 | } | |
fb5ccc98 | 6950 | |
354420f7 | 6951 | trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data, |
0cdaf760 | 6952 | true, io_async_submit(ctx)); |
f13fad7b | 6953 | err = io_submit_sqe(req, sqe, &link, &state.comp); |
1d4240cc PB |
6954 | if (err) |
6955 | goto fail_req; | |
6c271ce2 JA |
6956 | } |
6957 | ||
9466f437 PB |
6958 | if (unlikely(submitted != nr)) { |
6959 | int ref_used = (submitted == -EAGAIN) ? 0 : submitted; | |
d8a6df10 JA |
6960 | struct io_uring_task *tctx = current->io_uring; |
6961 | int unused = nr - ref_used; | |
9466f437 | 6962 | |
d8a6df10 JA |
6963 | percpu_ref_put_many(&ctx->refs, unused); |
6964 | percpu_counter_sub(&tctx->inflight, unused); | |
6965 | put_task_struct_many(current, unused); | |
9466f437 | 6966 | } |
863e0560 PB |
6967 | if (link.head) |
6968 | io_queue_link_head(link.head, &state.comp); | |
ac8691c4 | 6969 | io_submit_state_end(&state); |
6c271ce2 | 6970 | |
ae9428ca PB |
6971 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
6972 | io_commit_sqring(ctx); | |
6973 | ||
6c271ce2 JA |
6974 | return submitted; |
6975 | } | |
6976 | ||
23b3628e XW |
6977 | static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx) |
6978 | { | |
6979 | /* Tell userspace we may need a wakeup call */ | |
6980 | spin_lock_irq(&ctx->completion_lock); | |
6981 | ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP; | |
6982 | spin_unlock_irq(&ctx->completion_lock); | |
6983 | } | |
6984 | ||
6985 | static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx) | |
6986 | { | |
6987 | spin_lock_irq(&ctx->completion_lock); | |
6988 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; | |
6989 | spin_unlock_irq(&ctx->completion_lock); | |
6990 | } | |
6991 | ||
08369246 | 6992 | static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries) |
6c271ce2 | 6993 | { |
c8d1ba58 | 6994 | unsigned int to_submit; |
bdcd3eab | 6995 | int ret = 0; |
6c271ce2 | 6996 | |
c8d1ba58 | 6997 | to_submit = io_sqring_entries(ctx); |
e95eee2d JA |
6998 | /* if we're handling multiple rings, cap submit size for fairness */ |
6999 | if (cap_entries && to_submit > 8) | |
7000 | to_submit = 8; | |
7001 | ||
906a3c6f | 7002 | if (!list_empty(&ctx->iopoll_list) || to_submit) { |
c8d1ba58 | 7003 | unsigned nr_events = 0; |
a4c0b3de | 7004 | |
c8d1ba58 | 7005 | mutex_lock(&ctx->uring_lock); |
906a3c6f | 7006 | if (!list_empty(&ctx->iopoll_list)) |
c8d1ba58 | 7007 | io_do_iopoll(ctx, &nr_events, 0); |
906a3c6f | 7008 | |
d9d05217 PB |
7009 | if (to_submit && !ctx->sqo_dead && |
7010 | likely(!percpu_ref_is_dying(&ctx->refs))) | |
08369246 | 7011 | ret = io_submit_sqes(ctx, to_submit); |
c8d1ba58 JA |
7012 | mutex_unlock(&ctx->uring_lock); |
7013 | } | |
6c271ce2 | 7014 | |
90554200 JA |
7015 | if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait)) |
7016 | wake_up(&ctx->sqo_sq_wait); | |
6c271ce2 | 7017 | |
08369246 XW |
7018 | return ret; |
7019 | } | |
6c271ce2 | 7020 | |
08369246 XW |
7021 | static void io_sqd_update_thread_idle(struct io_sq_data *sqd) |
7022 | { | |
7023 | struct io_ring_ctx *ctx; | |
7024 | unsigned sq_thread_idle = 0; | |
6c271ce2 | 7025 | |
08369246 XW |
7026 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) { |
7027 | if (sq_thread_idle < ctx->sq_thread_idle) | |
7028 | sq_thread_idle = ctx->sq_thread_idle; | |
c8d1ba58 | 7029 | } |
c1edbf5f | 7030 | |
08369246 | 7031 | sqd->sq_thread_idle = sq_thread_idle; |
c8d1ba58 | 7032 | } |
6c271ce2 | 7033 | |
69fb2131 JA |
7034 | static void io_sqd_init_new(struct io_sq_data *sqd) |
7035 | { | |
7036 | struct io_ring_ctx *ctx; | |
7037 | ||
7038 | while (!list_empty(&sqd->ctx_new_list)) { | |
7039 | ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list); | |
69fb2131 JA |
7040 | list_move_tail(&ctx->sqd_list, &sqd->ctx_list); |
7041 | complete(&ctx->sq_thread_comp); | |
7042 | } | |
08369246 XW |
7043 | |
7044 | io_sqd_update_thread_idle(sqd); | |
69fb2131 JA |
7045 | } |
7046 | ||
c8d1ba58 JA |
7047 | static int io_sq_thread(void *data) |
7048 | { | |
91d8f519 | 7049 | struct cgroup_subsys_state *cur_css = NULL; |
28cea78a JA |
7050 | struct files_struct *old_files = current->files; |
7051 | struct nsproxy *old_nsproxy = current->nsproxy; | |
69fb2131 JA |
7052 | const struct cred *old_cred = NULL; |
7053 | struct io_sq_data *sqd = data; | |
7054 | struct io_ring_ctx *ctx; | |
a0d9205f | 7055 | unsigned long timeout = 0; |
08369246 | 7056 | DEFINE_WAIT(wait); |
6c271ce2 | 7057 | |
28cea78a JA |
7058 | task_lock(current); |
7059 | current->files = NULL; | |
7060 | current->nsproxy = NULL; | |
7061 | task_unlock(current); | |
6c271ce2 | 7062 | |
69fb2131 | 7063 | while (!kthread_should_stop()) { |
08369246 XW |
7064 | int ret; |
7065 | bool cap_entries, sqt_spin, needs_sched; | |
c1edbf5f JA |
7066 | |
7067 | /* | |
69fb2131 JA |
7068 | * Any changes to the sqd lists are synchronized through the |
7069 | * kthread parking. This synchronizes the thread vs users, | |
7070 | * the users are synchronized on the sqd->ctx_lock. | |
c1edbf5f | 7071 | */ |
65b2b213 | 7072 | if (kthread_should_park()) { |
69fb2131 | 7073 | kthread_parkme(); |
65b2b213 XW |
7074 | /* |
7075 | * When sq thread is unparked, in case the previous park operation | |
7076 | * comes from io_put_sq_data(), which means that sq thread is going | |
7077 | * to be stopped, so here needs to have a check. | |
7078 | */ | |
7079 | if (kthread_should_stop()) | |
7080 | break; | |
7081 | } | |
7143b5ac | 7082 | |
08369246 | 7083 | if (unlikely(!list_empty(&sqd->ctx_new_list))) { |
69fb2131 | 7084 | io_sqd_init_new(sqd); |
08369246 XW |
7085 | timeout = jiffies + sqd->sq_thread_idle; |
7086 | } | |
6c271ce2 | 7087 | |
08369246 | 7088 | sqt_spin = false; |
e95eee2d | 7089 | cap_entries = !list_is_singular(&sqd->ctx_list); |
69fb2131 JA |
7090 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) { |
7091 | if (current->cred != ctx->creds) { | |
7092 | if (old_cred) | |
7093 | revert_creds(old_cred); | |
7094 | old_cred = override_creds(ctx->creds); | |
bdcd3eab | 7095 | } |
91d8f519 | 7096 | io_sq_thread_associate_blkcg(ctx, &cur_css); |
4ea33a97 JA |
7097 | #ifdef CONFIG_AUDIT |
7098 | current->loginuid = ctx->loginuid; | |
7099 | current->sessionid = ctx->sessionid; | |
7100 | #endif | |
bdcd3eab | 7101 | |
08369246 XW |
7102 | ret = __io_sq_thread(ctx, cap_entries); |
7103 | if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list))) | |
7104 | sqt_spin = true; | |
6c271ce2 | 7105 | |
28cea78a | 7106 | io_sq_thread_drop_mm_files(); |
69fb2131 | 7107 | } |
6c271ce2 | 7108 | |
08369246 | 7109 | if (sqt_spin || !time_after(jiffies, timeout)) { |
c8d1ba58 | 7110 | io_run_task_work(); |
d434ab6d | 7111 | io_sq_thread_drop_mm_files(); |
c8d1ba58 | 7112 | cond_resched(); |
08369246 XW |
7113 | if (sqt_spin) |
7114 | timeout = jiffies + sqd->sq_thread_idle; | |
7115 | continue; | |
7116 | } | |
7117 | ||
7118 | if (kthread_should_park()) | |
7119 | continue; | |
7120 | ||
7121 | needs_sched = true; | |
7122 | prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE); | |
7123 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) { | |
7124 | if ((ctx->flags & IORING_SETUP_IOPOLL) && | |
7125 | !list_empty_careful(&ctx->iopoll_list)) { | |
7126 | needs_sched = false; | |
7127 | break; | |
7128 | } | |
7129 | if (io_sqring_entries(ctx)) { | |
7130 | needs_sched = false; | |
7131 | break; | |
7132 | } | |
7133 | } | |
7134 | ||
7135 | if (needs_sched) { | |
69fb2131 JA |
7136 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) |
7137 | io_ring_set_wakeup_flag(ctx); | |
08369246 | 7138 | |
69fb2131 | 7139 | schedule(); |
69fb2131 JA |
7140 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) |
7141 | io_ring_clear_wakeup_flag(ctx); | |
6c271ce2 | 7142 | } |
08369246 XW |
7143 | |
7144 | finish_wait(&sqd->wait, &wait); | |
7145 | timeout = jiffies + sqd->sq_thread_idle; | |
6c271ce2 JA |
7146 | } |
7147 | ||
4c6e277c | 7148 | io_run_task_work(); |
d434ab6d | 7149 | io_sq_thread_drop_mm_files(); |
b41e9852 | 7150 | |
91d8f519 DZ |
7151 | if (cur_css) |
7152 | io_sq_thread_unassociate_blkcg(); | |
69fb2131 JA |
7153 | if (old_cred) |
7154 | revert_creds(old_cred); | |
06058632 | 7155 | |
28cea78a JA |
7156 | task_lock(current); |
7157 | current->files = old_files; | |
7158 | current->nsproxy = old_nsproxy; | |
7159 | task_unlock(current); | |
7160 | ||
2bbcd6d3 | 7161 | kthread_parkme(); |
06058632 | 7162 | |
6c271ce2 JA |
7163 | return 0; |
7164 | } | |
7165 | ||
bda52162 JA |
7166 | struct io_wait_queue { |
7167 | struct wait_queue_entry wq; | |
7168 | struct io_ring_ctx *ctx; | |
7169 | unsigned to_wait; | |
7170 | unsigned nr_timeouts; | |
7171 | }; | |
7172 | ||
6c503150 | 7173 | static inline bool io_should_wake(struct io_wait_queue *iowq) |
bda52162 JA |
7174 | { |
7175 | struct io_ring_ctx *ctx = iowq->ctx; | |
7176 | ||
7177 | /* | |
d195a66e | 7178 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
7179 | * started waiting. For timeouts, we always want to return to userspace, |
7180 | * regardless of event count. | |
7181 | */ | |
6c503150 | 7182 | return io_cqring_events(ctx) >= iowq->to_wait || |
bda52162 JA |
7183 | atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
7184 | } | |
7185 | ||
7186 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
7187 | int wake_flags, void *key) | |
7188 | { | |
7189 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
7190 | wq); | |
7191 | ||
6c503150 PB |
7192 | /* |
7193 | * Cannot safely flush overflowed CQEs from here, ensure we wake up | |
7194 | * the task, and the next invocation will do it. | |
7195 | */ | |
7196 | if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow)) | |
7197 | return autoremove_wake_function(curr, mode, wake_flags, key); | |
7198 | return -1; | |
bda52162 JA |
7199 | } |
7200 | ||
af9c1a44 JA |
7201 | static int io_run_task_work_sig(void) |
7202 | { | |
7203 | if (io_run_task_work()) | |
7204 | return 1; | |
7205 | if (!signal_pending(current)) | |
7206 | return 0; | |
792ee0f6 JA |
7207 | if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL)) |
7208 | return -ERESTARTSYS; | |
af9c1a44 JA |
7209 | return -EINTR; |
7210 | } | |
7211 | ||
2b188cc1 JA |
7212 | /* |
7213 | * Wait until events become available, if we don't already have some. The | |
7214 | * application must reap them itself, as they reside on the shared cq ring. | |
7215 | */ | |
7216 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
c73ebb68 HX |
7217 | const sigset_t __user *sig, size_t sigsz, |
7218 | struct __kernel_timespec __user *uts) | |
2b188cc1 | 7219 | { |
bda52162 JA |
7220 | struct io_wait_queue iowq = { |
7221 | .wq = { | |
7222 | .private = current, | |
7223 | .func = io_wake_function, | |
7224 | .entry = LIST_HEAD_INIT(iowq.wq.entry), | |
7225 | }, | |
7226 | .ctx = ctx, | |
7227 | .to_wait = min_events, | |
7228 | }; | |
75b28aff | 7229 | struct io_rings *rings = ctx->rings; |
c73ebb68 HX |
7230 | struct timespec64 ts; |
7231 | signed long timeout = 0; | |
e9ffa5c2 | 7232 | int ret = 0; |
2b188cc1 | 7233 | |
b41e9852 | 7234 | do { |
6c503150 PB |
7235 | io_cqring_overflow_flush(ctx, false, NULL, NULL); |
7236 | if (io_cqring_events(ctx) >= min_events) | |
b41e9852 | 7237 | return 0; |
4c6e277c | 7238 | if (!io_run_task_work()) |
b41e9852 | 7239 | break; |
b41e9852 | 7240 | } while (1); |
2b188cc1 JA |
7241 | |
7242 | if (sig) { | |
9e75ad5d AB |
7243 | #ifdef CONFIG_COMPAT |
7244 | if (in_compat_syscall()) | |
7245 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 7246 | sigsz); |
9e75ad5d AB |
7247 | else |
7248 | #endif | |
b772434b | 7249 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 7250 | |
2b188cc1 JA |
7251 | if (ret) |
7252 | return ret; | |
7253 | } | |
7254 | ||
c73ebb68 HX |
7255 | if (uts) { |
7256 | if (get_timespec64(&ts, uts)) | |
7257 | return -EFAULT; | |
7258 | timeout = timespec64_to_jiffies(&ts); | |
7259 | } | |
7260 | ||
bda52162 | 7261 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
c826bd7a | 7262 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 | 7263 | do { |
6c503150 | 7264 | io_cqring_overflow_flush(ctx, false, NULL, NULL); |
bda52162 JA |
7265 | prepare_to_wait_exclusive(&ctx->wait, &iowq.wq, |
7266 | TASK_INTERRUPTIBLE); | |
ce593a6c | 7267 | /* make sure we run task_work before checking for signals */ |
af9c1a44 JA |
7268 | ret = io_run_task_work_sig(); |
7269 | if (ret > 0) | |
4c6e277c | 7270 | continue; |
af9c1a44 | 7271 | else if (ret < 0) |
bda52162 | 7272 | break; |
6c503150 | 7273 | if (io_should_wake(&iowq)) |
ce593a6c | 7274 | break; |
6c503150 PB |
7275 | if (test_bit(0, &ctx->cq_check_overflow)) |
7276 | continue; | |
c73ebb68 HX |
7277 | if (uts) { |
7278 | timeout = schedule_timeout(timeout); | |
7279 | if (timeout == 0) { | |
7280 | ret = -ETIME; | |
7281 | break; | |
7282 | } | |
7283 | } else { | |
7284 | schedule(); | |
7285 | } | |
bda52162 JA |
7286 | } while (1); |
7287 | finish_wait(&ctx->wait, &iowq.wq); | |
7288 | ||
b7db41c9 | 7289 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 7290 | |
75b28aff | 7291 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
7292 | } |
7293 | ||
6b06314c JA |
7294 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
7295 | { | |
7296 | #if defined(CONFIG_UNIX) | |
7297 | if (ctx->ring_sock) { | |
7298 | struct sock *sock = ctx->ring_sock->sk; | |
7299 | struct sk_buff *skb; | |
7300 | ||
7301 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
7302 | kfree_skb(skb); | |
7303 | } | |
7304 | #else | |
7305 | int i; | |
7306 | ||
65e19f54 JA |
7307 | for (i = 0; i < ctx->nr_user_files; i++) { |
7308 | struct file *file; | |
7309 | ||
7310 | file = io_file_from_index(ctx, i); | |
7311 | if (file) | |
7312 | fput(file); | |
7313 | } | |
6b06314c JA |
7314 | #endif |
7315 | } | |
7316 | ||
05f3fb3c JA |
7317 | static void io_file_ref_kill(struct percpu_ref *ref) |
7318 | { | |
7319 | struct fixed_file_data *data; | |
7320 | ||
7321 | data = container_of(ref, struct fixed_file_data, refs); | |
7322 | complete(&data->done); | |
7323 | } | |
7324 | ||
1642b445 PB |
7325 | static void io_sqe_files_set_node(struct fixed_file_data *file_data, |
7326 | struct fixed_file_ref_node *ref_node) | |
7327 | { | |
7328 | spin_lock_bh(&file_data->lock); | |
7329 | file_data->node = ref_node; | |
7330 | list_add_tail(&ref_node->node, &file_data->ref_list); | |
7331 | spin_unlock_bh(&file_data->lock); | |
7332 | percpu_ref_get(&file_data->refs); | |
7333 | } | |
7334 | ||
6b06314c JA |
7335 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) |
7336 | { | |
05f3fb3c | 7337 | struct fixed_file_data *data = ctx->file_data; |
1ffc5422 | 7338 | struct fixed_file_ref_node *backup_node, *ref_node = NULL; |
65e19f54 | 7339 | unsigned nr_tables, i; |
