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); |
907d1df3 | 1029 | static void io_req_task_queue(struct io_kiocb *req); |
de0617e4 | 1030 | |
2b188cc1 JA |
1031 | static struct kmem_cache *req_cachep; |
1032 | ||
0918682b | 1033 | static const struct file_operations io_uring_fops; |
2b188cc1 JA |
1034 | |
1035 | struct sock *io_uring_get_socket(struct file *file) | |
1036 | { | |
1037 | #if defined(CONFIG_UNIX) | |
1038 | if (file->f_op == &io_uring_fops) { | |
1039 | struct io_ring_ctx *ctx = file->private_data; | |
1040 | ||
1041 | return ctx->ring_sock->sk; | |
1042 | } | |
1043 | #endif | |
1044 | return NULL; | |
1045 | } | |
1046 | EXPORT_SYMBOL(io_uring_get_socket); | |
1047 | ||
f2f87370 PB |
1048 | #define io_for_each_link(pos, head) \ |
1049 | for (pos = (head); pos; pos = pos->link) | |
1050 | ||
3ca405eb PB |
1051 | static inline void io_clean_op(struct io_kiocb *req) |
1052 | { | |
9d5c8190 | 1053 | if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED)) |
3ca405eb PB |
1054 | __io_clean_op(req); |
1055 | } | |
1056 | ||
36f72fe2 PB |
1057 | static inline void io_set_resource_node(struct io_kiocb *req) |
1058 | { | |
1059 | struct io_ring_ctx *ctx = req->ctx; | |
1060 | ||
1061 | if (!req->fixed_file_refs) { | |
1062 | req->fixed_file_refs = &ctx->file_data->node->refs; | |
1063 | percpu_ref_get(req->fixed_file_refs); | |
1064 | } | |
1065 | } | |
1066 | ||
08d23634 PB |
1067 | static bool io_match_task(struct io_kiocb *head, |
1068 | struct task_struct *task, | |
1069 | struct files_struct *files) | |
1070 | { | |
1071 | struct io_kiocb *req; | |
1072 | ||
84965ff8 JA |
1073 | if (task && head->task != task) { |
1074 | /* in terms of cancelation, always match if req task is dead */ | |
1075 | if (head->task->flags & PF_EXITING) | |
1076 | return true; | |
08d23634 | 1077 | return false; |
84965ff8 | 1078 | } |
08d23634 PB |
1079 | if (!files) |
1080 | return true; | |
1081 | ||
1082 | io_for_each_link(req, head) { | |
02a13674 JA |
1083 | if (!(req->flags & REQ_F_WORK_INITIALIZED)) |
1084 | continue; | |
1085 | if (req->file && req->file->f_op == &io_uring_fops) | |
1086 | return true; | |
1087 | if ((req->work.flags & IO_WQ_WORK_FILES) && | |
08d23634 PB |
1088 | req->work.identity->files == files) |
1089 | return true; | |
1090 | } | |
1091 | return false; | |
1092 | } | |
1093 | ||
28cea78a | 1094 | static void io_sq_thread_drop_mm_files(void) |
c40f6379 | 1095 | { |
28cea78a | 1096 | struct files_struct *files = current->files; |
c40f6379 JA |
1097 | struct mm_struct *mm = current->mm; |
1098 | ||
1099 | if (mm) { | |
1100 | kthread_unuse_mm(mm); | |
1101 | mmput(mm); | |
4b70cf9d | 1102 | current->mm = NULL; |
c40f6379 | 1103 | } |
28cea78a JA |
1104 | if (files) { |
1105 | struct nsproxy *nsproxy = current->nsproxy; | |
1106 | ||
1107 | task_lock(current); | |
1108 | current->files = NULL; | |
1109 | current->nsproxy = NULL; | |
1110 | task_unlock(current); | |
1111 | put_files_struct(files); | |
1112 | put_nsproxy(nsproxy); | |
1113 | } | |
1114 | } | |
1115 | ||
1a38ffc9 | 1116 | static int __io_sq_thread_acquire_files(struct io_ring_ctx *ctx) |
28cea78a | 1117 | { |
621fadc2 PB |
1118 | if (current->flags & PF_EXITING) |
1119 | return -EFAULT; | |
1120 | ||
28cea78a JA |
1121 | if (!current->files) { |
1122 | struct files_struct *files; | |
1123 | struct nsproxy *nsproxy; | |
1124 | ||
1125 | task_lock(ctx->sqo_task); | |
1126 | files = ctx->sqo_task->files; | |
1127 | if (!files) { | |
1128 | task_unlock(ctx->sqo_task); | |
1a38ffc9 | 1129 | return -EOWNERDEAD; |
28cea78a JA |
1130 | } |
1131 | atomic_inc(&files->count); | |
1132 | get_nsproxy(ctx->sqo_task->nsproxy); | |
1133 | nsproxy = ctx->sqo_task->nsproxy; | |
1134 | task_unlock(ctx->sqo_task); | |
1135 | ||
1136 | task_lock(current); | |
1137 | current->files = files; | |
1138 | current->nsproxy = nsproxy; | |
1139 | task_unlock(current); | |
1140 | } | |
1a38ffc9 | 1141 | return 0; |
c40f6379 JA |
1142 | } |
1143 | ||
1144 | static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx) | |
1145 | { | |
4b70cf9d JA |
1146 | struct mm_struct *mm; |
1147 | ||
621fadc2 PB |
1148 | if (current->flags & PF_EXITING) |
1149 | return -EFAULT; | |
4b70cf9d JA |
1150 | if (current->mm) |
1151 | return 0; | |
1152 | ||
1153 | /* Should never happen */ | |
1154 | if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL))) | |
1155 | return -EFAULT; | |
1156 | ||
1157 | task_lock(ctx->sqo_task); | |
1158 | mm = ctx->sqo_task->mm; | |
1159 | if (unlikely(!mm || !mmget_not_zero(mm))) | |
1160 | mm = NULL; | |
1161 | task_unlock(ctx->sqo_task); | |
1162 | ||
1163 | if (mm) { | |
1164 | kthread_use_mm(mm); | |
1165 | return 0; | |
c40f6379 JA |
1166 | } |
1167 | ||
4b70cf9d | 1168 | return -EFAULT; |
c40f6379 JA |
1169 | } |
1170 | ||
28cea78a JA |
1171 | static int io_sq_thread_acquire_mm_files(struct io_ring_ctx *ctx, |
1172 | struct io_kiocb *req) | |
c40f6379 | 1173 | { |
28cea78a | 1174 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
1a38ffc9 | 1175 | int ret; |
28cea78a JA |
1176 | |
1177 | if (def->work_flags & IO_WQ_WORK_MM) { | |
1a38ffc9 | 1178 | ret = __io_sq_thread_acquire_mm(ctx); |
28cea78a JA |
1179 | if (unlikely(ret)) |
1180 | return ret; | |
1181 | } | |
1182 | ||
1a38ffc9 PB |
1183 | if (def->needs_file || (def->work_flags & IO_WQ_WORK_FILES)) { |
1184 | ret = __io_sq_thread_acquire_files(ctx); | |
1185 | if (unlikely(ret)) | |
1186 | return ret; | |
1187 | } | |
28cea78a JA |
1188 | |
1189 | return 0; | |
c40f6379 JA |
1190 | } |
1191 | ||
91d8f519 DZ |
1192 | static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx, |
1193 | struct cgroup_subsys_state **cur_css) | |
1194 | ||
1195 | { | |
1196 | #ifdef CONFIG_BLK_CGROUP | |
1197 | /* puts the old one when swapping */ | |
1198 | if (*cur_css != ctx->sqo_blkcg_css) { | |
1199 | kthread_associate_blkcg(ctx->sqo_blkcg_css); | |
1200 | *cur_css = ctx->sqo_blkcg_css; | |
1201 | } | |
1202 | #endif | |
1203 | } | |
1204 | ||
1205 | static void io_sq_thread_unassociate_blkcg(void) | |
1206 | { | |
1207 | #ifdef CONFIG_BLK_CGROUP | |
1208 | kthread_associate_blkcg(NULL); | |
1209 | #endif | |
1210 | } | |
1211 | ||
c40f6379 JA |
1212 | static inline void req_set_fail_links(struct io_kiocb *req) |
1213 | { | |
1214 | if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK) | |
1215 | req->flags |= REQ_F_FAIL_LINK; | |
1216 | } | |
4a38aed2 | 1217 | |
1e6fa521 JA |
1218 | /* |
1219 | * None of these are dereferenced, they are simply used to check if any of | |
1220 | * them have changed. If we're under current and check they are still the | |
1221 | * same, we're fine to grab references to them for actual out-of-line use. | |
1222 | */ | |
1223 | static void io_init_identity(struct io_identity *id) | |
1224 | { | |
1225 | id->files = current->files; | |
1226 | id->mm = current->mm; | |
1227 | #ifdef CONFIG_BLK_CGROUP | |
1228 | rcu_read_lock(); | |
1229 | id->blkcg_css = blkcg_css(); | |
1230 | rcu_read_unlock(); | |
1231 | #endif | |
1232 | id->creds = current_cred(); | |
1233 | id->nsproxy = current->nsproxy; | |
1234 | id->fs = current->fs; | |
1235 | id->fsize = rlimit(RLIMIT_FSIZE); | |
4ea33a97 JA |
1236 | #ifdef CONFIG_AUDIT |
1237 | id->loginuid = current->loginuid; | |
1238 | id->sessionid = current->sessionid; | |
1239 | #endif | |
1e6fa521 JA |
1240 | refcount_set(&id->count, 1); |
1241 | } | |
1242 | ||
ec99ca6c PB |
1243 | static inline void __io_req_init_async(struct io_kiocb *req) |
1244 | { | |
1245 | memset(&req->work, 0, sizeof(req->work)); | |
1246 | req->flags |= REQ_F_WORK_INITIALIZED; | |
1247 | } | |
1248 | ||
7cdaf587 XW |
1249 | /* |
1250 | * Note: must call io_req_init_async() for the first time you | |
1251 | * touch any members of io_wq_work. | |
1252 | */ | |
1253 | static inline void io_req_init_async(struct io_kiocb *req) | |
1254 | { | |
500a373d JA |
1255 | struct io_uring_task *tctx = current->io_uring; |
1256 | ||
7cdaf587 XW |
1257 | if (req->flags & REQ_F_WORK_INITIALIZED) |
1258 | return; | |
1259 | ||
ec99ca6c | 1260 | __io_req_init_async(req); |
500a373d JA |
1261 | |
1262 | /* Grab a ref if this isn't our static identity */ | |
1263 | req->work.identity = tctx->identity; | |
1264 | if (tctx->identity != &tctx->__identity) | |
1265 | refcount_inc(&req->work.identity->count); | |
7cdaf587 XW |
1266 | } |
1267 | ||
0cdaf760 PB |
1268 | static inline bool io_async_submit(struct io_ring_ctx *ctx) |
1269 | { | |
1270 | return ctx->flags & IORING_SETUP_SQPOLL; | |
1271 | } | |
1272 | ||
2b188cc1 JA |
1273 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) |
1274 | { | |
1275 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
1276 | ||
0f158b4c | 1277 | complete(&ctx->ref_comp); |