1ffc5422 | 7340 | int ret; |
65e19f54 | 7341 | |
05f3fb3c | 7342 | if (!data) |
6b06314c | 7343 | return -ENXIO; |
1ffc5422 PB |
7344 | backup_node = alloc_fixed_file_ref_node(ctx); |
7345 | if (!backup_node) | |
7346 | return -ENOMEM; | |
6b06314c | 7347 | |
ac0648a5 | 7348 | spin_lock_bh(&data->lock); |
1e5d770b | 7349 | ref_node = data->node; |
ac0648a5 | 7350 | spin_unlock_bh(&data->lock); |
05589553 XW |
7351 | if (ref_node) |
7352 | percpu_ref_kill(&ref_node->refs); | |
7353 | ||
7354 | percpu_ref_kill(&data->refs); | |
7355 | ||
7356 | /* wait for all refs nodes to complete */ | |
4a38aed2 | 7357 | flush_delayed_work(&ctx->file_put_work); |
1ffc5422 PB |
7358 | do { |
7359 | ret = wait_for_completion_interruptible(&data->done); | |
7360 | if (!ret) | |
7361 | break; | |
7362 | ret = io_run_task_work_sig(); | |
7363 | if (ret < 0) { | |
7364 | percpu_ref_resurrect(&data->refs); | |
7365 | reinit_completion(&data->done); | |
7366 | io_sqe_files_set_node(data, backup_node); | |
7367 | return ret; | |
7368 | } | |
7369 | } while (1); | |
05f3fb3c | 7370 | |
6b06314c | 7371 | __io_sqe_files_unregister(ctx); |
65e19f54 JA |
7372 | nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE); |
7373 | for (i = 0; i < nr_tables; i++) | |
05f3fb3c JA |
7374 | kfree(data->table[i].files); |
7375 | kfree(data->table); | |
05589553 XW |
7376 | percpu_ref_exit(&data->refs); |
7377 | kfree(data); | |
05f3fb3c | 7378 | ctx->file_data = NULL; |
6b06314c | 7379 | ctx->nr_user_files = 0; |
1ffc5422 | 7380 | destroy_fixed_file_ref_node(backup_node); |
6b06314c JA |
7381 | return 0; |
7382 | } | |
7383 | ||
534ca6d6 | 7384 | static void io_put_sq_data(struct io_sq_data *sqd) |
6c271ce2 | 7385 | { |
534ca6d6 | 7386 | if (refcount_dec_and_test(&sqd->refs)) { |
2bbcd6d3 RP |
7387 | /* |
7388 | * The park is a bit of a work-around, without it we get | |
7389 | * warning spews on shutdown with SQPOLL set and affinity | |
7390 | * set to a single CPU. | |
7391 | */ | |
534ca6d6 JA |
7392 | if (sqd->thread) { |
7393 | kthread_park(sqd->thread); | |
7394 | kthread_stop(sqd->thread); | |
7395 | } | |
7396 | ||
7397 | kfree(sqd); | |
7398 | } | |
7399 | } | |
7400 | ||
aa06165d JA |
7401 | static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p) |
7402 | { | |
7403 | struct io_ring_ctx *ctx_attach; | |
7404 | struct io_sq_data *sqd; | |
7405 | struct fd f; | |
7406 | ||
7407 | f = fdget(p->wq_fd); | |
7408 | if (!f.file) | |
7409 | return ERR_PTR(-ENXIO); | |
7410 | if (f.file->f_op != &io_uring_fops) { | |
7411 | fdput(f); | |
7412 | return ERR_PTR(-EINVAL); | |
7413 | } | |
7414 | ||
7415 | ctx_attach = f.file->private_data; | |
7416 | sqd = ctx_attach->sq_data; | |
7417 | if (!sqd) { | |
7418 | fdput(f); | |
7419 | return ERR_PTR(-EINVAL); | |
7420 | } | |
7421 | ||
7422 | refcount_inc(&sqd->refs); | |
7423 | fdput(f); | |
7424 | return sqd; | |
7425 | } | |
7426 | ||
534ca6d6 JA |
7427 | static struct io_sq_data *io_get_sq_data(struct io_uring_params *p) |
7428 | { | |
7429 | struct io_sq_data *sqd; | |
7430 | ||
aa06165d JA |
7431 | if (p->flags & IORING_SETUP_ATTACH_WQ) |
7432 | return io_attach_sq_data(p); | |
7433 | ||
534ca6d6 JA |
7434 | sqd = kzalloc(sizeof(*sqd), GFP_KERNEL); |
7435 | if (!sqd) | |
7436 | return ERR_PTR(-ENOMEM); | |
7437 | ||
7438 | refcount_set(&sqd->refs, 1); | |
69fb2131 JA |
7439 | INIT_LIST_HEAD(&sqd->ctx_list); |
7440 | INIT_LIST_HEAD(&sqd->ctx_new_list); | |
7441 | mutex_init(&sqd->ctx_lock); | |
7442 | mutex_init(&sqd->lock); | |
534ca6d6 JA |
7443 | init_waitqueue_head(&sqd->wait); |
7444 | return sqd; | |
7445 | } | |
7446 | ||
69fb2131 JA |
7447 | static void io_sq_thread_unpark(struct io_sq_data *sqd) |
7448 | __releases(&sqd->lock) | |
7449 | { | |
7450 | if (!sqd->thread) | |
7451 | return; | |
7452 | kthread_unpark(sqd->thread); | |
7453 | mutex_unlock(&sqd->lock); | |
7454 | } | |
7455 | ||
7456 | static void io_sq_thread_park(struct io_sq_data *sqd) | |
7457 | __acquires(&sqd->lock) | |
7458 | { | |
7459 | if (!sqd->thread) | |
7460 | return; | |
7461 | mutex_lock(&sqd->lock); | |
7462 | kthread_park(sqd->thread); | |
7463 | } | |
7464 | ||
534ca6d6 JA |
7465 | static void io_sq_thread_stop(struct io_ring_ctx *ctx) |
7466 | { | |
7467 | struct io_sq_data *sqd = ctx->sq_data; | |
7468 | ||
7469 | if (sqd) { | |
7470 | if (sqd->thread) { | |
7471 | /* | |
7472 | * We may arrive here from the error branch in | |
7473 | * io_sq_offload_create() where the kthread is created | |
7474 | * without being waked up, thus wake it up now to make | |
7475 | * sure the wait will complete. | |
7476 | */ | |
7477 | wake_up_process(sqd->thread); | |
7478 | wait_for_completion(&ctx->sq_thread_comp); | |
69fb2131 JA |
7479 | |
7480 | io_sq_thread_park(sqd); | |
7481 | } | |
7482 | ||
7483 | mutex_lock(&sqd->ctx_lock); | |
7484 | list_del(&ctx->sqd_list); | |
08369246 | 7485 | io_sqd_update_thread_idle(sqd); |
69fb2131 JA |
7486 | mutex_unlock(&sqd->ctx_lock); |
7487 | ||
08369246 | 7488 | if (sqd->thread) |
69fb2131 | 7489 | io_sq_thread_unpark(sqd); |
534ca6d6 JA |
7490 | |
7491 | io_put_sq_data(sqd); | |
7492 | ctx->sq_data = NULL; | |
6c271ce2 JA |
7493 | } |
7494 | } | |
7495 | ||
6b06314c JA |
7496 | static void io_finish_async(struct io_ring_ctx *ctx) |
7497 | { | |
6c271ce2 JA |
7498 | io_sq_thread_stop(ctx); |
7499 | ||
561fb04a JA |
7500 | if (ctx->io_wq) { |
7501 | io_wq_destroy(ctx->io_wq); | |
7502 | ctx->io_wq = NULL; | |
6b06314c JA |
7503 | } |
7504 | } | |
7505 | ||
7506 | #if defined(CONFIG_UNIX) | |
6b06314c JA |
7507 | /* |
7508 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
7509 | * the io_uring can be safely unregistered on process exit, even if we have | |
7510 | * loops in the file referencing. | |
7511 | */ | |
7512 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
7513 | { | |
7514 | struct sock *sk = ctx->ring_sock->sk; | |
7515 | struct scm_fp_list *fpl; | |
7516 | struct sk_buff *skb; | |
08a45173 | 7517 | int i, nr_files; |
6b06314c | 7518 | |
6b06314c JA |
7519 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); |
7520 | if (!fpl) | |
7521 | return -ENOMEM; | |
7522 | ||
7523 | skb = alloc_skb(0, GFP_KERNEL); | |
7524 | if (!skb) { | |
7525 | kfree(fpl); | |
7526 | return -ENOMEM; | |
7527 | } | |
7528 | ||
7529 | skb->sk = sk; | |
6b06314c | 7530 | |
08a45173 | 7531 | nr_files = 0; |
6b06314c JA |
7532 | fpl->user = get_uid(ctx->user); |
7533 | for (i = 0; i < nr; i++) { | |
65e19f54 JA |
7534 | struct file *file = io_file_from_index(ctx, i + offset); |
7535 | ||
7536 | if (!file) | |
08a45173 | 7537 | continue; |
65e19f54 | 7538 | fpl->fp[nr_files] = get_file(file); |
08a45173 JA |
7539 | unix_inflight(fpl->user, fpl->fp[nr_files]); |
7540 | nr_files++; | |
6b06314c JA |
7541 | } |
7542 | ||
08a45173 JA |
7543 | if (nr_files) { |
7544 | fpl->max = SCM_MAX_FD; | |
7545 | fpl->count = nr_files; | |
7546 | UNIXCB(skb).fp = fpl; | |
05f3fb3c | 7547 | skb->destructor = unix_destruct_scm; |
08a45173 JA |
7548 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); |
7549 | skb_queue_head(&sk->sk_receive_queue, skb); | |
6b06314c | 7550 | |
08a45173 JA |
7551 | for (i = 0; i < nr_files; i++) |
7552 | fput(fpl->fp[i]); | |
7553 | } else { | |
7554 | kfree_skb(skb); | |
7555 | kfree(fpl); | |
7556 | } | |
6b06314c JA |
7557 | |
7558 | return 0; | |
7559 | } | |
7560 | ||
7561 | /* | |
7562 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
7563 | * causes regular reference counting to break down. We rely on the UNIX | |
7564 | * garbage collection to take care of this problem for us. | |
7565 | */ | |
7566 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
7567 | { | |
7568 | unsigned left, total; | |
7569 | int ret = 0; | |
7570 | ||
7571 | total = 0; | |
7572 | left = ctx->nr_user_files; | |
7573 | while (left) { | |
7574 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
7575 | |
7576 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
7577 | if (ret) | |
7578 | break; | |
7579 | left -= this_files; | |
7580 | total += this_files; | |
7581 | } | |
7582 | ||
7583 | if (!ret) | |
7584 | return 0; | |
7585 | ||
7586 | while (total < ctx->nr_user_files) { | |
65e19f54 JA |
7587 | struct file *file = io_file_from_index(ctx, total); |
7588 | ||
7589 | if (file) | |
7590 | fput(file); | |
6b06314c JA |
7591 | total++; |
7592 | } | |
7593 | ||
7594 | return ret; | |
7595 | } | |
7596 | #else | |
7597 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
7598 | { | |
7599 | return 0; | |
7600 | } | |
7601 | #endif | |
7602 | ||
5398ae69 PB |
7603 | static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data, |
7604 | unsigned nr_tables, unsigned nr_files) | |
65e19f54 JA |
7605 | { |
7606 | int i; | |
7607 | ||
7608 | for (i = 0; i < nr_tables; i++) { | |
5398ae69 | 7609 | struct fixed_file_table *table = &file_data->table[i]; |
65e19f54 JA |
7610 | unsigned this_files; |
7611 | ||
7612 | this_files = min(nr_files, IORING_MAX_FILES_TABLE); | |
7613 | table->files = kcalloc(this_files, sizeof(struct file *), | |
7614 | GFP_KERNEL); | |
7615 | if (!table->files) | |
7616 | break; | |
7617 | nr_files -= this_files; | |
7618 | } | |
7619 | ||
7620 | if (i == nr_tables) | |
7621 | return 0; | |
7622 | ||
7623 | for (i = 0; i < nr_tables; i++) { | |
5398ae69 | 7624 | struct fixed_file_table *table = &file_data->table[i]; |
65e19f54 JA |
7625 | kfree(table->files); |
7626 | } | |
7627 | return 1; | |
7628 | } | |
7629 | ||
05f3fb3c JA |
7630 | static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file) |
7631 | { | |
7632 | #if defined(CONFIG_UNIX) | |
7633 | struct sock *sock = ctx->ring_sock->sk; | |
7634 | struct sk_buff_head list, *head = &sock->sk_receive_queue; | |
7635 | struct sk_buff *skb; | |
7636 | int i; | |
7637 | ||
7638 | __skb_queue_head_init(&list); | |
7639 | ||
7640 | /* | |
7641 | * Find the skb that holds this file in its SCM_RIGHTS. When found, | |
7642 | * remove this entry and rearrange the file array. | |
7643 | */ | |
7644 | skb = skb_dequeue(head); | |
7645 | while (skb) { | |
7646 | struct scm_fp_list *fp; | |
7647 | ||
7648 | fp = UNIXCB(skb).fp; | |
7649 | for (i = 0; i < fp->count; i++) { | |
7650 | int left; | |
7651 | ||
7652 | if (fp->fp[i] != file) | |
7653 | continue; | |
7654 | ||
7655 | unix_notinflight(fp->user, fp->fp[i]); | |
7656 | left = fp->count - 1 - i; | |
7657 | if (left) { | |
7658 | memmove(&fp->fp[i], &fp->fp[i + 1], | |
7659 | left * sizeof(struct file *)); | |
7660 | } | |
7661 | fp->count--; | |
7662 | if (!fp->count) { | |
7663 | kfree_skb(skb); | |
7664 | skb = NULL; | |
7665 | } else { | |
7666 | __skb_queue_tail(&list, skb); | |
7667 | } | |
7668 | fput(file); | |
7669 | file = NULL; | |
7670 | break; | |
7671 | } | |
7672 | ||
7673 | if (!file) | |
7674 | break; | |
7675 | ||
7676 | __skb_queue_tail(&list, skb); | |
7677 | ||
7678 | skb = skb_dequeue(head); | |
7679 | } | |
7680 | ||
7681 | if (skb_peek(&list)) { | |
7682 | spin_lock_irq(&head->lock); | |
7683 | while ((skb = __skb_dequeue(&list)) != NULL) | |
7684 | __skb_queue_tail(head, skb); | |
7685 | spin_unlock_irq(&head->lock); | |
7686 | } | |
7687 | #else | |
7688 | fput(file); | |
7689 | #endif | |
7690 | } | |
7691 | ||
7692 | struct io_file_put { | |
05589553 | 7693 | struct list_head list; |
05f3fb3c | 7694 | struct file *file; |
05f3fb3c JA |
7695 | }; |
7696 | ||
4a38aed2 | 7697 | static void __io_file_put_work(struct fixed_file_ref_node *ref_node) |
65e19f54 | 7698 | { |
4a38aed2 JA |
7699 | struct fixed_file_data *file_data = ref_node->file_data; |
7700 | struct io_ring_ctx *ctx = file_data->ctx; | |
05f3fb3c | 7701 | struct io_file_put *pfile, *tmp; |
05589553 XW |
7702 | |
7703 | list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) { | |
6a4d07cd | 7704 | list_del(&pfile->list); |
05589553 XW |
7705 | io_ring_file_put(ctx, pfile->file); |
7706 | kfree(pfile); | |
65e19f54 | 7707 | } |
05589553 | 7708 | |
05589553 XW |
7709 | percpu_ref_exit(&ref_node->refs); |
7710 | kfree(ref_node); | |
7711 | percpu_ref_put(&file_data->refs); | |
2faf852d | 7712 | } |
65e19f54 | 7713 | |
4a38aed2 JA |
7714 | static void io_file_put_work(struct work_struct *work) |
7715 | { | |
7716 | struct io_ring_ctx *ctx; | |
7717 | struct llist_node *node; | |
7718 | ||
7719 | ctx = container_of(work, struct io_ring_ctx, file_put_work.work); | |
7720 | node = llist_del_all(&ctx->file_put_llist); | |
7721 | ||
7722 | while (node) { | |
7723 | struct fixed_file_ref_node *ref_node; | |
7724 | struct llist_node *next = node->next; | |
7725 | ||
7726 | ref_node = llist_entry(node, struct fixed_file_ref_node, llist); | |
7727 | __io_file_put_work(ref_node); | |
7728 | node = next; | |
7729 | } | |
7730 | } | |
7731 | ||
05589553 | 7732 | static void io_file_data_ref_zero(struct percpu_ref *ref) |
2faf852d | 7733 | { |
05589553 | 7734 | struct fixed_file_ref_node *ref_node; |
e297822b | 7735 | struct fixed_file_data *data; |
4a38aed2 | 7736 | struct io_ring_ctx *ctx; |
e297822b | 7737 | bool first_add = false; |
4a38aed2 | 7738 | int delay = HZ; |
65e19f54 | 7739 | |
05589553 | 7740 | ref_node = container_of(ref, struct fixed_file_ref_node, refs); |
e297822b PB |
7741 | data = ref_node->file_data; |
7742 | ctx = data->ctx; | |
7743 | ||
ac0648a5 | 7744 | spin_lock_bh(&data->lock); |
e297822b PB |
7745 | ref_node->done = true; |
7746 | ||
7747 | while (!list_empty(&data->ref_list)) { | |
7748 | ref_node = list_first_entry(&data->ref_list, | |
7749 | struct fixed_file_ref_node, node); | |
7750 | /* recycle ref nodes in order */ | |
7751 | if (!ref_node->done) | |
7752 | break; | |
7753 | list_del(&ref_node->node); | |
7754 | first_add |= llist_add(&ref_node->llist, &ctx->file_put_llist); | |
7755 | } | |
ac0648a5 | 7756 | spin_unlock_bh(&data->lock); |
05589553 | 7757 | |
e297822b | 7758 | if (percpu_ref_is_dying(&data->refs)) |
4a38aed2 | 7759 | delay = 0; |
05589553 | 7760 | |
4a38aed2 JA |
7761 | if (!delay) |
7762 | mod_delayed_work(system_wq, &ctx->file_put_work, 0); | |
7763 | else if (first_add) | |
7764 | queue_delayed_work(system_wq, &ctx->file_put_work, delay); | |
05f3fb3c | 7765 | } |
65e19f54 | 7766 | |
05589553 XW |
7767 | static struct fixed_file_ref_node *alloc_fixed_file_ref_node( |
7768 | struct io_ring_ctx *ctx) | |
05f3fb3c | 7769 | { |
05589553 | 7770 | struct fixed_file_ref_node *ref_node; |
05f3fb3c | 7771 | |
05589553 XW |
7772 | ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL); |
7773 | if (!ref_node) | |
3e2224c5 | 7774 | return NULL; |
05f3fb3c | 7775 | |
05589553 XW |
7776 | if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero, |
7777 | 0, GFP_KERNEL)) { | |
7778 | kfree(ref_node); | |
3e2224c5 | 7779 | return NULL; |
05589553 XW |
7780 | } |
7781 | INIT_LIST_HEAD(&ref_node->node); | |
7782 | INIT_LIST_HEAD(&ref_node->file_list); | |
05589553 | 7783 | ref_node->file_data = ctx->file_data; |
e297822b | 7784 | ref_node->done = false; |
05589553 | 7785 | return ref_node; |
05589553 XW |
7786 | } |
7787 | ||
7788 | static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node) | |
7789 | { | |
7790 | percpu_ref_exit(&ref_node->refs); | |
7791 | kfree(ref_node); | |
65e19f54 JA |
7792 | } |
7793 | ||
6b06314c JA |
7794 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, |
7795 | unsigned nr_args) | |
7796 | { | |
7797 | __s32 __user *fds = (__s32 __user *) arg; | |
600cf3f8 | 7798 | unsigned nr_tables, i; |
05f3fb3c | 7799 | struct file *file; |
600cf3f8 | 7800 | int fd, ret = -ENOMEM; |
05589553 | 7801 | struct fixed_file_ref_node *ref_node; |
5398ae69 | 7802 | struct fixed_file_data *file_data; |
6b06314c | 7803 | |
05f3fb3c | 7804 | if (ctx->file_data) |
6b06314c JA |
7805 | return -EBUSY; |
7806 | if (!nr_args) | |
7807 | return -EINVAL; | |
7808 | if (nr_args > IORING_MAX_FIXED_FILES) | |
7809 | return -EMFILE; | |
7810 | ||
5398ae69 PB |
7811 | file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL); |
7812 | if (!file_data) | |
05f3fb3c | 7813 | return -ENOMEM; |
5398ae69 PB |
7814 | file_data->ctx = ctx; |
7815 | init_completion(&file_data->done); | |
7816 | INIT_LIST_HEAD(&file_data->ref_list); | |
7817 | spin_lock_init(&file_data->lock); | |
05f3fb3c | 7818 | |
65e19f54 | 7819 | nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE); |
035fbafc | 7820 | file_data->table = kcalloc(nr_tables, sizeof(*file_data->table), |
5398ae69 | 7821 | GFP_KERNEL); |
600cf3f8 PB |
7822 | if (!file_data->table) |
7823 | goto out_free; | |
05f3fb3c | 7824 | |
5398ae69 | 7825 | if (percpu_ref_init(&file_data->refs, io_file_ref_kill, |
600cf3f8 PB |
7826 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) |
7827 | goto out_free; | |
6b06314c | 7828 | |
600cf3f8 PB |
7829 | if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args)) |
7830 | goto out_ref; | |
55cbc256 | 7831 | ctx->file_data = file_data; |
65e19f54 | 7832 | |
08a45173 | 7833 | for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { |
65e19f54 JA |
7834 | struct fixed_file_table *table; |
7835 | unsigned index; | |
7836 | ||
600cf3f8 PB |
7837 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) { |
7838 | ret = -EFAULT; | |
7839 | goto out_fput; | |
7840 | } | |
08a45173 | 7841 | /* allow sparse sets */ |
600cf3f8 | 7842 | if (fd == -1) |
08a45173 | 7843 | continue; |
6b06314c | 7844 | |
05f3fb3c | 7845 | file = fget(fd); |
6b06314c | 7846 | ret = -EBADF; |
05f3fb3c | 7847 | if (!file) |
600cf3f8 | 7848 | goto out_fput; |
05f3fb3c | 7849 | |
6b06314c JA |
7850 | /* |
7851 | * Don't allow io_uring instances to be registered. If UNIX | |
7852 | * isn't enabled, then this causes a reference cycle and this | |
7853 | * instance can never get freed. If UNIX is enabled we'll | |
7854 | * handle it just fine, but there's still no point in allowing | |
7855 | * a ring fd as it doesn't support regular read/write anyway. | |
7856 | */ | |
05f3fb3c JA |
7857 | if (file->f_op == &io_uring_fops) { |
7858 | fput(file); | |
600cf3f8 | 7859 | goto out_fput; |
6b06314c | 7860 | } |
600cf3f8 PB |
7861 | table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT]; |
7862 | index = i & IORING_FILE_TABLE_MASK; | |
05f3fb3c | 7863 | table->files[index] = file; |
6b06314c JA |
7864 | } |
7865 | ||
6b06314c | 7866 | ret = io_sqe_files_scm(ctx); |
05589553 | 7867 | if (ret) { |
6b06314c | 7868 | io_sqe_files_unregister(ctx); |
05589553 XW |
7869 | return ret; |
7870 | } | |
6b06314c | 7871 | |
05589553 | 7872 | ref_node = alloc_fixed_file_ref_node(ctx); |
3e2224c5 | 7873 | if (!ref_node) { |
05589553 | 7874 | io_sqe_files_unregister(ctx); |
3e2224c5 | 7875 | return -ENOMEM; |
05589553 XW |
7876 | } |
7877 | ||
1642b445 | 7878 | io_sqe_files_set_node(file_data, ref_node); |
6b06314c | 7879 | return ret; |
600cf3f8 PB |
7880 | out_fput: |
7881 | for (i = 0; i < ctx->nr_user_files; i++) { | |
7882 | file = io_file_from_index(ctx, i); | |
7883 | if (file) | |
7884 | fput(file); | |
7885 | } | |
7886 | for (i = 0; i < nr_tables; i++) | |
7887 | kfree(file_data->table[i].files); | |
7888 | ctx->nr_user_files = 0; | |
7889 | out_ref: | |
7890 | percpu_ref_exit(&file_data->refs); | |
7891 | out_free: | |
7892 | kfree(file_data->table); | |
7893 | kfree(file_data); | |
55cbc256 | 7894 | ctx->file_data = NULL; |
6b06314c JA |
7895 | return ret; |
7896 | } | |
7897 | ||
c3a31e60 JA |
7898 | static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file, |
7899 | int index) | |
7900 | { | |
7901 | #if defined(CONFIG_UNIX) | |
7902 | struct sock *sock = ctx->ring_sock->sk; | |
7903 | struct sk_buff_head *head = &sock->sk_receive_queue; | |
7904 | struct sk_buff *skb; | |
7905 | ||
7906 | /* | |
7907 | * See if we can merge this file into an existing skb SCM_RIGHTS | |
7908 | * file set. If there's no room, fall back to allocating a new skb | |
7909 | * and filling it in. | |
7910 | */ | |
7911 | spin_lock_irq(&head->lock); | |
7912 | skb = skb_peek(head); | |
7913 | if (skb) { | |
7914 | struct scm_fp_list *fpl = UNIXCB(skb).fp; | |
7915 | ||
7916 | if (fpl->count < SCM_MAX_FD) { | |
7917 | __skb_unlink(skb, head); | |
7918 | spin_unlock_irq(&head->lock); | |
7919 | fpl->fp[fpl->count] = get_file(file); | |
7920 | unix_inflight(fpl->user, fpl->fp[fpl->count]); | |
7921 | fpl->count++; | |
7922 | spin_lock_irq(&head->lock); | |
7923 | __skb_queue_head(head, skb); | |
7924 | } else { | |
7925 | skb = NULL; | |
7926 | } | |
7927 | } | |
7928 | spin_unlock_irq(&head->lock); | |
7929 | ||
7930 | if (skb) { | |
7931 | fput(file); | |
7932 | return 0; | |
7933 | } | |
7934 | ||
7935 | return __io_sqe_files_scm(ctx, 1, index); | |
7936 | #else | |
7937 | return 0; | |
7938 | #endif | |
7939 | } | |
7940 | ||
a5318d3c | 7941 | static int io_queue_file_removal(struct fixed_file_data *data, |
05589553 | 7942 | struct file *file) |
05f3fb3c | 7943 | { |
a5318d3c | 7944 | struct io_file_put *pfile; |
b2e96852 | 7945 | struct fixed_file_ref_node *ref_node = data->node; |
05f3fb3c | 7946 | |
05f3fb3c | 7947 | pfile = kzalloc(sizeof(*pfile), GFP_KERNEL); |
a5318d3c HD |
7948 | if (!pfile) |
7949 | return -ENOMEM; | |
05f3fb3c JA |
7950 | |
7951 | pfile->file = file; | |
05589553 XW |
7952 | list_add(&pfile->list, &ref_node->file_list); |
7953 | ||
a5318d3c | 7954 | return 0; |
05f3fb3c JA |
7955 | } |
7956 | ||
7957 | static int __io_sqe_files_update(struct io_ring_ctx *ctx, | |
7958 | struct io_uring_files_update *up, | |
7959 | unsigned nr_args) | |
7960 | { | |
7961 | struct fixed_file_data *data = ctx->file_data; | |
05589553 | 7962 | struct fixed_file_ref_node *ref_node; |
05f3fb3c | 7963 | struct file *file; |
c3a31e60 JA |
7964 | __s32 __user *fds; |
7965 | int fd, i, err; | |
7966 | __u32 done; | |
05589553 | 7967 | bool needs_switch = false; |
c3a31e60 | 7968 | |
05f3fb3c | 7969 | if (check_add_overflow(up->offset, nr_args, &done)) |
c3a31e60 JA |
7970 | return -EOVERFLOW; |
7971 | if (done > ctx->nr_user_files) | |
7972 | return -EINVAL; | |
7973 | ||
05589553 | 7974 | ref_node = alloc_fixed_file_ref_node(ctx); |
3e2224c5 MWO |
7975 | if (!ref_node) |
7976 | return -ENOMEM; | |
05589553 | 7977 | |
c3a31e60 | 7978 | done = 0; |
05f3fb3c | 7979 | fds = u64_to_user_ptr(up->fds); |
c3a31e60 | 7980 | while (nr_args) { |
65e19f54 JA |
7981 | struct fixed_file_table *table; |
7982 | unsigned index; | |
7983 | ||
c3a31e60 JA |
7984 | err = 0; |
7985 | if (copy_from_user(&fd, &fds[done], sizeof(fd))) { | |
7986 | err = -EFAULT; | |
7987 | break; | |
7988 | } | |
05f3fb3c JA |
7989 | i = array_index_nospec(up->offset, ctx->nr_user_files); |
7990 | table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT]; | |
65e19f54 JA |
7991 | index = i & IORING_FILE_TABLE_MASK; |
7992 | if (table->files[index]) { | |
98dfd502 | 7993 | file = table->files[index]; |
a5318d3c HD |
7994 | err = io_queue_file_removal(data, file); |
7995 | if (err) | |
7996 | break; | |
65e19f54 | 7997 | table->files[index] = NULL; |
05589553 | 7998 | needs_switch = true; |
c3a31e60 JA |
7999 | } |
8000 | if (fd != -1) { | |
c3a31e60 JA |
8001 | file = fget(fd); |
8002 | if (!file) { | |
8003 | err = -EBADF; | |
8004 | break; | |
8005 | } | |
8006 | /* | |
8007 | * Don't allow io_uring instances to be registered. If | |
8008 | * UNIX isn't enabled, then this causes a reference | |
8009 | * cycle and this instance can never get freed. If UNIX | |
8010 | * is enabled we'll handle it just fine, but there's | |
8011 | * still no point in allowing a ring fd as it doesn't | |
8012 | * support regular read/write anyway. | |
8013 | */ | |
8014 | if (file->f_op == &io_uring_fops) { | |
8015 | fput(file); | |
8016 | err = -EBADF; | |
8017 | break; | |
8018 | } | |
65e19f54 | 8019 | table->files[index] = file; |
c3a31e60 | 8020 | err = io_sqe_file_register(ctx, file, i); |
f3bd9dae | 8021 | if (err) { |
95d1c8e5 | 8022 | table->files[index] = NULL; |
f3bd9dae | 8023 | fput(file); |
c3a31e60 | 8024 | break; |
f3bd9dae | 8025 | } |
c3a31e60 JA |
8026 | } |
8027 | nr_args--; | |
8028 | done++; | |
05f3fb3c JA |
8029 | up->offset++; |
8030 | } | |
8031 | ||
05589553 | 8032 | if (needs_switch) { |
b2e96852 | 8033 | percpu_ref_kill(&data->node->refs); |
1642b445 | 8034 | io_sqe_files_set_node(data, ref_node); |
05589553 XW |
8035 | } else |
8036 | destroy_fixed_file_ref_node(ref_node); | |
c3a31e60 JA |
8037 | |
8038 | return done ? done : err; | |
8039 | } | |
05589553 | 8040 | |
05f3fb3c JA |
8041 | static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg, |
8042 | unsigned nr_args) | |
8043 | { | |
8044 | struct io_uring_files_update up; | |
8045 | ||
8046 | if (!ctx->file_data) | |
8047 | return -ENXIO; | |
8048 | if (!nr_args) | |
8049 | return -EINVAL; | |
8050 | if (copy_from_user(&up, arg, sizeof(up))) | |
8051 | return -EFAULT; | |
8052 | if (up.resv) | |
8053 | return -EINVAL; | |
8054 | ||
8055 | return __io_sqe_files_update(ctx, &up, nr_args); | |
8056 | } | |
c3a31e60 | 8057 | |
e9fd9396 | 8058 | static void io_free_work(struct io_wq_work *work) |
7d723065 JA |
8059 | { |
8060 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
8061 | ||
e9fd9396 | 8062 | /* Consider that io_steal_work() relies on this ref */ |
7d723065 JA |
8063 | io_put_req(req); |
8064 | } | |
8065 | ||
24369c2e PB |
8066 | static int io_init_wq_offload(struct io_ring_ctx *ctx, |
8067 | struct io_uring_params *p) | |
8068 | { | |
8069 | struct io_wq_data data; | |
8070 | struct fd f; | |
8071 | struct io_ring_ctx *ctx_attach; | |
8072 | unsigned int concurrency; | |
8073 | int ret = 0; | |
8074 | ||
8075 | data.user = ctx->user; | |
e9fd9396 | 8076 | data.free_work = io_free_work; |
f5fa38c5 | 8077 | data.do_work = io_wq_submit_work; |
24369c2e PB |
8078 | |
8079 | if (!(p->flags & IORING_SETUP_ATTACH_WQ)) { | |
8080 | /* Do QD, or 4 * CPUS, whatever is smallest */ | |
8081 | concurrency = min(ctx->sq_entries, 4 * num_online_cpus()); | |
8082 | ||
8083 | ctx->io_wq = io_wq_create(concurrency, &data); | |
8084 | if (IS_ERR(ctx->io_wq)) { | |
8085 | ret = PTR_ERR(ctx->io_wq); | |
8086 | ctx->io_wq = NULL; | |
8087 | } | |
8088 | return ret; | |
8089 | } | |
8090 | ||
8091 | f = fdget(p->wq_fd); | |
8092 | if (!f.file) | |
8093 | return -EBADF; | |
8094 | ||
8095 | if (f.file->f_op != &io_uring_fops) { | |
8096 | ret = -EINVAL; | |
8097 | goto out_fput; | |
8098 | } | |
8099 | ||
8100 | ctx_attach = f.file->private_data; | |
8101 | /* @io_wq is protected by holding the fd */ | |
8102 | if (!io_wq_get(ctx_attach->io_wq, &data)) { | |
8103 | ret = -EINVAL; | |
8104 | goto out_fput; | |
8105 | } | |
8106 | ||
8107 | ctx->io_wq = ctx_attach->io_wq; | |
8108 | out_fput: | |
8109 | fdput(f); | |
8110 | return ret; | |
8111 | } | |
8112 | ||
0f212204 JA |
8113 | static int io_uring_alloc_task_context(struct task_struct *task) |
8114 | { | |
8115 | struct io_uring_task *tctx; | |
d8a6df10 | 8116 | int ret; |
0f212204 JA |
8117 | |
8118 | tctx = kmalloc(sizeof(*tctx), GFP_KERNEL); | |
8119 | if (unlikely(!tctx)) | |
8120 | return -ENOMEM; | |
8121 | ||
d8a6df10 JA |
8122 | ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL); |
8123 | if (unlikely(ret)) { | |
8124 | kfree(tctx); | |
8125 | return ret; | |
8126 | } | |
8127 | ||
0f212204 JA |
8128 | xa_init(&tctx->xa); |
8129 | init_waitqueue_head(&tctx->wait); | |
8130 | tctx->last = NULL; | |
fdaf083c JA |
8131 | atomic_set(&tctx->in_idle, 0); |
8132 | tctx->sqpoll = false; | |
500a373d JA |
8133 | io_init_identity(&tctx->__identity); |
8134 | tctx->identity = &tctx->__identity; | |
0f212204 JA |
8135 | task->io_uring = tctx; |
8136 | return 0; | |
8137 | } | |
8138 | ||
8139 | void __io_uring_free(struct task_struct *tsk) | |
8140 | { | |
8141 | struct io_uring_task *tctx = tsk->io_uring; | |
8142 | ||
8143 | WARN_ON_ONCE(!xa_empty(&tctx->xa)); | |
500a373d JA |
8144 | WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1); |
8145 | if (tctx->identity != &tctx->__identity) | |
8146 | kfree(tctx->identity); | |
d8a6df10 | 8147 | percpu_counter_destroy(&tctx->inflight); |
0f212204 JA |
8148 | kfree(tctx); |
8149 | tsk->io_uring = NULL; | |
8150 | } | |
8151 | ||
7e84e1c7 SG |
8152 | static int io_sq_offload_create(struct io_ring_ctx *ctx, |
8153 | struct io_uring_params *p) | |
2b188cc1 JA |
8154 | { |
8155 | int ret; | |
8156 | ||
6c271ce2 | 8157 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
534ca6d6 JA |
8158 | struct io_sq_data *sqd; |
8159 | ||
3ec482d1 | 8160 | ret = -EPERM; |
ce59fc69 | 8161 | if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE)) |
3ec482d1 JA |
8162 | goto err; |
8163 | ||
534ca6d6 JA |
8164 | sqd = io_get_sq_data(p); |
8165 | if (IS_ERR(sqd)) { | |
8166 | ret = PTR_ERR(sqd); | |
8167 | goto err; | |
8168 | } | |
69fb2131 | 8169 | |
534ca6d6 | 8170 | ctx->sq_data = sqd; |
69fb2131 JA |
8171 | io_sq_thread_park(sqd); |
8172 | mutex_lock(&sqd->ctx_lock); | |
8173 | list_add(&ctx->sqd_list, &sqd->ctx_new_list); | |
8174 | mutex_unlock(&sqd->ctx_lock); | |
8175 | io_sq_thread_unpark(sqd); | |
534ca6d6 | 8176 | |
917257da JA |
8177 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
8178 | if (!ctx->sq_thread_idle) | |
8179 | ctx->sq_thread_idle = HZ; | |
8180 | ||
aa06165d JA |
8181 | if (sqd->thread) |
8182 | goto done; | |
8183 | ||
6c271ce2 | 8184 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 8185 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 8186 | |
917257da | 8187 | ret = -EINVAL; |
44a9bd18 JA |
8188 | if (cpu >= nr_cpu_ids) |
8189 | goto err; | |
7889f44d | 8190 | if (!cpu_online(cpu)) |
917257da JA |
8191 | goto err; |
8192 | ||
69fb2131 | 8193 | sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd, |