2b188cc1 JA |
1278 | } |
1279 | ||
8eb7e2d0 PB |
1280 | static inline bool io_is_timeout_noseq(struct io_kiocb *req) |
1281 | { | |
1282 | return !req->timeout.off; | |
1283 | } | |
1284 | ||
2b188cc1 JA |
1285 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) |
1286 | { | |
1287 | struct io_ring_ctx *ctx; | |
78076bb6 | 1288 | int hash_bits; |
2b188cc1 JA |
1289 | |
1290 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
1291 | if (!ctx) | |
1292 | return NULL; | |
1293 | ||
0ddf92e8 JA |
1294 | ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL); |
1295 | if (!ctx->fallback_req) | |
1296 | goto err; | |
1297 | ||
78076bb6 JA |
1298 | /* |
1299 | * Use 5 bits less than the max cq entries, that should give us around | |
1300 | * 32 entries per hash list if totally full and uniformly spread. | |
1301 | */ | |
1302 | hash_bits = ilog2(p->cq_entries); | |
1303 | hash_bits -= 5; | |
1304 | if (hash_bits <= 0) | |
1305 | hash_bits = 1; | |
1306 | ctx->cancel_hash_bits = hash_bits; | |
1307 | ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head), | |
1308 | GFP_KERNEL); | |
1309 | if (!ctx->cancel_hash) | |
1310 | goto err; | |
1311 | __hash_init(ctx->cancel_hash, 1U << hash_bits); | |
1312 | ||
21482896 | 1313 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
206aefde JA |
1314 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) |
1315 | goto err; | |
2b188cc1 JA |
1316 | |
1317 | ctx->flags = p->flags; | |
90554200 | 1318 | init_waitqueue_head(&ctx->sqo_sq_wait); |
69fb2131 | 1319 | INIT_LIST_HEAD(&ctx->sqd_list); |
2b188cc1 | 1320 | init_waitqueue_head(&ctx->cq_wait); |
1d7bb1d5 | 1321 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
0f158b4c JA |
1322 | init_completion(&ctx->ref_comp); |
1323 | init_completion(&ctx->sq_thread_comp); | |
5a2e745d | 1324 | idr_init(&ctx->io_buffer_idr); |
071698e1 | 1325 | idr_init(&ctx->personality_idr); |
2b188cc1 JA |
1326 | mutex_init(&ctx->uring_lock); |
1327 | init_waitqueue_head(&ctx->wait); | |
1328 | spin_lock_init(&ctx->completion_lock); | |
540e32a0 | 1329 | INIT_LIST_HEAD(&ctx->iopoll_list); |
de0617e4 | 1330 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 1331 | INIT_LIST_HEAD(&ctx->timeout_list); |
fcb323cc JA |
1332 | spin_lock_init(&ctx->inflight_lock); |
1333 | INIT_LIST_HEAD(&ctx->inflight_list); | |
4a38aed2 JA |
1334 | INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work); |
1335 | init_llist_head(&ctx->file_put_llist); | |
2b188cc1 | 1336 | return ctx; |
206aefde | 1337 | err: |
0ddf92e8 JA |
1338 | if (ctx->fallback_req) |
1339 | kmem_cache_free(req_cachep, ctx->fallback_req); | |
78076bb6 | 1340 | kfree(ctx->cancel_hash); |
206aefde JA |
1341 | kfree(ctx); |
1342 | return NULL; | |
2b188cc1 JA |
1343 | } |
1344 | ||
9cf7c104 | 1345 | static bool req_need_defer(struct io_kiocb *req, u32 seq) |
7adf4eaf | 1346 | { |
2bc9930e JA |
1347 | if (unlikely(req->flags & REQ_F_IO_DRAIN)) { |
1348 | struct io_ring_ctx *ctx = req->ctx; | |
a197f664 | 1349 | |
9cf7c104 | 1350 | return seq != ctx->cached_cq_tail |
2c3bac6d | 1351 | + READ_ONCE(ctx->cached_cq_overflow); |
2bc9930e | 1352 | } |
de0617e4 | 1353 | |
9d858b21 | 1354 | return false; |
de0617e4 JA |
1355 | } |
1356 | ||
de0617e4 | 1357 | static void __io_commit_cqring(struct io_ring_ctx *ctx) |
2b188cc1 | 1358 | { |
75b28aff | 1359 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 1360 | |
07910158 PB |
1361 | /* order cqe stores with ring update */ |
1362 | smp_store_release(&rings->cq.tail, ctx->cached_cq_tail); | |
2b188cc1 JA |
1363 | } |
1364 | ||
5c3462cf | 1365 | static void io_put_identity(struct io_uring_task *tctx, struct io_kiocb *req) |
1e6fa521 | 1366 | { |
500a373d | 1367 | if (req->work.identity == &tctx->__identity) |
1e6fa521 JA |
1368 | return; |
1369 | if (refcount_dec_and_test(&req->work.identity->count)) | |
1370 | kfree(req->work.identity); | |
1371 | } | |
1372 | ||
4edf20f9 | 1373 | static void io_req_clean_work(struct io_kiocb *req) |
18d9be1a | 1374 | { |
7cdaf587 | 1375 | if (!(req->flags & REQ_F_WORK_INITIALIZED)) |
4edf20f9 | 1376 | return; |
51a4cc11 JA |
1377 | |
1378 | req->flags &= ~REQ_F_WORK_INITIALIZED; | |
7cdaf587 | 1379 | |
dfead8a8 | 1380 | if (req->work.flags & IO_WQ_WORK_MM) { |
98447d65 | 1381 | mmdrop(req->work.identity->mm); |
dfead8a8 | 1382 | req->work.flags &= ~IO_WQ_WORK_MM; |
cccf0ee8 | 1383 | } |
91d8f519 | 1384 | #ifdef CONFIG_BLK_CGROUP |
dfead8a8 | 1385 | if (req->work.flags & IO_WQ_WORK_BLKCG) { |
98447d65 | 1386 | css_put(req->work.identity->blkcg_css); |
dfead8a8 JA |
1387 | req->work.flags &= ~IO_WQ_WORK_BLKCG; |
1388 | } | |
91d8f519 | 1389 | #endif |
dfead8a8 | 1390 | if (req->work.flags & IO_WQ_WORK_CREDS) { |
98447d65 | 1391 | put_cred(req->work.identity->creds); |
dfead8a8 | 1392 | req->work.flags &= ~IO_WQ_WORK_CREDS; |
cccf0ee8 | 1393 | } |
dfead8a8 | 1394 | if (req->work.flags & IO_WQ_WORK_FS) { |
98447d65 | 1395 | struct fs_struct *fs = req->work.identity->fs; |
51a4cc11 | 1396 | |
98447d65 | 1397 | spin_lock(&req->work.identity->fs->lock); |
ff002b30 JA |
1398 | if (--fs->users) |
1399 | fs = NULL; | |
98447d65 | 1400 | spin_unlock(&req->work.identity->fs->lock); |
ff002b30 JA |
1401 | if (fs) |
1402 | free_fs_struct(fs); | |
dfead8a8 | 1403 | req->work.flags &= ~IO_WQ_WORK_FS; |
ff002b30 | 1404 | } |
9d5c8190 PB |
1405 | if (req->flags & REQ_F_INFLIGHT) |
1406 | io_req_drop_files(req); | |
51a4cc11 | 1407 | |
5c3462cf | 1408 | io_put_identity(req->task->io_uring, req); |
561fb04a JA |
1409 | } |
1410 | ||
1e6fa521 JA |
1411 | /* |
1412 | * Create a private copy of io_identity, since some fields don't match | |
1413 | * the current context. | |
1414 | */ | |
1415 | static bool io_identity_cow(struct io_kiocb *req) | |
1416 | { | |
5c3462cf | 1417 | struct io_uring_task *tctx = current->io_uring; |
1e6fa521 JA |
1418 | const struct cred *creds = NULL; |
1419 | struct io_identity *id; | |
1420 | ||
1421 | if (req->work.flags & IO_WQ_WORK_CREDS) | |
1422 | creds = req->work.identity->creds; | |
1423 | ||
1424 | id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL); | |
1425 | if (unlikely(!id)) { | |
1426 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
1427 | return false; | |
1428 | } | |
1429 | ||
1430 | /* | |
1431 | * We can safely just re-init the creds we copied Either the field | |
1432 | * matches the current one, or we haven't grabbed it yet. The only | |
1433 | * exception is ->creds, through registered personalities, so handle | |
1434 | * that one separately. | |
1435 | */ | |
1436 | io_init_identity(id); | |
1437 | if (creds) | |
e8c954df | 1438 | id->creds = creds; |
1e6fa521 JA |
1439 | |
1440 | /* add one for this request */ | |
1441 | refcount_inc(&id->count); | |
1442 | ||
cb8a8ae3 JA |
1443 | /* drop tctx and req identity references, if needed */ |
1444 | if (tctx->identity != &tctx->__identity && | |
1445 | refcount_dec_and_test(&tctx->identity->count)) | |
1446 | kfree(tctx->identity); | |
1447 | if (req->work.identity != &tctx->__identity && | |
1448 | refcount_dec_and_test(&req->work.identity->count)) | |
1e6fa521 JA |
1449 | kfree(req->work.identity); |
1450 | ||
1451 | req->work.identity = id; | |
500a373d | 1452 | tctx->identity = id; |
1e6fa521 JA |
1453 | return true; |
1454 | } | |
1455 | ||
1456 | static bool io_grab_identity(struct io_kiocb *req) | |
18d9be1a | 1457 | { |
d3656344 | 1458 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
5c3462cf | 1459 | struct io_identity *id = req->work.identity; |
23329513 | 1460 | struct io_ring_ctx *ctx = req->ctx; |
54a91f3b | 1461 | |
69228338 JA |
1462 | if (def->work_flags & IO_WQ_WORK_FSIZE) { |
1463 | if (id->fsize != rlimit(RLIMIT_FSIZE)) | |
1464 | return false; | |
1465 | req->work.flags |= IO_WQ_WORK_FSIZE; | |
1466 | } | |
91d8f519 | 1467 | #ifdef CONFIG_BLK_CGROUP |
dfead8a8 JA |
1468 | if (!(req->work.flags & IO_WQ_WORK_BLKCG) && |
1469 | (def->work_flags & IO_WQ_WORK_BLKCG)) { | |
91d8f519 | 1470 | rcu_read_lock(); |
1e6fa521 JA |
1471 | if (id->blkcg_css != blkcg_css()) { |
1472 | rcu_read_unlock(); | |
1473 | return false; | |
1474 | } | |
91d8f519 DZ |
1475 | /* |
1476 | * This should be rare, either the cgroup is dying or the task | |
1477 | * is moving cgroups. Just punt to root for the handful of ios. | |
1478 | */ | |
1e6fa521 | 1479 | if (css_tryget_online(id->blkcg_css)) |
dfead8a8 | 1480 | req->work.flags |= IO_WQ_WORK_BLKCG; |
91d8f519 DZ |
1481 | rcu_read_unlock(); |
1482 | } | |
1483 | #endif | |
dfead8a8 | 1484 | if (!(req->work.flags & IO_WQ_WORK_CREDS)) { |
1e6fa521 JA |
1485 | if (id->creds != current_cred()) |
1486 | return false; | |
1487 | get_cred(id->creds); | |
dfead8a8 JA |
1488 | req->work.flags |= IO_WQ_WORK_CREDS; |
1489 | } | |
4ea33a97 JA |
1490 | #ifdef CONFIG_AUDIT |
1491 | if (!uid_eq(current->loginuid, id->loginuid) || | |
1492 | current->sessionid != id->sessionid) | |
1493 | return false; | |
1494 | #endif | |
dfead8a8 JA |
1495 | if (!(req->work.flags & IO_WQ_WORK_FS) && |
1496 | (def->work_flags & IO_WQ_WORK_FS)) { | |
1e6fa521 JA |
1497 | if (current->fs != id->fs) |
1498 | return false; | |
1499 | spin_lock(&id->fs->lock); | |
1500 | if (!id->fs->in_exec) { | |
1501 | id->fs->users++; | |
dfead8a8 | 1502 | req->work.flags |= IO_WQ_WORK_FS; |
dca9cf8b PB |
1503 | } else { |
1504 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
1505 | } | |
1506 | spin_unlock(¤t->fs->lock); | |
1507 | } | |
af604703 PB |
1508 | if (!(req->work.flags & IO_WQ_WORK_FILES) && |
1509 | (def->work_flags & IO_WQ_WORK_FILES) && | |
1510 | !(req->flags & REQ_F_NO_FILE_TABLE)) { | |
1511 | if (id->files != current->files || | |
1512 | id->nsproxy != current->nsproxy) | |
1513 | return false; | |
1514 | atomic_inc(&id->files->count); | |
1515 | get_nsproxy(id->nsproxy); | |
af604703 | 1516 | |
02a13674 JA |
1517 | if (!(req->flags & REQ_F_INFLIGHT)) { |
1518 | req->flags |= REQ_F_INFLIGHT; | |
1519 | ||
1520 | spin_lock_irq(&ctx->inflight_lock); | |
1521 | list_add(&req->inflight_entry, &ctx->inflight_list); | |
1522 | spin_unlock_irq(&ctx->inflight_lock); | |
1523 | } | |
af604703 PB |
1524 | req->work.flags |= IO_WQ_WORK_FILES; |
1525 | } | |
77788775 JA |
1526 | if (!(req->work.flags & IO_WQ_WORK_MM) && |
1527 | (def->work_flags & IO_WQ_WORK_MM)) { | |
1528 | if (id->mm != current->mm) | |
1529 | return false; | |
1530 | mmgrab(id->mm); | |
1531 | req->work.flags |= IO_WQ_WORK_MM; | |
1532 | } | |
1e6fa521 JA |
1533 | |
1534 | return true; | |
1535 | } | |
1536 | ||
1537 | static void io_prep_async_work(struct io_kiocb *req) | |
1538 | { | |
1539 | const struct io_op_def *def = &io_op_defs[req->opcode]; | |
1e6fa521 JA |
1540 | struct io_ring_ctx *ctx = req->ctx; |
1541 | ||
1542 | io_req_init_async(req); | |
1543 | ||
feaadc4f PB |
1544 | if (req->flags & REQ_F_FORCE_ASYNC) |
1545 | req->work.flags |= IO_WQ_WORK_CONCURRENT; | |
1546 | ||
1e6fa521 JA |
1547 | if (req->flags & REQ_F_ISREG) { |
1548 | if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL)) | |
1549 | io_wq_hash_work(&req->work, file_inode(req->file)); | |
1550 | } else { | |
1551 | if (def->unbound_nonreg_file) | |
1552 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
1553 | } | |
1554 | ||
1e6fa521 JA |
1555 | /* if we fail grabbing identity, we must COW, regrab, and retry */ |
1556 | if (io_grab_identity(req)) | |
1557 | return; | |
1558 | ||
1559 | if (!io_identity_cow(req)) | |
1560 | return; | |
1561 | ||
1562 | /* can't fail at this point */ | |
1563 | if (!io_grab_identity(req)) | |
1564 | WARN_ON(1); | |
561fb04a | 1565 | } |
cccf0ee8 | 1566 | |
cbdcb435 | 1567 | static void io_prep_async_link(struct io_kiocb *req) |
561fb04a | 1568 | { |
cbdcb435 | 1569 | struct io_kiocb *cur; |
54a91f3b | 1570 | |
f2f87370 PB |
1571 | io_for_each_link(cur, req) |
1572 | io_prep_async_work(cur); | |
561fb04a JA |
1573 | } |
1574 | ||
7271ef3a | 1575 | static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req) |
561fb04a | 1576 | { |
a197f664 | 1577 | struct io_ring_ctx *ctx = req->ctx; |
cbdcb435 | 1578 | struct io_kiocb *link = io_prep_linked_timeout(req); |
561fb04a | 1579 | |
8766dd51 PB |
1580 | trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req, |
1581 | &req->work, req->flags); | |
1582 | io_wq_enqueue(ctx->io_wq, &req->work); | |
7271ef3a | 1583 | return link; |
18d9be1a JA |
1584 | } |
1585 | ||
cbdcb435 PB |
1586 | static void io_queue_async_work(struct io_kiocb *req) |
1587 | { | |
7271ef3a JA |
1588 | struct io_kiocb *link; |
1589 | ||
cbdcb435 PB |
1590 | /* init ->work of the whole link before punting */ |
1591 | io_prep_async_link(req); | |
7271ef3a JA |
1592 | link = __io_queue_async_work(req); |
1593 | ||
1594 | if (link) | |
1595 | io_queue_linked_timeout(link); | |
cbdcb435 PB |
1596 | } |
1597 | ||
5262f567 JA |
1598 | static void io_kill_timeout(struct io_kiocb *req) |
1599 | { | |
e8c2bc1f | 1600 | struct io_timeout_data *io = req->async_data; |
5262f567 JA |
1601 | int ret; |
1602 | ||
e8c2bc1f | 1603 | ret = hrtimer_try_to_cancel(&io->timer); |
5262f567 | 1604 | if (ret != -1) { |
01cec8c1 PB |
1605 | atomic_set(&req->ctx->cq_timeouts, |
1606 | atomic_read(&req->ctx->cq_timeouts) + 1); | |
135fcde8 | 1607 | list_del_init(&req->timeout.list); |
78e19bbe | 1608 | io_cqring_fill_event(req, 0); |
216578e5 | 1609 | io_put_req_deferred(req, 1); |
5262f567 JA |
1610 | } |
1611 | } | |
1612 | ||
76e1b642 JA |
1613 | /* |
1614 | * Returns true if we found and killed one or more timeouts | |
1615 | */ | |
6b81928d PB |
1616 | static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk, |
1617 | struct files_struct *files) | |
5262f567 JA |
1618 | { |
1619 | struct io_kiocb *req, *tmp; | |
76e1b642 | 1620 | int canceled = 0; |
5262f567 JA |
1621 | |
1622 | spin_lock_irq(&ctx->completion_lock); | |
f3606e3a | 1623 | list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) { |
6b81928d | 1624 | if (io_match_task(req, tsk, files)) { |
f3606e3a | 1625 | io_kill_timeout(req); |
76e1b642 JA |
1626 | canceled++; |
1627 | } | |
f3606e3a | 1628 | } |
5262f567 | 1629 | spin_unlock_irq(&ctx->completion_lock); |
76e1b642 | 1630 | return canceled != 0; |
5262f567 JA |
1631 | } |
1632 | ||
04518945 | 1633 | static void __io_queue_deferred(struct io_ring_ctx *ctx) |
de0617e4 | 1634 | { |
04518945 | 1635 | do { |
27dc8338 PB |
1636 | struct io_defer_entry *de = list_first_entry(&ctx->defer_list, |
1637 | struct io_defer_entry, list); | |
de0617e4 | 1638 | |
9cf7c104 | 1639 | if (req_need_defer(de->req, de->seq)) |
04518945 | 1640 | break; |
27dc8338 | 1641 | list_del_init(&de->list); |
907d1df3 | 1642 | io_req_task_queue(de->req); |
27dc8338 | 1643 | kfree(de); |
04518945 PB |
1644 | } while (!list_empty(&ctx->defer_list)); |
1645 | } | |
1646 | ||
360428f8 | 1647 | static void io_flush_timeouts(struct io_ring_ctx *ctx) |
de0617e4 | 1648 | { |
f010505b MDG |
1649 | u32 seq; |
1650 | ||
1651 | if (list_empty(&ctx->timeout_list)) | |
1652 | return; | |
1653 | ||
1654 | seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts); | |
1655 | ||
1656 | do { | |
1657 | u32 events_needed, events_got; | |
360428f8 | 1658 | struct io_kiocb *req = list_first_entry(&ctx->timeout_list, |
135fcde8 | 1659 | struct io_kiocb, timeout.list); |
de0617e4 | 1660 | |
8eb7e2d0 | 1661 | if (io_is_timeout_noseq(req)) |
360428f8 | 1662 | break; |
f010505b MDG |
1663 | |
1664 | /* | |
1665 | * Since seq can easily wrap around over time, subtract | |
1666 | * the last seq at which timeouts were flushed before comparing. | |
1667 | * Assuming not more than 2^31-1 events have happened since, | |
1668 | * these subtractions won't have wrapped, so we can check if | |
1669 | * target is in [last_seq, current_seq] by comparing the two. | |
1670 | */ | |
1671 | events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush; | |
1672 | events_got = seq - ctx->cq_last_tm_flush; | |
1673 | if (events_got < events_needed) | |
360428f8 | 1674 | break; |
bfe68a22 | 1675 | |
135fcde8 | 1676 | list_del_init(&req->timeout.list); |
5262f567 | 1677 | io_kill_timeout(req); |
f010505b MDG |
1678 | } while (!list_empty(&ctx->timeout_list)); |
1679 | ||
1680 | ctx->cq_last_tm_flush = seq; | |
360428f8 | 1681 | } |
5262f567 | 1682 | |
360428f8 PB |
1683 | static void io_commit_cqring(struct io_ring_ctx *ctx) |
1684 | { | |
1685 | io_flush_timeouts(ctx); | |
de0617e4 JA |
1686 | __io_commit_cqring(ctx); |
1687 | ||
04518945 PB |
1688 | if (unlikely(!list_empty(&ctx->defer_list))) |
1689 | __io_queue_deferred(ctx); | |
de0617e4 JA |
1690 | } |
1691 | ||
90554200 JA |
1692 | static inline bool io_sqring_full(struct io_ring_ctx *ctx) |
1693 | { | |
1694 | struct io_rings *r = ctx->rings; | |
1695 | ||
1696 | return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries; | |
1697 | } | |
1698 | ||
2b188cc1 JA |
1699 | static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx) |
1700 | { | |
75b28aff | 1701 | struct io_rings *rings = ctx->rings; |
2b188cc1 JA |
1702 | unsigned tail; |
1703 | ||
1704 | tail = ctx->cached_cq_tail; | |
115e12e5 SB |
1705 | /* |
1706 | * writes to the cq entry need to come after reading head; the | |
1707 | * control dependency is enough as we're using WRITE_ONCE to | |
1708 | * fill the cq entry | |
1709 | */ | |
75b28aff | 1710 | if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries) |
2b188cc1 JA |
1711 | return NULL; |
1712 | ||
1713 | ctx->cached_cq_tail++; | |
75b28aff | 1714 | return &rings->cqes[tail & ctx->cq_mask]; |