534ca6d6 | 8194 | cpu, "io_uring-sq"); |
6c271ce2 | 8195 | } else { |
69fb2131 | 8196 | sqd->thread = kthread_create(io_sq_thread, sqd, |
6c271ce2 JA |
8197 | "io_uring-sq"); |
8198 | } | |
534ca6d6 JA |
8199 | if (IS_ERR(sqd->thread)) { |
8200 | ret = PTR_ERR(sqd->thread); | |
8201 | sqd->thread = NULL; | |
6c271ce2 JA |
8202 | goto err; |
8203 | } | |
534ca6d6 | 8204 | ret = io_uring_alloc_task_context(sqd->thread); |
0f212204 JA |
8205 | if (ret) |
8206 | goto err; | |
6c271ce2 JA |
8207 | } else if (p->flags & IORING_SETUP_SQ_AFF) { |
8208 | /* Can't have SQ_AFF without SQPOLL */ | |
8209 | ret = -EINVAL; | |
8210 | goto err; | |
8211 | } | |
8212 | ||
aa06165d | 8213 | done: |
24369c2e PB |
8214 | ret = io_init_wq_offload(ctx, p); |
8215 | if (ret) | |
2b188cc1 | 8216 | goto err; |
2b188cc1 JA |
8217 | |
8218 | return 0; | |
8219 | err: | |
54a91f3b | 8220 | io_finish_async(ctx); |
2b188cc1 JA |
8221 | return ret; |
8222 | } | |
8223 | ||
7e84e1c7 SG |
8224 | static void io_sq_offload_start(struct io_ring_ctx *ctx) |
8225 | { | |
534ca6d6 JA |
8226 | struct io_sq_data *sqd = ctx->sq_data; |
8227 | ||
8228 | if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread) | |
8229 | wake_up_process(sqd->thread); | |
7e84e1c7 SG |
8230 | } |
8231 | ||
a087e2b5 BM |
8232 | static inline void __io_unaccount_mem(struct user_struct *user, |
8233 | unsigned long nr_pages) | |
2b188cc1 JA |
8234 | { |
8235 | atomic_long_sub(nr_pages, &user->locked_vm); | |
8236 | } | |
8237 | ||
a087e2b5 BM |
8238 | static inline int __io_account_mem(struct user_struct *user, |
8239 | unsigned long nr_pages) | |
2b188cc1 JA |
8240 | { |
8241 | unsigned long page_limit, cur_pages, new_pages; | |
8242 | ||
8243 | /* Don't allow more pages than we can safely lock */ | |
8244 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
8245 | ||
8246 | do { | |
8247 | cur_pages = atomic_long_read(&user->locked_vm); | |
8248 | new_pages = cur_pages + nr_pages; | |
8249 | if (new_pages > page_limit) | |
8250 | return -ENOMEM; | |
8251 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
8252 | new_pages) != cur_pages); | |
8253 | ||
8254 | return 0; | |
8255 | } | |
8256 | ||
2e0464d4 BM |
8257 | static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages, |
8258 | enum io_mem_account acct) | |
a087e2b5 | 8259 | { |
aad5d8da | 8260 | if (ctx->limit_mem) |
a087e2b5 | 8261 | __io_unaccount_mem(ctx->user, nr_pages); |
30975825 | 8262 | |
2aede0e4 | 8263 | if (ctx->mm_account) { |
4bc4a912 JA |
8264 | if (acct == ACCT_LOCKED) { |
8265 | mmap_write_lock(ctx->mm_account); | |
2aede0e4 | 8266 | ctx->mm_account->locked_vm -= nr_pages; |
4bc4a912 JA |
8267 | mmap_write_unlock(ctx->mm_account); |
8268 | }else if (acct == ACCT_PINNED) { | |
2aede0e4 | 8269 | atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm); |
4bc4a912 | 8270 | } |
2e0464d4 | 8271 | } |
a087e2b5 BM |
8272 | } |
8273 | ||
2e0464d4 BM |
8274 | static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages, |
8275 | enum io_mem_account acct) | |
a087e2b5 | 8276 | { |
30975825 BM |
8277 | int ret; |
8278 | ||
8279 | if (ctx->limit_mem) { | |
8280 | ret = __io_account_mem(ctx->user, nr_pages); | |
8281 | if (ret) | |
8282 | return ret; | |
8283 | } | |
8284 | ||
2aede0e4 | 8285 | if (ctx->mm_account) { |
4bc4a912 JA |
8286 | if (acct == ACCT_LOCKED) { |
8287 | mmap_write_lock(ctx->mm_account); | |
2aede0e4 | 8288 | ctx->mm_account->locked_vm += nr_pages; |
4bc4a912 JA |
8289 | mmap_write_unlock(ctx->mm_account); |
8290 | } else if (acct == ACCT_PINNED) { | |
2aede0e4 | 8291 | atomic64_add(nr_pages, &ctx->mm_account->pinned_vm); |
4bc4a912 | 8292 | } |
2e0464d4 | 8293 | } |
a087e2b5 BM |
8294 | |
8295 | return 0; | |
8296 | } | |
8297 | ||
2b188cc1 JA |
8298 | static void io_mem_free(void *ptr) |
8299 | { | |
52e04ef4 MR |
8300 | struct page *page; |
8301 | ||
8302 | if (!ptr) | |
8303 | return; | |
2b188cc1 | 8304 | |
52e04ef4 | 8305 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
8306 | if (put_page_testzero(page)) |
8307 | free_compound_page(page); | |
8308 | } | |
8309 | ||
8310 | static void *io_mem_alloc(size_t size) | |
8311 | { | |
8312 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
8313 | __GFP_NORETRY; | |
8314 | ||
8315 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
8316 | } | |
8317 | ||
75b28aff HV |
8318 | static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries, |
8319 | size_t *sq_offset) | |
8320 | { | |
8321 | struct io_rings *rings; | |
8322 | size_t off, sq_array_size; | |
8323 | ||
8324 | off = struct_size(rings, cqes, cq_entries); | |
8325 | if (off == SIZE_MAX) | |
8326 | return SIZE_MAX; | |
8327 | ||
8328 | #ifdef CONFIG_SMP | |
8329 | off = ALIGN(off, SMP_CACHE_BYTES); | |
8330 | if (off == 0) | |
8331 | return SIZE_MAX; | |
8332 | #endif | |
8333 | ||
b36200f5 DV |
8334 | if (sq_offset) |
8335 | *sq_offset = off; | |
8336 | ||
75b28aff HV |
8337 | sq_array_size = array_size(sizeof(u32), sq_entries); |
8338 | if (sq_array_size == SIZE_MAX) | |
8339 | return SIZE_MAX; | |
8340 | ||
8341 | if (check_add_overflow(off, sq_array_size, &off)) | |
8342 | return SIZE_MAX; | |
8343 | ||
75b28aff HV |
8344 | return off; |
8345 | } | |
8346 | ||
2b188cc1 JA |
8347 | static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries) |
8348 | { | |
75b28aff | 8349 | size_t pages; |
2b188cc1 | 8350 | |
75b28aff HV |
8351 | pages = (size_t)1 << get_order( |
8352 | rings_size(sq_entries, cq_entries, NULL)); | |
8353 | pages += (size_t)1 << get_order( | |
8354 | array_size(sizeof(struct io_uring_sqe), sq_entries)); | |
2b188cc1 | 8355 | |
75b28aff | 8356 | return pages; |
2b188cc1 JA |
8357 | } |
8358 | ||
edafccee JA |
8359 | static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx) |
8360 | { | |
8361 | int i, j; | |
8362 | ||
8363 | if (!ctx->user_bufs) | |
8364 | return -ENXIO; | |
8365 | ||
8366 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
8367 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
8368 | ||
8369 | for (j = 0; j < imu->nr_bvecs; j++) | |
f1f6a7dd | 8370 | unpin_user_page(imu->bvec[j].bv_page); |
edafccee | 8371 | |
de293938 JA |
8372 | if (imu->acct_pages) |
8373 | io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED); | |
d4ef6475 | 8374 | kvfree(imu->bvec); |
edafccee JA |
8375 | imu->nr_bvecs = 0; |
8376 | } | |
8377 | ||
8378 | kfree(ctx->user_bufs); | |
8379 | ctx->user_bufs = NULL; | |
8380 | ctx->nr_user_bufs = 0; | |
8381 | return 0; | |
8382 | } | |
8383 | ||
8384 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
8385 | void __user *arg, unsigned index) | |
8386 | { | |
8387 | struct iovec __user *src; | |
8388 | ||
8389 | #ifdef CONFIG_COMPAT | |
8390 | if (ctx->compat) { | |
8391 | struct compat_iovec __user *ciovs; | |
8392 | struct compat_iovec ciov; | |
8393 | ||
8394 | ciovs = (struct compat_iovec __user *) arg; | |
8395 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
8396 | return -EFAULT; | |
8397 | ||
d55e5f5b | 8398 | dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base); |
edafccee JA |
8399 | dst->iov_len = ciov.iov_len; |
8400 | return 0; | |
8401 | } | |
8402 | #endif | |
8403 | src = (struct iovec __user *) arg; | |
8404 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
8405 | return -EFAULT; | |
8406 | return 0; | |
8407 | } | |
8408 | ||
de293938 JA |
8409 | /* |
8410 | * Not super efficient, but this is just a registration time. And we do cache | |
8411 | * the last compound head, so generally we'll only do a full search if we don't | |
8412 | * match that one. | |
8413 | * | |
8414 | * We check if the given compound head page has already been accounted, to | |
8415 | * avoid double accounting it. This allows us to account the full size of the | |
8416 | * page, not just the constituent pages of a huge page. | |
8417 | */ | |
8418 | static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages, | |
8419 | int nr_pages, struct page *hpage) | |
8420 | { | |
8421 | int i, j; | |
8422 | ||
8423 | /* check current page array */ | |
8424 | for (i = 0; i < nr_pages; i++) { | |
8425 | if (!PageCompound(pages[i])) | |
8426 | continue; | |
8427 | if (compound_head(pages[i]) == hpage) | |
8428 | return true; | |
8429 | } | |
8430 | ||
8431 | /* check previously registered pages */ | |
8432 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
8433 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
8434 | ||
8435 | for (j = 0; j < imu->nr_bvecs; j++) { | |
8436 | if (!PageCompound(imu->bvec[j].bv_page)) | |
8437 | continue; | |
8438 | if (compound_head(imu->bvec[j].bv_page) == hpage) | |
8439 | return true; | |
8440 | } | |
8441 | } | |
8442 | ||
8443 | return false; | |
8444 | } | |
8445 | ||
8446 | static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages, | |
8447 | int nr_pages, struct io_mapped_ubuf *imu, | |
8448 | struct page **last_hpage) | |
8449 | { | |
8450 | int i, ret; | |
8451 | ||
8452 | for (i = 0; i < nr_pages; i++) { | |
8453 | if (!PageCompound(pages[i])) { | |
8454 | imu->acct_pages++; | |
8455 | } else { | |
8456 | struct page *hpage; | |
8457 | ||
8458 | hpage = compound_head(pages[i]); | |
8459 | if (hpage == *last_hpage) | |
8460 | continue; | |
8461 | *last_hpage = hpage; | |
8462 | if (headpage_already_acct(ctx, pages, i, hpage)) | |
8463 | continue; | |
8464 | imu->acct_pages += page_size(hpage) >> PAGE_SHIFT; | |
8465 | } | |
8466 | } | |
8467 | ||
8468 | if (!imu->acct_pages) | |
8469 | return 0; | |
8470 | ||
8471 | ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED); | |
8472 | if (ret) | |
8473 | imu->acct_pages = 0; | |
8474 | return ret; | |
8475 | } | |
8476 | ||
edafccee JA |
8477 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, |
8478 | unsigned nr_args) | |
8479 | { | |
8480 | struct vm_area_struct **vmas = NULL; | |
8481 | struct page **pages = NULL; | |
de293938 | 8482 | struct page *last_hpage = NULL; |
edafccee JA |
8483 | int i, j, got_pages = 0; |
8484 | int ret = -EINVAL; | |
8485 | ||
8486 | if (ctx->user_bufs) | |
8487 | return -EBUSY; | |
8488 | if (!nr_args || nr_args > UIO_MAXIOV) | |
8489 | return -EINVAL; | |
8490 | ||
8491 | ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf), | |
8492 | GFP_KERNEL); | |
8493 | if (!ctx->user_bufs) | |
8494 | return -ENOMEM; | |
8495 | ||
8496 | for (i = 0; i < nr_args; i++) { | |
8497 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
8498 | unsigned long off, start, end, ubuf; | |
8499 | int pret, nr_pages; | |
8500 | struct iovec iov; | |
8501 | size_t size; | |
8502 | ||
8503 | ret = io_copy_iov(ctx, &iov, arg, i); | |
8504 | if (ret) | |
a278682d | 8505 | goto err; |
edafccee JA |
8506 | |
8507 | /* | |
8508 | * Don't impose further limits on the size and buffer | |
8509 | * constraints here, we'll -EINVAL later when IO is | |
8510 | * submitted if they are wrong. | |
8511 | */ | |
8512 | ret = -EFAULT; | |
8513 | if (!iov.iov_base || !iov.iov_len) | |
8514 | goto err; | |
8515 | ||
8516 | /* arbitrary limit, but we need something */ | |
8517 | if (iov.iov_len > SZ_1G) | |
8518 | goto err; | |
8519 | ||
8520 | ubuf = (unsigned long) iov.iov_base; | |
8521 | end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
8522 | start = ubuf >> PAGE_SHIFT; | |
8523 | nr_pages = end - start; | |
8524 | ||
edafccee JA |
8525 | ret = 0; |
8526 | if (!pages || nr_pages > got_pages) { | |
a8c73c1a DE |
8527 | kvfree(vmas); |
8528 | kvfree(pages); | |
d4ef6475 | 8529 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), |
edafccee | 8530 | GFP_KERNEL); |
d4ef6475 | 8531 | vmas = kvmalloc_array(nr_pages, |
edafccee JA |
8532 | sizeof(struct vm_area_struct *), |
8533 | GFP_KERNEL); | |
8534 | if (!pages || !vmas) { | |
8535 | ret = -ENOMEM; | |
edafccee JA |
8536 | goto err; |
8537 | } | |
8538 | got_pages = nr_pages; | |
8539 | } | |
8540 | ||
d4ef6475 | 8541 | imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec), |
edafccee JA |
8542 | GFP_KERNEL); |
8543 | ret = -ENOMEM; | |
de293938 | 8544 | if (!imu->bvec) |
edafccee | 8545 | goto err; |
edafccee JA |
8546 | |
8547 | ret = 0; | |
d8ed45c5 | 8548 | mmap_read_lock(current->mm); |
2113b05d | 8549 | pret = pin_user_pages(ubuf, nr_pages, |
932f4a63 IW |
8550 | FOLL_WRITE | FOLL_LONGTERM, |
8551 | pages, vmas); | |
edafccee JA |
8552 | if (pret == nr_pages) { |
8553 | /* don't support file backed memory */ | |
8554 | for (j = 0; j < nr_pages; j++) { | |
8555 | struct vm_area_struct *vma = vmas[j]; | |
8556 | ||
8557 | if (vma->vm_file && | |
8558 | !is_file_hugepages(vma->vm_file)) { | |
8559 | ret = -EOPNOTSUPP; | |
8560 | break; | |
8561 | } | |
8562 | } | |
8563 | } else { | |
8564 | ret = pret < 0 ? pret : -EFAULT; | |
8565 | } | |
d8ed45c5 | 8566 | mmap_read_unlock(current->mm); |
edafccee JA |
8567 | if (ret) { |
8568 | /* | |
8569 | * if we did partial map, or found file backed vmas, | |
8570 | * release any pages we did get | |
8571 | */ | |
27c4d3a3 | 8572 | if (pret > 0) |
f1f6a7dd | 8573 | unpin_user_pages(pages, pret); |
de293938 JA |
8574 | kvfree(imu->bvec); |
8575 | goto err; | |
8576 | } | |
8577 | ||
8578 | ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage); | |
8579 | if (ret) { | |
8580 | unpin_user_pages(pages, pret); | |
d4ef6475 | 8581 | kvfree(imu->bvec); |
edafccee JA |
8582 | goto err; |
8583 | } | |
8584 | ||
8585 | off = ubuf & ~PAGE_MASK; | |
8586 | size = iov.iov_len; | |
8587 | for (j = 0; j < nr_pages; j++) { | |
8588 | size_t vec_len; | |
8589 | ||
8590 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
8591 | imu->bvec[j].bv_page = pages[j]; | |
8592 | imu->bvec[j].bv_len = vec_len; | |
8593 | imu->bvec[j].bv_offset = off; | |
8594 | off = 0; | |
8595 | size -= vec_len; | |
8596 | } | |
8597 | /* store original address for later verification */ | |
8598 | imu->ubuf = ubuf; | |
8599 | imu->len = iov.iov_len; | |
8600 | imu->nr_bvecs = nr_pages; | |
8601 | ||
8602 | ctx->nr_user_bufs++; | |
8603 | } | |
d4ef6475 MR |
8604 | kvfree(pages); |
8605 | kvfree(vmas); | |
edafccee JA |
8606 | return 0; |
8607 | err: | |
d4ef6475 MR |
8608 | kvfree(pages); |
8609 | kvfree(vmas); | |
edafccee JA |
8610 | io_sqe_buffer_unregister(ctx); |
8611 | return ret; | |
8612 | } | |
8613 | ||
9b402849 JA |
8614 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
8615 | { | |
8616 | __s32 __user *fds = arg; | |
8617 | int fd; | |
8618 | ||
8619 | if (ctx->cq_ev_fd) | |
8620 | return -EBUSY; | |
8621 | ||
8622 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
8623 | return -EFAULT; | |
8624 | ||
8625 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
8626 | if (IS_ERR(ctx->cq_ev_fd)) { | |
8627 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
8628 | ctx->cq_ev_fd = NULL; | |
8629 | return ret; | |
8630 | } | |
8631 | ||
8632 | return 0; | |
8633 | } | |