2b188cc1 JA |
1715 | } |
1716 | ||
f2842ab5 JA |
1717 | static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx) |
1718 | { | |
f0b493e6 JA |
1719 | if (!ctx->cq_ev_fd) |
1720 | return false; | |
7e55a19c SG |
1721 | if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) |
1722 | return false; | |
f2842ab5 JA |
1723 | if (!ctx->eventfd_async) |
1724 | return true; | |
b41e9852 | 1725 | return io_wq_current_is_worker(); |
f2842ab5 JA |
1726 | } |
1727 | ||
e23de15f PB |
1728 | static inline unsigned __io_cqring_events(struct io_ring_ctx *ctx) |
1729 | { | |
1730 | return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head); | |
1731 | } | |
1732 | ||
b41e9852 | 1733 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
1d7bb1d5 | 1734 | { |
b1445e59 PB |
1735 | /* see waitqueue_active() comment */ |
1736 | smp_mb(); | |
1737 | ||
1d7bb1d5 JA |
1738 | if (waitqueue_active(&ctx->wait)) |
1739 | wake_up(&ctx->wait); | |
534ca6d6 JA |
1740 | if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait)) |
1741 | wake_up(&ctx->sq_data->wait); | |
b41e9852 | 1742 | if (io_should_trigger_evfd(ctx)) |
1d7bb1d5 | 1743 | eventfd_signal(ctx->cq_ev_fd, 1); |
b1445e59 | 1744 | if (waitqueue_active(&ctx->cq_wait)) { |
4aa84f2f PB |
1745 | wake_up_interruptible(&ctx->cq_wait); |
1746 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
1747 | } | |
1d7bb1d5 JA |
1748 | } |
1749 | ||
80c18e4a PB |
1750 | static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx) |
1751 | { | |
b1445e59 PB |
1752 | /* see waitqueue_active() comment */ |
1753 | smp_mb(); | |
1754 | ||
80c18e4a PB |
1755 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
1756 | if (waitqueue_active(&ctx->wait)) | |
1757 | wake_up(&ctx->wait); | |
1758 | } | |
1759 | if (io_should_trigger_evfd(ctx)) | |
1760 | eventfd_signal(ctx->cq_ev_fd, 1); | |
b1445e59 | 1761 | if (waitqueue_active(&ctx->cq_wait)) { |
4aa84f2f PB |
1762 | wake_up_interruptible(&ctx->cq_wait); |
1763 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
1764 | } | |
80c18e4a PB |
1765 | } |
1766 | ||
c4a2ed72 | 1767 | /* Returns true if there are no backlogged entries after the flush */ |
6c503150 PB |
1768 | static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force, |
1769 | struct task_struct *tsk, | |
1770 | struct files_struct *files) | |
1d7bb1d5 JA |
1771 | { |
1772 | struct io_rings *rings = ctx->rings; | |
e6c8aa9a | 1773 | struct io_kiocb *req, *tmp; |
1d7bb1d5 | 1774 | struct io_uring_cqe *cqe; |
1d7bb1d5 | 1775 | unsigned long flags; |
b18032bb | 1776 | bool all_flushed, posted; |
1d7bb1d5 JA |
1777 | LIST_HEAD(list); |
1778 | ||
e23de15f PB |
1779 | if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries) |
1780 | return false; | |
1d7bb1d5 | 1781 | |
b18032bb | 1782 | posted = false; |
1d7bb1d5 | 1783 | spin_lock_irqsave(&ctx->completion_lock, flags); |
e6c8aa9a | 1784 | list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) { |
08d23634 | 1785 | if (!io_match_task(req, tsk, files)) |
e6c8aa9a JA |
1786 | continue; |
1787 | ||
1d7bb1d5 JA |
1788 | cqe = io_get_cqring(ctx); |
1789 | if (!cqe && !force) | |
1790 | break; | |
1791 | ||
40d8ddd4 | 1792 | list_move(&req->compl.list, &list); |
1d7bb1d5 JA |
1793 | if (cqe) { |
1794 | WRITE_ONCE(cqe->user_data, req->user_data); | |
1795 | WRITE_ONCE(cqe->res, req->result); | |
0f7e466b | 1796 | WRITE_ONCE(cqe->flags, req->compl.cflags); |
1d7bb1d5 | 1797 | } else { |
2c3bac6d | 1798 | ctx->cached_cq_overflow++; |
1d7bb1d5 | 1799 | WRITE_ONCE(ctx->rings->cq_overflow, |
2c3bac6d | 1800 | ctx->cached_cq_overflow); |
1d7bb1d5 | 1801 | } |
b18032bb | 1802 | posted = true; |
1d7bb1d5 JA |
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 | |
b18032bb JA |
1812 | if (posted) |
1813 | io_commit_cqring(ctx); | |
1d7bb1d5 | 1814 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
b18032bb JA |
1815 | if (posted) |
1816 | io_cqring_ev_posted(ctx); | |
1d7bb1d5 JA |
1817 | |
1818 | while (!list_empty(&list)) { | |
40d8ddd4 PB |
1819 | req = list_first_entry(&list, struct io_kiocb, compl.list); |
1820 | list_del(&req->compl.list); | |
ec9c02ad | 1821 | io_put_req(req); |
1d7bb1d5 | 1822 | } |
c4a2ed72 | 1823 | |
09e88404 | 1824 | return all_flushed; |
1d7bb1d5 JA |
1825 | } |
1826 | ||
6c503150 PB |
1827 | static void io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force, |
1828 | struct task_struct *tsk, | |
1829 | struct files_struct *files) | |
1830 | { | |
1831 | if (test_bit(0, &ctx->cq_check_overflow)) { | |
1832 | /* iopoll syncs against uring_lock, not completion_lock */ | |
1833 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
1834 | mutex_lock(&ctx->uring_lock); | |
1835 | __io_cqring_overflow_flush(ctx, force, tsk, files); | |
1836 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
1837 | mutex_unlock(&ctx->uring_lock); | |
1838 | } | |
1839 | } | |
1840 | ||
bcda7baa | 1841 | static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags) |
2b188cc1 | 1842 | { |
78e19bbe | 1843 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
1844 | struct io_uring_cqe *cqe; |
1845 | ||
78e19bbe | 1846 | trace_io_uring_complete(ctx, req->user_data, res); |
51c3ff62 | 1847 | |
2b188cc1 JA |
1848 | /* |
1849 | * If we can't get a cq entry, userspace overflowed the | |
1850 | * submission (by quite a lot). Increment the overflow count in | |
1851 | * the ring. | |
1852 | */ | |
1853 | cqe = io_get_cqring(ctx); | |
1d7bb1d5 | 1854 | if (likely(cqe)) { |
78e19bbe | 1855 | WRITE_ONCE(cqe->user_data, req->user_data); |
2b188cc1 | 1856 | WRITE_ONCE(cqe->res, res); |
bcda7baa | 1857 | WRITE_ONCE(cqe->flags, cflags); |
fdaf083c JA |
1858 | } else if (ctx->cq_overflow_flushed || |
1859 | atomic_read(&req->task->io_uring->in_idle)) { | |
0f212204 JA |
1860 | /* |
1861 | * If we're in ring overflow flush mode, or in task cancel mode, | |
1862 | * then we cannot store the request for later flushing, we need | |
1863 | * to drop it on the floor. | |
1864 | */ | |
2c3bac6d PB |
1865 | ctx->cached_cq_overflow++; |
1866 | WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow); | |
1d7bb1d5 | 1867 | } else { |
ad3eb2c8 JA |
1868 | if (list_empty(&ctx->cq_overflow_list)) { |
1869 | set_bit(0, &ctx->sq_check_overflow); | |
1870 | set_bit(0, &ctx->cq_check_overflow); | |
6d5f9049 | 1871 | ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW; |
ad3eb2c8 | 1872 | } |
40d8ddd4 | 1873 | io_clean_op(req); |
1d7bb1d5 | 1874 | req->result = res; |
0f7e466b | 1875 | req->compl.cflags = cflags; |
40d8ddd4 PB |
1876 | refcount_inc(&req->refs); |
1877 | list_add_tail(&req->compl.list, &ctx->cq_overflow_list); | |
2b188cc1 JA |
1878 | } |
1879 | } | |
1880 | ||
bcda7baa JA |
1881 | static void io_cqring_fill_event(struct io_kiocb *req, long res) |
1882 | { | |
1883 | __io_cqring_fill_event(req, res, 0); | |
1884 | } | |
1885 | ||
e1e16097 | 1886 | static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags) |
2b188cc1 | 1887 | { |
78e19bbe | 1888 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
1889 | unsigned long flags; |
1890 | ||
1891 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
bcda7baa | 1892 | __io_cqring_fill_event(req, res, cflags); |
2b188cc1 JA |
1893 | io_commit_cqring(ctx); |
1894 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1895 | ||
8c838788 | 1896 | io_cqring_ev_posted(ctx); |
2b188cc1 JA |
1897 | } |
1898 | ||
229a7b63 | 1899 | static void io_submit_flush_completions(struct io_comp_state *cs) |
bcda7baa | 1900 | { |
229a7b63 JA |
1901 | struct io_ring_ctx *ctx = cs->ctx; |
1902 | ||
1903 | spin_lock_irq(&ctx->completion_lock); | |
1904 | while (!list_empty(&cs->list)) { | |
1905 | struct io_kiocb *req; | |
1906 | ||
3ca405eb PB |
1907 | req = list_first_entry(&cs->list, struct io_kiocb, compl.list); |
1908 | list_del(&req->compl.list); | |
0f7e466b | 1909 | __io_cqring_fill_event(req, req->result, req->compl.cflags); |
216578e5 PB |
1910 | |
1911 | /* | |
1912 | * io_free_req() doesn't care about completion_lock unless one | |
1913 | * of these flags is set. REQ_F_WORK_INITIALIZED is in the list | |
1914 | * because of a potential deadlock with req->work.fs->lock | |
1915 | */ | |
1916 | if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT | |
1917 | |REQ_F_WORK_INITIALIZED)) { | |
229a7b63 JA |
1918 | spin_unlock_irq(&ctx->completion_lock); |
1919 | io_put_req(req); | |
1920 | spin_lock_irq(&ctx->completion_lock); | |
216578e5 PB |
1921 | } else { |
1922 | io_put_req(req); | |
229a7b63 JA |
1923 | } |
1924 | } | |
1925 | io_commit_cqring(ctx); | |
1926 | spin_unlock_irq(&ctx->completion_lock); | |
1927 | ||
1928 | io_cqring_ev_posted(ctx); | |
1929 | cs->nr = 0; | |
1930 | } | |
1931 | ||
1932 | static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags, | |
1933 | struct io_comp_state *cs) | |
1934 | { | |
1935 | if (!cs) { | |
1936 | io_cqring_add_event(req, res, cflags); | |
1937 | io_put_req(req); | |
1938 | } else { | |
3ca405eb | 1939 | io_clean_op(req); |
229a7b63 | 1940 | req->result = res; |
0f7e466b | 1941 | req->compl.cflags = cflags; |
3ca405eb | 1942 | list_add_tail(&req->compl.list, &cs->list); |
229a7b63 JA |
1943 | if (++cs->nr >= 32) |
1944 | io_submit_flush_completions(cs); | |
1945 | } | |
e1e16097 JA |
1946 | } |
1947 | ||
1948 | static void io_req_complete(struct io_kiocb *req, long res) | |
bcda7baa | 1949 | { |
229a7b63 | 1950 | __io_req_complete(req, res, 0, NULL); |
bcda7baa JA |
1951 | } |
1952 | ||
0ddf92e8 JA |
1953 | static inline bool io_is_fallback_req(struct io_kiocb *req) |
1954 | { | |
1955 | return req == (struct io_kiocb *) | |
1956 | ((unsigned long) req->ctx->fallback_req & ~1UL); | |
1957 | } | |
1958 | ||
1959 | static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx) | |
1960 | { | |
1961 | struct io_kiocb *req; | |
1962 | ||
1963 | req = ctx->fallback_req; | |
dd461af6 | 1964 | if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req)) |
0ddf92e8 JA |
1965 | return req; |
1966 | ||
1967 | return NULL; | |
1968 | } | |
1969 | ||
0553b8bd PB |
1970 | static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx, |
1971 | struct io_submit_state *state) | |
2b188cc1 | 1972 | { |
f6b6c7d6 | 1973 | if (!state->free_reqs) { |
291b2821 | 1974 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
2579f913 JA |
1975 | size_t sz; |
1976 | int ret; | |
1977 | ||
1978 | sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs)); | |
fd6fab2c JA |
1979 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs); |
1980 | ||
1981 | /* | |
1982 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
1983 | * retry single alloc to be on the safe side. | |
1984 | */ | |
1985 | if (unlikely(ret <= 0)) { | |
1986 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
1987 | if (!state->reqs[0]) | |
0ddf92e8 | 1988 | goto fallback; |
fd6fab2c JA |
1989 | ret = 1; |
1990 | } | |
291b2821 | 1991 | state->free_reqs = ret; |
2b188cc1 JA |
1992 | } |
1993 | ||
291b2821 PB |
1994 | state->free_reqs--; |
1995 | return state->reqs[state->free_reqs]; | |
0ddf92e8 | 1996 | fallback: |
0553b8bd | 1997 | return io_get_fallback_req(ctx); |
2b188cc1 JA |
1998 | } |
1999 | ||
8da11c19 PB |
2000 | static inline void io_put_file(struct io_kiocb *req, struct file *file, |
2001 | bool fixed) | |
2002 | { | |
36f72fe2 | 2003 | if (!fixed) |
8da11c19 PB |
2004 | fput(file); |
2005 | } | |
2006 | ||
4edf20f9 | 2007 | static void io_dismantle_req(struct io_kiocb *req) |
2b188cc1 | 2008 | { |
3ca405eb | 2009 | io_clean_op(req); |
929a3af9 | 2010 | |
e8c2bc1f JA |
2011 | if (req->async_data) |
2012 | kfree(req->async_data); | |
8da11c19 PB |
2013 | if (req->file) |
2014 | io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE)); | |
36f72fe2 PB |
2015 | if (req->fixed_file_refs) |
2016 | percpu_ref_put(req->fixed_file_refs); | |
4edf20f9 | 2017 | io_req_clean_work(req); |
e65ef56d JA |
2018 | } |
2019 | ||
216578e5 | 2020 | static void __io_free_req(struct io_kiocb *req) |
c6ca97b3 | 2021 | { |
0f212204 | 2022 | struct io_uring_task *tctx = req->task->io_uring; |
51a4cc11 | 2023 | struct io_ring_ctx *ctx = req->ctx; |
c6ca97b3 | 2024 | |
216578e5 | 2025 | io_dismantle_req(req); |
c6ca97b3 | 2026 | |
d8a6df10 | 2027 | percpu_counter_dec(&tctx->inflight); |
fdaf083c | 2028 | if (atomic_read(&tctx->in_idle)) |
0f212204 | 2029 | wake_up(&tctx->wait); |
e3bc8e9d JA |
2030 | put_task_struct(req->task); |
2031 | ||
b1e50e54 PB |
2032 | if (likely(!io_is_fallback_req(req))) |
2033 | kmem_cache_free(req_cachep, req); | |
2034 | else | |
ecfc5177 PB |
2035 | clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req); |
2036 | percpu_ref_put(&ctx->refs); | |
e65ef56d JA |
2037 | } |
2038 | ||
f2f87370 PB |
2039 | static inline void io_remove_next_linked(struct io_kiocb *req) |
2040 | { | |
2041 | struct io_kiocb *nxt = req->link; | |
2042 | ||
2043 | req->link = nxt->link; | |
2044 | nxt->link = NULL; | |
2045 | } | |
2046 | ||
c9abd7ad | 2047 | static void io_kill_linked_timeout(struct io_kiocb *req) |
2665abfd | 2048 | { |
a197f664 | 2049 | struct io_ring_ctx *ctx = req->ctx; |
7c86ffee | 2050 | struct io_kiocb *link; |
c9abd7ad PB |
2051 | bool cancelled = false; |
2052 | unsigned long flags; | |
7c86ffee | 2053 | |
c9abd7ad | 2054 | spin_lock_irqsave(&ctx->completion_lock, flags); |
f2f87370 PB |
2055 | link = req->link; |
2056 | ||
900fad45 PB |
2057 | /* |
2058 | * Can happen if a linked timeout fired and link had been like | |
2059 | * req -> link t-out -> link t-out [-> ...] | |
2060 | */ | |
c9abd7ad PB |
2061 | if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) { |
2062 | struct io_timeout_data *io = link->async_data; | |
2063 | int ret; | |
7c86ffee | 2064 | |
f2f87370 | 2065 | io_remove_next_linked(req); |
90cd7e42 | 2066 | link->timeout.head = NULL; |
c9abd7ad PB |
2067 | ret = hrtimer_try_to_cancel(&io->timer); |
2068 | if (ret != -1) { | |
2069 | io_cqring_fill_event(link, -ECANCELED); | |
2070 | io_commit_cqring(ctx); | |
2071 | cancelled = true; | |
2072 | } | |
2073 | } | |
7c86ffee | 2074 | req->flags &= ~REQ_F_LINK_TIMEOUT; |
216578e5 | 2075 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
ab0b6451 | 2076 | |
c9abd7ad | 2077 | if (cancelled) { |
7c86ffee | 2078 | io_cqring_ev_posted(ctx); |
c9abd7ad PB |
2079 | io_put_req(link); |
2080 | } | |
7c86ffee PB |
2081 | } |
2082 | ||
9e645e11 | 2083 | |
d148ca4b | 2084 | static void io_fail_links(struct io_kiocb *req) |
9e645e11 | 2085 | { |
f2f87370 | 2086 | struct io_kiocb *link, *nxt; |
2665abfd | 2087 | struct io_ring_ctx *ctx = req->ctx; |
d148ca4b | 2088 | unsigned long flags; |
9e645e11 | 2089 | |
d148ca4b | 2090 | spin_lock_irqsave(&ctx->completion_lock, flags); |
f2f87370 PB |
2091 | link = req->link; |
2092 | req->link = NULL; | |
9e645e11 | 2093 | |
f2f87370 PB |
2094 | while (link) { |
2095 | nxt = link->link; | |
2096 | link->link = NULL; | |
2665abfd | 2097 | |
f2f87370 | 2098 | trace_io_uring_fail_link(req, link); |
7c86ffee | 2099 | io_cqring_fill_event(link, -ECANCELED); |
216578e5 PB |
2100 | |
2101 | /* | |
2102 | * It's ok to free under spinlock as they're not linked anymore, | |
2103 | * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on | |
2104 | * work.fs->lock. | |
2105 | */ | |
2106 | if (link->flags & REQ_F_WORK_INITIALIZED) | |
2107 | io_put_req_deferred(link, 2); | |
2108 | else | |
2109 | io_double_put_req(link); | |
f2f87370 | 2110 | link = nxt; |
9e645e11 | 2111 | } |
2665abfd | 2112 | io_commit_cqring(ctx); |
216578e5 | 2113 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
9e645e11 | 2114 | |
2665abfd | 2115 | io_cqring_ev_posted(ctx); |
9e645e11 JA |
2116 | } |
2117 | ||
3fa5e0f3 | 2118 | static struct io_kiocb *__io_req_find_next(struct io_kiocb *req) |
c69f8dbe | 2119 | { |
7c86ffee PB |
2120 | if (req->flags & REQ_F_LINK_TIMEOUT) |
2121 | io_kill_linked_timeout(req); | |
944e58bf | 2122 | |
9e645e11 JA |
2123 | /* |
2124 | * If LINK is set, we have dependent requests in this chain. If we | |
2125 | * didn't fail this request, queue the first one up, moving any other | |
2126 | * dependencies to the next request. In case of failure, fail the rest | |
2127 | * of the chain. | |
2128 | */ | |
f2f87370 PB |
2129 | if (likely(!(req->flags & REQ_F_FAIL_LINK))) { |
2130 | struct io_kiocb *nxt = req->link; | |
2131 | ||
2132 | req->link = NULL; | |
2133 | return nxt; | |
2134 | } | |
9b5f7bd9 PB |
2135 | io_fail_links(req); |
2136 | return NULL; | |
4d7dd462 | 2137 | } |
9e645e11 | 2138 | |
f2f87370 | 2139 | static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req) |
3fa5e0f3 | 2140 | { |
f2f87370 | 2141 | if (likely(!(req->link) && !(req->flags & REQ_F_LINK_TIMEOUT))) |
3fa5e0f3 PB |
2142 | return NULL; |
2143 | return __io_req_find_next(req); | |
2144 | } | |
2145 | ||
355fb9e2 | 2146 | static int io_req_task_work_add(struct io_kiocb *req) |
c2c4c83c JA |
2147 | { |
2148 | struct task_struct *tsk = req->task; | |
2149 | struct io_ring_ctx *ctx = req->ctx; | |
91989c70 JA |
2150 | enum task_work_notify_mode notify; |
2151 | int ret; | |
c2c4c83c | 2152 | |
6200b0ae JA |
2153 | if (tsk->flags & PF_EXITING) |
2154 | return -ESRCH; | |
2155 | ||
c2c4c83c | 2156 | /* |
0ba9c9ed JA |
2157 | * SQPOLL kernel thread doesn't need notification, just a wakeup. For |
2158 | * all other cases, use TWA_SIGNAL unconditionally to ensure we're | |
2159 | * processing task_work. There's no reliable way to tell if TWA_RESUME | |
2160 | * will do the job. | |
c2c4c83c | 2161 | */ |