8634 | ||
8635 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
8636 | { | |
8637 | if (ctx->cq_ev_fd) { | |
8638 | eventfd_ctx_put(ctx->cq_ev_fd); | |
8639 | ctx->cq_ev_fd = NULL; | |
8640 | return 0; | |
8641 | } | |
8642 | ||
8643 | return -ENXIO; | |
8644 | } | |
8645 | ||
5a2e745d JA |
8646 | static int __io_destroy_buffers(int id, void *p, void *data) |
8647 | { | |
8648 | struct io_ring_ctx *ctx = data; | |
8649 | struct io_buffer *buf = p; | |
8650 | ||
067524e9 | 8651 | __io_remove_buffers(ctx, buf, id, -1U); |
5a2e745d JA |
8652 | return 0; |
8653 | } | |
8654 | ||
8655 | static void io_destroy_buffers(struct io_ring_ctx *ctx) | |
8656 | { | |
8657 | idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx); | |
8658 | idr_destroy(&ctx->io_buffer_idr); | |
8659 | } | |
8660 | ||
2b188cc1 JA |
8661 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
8662 | { | |
6b06314c | 8663 | io_finish_async(ctx); |
5dbcad51 | 8664 | io_sqe_buffer_unregister(ctx); |
2aede0e4 JA |
8665 | |
8666 | if (ctx->sqo_task) { | |
8667 | put_task_struct(ctx->sqo_task); | |
8668 | ctx->sqo_task = NULL; | |
8669 | mmdrop(ctx->mm_account); | |
8670 | ctx->mm_account = NULL; | |
30975825 | 8671 | } |
def596e9 | 8672 | |
91d8f519 DZ |
8673 | #ifdef CONFIG_BLK_CGROUP |
8674 | if (ctx->sqo_blkcg_css) | |
8675 | css_put(ctx->sqo_blkcg_css); | |
8676 | #endif | |
8677 | ||
6b06314c | 8678 | io_sqe_files_unregister(ctx); |
9b402849 | 8679 | io_eventfd_unregister(ctx); |
5a2e745d | 8680 | io_destroy_buffers(ctx); |
41726c9a | 8681 | idr_destroy(&ctx->personality_idr); |
def596e9 | 8682 | |
2b188cc1 | 8683 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
8684 | if (ctx->ring_sock) { |
8685 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 8686 | sock_release(ctx->ring_sock); |
355e8d26 | 8687 | } |
2b188cc1 JA |
8688 | #endif |
8689 | ||
75b28aff | 8690 | io_mem_free(ctx->rings); |
2b188cc1 | 8691 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
8692 | |
8693 | percpu_ref_exit(&ctx->refs); | |
2b188cc1 | 8694 | free_uid(ctx->user); |
181e448d | 8695 | put_cred(ctx->creds); |
78076bb6 | 8696 | kfree(ctx->cancel_hash); |
0ddf92e8 | 8697 | kmem_cache_free(req_cachep, ctx->fallback_req); |
2b188cc1 JA |
8698 | kfree(ctx); |
8699 | } | |
8700 | ||
8701 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
8702 | { | |
8703 | struct io_ring_ctx *ctx = file->private_data; | |
8704 | __poll_t mask = 0; | |
8705 | ||
8706 | poll_wait(file, &ctx->cq_wait, wait); | |
4f7067c3 SB |
8707 | /* |
8708 | * synchronizes with barrier from wq_has_sleeper call in | |
8709 | * io_commit_cqring | |
8710 | */ | |
2b188cc1 | 8711 | smp_rmb(); |
90554200 | 8712 | if (!io_sqring_full(ctx)) |
2b188cc1 | 8713 | mask |= EPOLLOUT | EPOLLWRNORM; |
6c503150 PB |
8714 | io_cqring_overflow_flush(ctx, false, NULL, NULL); |
8715 | if (io_cqring_events(ctx)) | |
2b188cc1 JA |
8716 | mask |= EPOLLIN | EPOLLRDNORM; |
8717 | ||
8718 | return mask; | |
8719 | } | |
8720 | ||
8721 | static int io_uring_fasync(int fd, struct file *file, int on) | |
8722 | { | |
8723 | struct io_ring_ctx *ctx = file->private_data; | |
8724 | ||
8725 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
8726 | } | |
8727 | ||
071698e1 JA |
8728 | static int io_remove_personalities(int id, void *p, void *data) |
8729 | { | |
8730 | struct io_ring_ctx *ctx = data; | |
1e6fa521 | 8731 | struct io_identity *iod; |
071698e1 | 8732 | |
1e6fa521 JA |
8733 | iod = idr_remove(&ctx->personality_idr, id); |
8734 | if (iod) { | |
8735 | put_cred(iod->creds); | |
8736 | if (refcount_dec_and_test(&iod->count)) | |
8737 | kfree(iod); | |
8738 | } | |
071698e1 JA |
8739 | return 0; |
8740 | } | |
8741 | ||
85faa7b8 JA |
8742 | static void io_ring_exit_work(struct work_struct *work) |
8743 | { | |
b2edc0a7 PB |
8744 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, |
8745 | exit_work); | |
85faa7b8 | 8746 | |
56952e91 JA |
8747 | /* |
8748 | * If we're doing polled IO and end up having requests being | |
8749 | * submitted async (out-of-line), then completions can come in while | |
8750 | * we're waiting for refs to drop. We need to reap these manually, | |
8751 | * as nobody else will be looking for them. | |
8752 | */ | |
b2edc0a7 | 8753 | do { |
90df0853 | 8754 | __io_uring_cancel_task_requests(ctx, NULL); |
b2edc0a7 | 8755 | } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20)); |
85faa7b8 JA |
8756 | io_ring_ctx_free(ctx); |
8757 | } | |
8758 | ||
00c18640 JA |
8759 | static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data) |
8760 | { | |
8761 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
8762 | ||
8763 | return req->ctx == data; | |
8764 | } | |
8765 | ||
2b188cc1 JA |
8766 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
8767 | { | |
8768 | mutex_lock(&ctx->uring_lock); | |
8769 | percpu_ref_kill(&ctx->refs); | |
d9d05217 PB |
8770 | |
8771 | if (WARN_ON_ONCE((ctx->flags & IORING_SETUP_SQPOLL) && !ctx->sqo_dead)) | |
8772 | ctx->sqo_dead = 1; | |
8773 | ||
cda286f0 PB |
8774 | /* if force is set, the ring is going away. always drop after that */ |
8775 | ctx->cq_overflow_flushed = 1; | |
634578f8 | 8776 | if (ctx->rings) |
6c503150 | 8777 | __io_cqring_overflow_flush(ctx, true, NULL, NULL); |
2b188cc1 JA |
8778 | mutex_unlock(&ctx->uring_lock); |
8779 | ||
6b81928d PB |
8780 | io_kill_timeouts(ctx, NULL, NULL); |
8781 | io_poll_remove_all(ctx, NULL, NULL); | |
561fb04a JA |
8782 | |
8783 | if (ctx->io_wq) | |
00c18640 | 8784 | io_wq_cancel_cb(ctx->io_wq, io_cancel_ctx_cb, ctx, true); |
561fb04a | 8785 | |
15dff286 | 8786 | /* if we failed setting up the ctx, we might not have any rings */ |
b2edc0a7 | 8787 | io_iopoll_try_reap_events(ctx); |
071698e1 | 8788 | idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx); |
309fc03a JA |
8789 | |
8790 | /* | |
8791 | * Do this upfront, so we won't have a grace period where the ring | |
8792 | * is closed but resources aren't reaped yet. This can cause | |
8793 | * spurious failure in setting up a new ring. | |
8794 | */ | |
760618f7 JA |
8795 | io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries), |
8796 | ACCT_LOCKED); | |
309fc03a | 8797 | |
85faa7b8 | 8798 | INIT_WORK(&ctx->exit_work, io_ring_exit_work); |
fc666777 JA |
8799 | /* |
8800 | * Use system_unbound_wq to avoid spawning tons of event kworkers | |
8801 | * if we're exiting a ton of rings at the same time. It just adds | |
8802 | * noise and overhead, there's no discernable change in runtime | |
8803 | * over using system_wq. | |
8804 | */ | |
8805 | queue_work(system_unbound_wq, &ctx->exit_work); | |
2b188cc1 JA |
8806 | } |
8807 | ||
8808 | static int io_uring_release(struct inode *inode, struct file *file) | |
8809 | { | |
8810 | struct io_ring_ctx *ctx = file->private_data; | |
8811 | ||
8812 | file->private_data = NULL; | |
8813 | io_ring_ctx_wait_and_kill(ctx); | |
8814 | return 0; | |
8815 | } | |
8816 | ||
f6edbabb PB |
8817 | struct io_task_cancel { |
8818 | struct task_struct *task; | |
8819 | struct files_struct *files; | |
8820 | }; | |
f254ac04 | 8821 | |
f6edbabb | 8822 | static bool io_cancel_task_cb(struct io_wq_work *work, void *data) |
b711d4ea | 8823 | { |
9a472ef7 | 8824 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
f6edbabb | 8825 | struct io_task_cancel *cancel = data; |
9a472ef7 PB |
8826 | bool ret; |
8827 | ||
f6edbabb | 8828 | if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) { |
9a472ef7 PB |
8829 | unsigned long flags; |
8830 | struct io_ring_ctx *ctx = req->ctx; | |
8831 | ||
8832 | /* protect against races with linked timeouts */ | |
8833 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
f6edbabb | 8834 | ret = io_match_task(req, cancel->task, cancel->files); |
9a472ef7 PB |
8835 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
8836 | } else { | |
f6edbabb | 8837 | ret = io_match_task(req, cancel->task, cancel->files); |
9a472ef7 PB |
8838 | } |
8839 | return ret; | |
b711d4ea JA |
8840 | } |
8841 | ||
b7ddce3c | 8842 | static void io_cancel_defer_files(struct io_ring_ctx *ctx, |
ef9865a4 | 8843 | struct task_struct *task, |
b7ddce3c PB |
8844 | struct files_struct *files) |
8845 | { | |
8846 | struct io_defer_entry *de = NULL; | |
8847 | LIST_HEAD(list); | |
8848 | ||
8849 | spin_lock_irq(&ctx->completion_lock); | |
8850 | list_for_each_entry_reverse(de, &ctx->defer_list, list) { | |
08d23634 | 8851 | if (io_match_task(de->req, task, files)) { |
b7ddce3c PB |
8852 | list_cut_position(&list, &ctx->defer_list, &de->list); |
8853 | break; | |
8854 | } | |
8855 | } | |
8856 | spin_unlock_irq(&ctx->completion_lock); | |
8857 | ||
8858 | while (!list_empty(&list)) { | |
8859 | de = list_first_entry(&list, struct io_defer_entry, list); | |
8860 | list_del_init(&de->list); | |
8861 | req_set_fail_links(de->req); | |
8862 | io_put_req(de->req); | |
8863 | io_req_complete(de->req, -ECANCELED); | |
8864 | kfree(de); | |
8865 | } | |
8866 | } | |
8867 | ||
b52fda00 | 8868 | static void io_uring_cancel_files(struct io_ring_ctx *ctx, |
df9923f9 | 8869 | struct task_struct *task, |
fcb323cc JA |
8870 | struct files_struct *files) |
8871 | { | |
fcb323cc | 8872 | while (!list_empty_careful(&ctx->inflight_list)) { |
bee749b1 | 8873 | struct io_task_cancel cancel = { .task = task, .files = files }; |
f6edbabb | 8874 | struct io_kiocb *req; |
d8f1b971 | 8875 | DEFINE_WAIT(wait); |
f6edbabb | 8876 | bool found = false; |
fcb323cc JA |
8877 | |
8878 | spin_lock_irq(&ctx->inflight_lock); | |
8879 | list_for_each_entry(req, &ctx->inflight_list, inflight_entry) { | |
02a13674 | 8880 | if (!io_match_task(req, task, files)) |
768134d4 | 8881 | continue; |
f6edbabb | 8882 | found = true; |
768134d4 | 8883 | break; |
fcb323cc | 8884 | } |
f6edbabb | 8885 | if (found) |
c98de08c PB |
8886 | prepare_to_wait(&task->io_uring->wait, &wait, |
8887 | TASK_UNINTERRUPTIBLE); | |
fcb323cc JA |
8888 | spin_unlock_irq(&ctx->inflight_lock); |
8889 | ||
768134d4 | 8890 | /* We need to keep going until we don't find a matching req */ |
f6edbabb | 8891 | if (!found) |
fcb323cc | 8892 | break; |
f6edbabb PB |
8893 | |
8894 | io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, &cancel, true); | |
8895 | io_poll_remove_all(ctx, task, files); | |
8896 | io_kill_timeouts(ctx, task, files); | |
9d5c8190 | 8897 | io_cqring_overflow_flush(ctx, true, task, files); |
6200b0ae JA |
8898 | /* cancellations _may_ trigger task work */ |
8899 | io_run_task_work(); | |
fcb323cc | 8900 | schedule(); |
c98de08c | 8901 | finish_wait(&task->io_uring->wait, &wait); |
fcb323cc | 8902 | } |
44e728b8 PB |
8903 | } |
8904 | ||
b52fda00 PB |
8905 | static void __io_uring_cancel_task_requests(struct io_ring_ctx *ctx, |
8906 | struct task_struct *task) | |
0f212204 | 8907 | { |
b52fda00 | 8908 | while (1) { |
f6edbabb | 8909 | struct io_task_cancel cancel = { .task = task, .files = NULL, }; |
0f212204 | 8910 | enum io_wq_cancel cret; |
b52fda00 | 8911 | bool ret = false; |
0f212204 | 8912 | |
90df0853 PB |
8913 | if (ctx->io_wq) { |
8914 | cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, | |
8915 | &cancel, true); | |
8916 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
8917 | } | |
0f212204 JA |
8918 | |
8919 | /* SQPOLL thread does its own polling */ | |
8920 | if (!(ctx->flags & IORING_SETUP_SQPOLL)) { | |
8921 | while (!list_empty_careful(&ctx->iopoll_list)) { | |
8922 | io_iopoll_try_reap_events(ctx); | |
8923 | ret = true; | |
8924 | } | |
8925 | } | |
8926 | ||
6b81928d PB |
8927 | ret |= io_poll_remove_all(ctx, task, NULL); |
8928 | ret |= io_kill_timeouts(ctx, task, NULL); | |
55583d72 | 8929 | ret |= io_run_task_work(); |
b52fda00 PB |
8930 | if (!ret) |
8931 | break; | |
b52fda00 | 8932 | cond_resched(); |
0f212204 | 8933 | } |
0f212204 JA |
8934 | } |
8935 | ||
d9d05217 PB |
8936 | static void io_disable_sqo_submit(struct io_ring_ctx *ctx) |
8937 | { | |
d9d05217 PB |
8938 | mutex_lock(&ctx->uring_lock); |
8939 | ctx->sqo_dead = 1; | |
8940 | mutex_unlock(&ctx->uring_lock); | |
8941 | ||
8942 | /* make sure callers enter the ring to get error */ | |
b4411616 PB |
8943 | if (ctx->rings) |
8944 | io_ring_set_wakeup_flag(ctx); | |
d9d05217 PB |
8945 | } |
8946 | ||
0f212204 JA |
8947 | /* |
8948 | * We need to iteratively cancel requests, in case a request has dependent | |
8949 | * hard links. These persist even for failure of cancelations, hence keep | |
8950 | * looping until none are found. | |
8951 | */ | |
8952 | static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx, | |
8953 | struct files_struct *files) | |
8954 | { | |
8955 | struct task_struct *task = current; | |
8956 | ||
fdaf083c | 8957 | if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) { |
d9d05217 | 8958 | /* for SQPOLL only sqo_task has task notes */ |
6b393a1f | 8959 | WARN_ON_ONCE(ctx->sqo_task != current); |
d9d05217 | 8960 | io_disable_sqo_submit(ctx); |
534ca6d6 | 8961 | task = ctx->sq_data->thread; |
fdaf083c JA |
8962 | atomic_inc(&task->io_uring->in_idle); |
8963 | io_sq_thread_park(ctx->sq_data); | |
8964 | } | |
0f212204 | 8965 | |
df9923f9 | 8966 | io_cancel_defer_files(ctx, task, files); |
0f212204 JA |
8967 | io_cqring_overflow_flush(ctx, true, task, files); |
8968 | ||
b52fda00 PB |
8969 | if (!files) |
8970 | __io_uring_cancel_task_requests(ctx, task); | |
bee749b1 PB |
8971 | else |
8972 | io_uring_cancel_files(ctx, task, files); | |
fdaf083c JA |
8973 | |
8974 | if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) { | |
8975 | atomic_dec(&task->io_uring->in_idle); | |
8976 | /* | |
8977 | * If the files that are going away are the ones in the thread | |
8978 | * identity, clear them out. | |
8979 | */ | |
8980 | if (task->io_uring->identity->files == files) | |
8981 | task->io_uring->identity->files = NULL; | |
8982 | io_sq_thread_unpark(ctx->sq_data); | |
8983 | } | |
0f212204 JA |
8984 | } |
8985 | ||
8986 | /* | |
8987 | * Note that this task has used io_uring. We use it for cancelation purposes. | |
8988 | */ | |
fdaf083c | 8989 | static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file) |
0f212204 | 8990 | { |
236434c3 | 8991 | struct io_uring_task *tctx = current->io_uring; |
a528b04e | 8992 | int ret; |
236434c3 MWO |
8993 | |
8994 | if (unlikely(!tctx)) { | |
0f212204 JA |
8995 | ret = io_uring_alloc_task_context(current); |
8996 | if (unlikely(ret)) | |
8997 | return ret; | |
236434c3 | 8998 | tctx = current->io_uring; |
0f212204 | 8999 | } |
236434c3 MWO |
9000 | if (tctx->last != file) { |
9001 | void *old = xa_load(&tctx->xa, (unsigned long)file); | |
0f212204 | 9002 | |
236434c3 | 9003 | if (!old) { |
0f212204 | 9004 | get_file(file); |
a528b04e PB |
9005 | ret = xa_err(xa_store(&tctx->xa, (unsigned long)file, |
9006 | file, GFP_KERNEL)); | |
9007 | if (ret) { | |