91989c70 | 2162 | notify = TWA_NONE; |
355fb9e2 | 2163 | if (!(ctx->flags & IORING_SETUP_SQPOLL)) |
c2c4c83c JA |
2164 | notify = TWA_SIGNAL; |
2165 | ||
87c4311f | 2166 | ret = task_work_add(tsk, &req->task_work, notify); |
c2c4c83c JA |
2167 | if (!ret) |
2168 | wake_up_process(tsk); | |
0ba9c9ed | 2169 | |
c2c4c83c JA |
2170 | return ret; |
2171 | } | |
2172 | ||
c40f6379 JA |
2173 | static void __io_req_task_cancel(struct io_kiocb *req, int error) |
2174 | { | |
2175 | struct io_ring_ctx *ctx = req->ctx; | |
2176 | ||
2177 | spin_lock_irq(&ctx->completion_lock); | |
2178 | io_cqring_fill_event(req, error); | |
2179 | io_commit_cqring(ctx); | |
2180 | spin_unlock_irq(&ctx->completion_lock); | |
2181 | ||
2182 | io_cqring_ev_posted(ctx); | |
2183 | req_set_fail_links(req); | |
2184 | io_double_put_req(req); | |
2185 | } | |
2186 | ||
2187 | static void io_req_task_cancel(struct callback_head *cb) | |
2188 | { | |
2189 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); | |
87ceb6a6 | 2190 | struct io_ring_ctx *ctx = req->ctx; |
c40f6379 JA |
2191 | |
2192 | __io_req_task_cancel(req, -ECANCELED); | |
87ceb6a6 | 2193 | percpu_ref_put(&ctx->refs); |
c40f6379 JA |
2194 | } |
2195 | ||
2196 | static void __io_req_task_submit(struct io_kiocb *req) | |
2197 | { | |
2198 | struct io_ring_ctx *ctx = req->ctx; | |
2199 | ||
81b6d05c | 2200 | mutex_lock(&ctx->uring_lock); |
d9d05217 PB |
2201 | if (!ctx->sqo_dead && |
2202 | !__io_sq_thread_acquire_mm(ctx) && | |
2203 | !__io_sq_thread_acquire_files(ctx)) | |
c1379e24 | 2204 | __io_queue_sqe(req, NULL); |
81b6d05c | 2205 | else |
c40f6379 | 2206 | __io_req_task_cancel(req, -EFAULT); |
81b6d05c | 2207 | mutex_unlock(&ctx->uring_lock); |
c40f6379 JA |
2208 | } |
2209 | ||
2210 | static void io_req_task_submit(struct callback_head *cb) | |
2211 | { | |
2212 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); | |
6d816e08 | 2213 | struct io_ring_ctx *ctx = req->ctx; |
c40f6379 JA |
2214 | |
2215 | __io_req_task_submit(req); | |
6d816e08 | 2216 | percpu_ref_put(&ctx->refs); |
c40f6379 JA |
2217 | } |
2218 | ||
2219 | static void io_req_task_queue(struct io_kiocb *req) | |
2220 | { | |
c40f6379 JA |
2221 | int ret; |
2222 | ||
2223 | init_task_work(&req->task_work, io_req_task_submit); | |
6d816e08 | 2224 | percpu_ref_get(&req->ctx->refs); |
c40f6379 | 2225 | |
355fb9e2 | 2226 | ret = io_req_task_work_add(req); |
c40f6379 | 2227 | if (unlikely(ret)) { |
c2c4c83c JA |
2228 | struct task_struct *tsk; |
2229 | ||
c40f6379 JA |
2230 | init_task_work(&req->task_work, io_req_task_cancel); |
2231 | tsk = io_wq_get_task(req->ctx->io_wq); | |
91989c70 | 2232 | task_work_add(tsk, &req->task_work, TWA_NONE); |
c2c4c83c | 2233 | wake_up_process(tsk); |
c40f6379 | 2234 | } |
c40f6379 JA |
2235 | } |
2236 | ||
f2f87370 | 2237 | static inline void io_queue_next(struct io_kiocb *req) |
c69f8dbe | 2238 | { |
9b5f7bd9 | 2239 | struct io_kiocb *nxt = io_req_find_next(req); |
944e58bf PB |
2240 | |
2241 | if (nxt) | |
906a8c3f | 2242 | io_req_task_queue(nxt); |
c69f8dbe JL |
2243 | } |
2244 | ||
c3524383 | 2245 | static void io_free_req(struct io_kiocb *req) |
7a743e22 | 2246 | { |
c3524383 PB |
2247 | io_queue_next(req); |
2248 | __io_free_req(req); | |
2249 | } | |
8766dd51 | 2250 | |
2d6500d4 PB |
2251 | struct req_batch { |
2252 | void *reqs[IO_IOPOLL_BATCH]; | |
2253 | int to_free; | |
7a743e22 | 2254 | |
5af1d13e PB |
2255 | struct task_struct *task; |
2256 | int task_refs; | |
2d6500d4 PB |
2257 | }; |
2258 | ||
5af1d13e PB |
2259 | static inline void io_init_req_batch(struct req_batch *rb) |
2260 | { | |
2261 | rb->to_free = 0; | |
2262 | rb->task_refs = 0; | |
2263 | rb->task = NULL; | |
2264 | } | |
2265 | ||
2d6500d4 PB |
2266 | static void __io_req_free_batch_flush(struct io_ring_ctx *ctx, |
2267 | struct req_batch *rb) | |
2268 | { | |
2269 | kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs); | |
2270 | percpu_ref_put_many(&ctx->refs, rb->to_free); | |
2271 | rb->to_free = 0; | |
2272 | } | |
2273 | ||
2274 | static void io_req_free_batch_finish(struct io_ring_ctx *ctx, | |
2275 | struct req_batch *rb) | |
2276 | { | |
2277 | if (rb->to_free) | |
2278 | __io_req_free_batch_flush(ctx, rb); | |
5af1d13e | 2279 | if (rb->task) { |
d8a6df10 JA |
2280 | struct io_uring_task *tctx = rb->task->io_uring; |
2281 | ||
2282 | percpu_counter_sub(&tctx->inflight, rb->task_refs); | |
c93cc9e1 JA |
2283 | if (atomic_read(&tctx->in_idle)) |
2284 | wake_up(&tctx->wait); | |
5af1d13e PB |
2285 | put_task_struct_many(rb->task, rb->task_refs); |
2286 | rb->task = NULL; | |
2287 | } | |
2d6500d4 PB |
2288 | } |
2289 | ||
2290 | static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req) | |
2291 | { | |
2292 | if (unlikely(io_is_fallback_req(req))) { | |
2293 | io_free_req(req); | |
2294 | return; | |
2295 | } | |
f2f87370 | 2296 | io_queue_next(req); |
2d6500d4 | 2297 | |
e3bc8e9d | 2298 | if (req->task != rb->task) { |
0f212204 | 2299 | if (rb->task) { |
d8a6df10 JA |
2300 | struct io_uring_task *tctx = rb->task->io_uring; |
2301 | ||
2302 | percpu_counter_sub(&tctx->inflight, rb->task_refs); | |
c93cc9e1 JA |
2303 | if (atomic_read(&tctx->in_idle)) |
2304 | wake_up(&tctx->wait); | |
e3bc8e9d | 2305 | put_task_struct_many(rb->task, rb->task_refs); |
5af1d13e | 2306 | } |
e3bc8e9d JA |
2307 | rb->task = req->task; |
2308 | rb->task_refs = 0; | |
5af1d13e | 2309 | } |
e3bc8e9d | 2310 | rb->task_refs++; |
5af1d13e | 2311 | |
4edf20f9 | 2312 | io_dismantle_req(req); |
2d6500d4 PB |
2313 | rb->reqs[rb->to_free++] = req; |
2314 | if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs))) | |
2315 | __io_req_free_batch_flush(req->ctx, rb); | |
7a743e22 PB |
2316 | } |
2317 | ||
ba816ad6 JA |
2318 | /* |
2319 | * Drop reference to request, return next in chain (if there is one) if this | |
2320 | * was the last reference to this request. | |
2321 | */ | |
9b5f7bd9 | 2322 | static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req) |
e65ef56d | 2323 | { |
9b5f7bd9 PB |
2324 | struct io_kiocb *nxt = NULL; |
2325 | ||
2a44f467 | 2326 | if (refcount_dec_and_test(&req->refs)) { |
9b5f7bd9 | 2327 | nxt = io_req_find_next(req); |
4d7dd462 | 2328 | __io_free_req(req); |
2a44f467 | 2329 | } |
9b5f7bd9 | 2330 | return nxt; |
2b188cc1 JA |
2331 | } |
2332 | ||
e65ef56d JA |
2333 | static void io_put_req(struct io_kiocb *req) |
2334 | { | |
2335 | if (refcount_dec_and_test(&req->refs)) | |
2336 | io_free_req(req); | |
2b188cc1 JA |
2337 | } |
2338 | ||
216578e5 PB |
2339 | static void io_put_req_deferred_cb(struct callback_head *cb) |
2340 | { | |
2341 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); | |
2342 | ||
2343 | io_free_req(req); | |
2344 | } | |
2345 | ||
2346 | static void io_free_req_deferred(struct io_kiocb *req) | |
2347 | { | |
2348 | int ret; | |
2349 | ||
2350 | init_task_work(&req->task_work, io_put_req_deferred_cb); | |
355fb9e2 | 2351 | ret = io_req_task_work_add(req); |
216578e5 PB |
2352 | if (unlikely(ret)) { |
2353 | struct task_struct *tsk; | |
2354 | ||
2355 | tsk = io_wq_get_task(req->ctx->io_wq); | |
91989c70 | 2356 | task_work_add(tsk, &req->task_work, TWA_NONE); |
216578e5 PB |
2357 | wake_up_process(tsk); |
2358 | } | |
2359 | } | |
2360 | ||
2361 | static inline void io_put_req_deferred(struct io_kiocb *req, int refs) | |
2362 | { | |
2363 | if (refcount_sub_and_test(refs, &req->refs)) | |
2364 | io_free_req_deferred(req); | |
2365 | } | |
2366 | ||
f4db7182 | 2367 | static struct io_wq_work *io_steal_work(struct io_kiocb *req) |
7a743e22 | 2368 | { |
6df1db6b | 2369 | struct io_kiocb *nxt; |
f4db7182 | 2370 | |
7a743e22 | 2371 | /* |
f4db7182 PB |
2372 | * A ref is owned by io-wq in which context we're. So, if that's the |
2373 | * last one, it's safe to steal next work. False negatives are Ok, | |
2374 | * it just will be re-punted async in io_put_work() | |
7a743e22 | 2375 | */ |
f4db7182 PB |
2376 | if (refcount_read(&req->refs) != 1) |
2377 | return NULL; | |
7a743e22 | 2378 | |
9b5f7bd9 | 2379 | nxt = io_req_find_next(req); |
6df1db6b | 2380 | return nxt ? &nxt->work : NULL; |
7a743e22 PB |
2381 | } |
2382 | ||
978db57e JA |
2383 | static void io_double_put_req(struct io_kiocb *req) |
2384 | { | |
2385 | /* drop both submit and complete references */ | |
2386 | if (refcount_sub_and_test(2, &req->refs)) | |
2387 | io_free_req(req); | |
2388 | } | |
2389 | ||
6c503150 | 2390 | static unsigned io_cqring_events(struct io_ring_ctx *ctx) |
a3a0e43f JA |
2391 | { |
2392 | /* See comment at the top of this file */ | |
2393 | smp_rmb(); | |
e23de15f | 2394 | return __io_cqring_events(ctx); |
a3a0e43f JA |
2395 | } |
2396 | ||
fb5ccc98 PB |
2397 | static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) |
2398 | { | |
2399 | struct io_rings *rings = ctx->rings; | |
2400 | ||
2401 | /* make sure SQ entry isn't read before tail */ | |
2402 | return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; | |
2403 | } | |
2404 | ||
8ff069bf | 2405 | static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf) |
e94f141b | 2406 | { |
8ff069bf | 2407 | unsigned int cflags; |
e94f141b | 2408 | |
bcda7baa JA |
2409 | cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT; |
2410 | cflags |= IORING_CQE_F_BUFFER; | |
0e1b6fe3 | 2411 | req->flags &= ~REQ_F_BUFFER_SELECTED; |
bcda7baa JA |
2412 | kfree(kbuf); |
2413 | return cflags; | |
e94f141b JA |
2414 | } |
2415 | ||
8ff069bf | 2416 | static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req) |
bcda7baa | 2417 | { |
4d954c25 | 2418 | struct io_buffer *kbuf; |
bcda7baa | 2419 | |
4d954c25 | 2420 | kbuf = (struct io_buffer *) (unsigned long) req->rw.addr; |
8ff069bf PB |
2421 | return io_put_kbuf(req, kbuf); |
2422 | } | |
2423 | ||
4c6e277c JA |
2424 | static inline bool io_run_task_work(void) |
2425 | { | |
6200b0ae JA |
2426 | /* |
2427 | * Not safe to run on exiting task, and the task_work handling will | |
2428 | * not add work to such a task. | |
2429 | */ | |
2430 | if (unlikely(current->flags & PF_EXITING)) | |
2431 | return false; | |
4c6e277c JA |
2432 | if (current->task_works) { |
2433 | __set_current_state(TASK_RUNNING); | |
2434 | task_work_run(); | |
2435 | return true; | |
2436 | } | |
2437 | ||
2438 | return false; | |
bcda7baa JA |
2439 | } |
2440 | ||
bbde017a XW |
2441 | static void io_iopoll_queue(struct list_head *again) |
2442 | { | |
2443 | struct io_kiocb *req; | |
2444 | ||
2445 | do { | |
d21ffe7e PB |
2446 | req = list_first_entry(again, struct io_kiocb, inflight_entry); |
2447 | list_del(&req->inflight_entry); | |
81b68a5c | 2448 | __io_complete_rw(req, -EAGAIN, 0, NULL); |
bbde017a XW |
2449 | } while (!list_empty(again)); |
2450 | } | |
2451 | ||
def596e9 JA |
2452 | /* |
2453 | * Find and free completed poll iocbs | |
2454 | */ | |
2455 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
2456 | struct list_head *done) | |
2457 | { | |
8237e045 | 2458 | struct req_batch rb; |
def596e9 | 2459 | struct io_kiocb *req; |
bbde017a XW |
2460 | LIST_HEAD(again); |
2461 | ||
2462 | /* order with ->result store in io_complete_rw_iopoll() */ | |
2463 | smp_rmb(); | |
def596e9 | 2464 | |
5af1d13e | 2465 | io_init_req_batch(&rb); |
def596e9 | 2466 | while (!list_empty(done)) { |
bcda7baa JA |
2467 | int cflags = 0; |
2468 | ||
d21ffe7e | 2469 | req = list_first_entry(done, struct io_kiocb, inflight_entry); |
bbde017a | 2470 | if (READ_ONCE(req->result) == -EAGAIN) { |
56450c20 | 2471 | req->result = 0; |
bbde017a | 2472 | req->iopoll_completed = 0; |
d21ffe7e | 2473 | list_move_tail(&req->inflight_entry, &again); |
bbde017a XW |
2474 | continue; |
2475 | } | |
d21ffe7e | 2476 | list_del(&req->inflight_entry); |
def596e9 | 2477 | |
bcda7baa | 2478 | if (req->flags & REQ_F_BUFFER_SELECTED) |
8ff069bf | 2479 | cflags = io_put_rw_kbuf(req); |
bcda7baa JA |
2480 | |
2481 | __io_cqring_fill_event(req, req->result, cflags); | |
def596e9 JA |
2482 | (*nr_events)++; |
2483 | ||
c3524383 | 2484 | if (refcount_dec_and_test(&req->refs)) |
2d6500d4 | 2485 | io_req_free_batch(&rb, req); |
def596e9 | 2486 | } |
def596e9 | 2487 | |
09bb8394 | 2488 | io_commit_cqring(ctx); |
80c18e4a | 2489 | io_cqring_ev_posted_iopoll(ctx); |
2d6500d4 | 2490 | io_req_free_batch_finish(ctx, &rb); |
581f9810 | 2491 | |
bbde017a XW |
2492 | if (!list_empty(&again)) |
2493 | io_iopoll_queue(&again); | |
581f9810 BM |
2494 | } |
2495 | ||
def596e9 JA |
2496 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, |
2497 | long min) | |
2498 | { | |
2499 | struct io_kiocb *req, *tmp; | |
2500 | LIST_HEAD(done); | |
2501 | bool spin; | |
2502 | int ret; | |
2503 | ||
2504 | /* | |
2505 | * Only spin for completions if we don't have multiple devices hanging | |
2506 | * off our complete list, and we're under the requested amount. | |
2507 | */ | |
2508 | spin = !ctx->poll_multi_file && *nr_events < min; | |
2509 | ||
2510 | ret = 0; | |
d21ffe7e | 2511 | list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) { |
9adbd45d | 2512 | struct kiocb *kiocb = &req->rw.kiocb; |
def596e9 JA |
2513 | |
2514 | /* | |
581f9810 BM |
2515 | * Move completed and retryable entries to our local lists. |
2516 | * If we find a request that requires polling, break out | |
2517 | * and complete those lists first, if we have entries there. | |
def596e9 | 2518 | */ |
65a6543d | 2519 | if (READ_ONCE(req->iopoll_completed)) { |
d21ffe7e | 2520 | list_move_tail(&req->inflight_entry, &done); |
def596e9 JA |
2521 | continue; |
2522 | } | |
2523 | if (!list_empty(&done)) | |
2524 | break; | |
2525 | ||
2526 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
2527 | if (ret < 0) | |
2528 | break; | |
2529 | ||
3aadc23e PB |
2530 | /* iopoll may have completed current req */ |
2531 | if (READ_ONCE(req->iopoll_completed)) | |
d21ffe7e | 2532 | list_move_tail(&req->inflight_entry, &done); |
3aadc23e | 2533 | |
def596e9 JA |
2534 | if (ret && spin) |
2535 | spin = false; | |
2536 | ret = 0; | |
2537 | } | |
2538 | ||
2539 | if (!list_empty(&done)) | |
2540 | io_iopoll_complete(ctx, nr_events, &done); | |
2541 | ||
2542 | return ret; | |
2543 | } | |
2544 | ||
2545 | /* | |
d195a66e | 2546 | * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a |
def596e9 JA |
2547 | * non-spinning poll check - we'll still enter the driver poll loop, but only |
2548 | * as a non-spinning completion check. | |
2549 | */ | |
2550 | static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
2551 | long min) | |
2552 | { | |
540e32a0 | 2553 | while (!list_empty(&ctx->iopoll_list) && !need_resched()) { |
def596e9 JA |
2554 | int ret; |
2555 | ||
2556 | ret = io_do_iopoll(ctx, nr_events, min); | |
2557 | if (ret < 0) | |
2558 | return ret; | |
eba0a4dd | 2559 | if (*nr_events >= min) |
def596e9 JA |
2560 | return 0; |
2561 | } | |
2562 | ||
2563 | return 1; | |
2564 | } | |
2565 | ||
2566 | /* | |
2567 | * We can't just wait for polled events to come to us, we have to actively | |
2568 | * find and complete them. | |
2569 | */ | |
b2edc0a7 | 2570 | static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) |
def596e9 JA |
2571 | { |
2572 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
2573 | return; | |
2574 | ||
2575 | mutex_lock(&ctx->uring_lock); | |
540e32a0 | 2576 | while (!list_empty(&ctx->iopoll_list)) { |
def596e9 JA |
2577 | unsigned int nr_events = 0; |
2578 | ||
b2edc0a7 | 2579 | io_do_iopoll(ctx, &nr_events, 0); |
08f5439f | 2580 | |
b2edc0a7 PB |
2581 | /* let it sleep and repeat later if can't complete a request */ |
2582 | if (nr_events == 0) | |
2583 | break; | |
08f5439f JA |
2584 | /* |
2585 | * Ensure we allow local-to-the-cpu processing to take place, | |
2586 | * in this case we need to ensure that we reap all events. | |
3fcee5a6 | 2587 | * Also let task_work, etc. to progress by releasing the mutex |
08f5439f | 2588 | */ |
3fcee5a6 PB |
2589 | if (need_resched()) { |
2590 | mutex_unlock(&ctx->uring_lock); | |
2591 | cond_resched(); | |
2592 | mutex_lock(&ctx->uring_lock); | |
2593 | } | |
def596e9 JA |
2594 | } |
2595 | mutex_unlock(&ctx->uring_lock); | |
2596 | } | |
2597 | ||
7668b92a | 2598 | static int io_iopoll_check(struct io_ring_ctx *ctx, long min) |
def596e9 | 2599 | { |
7668b92a | 2600 | unsigned int nr_events = 0; |
2b2ed975 | 2601 | int iters = 0, ret = 0; |
500f9fba | 2602 | |
c7849be9 XW |
2603 | /* |
2604 | * We disallow the app entering submit/complete with polling, but we | |
2605 | * still need to lock the ring to prevent racing with polled issue | |
2606 | * that got punted to a workqueue. | |
2607 | */ | |
2608 | mutex_lock(&ctx->uring_lock); | |
def596e9 | 2609 | do { |
a3a0e43f JA |
2610 | /* |
2611 | * Don't enter poll loop if we already have events pending. | |
2612 | * If we do, we can potentially be spinning for commands that | |
2613 | * already triggered a CQE (eg in error). | |
2614 | */ | |
6c503150 PB |
2615 | if (test_bit(0, &ctx->cq_check_overflow)) |
2616 | __io_cqring_overflow_flush(ctx, false, NULL, NULL); | |
2617 | if (io_cqring_events(ctx)) | |
a3a0e43f JA |
2618 | break; |
2619 | ||
500f9fba JA |
2620 | /* |
2621 | * If a submit got punted to a workqueue, we can have the | |
2622 | * application entering polling for a command before it gets | |
2623 | * issued. That app will hold the uring_lock for the duration | |
2624 | * of the poll right here, so we need to take a breather every | |
2625 | * now and then to ensure that the issue has a chance to add | |
2626 | * the poll to the issued list. Otherwise we can spin here | |