9008 | fput(file); | |
9009 | return ret; | |
9010 | } | |
0f212204 | 9011 | } |
236434c3 | 9012 | tctx->last = file; |
0f212204 JA |
9013 | } |
9014 | ||
fdaf083c JA |
9015 | /* |
9016 | * This is race safe in that the task itself is doing this, hence it | |
9017 | * cannot be going through the exit/cancel paths at the same time. | |
9018 | * This cannot be modified while exit/cancel is running. | |
9019 | */ | |
9020 | if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL)) | |
9021 | tctx->sqpoll = true; | |
9022 | ||
0f212204 JA |
9023 | return 0; |
9024 | } | |
9025 | ||
9026 | /* | |
9027 | * Remove this io_uring_file -> task mapping. | |
9028 | */ | |
9029 | static void io_uring_del_task_file(struct file *file) | |
9030 | { | |
9031 | struct io_uring_task *tctx = current->io_uring; | |
0f212204 JA |
9032 | |
9033 | if (tctx->last == file) | |
9034 | tctx->last = NULL; | |
5e2ed8c4 | 9035 | file = xa_erase(&tctx->xa, (unsigned long)file); |
0f212204 JA |
9036 | if (file) |
9037 | fput(file); | |
9038 | } | |
9039 | ||
de7f1d9e PB |
9040 | static void io_uring_remove_task_files(struct io_uring_task *tctx) |
9041 | { | |
9042 | struct file *file; | |
9043 | unsigned long index; | |
9044 | ||
9045 | xa_for_each(&tctx->xa, index, file) | |
9046 | io_uring_del_task_file(file); | |
9047 | } | |
9048 | ||
0f212204 JA |
9049 | void __io_uring_files_cancel(struct files_struct *files) |
9050 | { | |
9051 | struct io_uring_task *tctx = current->io_uring; | |
ce765372 MWO |
9052 | struct file *file; |
9053 | unsigned long index; | |
0f212204 JA |
9054 | |
9055 | /* make sure overflow events are dropped */ | |
fdaf083c | 9056 | atomic_inc(&tctx->in_idle); |
de7f1d9e PB |
9057 | xa_for_each(&tctx->xa, index, file) |
9058 | io_uring_cancel_task_requests(file->private_data, files); | |
fdaf083c | 9059 | atomic_dec(&tctx->in_idle); |
de7f1d9e PB |
9060 | |
9061 | if (files) | |
9062 | io_uring_remove_task_files(tctx); | |
fdaf083c JA |
9063 | } |
9064 | ||
9065 | static s64 tctx_inflight(struct io_uring_task *tctx) | |
9066 | { | |
9067 | unsigned long index; | |
9068 | struct file *file; | |
9069 | s64 inflight; | |
9070 | ||
9071 | inflight = percpu_counter_sum(&tctx->inflight); | |
9072 | if (!tctx->sqpoll) | |
9073 | return inflight; | |
9074 | ||
9075 | /* | |
9076 | * If we have SQPOLL rings, then we need to iterate and find them, and | |
9077 | * add the pending count for those. | |
9078 | */ | |
9079 | xa_for_each(&tctx->xa, index, file) { | |
9080 | struct io_ring_ctx *ctx = file->private_data; | |
9081 | ||
9082 | if (ctx->flags & IORING_SETUP_SQPOLL) { | |
9083 | struct io_uring_task *__tctx = ctx->sqo_task->io_uring; | |
9084 | ||
9085 | inflight += percpu_counter_sum(&__tctx->inflight); | |
9086 | } | |
9087 | } | |
9088 | ||
9089 | return inflight; | |
0f212204 JA |
9090 | } |
9091 | ||
0f212204 JA |
9092 | /* |
9093 | * Find any io_uring fd that this task has registered or done IO on, and cancel | |
9094 | * requests. | |
9095 | */ | |
9096 | void __io_uring_task_cancel(void) | |
9097 | { | |
9098 | struct io_uring_task *tctx = current->io_uring; | |
9099 | DEFINE_WAIT(wait); | |
d8a6df10 | 9100 | s64 inflight; |
0f212204 JA |
9101 | |
9102 | /* make sure overflow events are dropped */ | |
fdaf083c | 9103 | atomic_inc(&tctx->in_idle); |
0f212204 | 9104 | |
0b5cd6c3 PB |
9105 | /* trigger io_disable_sqo_submit() */ |
9106 | if (tctx->sqpoll) | |
9107 | __io_uring_files_cancel(NULL); | |
9108 | ||
d8a6df10 | 9109 | do { |
0f212204 | 9110 | /* read completions before cancelations */ |
fdaf083c | 9111 | inflight = tctx_inflight(tctx); |
d8a6df10 JA |
9112 | if (!inflight) |
9113 | break; | |
0f212204 JA |
9114 | __io_uring_files_cancel(NULL); |
9115 | ||
9116 | prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE); | |
9117 | ||
9118 | /* | |
9119 | * If we've seen completions, retry. This avoids a race where | |
9120 | * a completion comes in before we did prepare_to_wait(). | |
9121 | */ | |
fdaf083c | 9122 | if (inflight != tctx_inflight(tctx)) |
0f212204 | 9123 | continue; |
0f212204 | 9124 | schedule(); |
f57555ed | 9125 | finish_wait(&tctx->wait, &wait); |
d8a6df10 | 9126 | } while (1); |
0f212204 | 9127 | |
a8d13dbc | 9128 | finish_wait(&tctx->wait, &wait); |
fdaf083c | 9129 | atomic_dec(&tctx->in_idle); |
de7f1d9e PB |
9130 | |
9131 | io_uring_remove_task_files(tctx); | |
44e728b8 PB |
9132 | } |
9133 | ||
fcb323cc JA |
9134 | static int io_uring_flush(struct file *file, void *data) |
9135 | { | |
6b5733eb | 9136 | struct io_uring_task *tctx = current->io_uring; |
d9d05217 | 9137 | struct io_ring_ctx *ctx = file->private_data; |
6b5733eb PB |
9138 | |
9139 | if (!tctx) | |
4f793dc4 PB |
9140 | return 0; |
9141 | ||
6b5733eb PB |
9142 | /* we should have cancelled and erased it before PF_EXITING */ |
9143 | WARN_ON_ONCE((current->flags & PF_EXITING) && | |
9144 | xa_load(&tctx->xa, (unsigned long)file)); | |
9145 | ||
4f793dc4 PB |
9146 | /* |
9147 | * fput() is pending, will be 2 if the only other ref is our potential | |
9148 | * task file note. If the task is exiting, drop regardless of count. | |
9149 | */ | |
6b5733eb PB |
9150 | if (atomic_long_read(&file->f_count) != 2) |
9151 | return 0; | |
4f793dc4 | 9152 | |
d9d05217 PB |
9153 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
9154 | /* there is only one file note, which is owned by sqo_task */ | |
4325cb49 PB |
9155 | WARN_ON_ONCE(ctx->sqo_task != current && |
9156 | xa_load(&tctx->xa, (unsigned long)file)); | |
9157 | /* sqo_dead check is for when this happens after cancellation */ | |
9158 | WARN_ON_ONCE(ctx->sqo_task == current && !ctx->sqo_dead && | |
d9d05217 PB |
9159 | !xa_load(&tctx->xa, (unsigned long)file)); |
9160 | ||
9161 | io_disable_sqo_submit(ctx); | |
9162 | } | |
9163 | ||
9164 | if (!(ctx->flags & IORING_SETUP_SQPOLL) || ctx->sqo_task == current) | |
9165 | io_uring_del_task_file(file); | |
fcb323cc JA |
9166 | return 0; |
9167 | } | |
9168 | ||
6c5c240e RP |
9169 | static void *io_uring_validate_mmap_request(struct file *file, |
9170 | loff_t pgoff, size_t sz) | |
2b188cc1 | 9171 | { |
2b188cc1 | 9172 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 9173 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
9174 | struct page *page; |
9175 | void *ptr; | |
9176 | ||
9177 | switch (offset) { | |
9178 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
9179 | case IORING_OFF_CQ_RING: |
9180 | ptr = ctx->rings; | |
2b188cc1 JA |
9181 | break; |
9182 | case IORING_OFF_SQES: | |
9183 | ptr = ctx->sq_sqes; | |
9184 | break; | |
2b188cc1 | 9185 | default: |
6c5c240e | 9186 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
9187 | } |
9188 | ||
9189 | page = virt_to_head_page(ptr); | |
a50b854e | 9190 | if (sz > page_size(page)) |
6c5c240e RP |
9191 | return ERR_PTR(-EINVAL); |
9192 | ||
9193 | return ptr; | |
9194 | } | |
9195 | ||
9196 | #ifdef CONFIG_MMU | |
9197 | ||
9198 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
9199 | { | |
9200 | size_t sz = vma->vm_end - vma->vm_start; | |
9201 | unsigned long pfn; | |
9202 | void *ptr; | |
9203 | ||
9204 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
9205 | if (IS_ERR(ptr)) | |
9206 | return PTR_ERR(ptr); | |
2b188cc1 JA |
9207 | |
9208 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
9209 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
9210 | } | |
9211 | ||
6c5c240e RP |
9212 | #else /* !CONFIG_MMU */ |
9213 | ||
9214 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
9215 | { | |
9216 | return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL; | |
9217 | } | |
9218 | ||
9219 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
9220 | { | |
9221 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
9222 | } | |
9223 | ||
9224 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
9225 | unsigned long addr, unsigned long len, | |
9226 | unsigned long pgoff, unsigned long flags) | |
9227 | { | |
9228 | void *ptr; | |
9229 | ||
9230 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
9231 | if (IS_ERR(ptr)) | |
9232 | return PTR_ERR(ptr); | |
9233 | ||
9234 | return (unsigned long) ptr; | |
9235 | } | |
9236 | ||
9237 | #endif /* !CONFIG_MMU */ | |
9238 | ||
d9d05217 | 9239 | static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx) |
90554200 | 9240 | { |
d9d05217 | 9241 | int ret = 0; |
90554200 JA |
9242 | DEFINE_WAIT(wait); |
9243 | ||
9244 | do { | |
9245 | if (!io_sqring_full(ctx)) | |
9246 | break; | |
9247 | ||
9248 | prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE); | |
9249 | ||
d9d05217 PB |
9250 | if (unlikely(ctx->sqo_dead)) { |
9251 | ret = -EOWNERDEAD; | |
9252 | goto out; | |
9253 | } | |
9254 | ||
90554200 JA |
9255 | if (!io_sqring_full(ctx)) |
9256 | break; | |
9257 | ||
9258 | schedule(); | |
9259 | } while (!signal_pending(current)); | |
9260 | ||
9261 | finish_wait(&ctx->sqo_sq_wait, &wait); | |
d9d05217 PB |
9262 | out: |
9263 | return ret; | |
90554200 JA |
9264 | } |
9265 | ||
c73ebb68 HX |
9266 | static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz, |
9267 | struct __kernel_timespec __user **ts, | |
9268 | const sigset_t __user **sig) | |
9269 | { | |
9270 | struct io_uring_getevents_arg arg; | |
9271 | ||
9272 | /* | |
9273 | * If EXT_ARG isn't set, then we have no timespec and the argp pointer | |
9274 | * is just a pointer to the sigset_t. | |
9275 | */ | |
9276 | if (!(flags & IORING_ENTER_EXT_ARG)) { | |
9277 | *sig = (const sigset_t __user *) argp; | |
9278 | *ts = NULL; | |
9279 | return 0; | |
9280 | } | |
9281 | ||
9282 | /* | |
9283 | * EXT_ARG is set - ensure we agree on the size of it and copy in our | |
9284 | * timespec and sigset_t pointers if good. | |
9285 | */ | |
9286 | if (*argsz != sizeof(arg)) | |
9287 | return -EINVAL; | |
9288 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
9289 | return -EFAULT; | |
9290 | *sig = u64_to_user_ptr(arg.sigmask); | |
9291 | *argsz = arg.sigmask_sz; | |
9292 | *ts = u64_to_user_ptr(arg.ts); | |
9293 | return 0; | |
9294 | } | |
9295 | ||
2b188cc1 | 9296 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
c73ebb68 HX |
9297 | u32, min_complete, u32, flags, const void __user *, argp, |
9298 | size_t, argsz) | |
2b188cc1 JA |
9299 | { |
9300 | struct io_ring_ctx *ctx; | |
9301 | long ret = -EBADF; | |
9302 | int submitted = 0; | |
9303 | struct fd f; | |
9304 | ||
4c6e277c | 9305 | io_run_task_work(); |
b41e9852 | 9306 | |
90554200 | 9307 | if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | |
c73ebb68 | 9308 | IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)) |
2b188cc1 JA |
9309 | return -EINVAL; |
9310 | ||
9311 | f = fdget(fd); | |
9312 | if (!f.file) | |
9313 | return -EBADF; | |
9314 | ||
9315 | ret = -EOPNOTSUPP; | |
9316 | if (f.file->f_op != &io_uring_fops) | |
9317 | goto out_fput; | |
9318 | ||
9319 | ret = -ENXIO; | |
9320 | ctx = f.file->private_data; | |
9321 | if (!percpu_ref_tryget(&ctx->refs)) | |
9322 | goto out_fput; | |
9323 | ||
7e84e1c7 SG |
9324 | ret = -EBADFD; |
9325 | if (ctx->flags & IORING_SETUP_R_DISABLED) | |
9326 | goto out; | |
9327 | ||
6c271ce2 JA |
9328 | /* |
9329 | * For SQ polling, the thread will do all submissions and completions. | |
9330 | * Just return the requested submit count, and wake the thread if | |
9331 | * we were asked to. | |
9332 | */ | |
b2a9eada | 9333 | ret = 0; |
6c271ce2 | 9334 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
6c503150 | 9335 | io_cqring_overflow_flush(ctx, false, NULL, NULL); |
89448c47 | 9336 | |
d9d05217 PB |
9337 | ret = -EOWNERDEAD; |
9338 | if (unlikely(ctx->sqo_dead)) | |
9339 | goto out; | |
6c271ce2 | 9340 | if (flags & IORING_ENTER_SQ_WAKEUP) |
534ca6d6 | 9341 | wake_up(&ctx->sq_data->wait); |
d9d05217 PB |
9342 | if (flags & IORING_ENTER_SQ_WAIT) { |
9343 | ret = io_sqpoll_wait_sq(ctx); | |
9344 | if (ret) | |
9345 | goto out; | |
9346 | } | |
6c271ce2 | 9347 | submitted = to_submit; |
b2a9eada | 9348 | } else if (to_submit) { |
fdaf083c | 9349 | ret = io_uring_add_task_file(ctx, f.file); |
0f212204 JA |
9350 | if (unlikely(ret)) |
9351 | goto out; | |
2b188cc1 | 9352 | mutex_lock(&ctx->uring_lock); |
0f212204 | 9353 | submitted = io_submit_sqes(ctx, to_submit); |
2b188cc1 | 9354 | mutex_unlock(&ctx->uring_lock); |
7c504e65 PB |
9355 | |
9356 | if (submitted != to_submit) | |
9357 | goto out; | |
2b188cc1 JA |
9358 | } |
9359 | if (flags & IORING_ENTER_GETEVENTS) { | |
c73ebb68 HX |
9360 | const sigset_t __user *sig; |
9361 | struct __kernel_timespec __user *ts; | |
9362 | ||
9363 | ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig); | |
9364 | if (unlikely(ret)) | |
9365 | goto out; | |
9366 | ||
2b188cc1 JA |
9367 | min_complete = min(min_complete, ctx->cq_entries); |
9368 | ||
32b2244a XW |
9369 | /* |
9370 | * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user | |
9371 | * space applications don't need to do io completion events | |
9372 | * polling again, they can rely on io_sq_thread to do polling | |
9373 | * work, which can reduce cpu usage and uring_lock contention. | |
9374 | */ | |
9375 | if (ctx->flags & IORING_SETUP_IOPOLL && | |
9376 | !(ctx->flags & IORING_SETUP_SQPOLL)) { | |
7668b92a | 9377 | ret = io_iopoll_check(ctx, min_complete); |
def596e9 | 9378 | } else { |
c73ebb68 | 9379 | ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts); |
def596e9 | 9380 | } |
2b188cc1 JA |
9381 | } |
9382 | ||
7c504e65 | 9383 | out: |
6805b32e | 9384 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
9385 | out_fput: |
9386 | fdput(f); | |
9387 | return submitted ? submitted : ret; | |
9388 | } | |
9389 | ||
bebdb65e | 9390 | #ifdef CONFIG_PROC_FS |
87ce955b JA |
9391 | static int io_uring_show_cred(int id, void *p, void *data) |
9392 | { | |
6b47ab81 JA |
9393 | struct io_identity *iod = p; |
9394 | const struct cred *cred = iod->creds; | |
87ce955b JA |
9395 | struct seq_file *m = data; |
9396 | struct user_namespace *uns = seq_user_ns(m); | |
9397 | struct group_info *gi; | |
9398 | kernel_cap_t cap; | |
9399 | unsigned __capi; | |
9400 | int g; | |
9401 | ||
9402 | seq_printf(m, "%5d\n", id); | |
9403 | seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid)); | |
9404 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid)); | |
9405 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid)); | |
9406 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid)); | |
9407 | seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid)); | |
9408 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid)); | |
9409 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid)); | |
9410 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid)); | |
9411 | seq_puts(m, "\n\tGroups:\t"); | |
9412 | gi = cred->group_info; | |
9413 | for (g = 0; g < gi->ngroups; g++) { | |
9414 | seq_put_decimal_ull(m, g ? " " : "", | |
9415 | from_kgid_munged(uns, gi->gid[g])); | |
9416 | } | |
9417 | seq_puts(m, "\n\tCapEff:\t"); | |
9418 | cap = cred->cap_effective; | |
9419 | CAP_FOR_EACH_U32(__capi) | |
9420 | seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8); | |
9421 | seq_putc(m, '\n'); | |
9422 | return 0; | |
9423 | } | |
9424 | ||
9425 | static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m) | |
9426 | { | |
dbbe9c64 | 9427 | struct io_sq_data *sq = NULL; |
fad8e0de | 9428 | bool has_lock; |
87ce955b JA |
9429 | int i; |
9430 | ||
fad8e0de JA |
9431 | /* |
9432 | * Avoid ABBA deadlock between the seq lock and the io_uring mutex, | |
9433 | * since fdinfo case grabs it in the opposite direction of normal use | |
9434 | * cases. If we fail to get the lock, we just don't iterate any | |
9435 | * structures that could be going away outside the io_uring mutex. | |
9436 | */ | |
9437 | has_lock = mutex_trylock(&ctx->uring_lock); | |
9438 | ||
dbbe9c64 JQ |
9439 | if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) |
9440 | sq = ctx->sq_data; | |
9441 | ||
9442 | seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1); | |
9443 | seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1); | |
87ce955b | 9444 | seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files); |
fad8e0de | 9445 | for (i = 0; has_lock && i < ctx->nr_user_files; i++) { |
87ce955b JA |
9446 | struct fixed_file_table *table; |
9447 | struct file *f; | |
9448 | ||
9449 | table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT]; | |
9450 | f = table->files[i & IORING_FILE_TABLE_MASK]; | |
9451 | if (f) | |
9452 | seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname); | |
9453 | else | |
9454 | seq_printf(m, "%5u: <none>\n", i); | |
9455 | } | |
9456 | seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs); | |
fad8e0de | 9457 | for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) { |
87ce955b JA |
9458 | struct io_mapped_ubuf *buf = &ctx->user_bufs[i]; |
9459 | ||
9460 | seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, | |
9461 | (unsigned int) buf->len); | |
9462 | } | |
fad8e0de | 9463 | if (has_lock && !idr_is_empty(&ctx->personality_idr)) { |
87ce955b JA |
9464 | seq_printf(m, "Personalities:\n"); |
9465 | idr_for_each(&ctx->personality_idr, io_uring_show_cred, m); | |
9466 | } | |
d7718a9d JA |
9467 | seq_printf(m, "PollList:\n"); |
9468 | spin_lock_irq(&ctx->completion_lock); | |
9469 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { | |
9470 | struct hlist_head *list = &ctx->cancel_hash[i]; | |
9471 | struct io_kiocb *req; | |
9472 | ||
9473 | hlist_for_each_entry(req, list, hash_node) | |
9474 | seq_printf(m, " op=%d, task_works=%d\n", req->opcode, | |
9475 | req->task->task_works != NULL); | |
9476 | } | |
9477 | spin_unlock_irq(&ctx->completion_lock); | |
fad8e0de JA |
9478 | if (has_lock) |
9479 | mutex_unlock(&ctx->uring_lock); | |
87ce955b JA |
9480 | } |
9481 | ||
9482 | static void io_uring_show_fdinfo(struct seq_file *m, struct file *f) | |
9483 | { | |
9484 | struct io_ring_ctx *ctx = f->private_data; | |
9485 | ||
9486 | if (percpu_ref_tryget(&ctx->refs)) { | |
9487 | __io_uring_show_fdinfo(ctx, m); | |
9488 | percpu_ref_put(&ctx->refs); | |
9489 | } | |
9490 | } | |
bebdb65e | 9491 | #endif |
87ce955b | 9492 | |
2b188cc1 JA |
9493 | static const struct file_operations io_uring_fops = { |
9494 | .release = io_uring_release, | |
fcb323cc | 9495 | .flush = io_uring_flush, |
2b188cc1 | 9496 | .mmap = io_uring_mmap, |
6c5c240e RP |
9497 | #ifndef CONFIG_MMU |
9498 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
9499 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
9500 | #endif | |
2b188cc1 JA |
9501 | .poll = io_uring_poll, |
9502 | .fasync = io_uring_fasync, | |
bebdb65e | 9503 | #ifdef CONFIG_PROC_FS |
87ce955b | 9504 | .show_fdinfo = io_uring_show_fdinfo, |
bebdb65e | 9505 | #endif |
2b188cc1 JA |
9506 | }; |
9507 | ||
9508 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
9509 | struct io_uring_params *p) | |
9510 | { | |
75b28aff HV |
9511 | struct io_rings *rings; |
9512 | size_t size, sq_array_offset; | |
2b188cc1 | 9513 | |
bd740481 JA |
9514 | /* make sure these are sane, as we already accounted them */ |
9515 | ctx->sq_entries = p->sq_entries; | |
9516 | ctx->cq_entries = p->cq_entries; | |
9517 | ||
75b28aff HV |
9518 | size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset); |
9519 | if (size == SIZE_MAX) | |
9520 | return -EOVERFLOW; | |
9521 | ||
9522 | rings = io_mem_alloc(size); | |
9523 | if (!rings) | |
2b188cc1 JA |
9524 | return -ENOMEM; |
9525 | ||
75b28aff HV |
9526 | ctx->rings = rings; |
9527 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
9528 | rings->sq_ring_mask = p->sq_entries - 1; | |
9529 | rings->cq_ring_mask = p->cq_entries - 1; | |
9530 | rings->sq_ring_entries = p->sq_entries; | |
9531 | rings->cq_ring_entries = p->cq_entries; | |
9532 | ctx->sq_mask = rings->sq_ring_mask; | |
9533 | ctx->cq_mask = rings->cq_ring_mask; | |
2b188cc1 JA |
9534 | |
9535 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 JA |
9536 | if (size == SIZE_MAX) { |
9537 | io_mem_free(ctx->rings); | |
9538 | ctx->rings = NULL; | |
2b188cc1 | 9539 | return -EOVERFLOW; |
eb065d30 | 9540 | } |
2b188cc1 JA |
9541 | |
9542 | ctx->sq_sqes = io_mem_alloc(size); | |
eb065d30 JA |
9543 | if (!ctx->sq_sqes) { |
9544 | io_mem_free(ctx->rings); | |
9545 | ctx->rings = NULL; | |
2b188cc1 | 9546 | return -ENOMEM; |
eb065d30 | 9547 | } |
2b188cc1 | 9548 | |
2b188cc1 JA |
9549 | return 0; |
9550 | } | |
9551 | ||
9faadcc8 PB |
9552 | static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file) |
9553 | { | |
9554 | int ret, fd; | |
9555 | ||
9556 | fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
9557 | if (fd < 0) | |
9558 | return fd; | |
9559 | ||
9560 | ret = io_uring_add_task_file(ctx, file); | |
9561 | if (ret) { | |
9562 | put_unused_fd(fd); | |
9563 | return ret; | |
9564 | } | |
9565 | fd_install(fd, file); | |
9566 | return fd; | |
9567 | } | |
9568 | ||
2b188cc1 JA |
9569 | /* |
9570 | * Allocate an anonymous fd, this is what constitutes the application | |
9571 | * visible backing of an io_uring instance. The application mmaps this | |
9572 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
9573 | * we have to tie this fd to a socket for file garbage collection purposes. | |
9574 | */ | |
9faadcc8 | 9575 | static struct file *io_uring_get_file(struct io_ring_ctx *ctx) |
2b188cc1 JA |
9576 | { |
9577 | struct file *file; | |
9faadcc8 | 9578 | #if defined(CONFIG_UNIX) |
2b188cc1 JA |
9579 | int ret; |
9580 | ||
2b188cc1 JA |
9581 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, |
9582 | &ctx->ring_sock); | |
9583 | if (ret) | |
9faadcc8 | 9584 | return ERR_PTR(ret); |
2b188cc1 JA |
9585 | #endif |
9586 | ||
2b188cc1 JA |
9587 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, |
9588 | O_RDWR | O_CLOEXEC); | |
2b188cc1 | 9589 | #if defined(CONFIG_UNIX) |
9faadcc8 PB |
9590 | if (IS_ERR(file)) { |
9591 | sock_release(ctx->ring_sock); | |
9592 | ctx->ring_sock = NULL; | |
9593 | } else { | |
9594 | ctx->ring_sock->file = file; | |
0f212204 | 9595 | } |
2b188cc1 | 9596 | #endif |
9faadcc8 | 9597 | return file; |
2b188cc1 JA |
9598 | } |
9599 | ||
7f13657d XW |
9600 | static int io_uring_create(unsigned entries, struct io_uring_params *p, |
9601 | struct io_uring_params __user *params) | |
2b188cc1 JA |
9602 | { |
9603 | struct user_struct *user = NULL; | |
9604 | struct io_ring_ctx *ctx; | |
9faadcc8 | 9605 | struct file *file; |
aad5d8da | 9606 | bool limit_mem; |
2b188cc1 JA |
9607 | int ret; |
9608 | ||
8110c1a6 | 9609 | if (!entries) |
2b188cc1 | 9610 | return -EINVAL; |
8110c1a6 JA |
9611 | if (entries > IORING_MAX_ENTRIES) { |
9612 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
9613 | return -EINVAL; | |
9614 | entries = IORING_MAX_ENTRIES; | |
9615 | } | |
2b188cc1 JA |
9616 | |
9617 | /* | |
9618 | * Use twice as many entries for the CQ ring. It's possible for the | |
9619 | * application to drive a higher depth than the size of the SQ ring, | |
9620 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
9621 | * some flexibility in overcommitting a bit. If the application has |
9622 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
9623 | * of CQ ring entries manually. | |
2b188cc1 JA |
9624 | */ |
9625 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
9626 | if (p->flags & IORING_SETUP_CQSIZE) { |
9627 | /* | |
9628 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
9629 | * to a power-of-two, if it isn't already. We do NOT impose | |
9630 | * any cq vs sq ring sizing. | |
9631 | */ | |
eb2667b3 | 9632 | if (!p->cq_entries) |
33a107f0 | 9633 | return -EINVAL; |
8110c1a6 JA |
9634 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
9635 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
9636 | return -EINVAL; | |
9637 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
9638 | } | |
eb2667b3 JQ |
9639 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
9640 | if (p->cq_entries < p->sq_entries) | |
9641 | return -EINVAL; | |
33a107f0 JA |
9642 | } else { |
9643 | p->cq_entries = 2 * p->sq_entries; | |
9644 | } | |
2b188cc1 JA |
9645 | |
9646 | user = get_uid(current_user()); | |
aad5d8da | 9647 | limit_mem = !capable(CAP_IPC_LOCK); |
2b188cc1 | 9648 | |
aad5d8da | 9649 | if (limit_mem) { |
a087e2b5 | 9650 | ret = __io_account_mem(user, |
2b188cc1 JA |
9651 | ring_pages(p->sq_entries, p->cq_entries)); |
9652 | if (ret) { | |
9653 | free_uid(user); | |
9654 | return ret; | |
9655 | } | |
9656 | } | |
9657 | ||
9658 | ctx = io_ring_ctx_alloc(p); | |
9659 | if (!ctx) { | |
aad5d8da | 9660 | if (limit_mem) |
a087e2b5 | 9661 | __io_unaccount_mem(user, ring_pages(p->sq_entries, |
2b188cc1 JA |
9662 | p->cq_entries)); |
9663 | free_uid(user); | |
9664 | return -ENOMEM; | |
9665 | } | |
9666 | ctx->compat = in_compat_syscall(); | |
2b188cc1 | 9667 | ctx->user = user; |
0b8c0ec7 | 9668 | ctx->creds = get_current_cred(); |
4ea33a97 JA |
9669 | #ifdef CONFIG_AUDIT |
9670 | ctx->loginuid = current->loginuid; | |
9671 | ctx->sessionid = current->sessionid; | |
9672 | #endif | |
2aede0e4 JA |
9673 | ctx->sqo_task = get_task_struct(current); |
9674 | ||
9675 | /* | |
9676 | * This is just grabbed for accounting purposes. When a process exits, | |
9677 | * the mm is exited and dropped before the files, hence we need to hang | |
9678 | * on to this mm purely for the purposes of being able to unaccount | |
9679 | * memory (locked/pinned vm). It's not used for anything else. | |
9680 | */ | |
6b7898eb | 9681 | mmgrab(current->mm); |
2aede0e4 | 9682 | ctx->mm_account = current->mm; |
6b7898eb | 9683 | |
91d8f519 DZ |
9684 | #ifdef CONFIG_BLK_CGROUP |
9685 | /* | |
9686 | * The sq thread will belong to the original cgroup it was inited in. | |
9687 | * If the cgroup goes offline (e.g. disabling the io controller), then | |
9688 | * issued bios will be associated with the closest cgroup later in the | |
9689 | * block layer. | |
9690 | */ | |
9691 | rcu_read_lock(); | |
9692 | ctx->sqo_blkcg_css = blkcg_css(); | |
9693 | ret = css_tryget_online(ctx->sqo_blkcg_css); | |
9694 | rcu_read_unlock(); | |
9695 | if (!ret) { | |
9696 | /* don't init against a dying cgroup, have the user try again */ | |
9697 | ctx->sqo_blkcg_css = NULL; | |
9698 | ret = -ENODEV; | |
9699 | goto err; | |
9700 | } | |
9701 | #endif | |
6b7898eb | 9702 | |
f74441e6 JA |
9703 | /* |
9704 | * Account memory _before_ installing the file descriptor. Once | |
9705 | * the descriptor is installed, it can get closed at any time. Also | |
9706 | * do this before hitting the general error path, as ring freeing | |
9707 | * will un-account as well. | |
9708 | */ | |
9709 | io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries), | |
9710 | ACCT_LOCKED); | |
9711 | ctx->limit_mem = limit_mem; | |
9712 | ||
2b188cc1 JA |
9713 | ret = io_allocate_scq_urings(ctx, p); |
9714 | if (ret) | |
9715 | goto err; | |
9716 | ||
7e84e1c7 | 9717 | ret = io_sq_offload_create(ctx, p); |
2b188cc1 JA |
9718 | if (ret) |
9719 | goto err; | |
9720 | ||
7e84e1c7 SG |
9721 | if (!(p->flags & IORING_SETUP_R_DISABLED)) |
9722 | io_sq_offload_start(ctx); | |
9723 | ||
2b188cc1 | 9724 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
9725 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
9726 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
9727 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
9728 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
9729 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
9730 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
9731 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
9732 | |
9733 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
9734 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
9735 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
9736 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
9737 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
9738 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
9739 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
0d9b5b3a | 9740 | p->cq_off.flags = offsetof(struct io_rings, cq_flags); |
ac90f249 | 9741 | |
7f13657d XW |
9742 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
9743 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | | |
5769a351 | 9744 | IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | |
c73ebb68 HX |
9745 | IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED | |
9746 | IORING_FEAT_EXT_ARG; | |
7f13657d XW |
9747 | |
9748 | if (copy_to_user(params, p, sizeof(*p))) { | |
9749 | ret = -EFAULT; | |
9750 | goto err; | |
9751 | } | |
d1719f70 | 9752 | |
9faadcc8 PB |
9753 | file = io_uring_get_file(ctx); |
9754 | if (IS_ERR(file)) { | |
9755 | ret = PTR_ERR(file); | |
9756 | goto err; | |
9757 | } | |
9758 | ||
044c1ab3 JA |
9759 | /* |
9760 | * Install ring fd as the very last thing, so we don't risk someone | |
9761 | * having closed it before we finish setup | |
9762 | */ | |
9faadcc8 PB |
9763 | ret = io_uring_install_fd(ctx, file); |
9764 | if (ret < 0) { | |
06585c49 | 9765 | io_disable_sqo_submit(ctx); |
9faadcc8 PB |
9766 | /* fput will clean it up */ |
9767 | fput(file); | |
9768 | return ret; | |
9769 | } | |
044c1ab3 | 9770 | |
c826bd7a | 9771 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
9772 | return ret; |
9773 | err: | |
d9d05217 | 9774 | io_disable_sqo_submit(ctx); |
2b188cc1 JA |
9775 | io_ring_ctx_wait_and_kill(ctx); |
9776 | return ret; | |
9777 | } | |
9778 | ||
9779 | /* | |
9780 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
9781 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
9782 | * params structure passed in. | |
9783 | */ | |
9784 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
9785 | { | |
9786 | struct io_uring_params p; | |
2b188cc1 JA |
9787 | int i; |
9788 | ||
9789 | if (copy_from_user(&p, params, sizeof(p))) | |
9790 | return -EFAULT; | |
9791 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
9792 | if (p.resv[i]) | |
9793 | return -EINVAL; | |
9794 | } | |
9795 | ||
6c271ce2 | 9796 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 9797 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
7e84e1c7 SG |
9798 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | |
9799 | IORING_SETUP_R_DISABLED)) | |
2b188cc1 JA |
9800 | return -EINVAL; |
9801 | ||
7f13657d | 9802 | return io_uring_create(entries, &p, params); |
2b188cc1 JA |
9803 | } |
9804 | ||
9805 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
9806 | struct io_uring_params __user *, params) | |
9807 | { | |
9808 | return io_uring_setup(entries, params); | |
9809 | } | |
9810 | ||
66f4af93 JA |
9811 | static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args) |
9812 | { | |
9813 | struct io_uring_probe *p; | |
9814 | size_t size; | |
9815 | int i, ret; | |
9816 | ||
9817 | size = struct_size(p, ops, nr_args); | |
9818 | if (size == SIZE_MAX) | |
9819 | return -EOVERFLOW; | |
9820 | p = kzalloc(size, GFP_KERNEL); | |
9821 | if (!p) | |
9822 | return -ENOMEM; | |
9823 | ||
9824 | ret = -EFAULT; | |
9825 | if (copy_from_user(p, arg, size)) | |
9826 | goto out; | |
9827 | ret = -EINVAL; | |
9828 | if (memchr_inv(p, 0, size)) | |
9829 | goto out; | |
9830 | ||
9831 | p->last_op = IORING_OP_LAST - 1; | |
9832 | if (nr_args > IORING_OP_LAST) | |
9833 | nr_args = IORING_OP_LAST; | |
9834 | ||
9835 | for (i = 0; i < nr_args; i++) { | |
9836 | p->ops[i].op = i; | |
9837 | if (!io_op_defs[i].not_supported) | |
9838 | p->ops[i].flags = IO_URING_OP_SUPPORTED; | |
9839 | } | |
9840 | p->ops_len = i; | |
9841 | ||
9842 | ret = 0; | |
9843 | if (copy_to_user(arg, p, size)) | |
9844 | ret = -EFAULT; | |
9845 | out: | |
9846 | kfree(p); | |
9847 | return ret; | |
9848 | } | |
9849 | ||
071698e1 JA |
9850 | static int io_register_personality(struct io_ring_ctx *ctx) |
9851 | { | |
1e6fa521 JA |
9852 | struct io_identity *id; |
9853 | int ret; | |
071698e1 | 9854 | |
1e6fa521 JA |
9855 | id = kmalloc(sizeof(*id), GFP_KERNEL); |
9856 | if (unlikely(!id)) | |
9857 | return -ENOMEM; | |
9858 | ||
9859 | io_init_identity(id); | |
9860 | id->creds = get_current_cred(); | |
9861 | ||
9862 | ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL); | |
9863 | if (ret < 0) { | |
9864 | put_cred(id->creds); | |
9865 | kfree(id); | |
9866 | } | |
9867 | return ret; | |
071698e1 JA |
9868 | } |
9869 | ||
9870 | static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) | |
9871 | { | |
1e6fa521 | 9872 | struct io_identity *iod; |
071698e1 | 9873 | |
1e6fa521 JA |
9874 | iod = idr_remove(&ctx->personality_idr, id); |
9875 | if (iod) { | |
9876 | put_cred(iod->creds); | |
9877 | if (refcount_dec_and_test(&iod->count)) | |
9878 | kfree(iod); | |
071698e1 JA |
9879 | return 0; |
9880 | } | |
9881 | ||
9882 | return -EINVAL; | |
9883 | } | |
9884 | ||
21b55dbc SG |
9885 | static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg, |
9886 | unsigned int nr_args) | |
9887 | { | |
9888 | struct io_uring_restriction *res; | |
9889 | size_t size; | |
9890 | int i, ret; | |
9891 | ||
7e84e1c7 SG |
9892 | /* Restrictions allowed only if rings started disabled */ |
9893 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
9894 | return -EBADFD; | |
9895 | ||
21b55dbc | 9896 | /* We allow only a single restrictions registration */ |
7e84e1c7 | 9897 | if (ctx->restrictions.registered) |
21b55dbc SG |
9898 | return -EBUSY; |
9899 | ||
9900 | if (!arg || nr_args > IORING_MAX_RESTRICTIONS) | |
9901 | return -EINVAL; | |
9902 | ||
9903 | size = array_size(nr_args, sizeof(*res)); | |
9904 | if (size == SIZE_MAX) | |
9905 | return -EOVERFLOW; | |
9906 | ||
9907 | res = memdup_user(arg, size); | |
9908 | if (IS_ERR(res)) | |
9909 | return PTR_ERR(res); | |
9910 | ||
9911 | ret = 0; | |
9912 | ||
9913 | for (i = 0; i < nr_args; i++) { | |
9914 | switch (res[i].opcode) { | |
9915 | case IORING_RESTRICTION_REGISTER_OP: | |
9916 | if (res[i].register_op >= IORING_REGISTER_LAST) { | |
9917 | ret = -EINVAL; | |
9918 | goto out; | |
9919 | } | |
9920 | ||
9921 | __set_bit(res[i].register_op, | |
9922 | ctx->restrictions.register_op); | |
9923 | break; | |
9924 | case IORING_RESTRICTION_SQE_OP: | |
9925 | if (res[i].sqe_op >= IORING_OP_LAST) { | |
9926 | ret = -EINVAL; | |
9927 | goto out; | |
9928 | } | |
9929 | ||
9930 | __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op); | |
9931 | break; | |
9932 | case IORING_RESTRICTION_SQE_FLAGS_ALLOWED: | |
9933 | ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags; | |
9934 | break; | |
9935 | case IORING_RESTRICTION_SQE_FLAGS_REQUIRED: | |
9936 | ctx->restrictions.sqe_flags_required = res[i].sqe_flags; | |
9937 | break; | |
9938 | default: | |
9939 | ret = -EINVAL; | |
9940 | goto out; | |
9941 | } | |
9942 | } | |
9943 | ||
9944 | out: | |
9945 | /* Reset all restrictions if an error happened */ | |
9946 | if (ret != 0) | |
9947 | memset(&ctx->restrictions, 0, sizeof(ctx->restrictions)); | |
9948 | else | |
7e84e1c7 | 9949 | ctx->restrictions.registered = true; |
21b55dbc SG |
9950 | |
9951 | kfree(res); | |
9952 | return ret; | |
9953 | } | |
9954 | ||
7e84e1c7 SG |
9955 | static int io_register_enable_rings(struct io_ring_ctx *ctx) |
9956 | { | |
9957 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
9958 | return -EBADFD; | |
9959 | ||
9960 | if (ctx->restrictions.registered) | |
9961 | ctx->restricted = 1; | |
9962 | ||
9963 | ctx->flags &= ~IORING_SETUP_R_DISABLED; | |
9964 | ||
9965 | io_sq_offload_start(ctx); | |
9966 | ||
9967 | return 0; | |
9968 | } | |
9969 | ||
071698e1 JA |
9970 | static bool io_register_op_must_quiesce(int op) |
9971 | { | |
9972 | switch (op) { | |
9973 | case IORING_UNREGISTER_FILES: | |
9974 | case IORING_REGISTER_FILES_UPDATE: | |
9975 | case IORING_REGISTER_PROBE: | |
9976 | case IORING_REGISTER_PERSONALITY: | |
9977 | case IORING_UNREGISTER_PERSONALITY: | |
9978 | return false; | |
9979 | default: | |
9980 | return true; | |
9981 | } | |
9982 | } | |
9983 | ||
edafccee JA |
9984 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
9985 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
9986 | __releases(ctx->uring_lock) |
9987 | __acquires(ctx->uring_lock) | |
edafccee JA |
9988 | { |
9989 | int ret; | |
9990 | ||
35fa71a0 JA |
9991 | /* |
9992 | * We're inside the ring mutex, if the ref is already dying, then | |
9993 | * someone else killed the ctx or is already going through | |
9994 | * io_uring_register(). | |
9995 | */ | |
9996 | if (percpu_ref_is_dying(&ctx->refs)) | |
9997 | return -ENXIO; | |
9998 | ||
071698e1 | 9999 | if (io_register_op_must_quiesce(opcode)) { |
05f3fb3c | 10000 | percpu_ref_kill(&ctx->refs); |
b19062a5 | 10001 | |
05f3fb3c JA |
10002 | /* |
10003 | * Drop uring mutex before waiting for references to exit. If | |
10004 | * another thread is currently inside io_uring_enter() it might | |
10005 | * need to grab the uring_lock to make progress. If we hold it | |
10006 | * here across the drain wait, then we can deadlock. It's safe | |
10007 | * to drop the mutex here, since no new references will come in | |
10008 | * after we've killed the percpu ref. | |
10009 | */ | |
10010 | mutex_unlock(&ctx->uring_lock); | |
af9c1a44 JA |
10011 | do { |
10012 | ret = wait_for_completion_interruptible(&ctx->ref_comp); | |
10013 | if (!ret) | |
10014 | break; | |
ed6930c9 JA |
10015 | ret = io_run_task_work_sig(); |
10016 | if (ret < 0) | |
10017 | break; | |
af9c1a44 JA |
10018 | } while (1); |
10019 | ||
05f3fb3c | 10020 | mutex_lock(&ctx->uring_lock); |
af9c1a44 | 10021 | |
c150368b JA |
10022 | if (ret) { |
10023 | percpu_ref_resurrect(&ctx->refs); | |
21b55dbc SG |
10024 | goto out_quiesce; |
10025 | } | |
10026 | } | |
10027 | ||
10028 | if (ctx->restricted) { | |
10029 | if (opcode >= IORING_REGISTER_LAST) { | |
10030 | ret = -EINVAL; | |
10031 | goto out; | |
10032 | } | |
10033 | ||
10034 | if (!test_bit(opcode, ctx->restrictions.register_op)) { | |
10035 | ret = -EACCES; | |
c150368b JA |
10036 | goto out; |
10037 | } | |
05f3fb3c | 10038 | } |
edafccee JA |
10039 | |
10040 | switch (opcode) { | |
10041 | case IORING_REGISTER_BUFFERS: | |
10042 | ret = io_sqe_buffer_register(ctx, arg, nr_args); | |
10043 | break; | |
10044 | case IORING_UNREGISTER_BUFFERS: | |
10045 | ret = -EINVAL; | |
10046 | if (arg || nr_args) | |
10047 | break; | |
10048 | ret = io_sqe_buffer_unregister(ctx); | |
10049 | break; | |
6b06314c JA |
10050 | case IORING_REGISTER_FILES: |
10051 | ret = io_sqe_files_register(ctx, arg, nr_args); | |
10052 | break; | |
10053 | case IORING_UNREGISTER_FILES: | |
10054 | ret = -EINVAL; | |
10055 | if (arg || nr_args) | |
10056 | break; | |
10057 | ret = io_sqe_files_unregister(ctx); | |
10058 | break; | |
c3a31e60 JA |
10059 | case IORING_REGISTER_FILES_UPDATE: |
10060 | ret = io_sqe_files_update(ctx, arg, nr_args); | |
10061 | break; | |
9b402849 | 10062 | case IORING_REGISTER_EVENTFD: |
f2842ab5 | 10063 | case IORING_REGISTER_EVENTFD_ASYNC: |
9b402849 JA |
10064 | ret = -EINVAL; |
10065 | if (nr_args != 1) | |
10066 | break; | |
10067 | ret = io_eventfd_register(ctx, arg); | |
f2842ab5 JA |
10068 | if (ret) |
10069 | break; | |
10070 | if (opcode == IORING_REGISTER_EVENTFD_ASYNC) | |
10071 | ctx->eventfd_async = 1; | |
10072 | else | |
10073 | ctx->eventfd_async = 0; | |
9b402849 JA |
10074 | break; |
10075 | case IORING_UNREGISTER_EVENTFD: | |
10076 | ret = -EINVAL; | |
10077 | if (arg || nr_args) | |
10078 | break; | |
10079 | ret = io_eventfd_unregister(ctx); | |
10080 | break; | |
66f4af93 JA |
10081 | case IORING_REGISTER_PROBE: |
10082 | ret = -EINVAL; | |
10083 | if (!arg || nr_args > 256) | |
10084 | break; | |
10085 | ret = io_probe(ctx, arg, nr_args); | |
10086 | break; | |
071698e1 JA |
10087 | case IORING_REGISTER_PERSONALITY: |
10088 | ret = -EINVAL; | |
10089 | if (arg || nr_args) | |
10090 | break; | |
10091 | ret = io_register_personality(ctx); | |
10092 | break; | |
10093 | case IORING_UNREGISTER_PERSONALITY: | |
10094 | ret = -EINVAL; | |
10095 | if (arg) | |
10096 | break; | |
10097 | ret = io_unregister_personality(ctx, nr_args); | |
10098 | break; | |
7e84e1c7 SG |
10099 | case IORING_REGISTER_ENABLE_RINGS: |
10100 | ret = -EINVAL; | |
10101 | if (arg || nr_args) | |
10102 | break; | |
10103 | ret = io_register_enable_rings(ctx); | |
10104 | break; | |
21b55dbc SG |
10105 | case IORING_REGISTER_RESTRICTIONS: |
10106 | ret = io_register_restrictions(ctx, arg, nr_args); | |
10107 | break; | |
edafccee JA |
10108 | default: |
10109 | ret = -EINVAL; | |
10110 | break; | |
10111 | } | |
10112 | ||
21b55dbc | 10113 | out: |
071698e1 | 10114 | if (io_register_op_must_quiesce(opcode)) { |
05f3fb3c | 10115 | /* bring the ctx back to life */ |
05f3fb3c | 10116 | percpu_ref_reinit(&ctx->refs); |
21b55dbc | 10117 | out_quiesce: |
0f158b4c | 10118 | reinit_completion(&ctx->ref_comp); |
05f3fb3c | 10119 | } |
edafccee JA |
10120 | return ret; |
10121 | } | |
10122 | ||
10123 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
10124 | void __user *, arg, unsigned int, nr_args) | |
10125 | { | |
10126 | struct io_ring_ctx *ctx; | |
10127 | long ret = -EBADF; | |
10128 | struct fd f; | |
10129 | ||
10130 | f = fdget(fd); | |
10131 | if (!f.file) | |
10132 | return -EBADF; | |
10133 | ||
10134 | ret = -EOPNOTSUPP; | |
10135 | if (f.file->f_op != &io_uring_fops) | |
10136 | goto out_fput; | |
10137 | ||
10138 | ctx = f.file->private_data; | |
10139 | ||
10140 | mutex_lock(&ctx->uring_lock); | |
10141 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
10142 | mutex_unlock(&ctx->uring_lock); | |
c826bd7a DD |
10143 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, |
10144 | ctx->cq_ev_fd != NULL, ret); | |
edafccee JA |
10145 | out_fput: |
10146 | fdput(f); | |
10147 | return ret; | |
10148 | } | |
10149 | ||
2b188cc1 JA |
10150 | static int __init io_uring_init(void) |
10151 | { | |
d7f62e82 SM |
10152 | #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \ |
10153 | BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ | |
10154 | BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \ | |
10155 | } while (0) | |
10156 | ||
10157 | #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ | |
10158 | __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename) | |
10159 | BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); | |
10160 | BUILD_BUG_SQE_ELEM(0, __u8, opcode); | |
10161 | BUILD_BUG_SQE_ELEM(1, __u8, flags); | |
10162 | BUILD_BUG_SQE_ELEM(2, __u16, ioprio); | |
10163 | BUILD_BUG_SQE_ELEM(4, __s32, fd); | |
10164 | BUILD_BUG_SQE_ELEM(8, __u64, off); | |
10165 | BUILD_BUG_SQE_ELEM(8, __u64, addr2); | |
10166 | BUILD_BUG_SQE_ELEM(16, __u64, addr); | |
7d67af2c | 10167 | BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); |
d7f62e82 SM |
10168 | BUILD_BUG_SQE_ELEM(24, __u32, len); |
10169 | BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); | |
10170 | BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); | |
10171 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); | |
10172 | BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); | |
5769a351 JX |
10173 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); |
10174 | BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); | |
d7f62e82 SM |
10175 | BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); |
10176 | BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); | |
10177 | BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); | |
10178 | BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); | |
10179 | BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); | |
10180 | BUILD_BUG_SQE_ELEM(28, __u32, open_flags); | |
10181 | BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); | |
10182 | BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); | |
7d67af2c | 10183 | BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); |
d7f62e82 SM |
10184 | BUILD_BUG_SQE_ELEM(32, __u64, user_data); |
10185 | BUILD_BUG_SQE_ELEM(40, __u16, buf_index); | |
10186 | BUILD_BUG_SQE_ELEM(42, __u16, personality); | |
7d67af2c | 10187 | BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); |
d7f62e82 | 10188 | |
d3656344 | 10189 | BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST); |
84557871 | 10190 | BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int)); |
2b188cc1 JA |
10191 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC); |
10192 | return 0; | |
10193 | }; | |
10194 | __initcall(io_uring_init); |