2627 | * forever, while the workqueue is stuck trying to acquire the | |
2628 | * very same mutex. | |
2629 | */ | |
2630 | if (!(++iters & 7)) { | |
2631 | mutex_unlock(&ctx->uring_lock); | |
4c6e277c | 2632 | io_run_task_work(); |
500f9fba JA |
2633 | mutex_lock(&ctx->uring_lock); |
2634 | } | |
2635 | ||
7668b92a | 2636 | ret = io_iopoll_getevents(ctx, &nr_events, min); |
def596e9 JA |
2637 | if (ret <= 0) |
2638 | break; | |
2639 | ret = 0; | |
7668b92a | 2640 | } while (min && !nr_events && !need_resched()); |
def596e9 | 2641 | |
500f9fba | 2642 | mutex_unlock(&ctx->uring_lock); |
def596e9 JA |
2643 | return ret; |
2644 | } | |
2645 | ||
491381ce | 2646 | static void kiocb_end_write(struct io_kiocb *req) |
2b188cc1 | 2647 | { |
491381ce JA |
2648 | /* |
2649 | * Tell lockdep we inherited freeze protection from submission | |
2650 | * thread. | |
2651 | */ | |
2652 | if (req->flags & REQ_F_ISREG) { | |
2653 | struct inode *inode = file_inode(req->file); | |
2b188cc1 | 2654 | |
491381ce | 2655 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); |
2b188cc1 | 2656 | } |
491381ce | 2657 | file_end_write(req->file); |
2b188cc1 JA |
2658 | } |
2659 | ||
a1d7c393 JA |
2660 | static void io_complete_rw_common(struct kiocb *kiocb, long res, |
2661 | struct io_comp_state *cs) | |
2b188cc1 | 2662 | { |
9adbd45d | 2663 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
bcda7baa | 2664 | int cflags = 0; |
2b188cc1 | 2665 | |
491381ce JA |
2666 | if (kiocb->ki_flags & IOCB_WRITE) |
2667 | kiocb_end_write(req); | |
2b188cc1 | 2668 | |
4e88d6e7 JA |
2669 | if (res != req->result) |
2670 | req_set_fail_links(req); | |
bcda7baa | 2671 | if (req->flags & REQ_F_BUFFER_SELECTED) |
8ff069bf | 2672 | cflags = io_put_rw_kbuf(req); |
a1d7c393 | 2673 | __io_req_complete(req, res, cflags, cs); |
ba816ad6 JA |
2674 | } |
2675 | ||
b63534c4 JA |
2676 | #ifdef CONFIG_BLOCK |
2677 | static bool io_resubmit_prep(struct io_kiocb *req, int error) | |
2678 | { | |
2679 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
2680 | ssize_t ret = -ECANCELED; | |
2681 | struct iov_iter iter; | |
2682 | int rw; | |
2683 | ||
2684 | if (error) { | |
2685 | ret = error; | |
2686 | goto end_req; | |
2687 | } | |
2688 | ||
2689 | switch (req->opcode) { | |
2690 | case IORING_OP_READV: | |
2691 | case IORING_OP_READ_FIXED: | |
2692 | case IORING_OP_READ: | |
2693 | rw = READ; | |
2694 | break; | |
2695 | case IORING_OP_WRITEV: | |
2696 | case IORING_OP_WRITE_FIXED: | |
2697 | case IORING_OP_WRITE: | |
2698 | rw = WRITE; | |
2699 | break; | |
2700 | default: | |
2701 | printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n", | |
2702 | req->opcode); | |
2703 | goto end_req; | |
2704 | } | |
2705 | ||
e8c2bc1f | 2706 | if (!req->async_data) { |
8f3d7496 JA |
2707 | ret = io_import_iovec(rw, req, &iovec, &iter, false); |
2708 | if (ret < 0) | |
2709 | goto end_req; | |
2710 | ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false); | |
2711 | if (!ret) | |
2712 | return true; | |
2713 | kfree(iovec); | |
2714 | } else { | |
b63534c4 | 2715 | return true; |
8f3d7496 | 2716 | } |
b63534c4 | 2717 | end_req: |
b63534c4 | 2718 | req_set_fail_links(req); |
b63534c4 JA |
2719 | return false; |
2720 | } | |
b63534c4 JA |
2721 | #endif |
2722 | ||
2723 | static bool io_rw_reissue(struct io_kiocb *req, long res) | |
2724 | { | |
2725 | #ifdef CONFIG_BLOCK | |
355afaeb | 2726 | umode_t mode = file_inode(req->file)->i_mode; |
b63534c4 JA |
2727 | int ret; |
2728 | ||
355afaeb JA |
2729 | if (!S_ISBLK(mode) && !S_ISREG(mode)) |
2730 | return false; | |
b63534c4 JA |
2731 | if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker()) |
2732 | return false; | |
2733 | ||
55e6ac1e PB |
2734 | lockdep_assert_held(&req->ctx->uring_lock); |
2735 | ||
28cea78a | 2736 | ret = io_sq_thread_acquire_mm_files(req->ctx, req); |
6d816e08 | 2737 | |
fdee946d JA |
2738 | if (io_resubmit_prep(req, ret)) { |
2739 | refcount_inc(&req->refs); | |
2740 | io_queue_async_work(req); | |
b63534c4 | 2741 | return true; |
fdee946d JA |
2742 | } |
2743 | ||
b63534c4 JA |
2744 | #endif |
2745 | return false; | |
2746 | } | |
2747 | ||
a1d7c393 JA |
2748 | static void __io_complete_rw(struct io_kiocb *req, long res, long res2, |
2749 | struct io_comp_state *cs) | |
2750 | { | |
2751 | if (!io_rw_reissue(req, res)) | |
2752 | io_complete_rw_common(&req->rw.kiocb, res, cs); | |
ba816ad6 JA |
2753 | } |
2754 | ||
2755 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
2756 | { | |
9adbd45d | 2757 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
ba816ad6 | 2758 | |
a1d7c393 | 2759 | __io_complete_rw(req, res, res2, NULL); |
2b188cc1 JA |
2760 | } |
2761 | ||
def596e9 JA |
2762 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
2763 | { | |
9adbd45d | 2764 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
def596e9 | 2765 | |
491381ce JA |
2766 | if (kiocb->ki_flags & IOCB_WRITE) |
2767 | kiocb_end_write(req); | |
def596e9 | 2768 | |
2d7d6792 | 2769 | if (res != -EAGAIN && res != req->result) |
4e88d6e7 | 2770 | req_set_fail_links(req); |
bbde017a XW |
2771 | |
2772 | WRITE_ONCE(req->result, res); | |
2773 | /* order with io_poll_complete() checking ->result */ | |
cd664b0e PB |
2774 | smp_wmb(); |
2775 | WRITE_ONCE(req->iopoll_completed, 1); | |
def596e9 JA |
2776 | } |
2777 | ||
2778 | /* | |
2779 | * After the iocb has been issued, it's safe to be found on the poll list. | |
2780 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
2781 | * find it from a io_iopoll_getevents() thread before the issuer is done | |
2782 | * accessing the kiocb cookie. | |
2783 | */ | |
2e9dbe90 | 2784 | static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async) |
def596e9 JA |
2785 | { |
2786 | struct io_ring_ctx *ctx = req->ctx; | |
2787 | ||
2788 | /* | |
2789 | * Track whether we have multiple files in our lists. This will impact | |
2790 | * how we do polling eventually, not spinning if we're on potentially | |
2791 | * different devices. | |
2792 | */ | |
540e32a0 | 2793 | if (list_empty(&ctx->iopoll_list)) { |
def596e9 JA |
2794 | ctx->poll_multi_file = false; |
2795 | } else if (!ctx->poll_multi_file) { | |
2796 | struct io_kiocb *list_req; | |
2797 | ||
540e32a0 | 2798 | list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb, |
d21ffe7e | 2799 | inflight_entry); |
9adbd45d | 2800 | if (list_req->file != req->file) |
def596e9 JA |
2801 | ctx->poll_multi_file = true; |
2802 | } | |
2803 | ||
2804 | /* | |
2805 | * For fast devices, IO may have already completed. If it has, add | |
2806 | * it to the front so we find it first. | |
2807 | */ | |
65a6543d | 2808 | if (READ_ONCE(req->iopoll_completed)) |
d21ffe7e | 2809 | list_add(&req->inflight_entry, &ctx->iopoll_list); |
def596e9 | 2810 | else |
d21ffe7e | 2811 | list_add_tail(&req->inflight_entry, &ctx->iopoll_list); |
bdcd3eab | 2812 | |
2e9dbe90 XW |
2813 | /* |
2814 | * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread | |
2815 | * task context or in io worker task context. If current task context is | |
2816 | * sq thread, we don't need to check whether should wake up sq thread. | |
2817 | */ | |
2818 | if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) && | |
534ca6d6 JA |
2819 | wq_has_sleeper(&ctx->sq_data->wait)) |
2820 | wake_up(&ctx->sq_data->wait); | |
def596e9 JA |
2821 | } |
2822 | ||
6e1271e6 | 2823 | static inline void __io_state_file_put(struct io_submit_state *state) |
9a56a232 | 2824 | { |
6e1271e6 PB |
2825 | fput_many(state->file, state->file_refs); |
2826 | state->file_refs = 0; | |
9f13c35b PB |
2827 | } |
2828 | ||
2829 | static inline void io_state_file_put(struct io_submit_state *state) | |
2830 | { | |
6e1271e6 | 2831 | if (state->file_refs) |
9f13c35b | 2832 | __io_state_file_put(state); |
9a56a232 JA |
2833 | } |
2834 | ||
2835 | /* | |
2836 | * Get as many references to a file as we have IOs left in this submission, | |
2837 | * assuming most submissions are for one file, or at least that each file | |
2838 | * has more than one submission. | |
2839 | */ | |
8da11c19 | 2840 | static struct file *__io_file_get(struct io_submit_state *state, int fd) |
9a56a232 JA |
2841 | { |
2842 | if (!state) | |
2843 | return fget(fd); | |
2844 | ||
6e1271e6 | 2845 | if (state->file_refs) { |
9a56a232 | 2846 | if (state->fd == fd) { |
6e1271e6 | 2847 | state->file_refs--; |
9a56a232 JA |
2848 | return state->file; |
2849 | } | |
9f13c35b | 2850 | __io_state_file_put(state); |
9a56a232 JA |
2851 | } |
2852 | state->file = fget_many(fd, state->ios_left); | |
6e1271e6 | 2853 | if (unlikely(!state->file)) |
9a56a232 JA |
2854 | return NULL; |
2855 | ||
2856 | state->fd = fd; | |
6e1271e6 | 2857 | state->file_refs = state->ios_left - 1; |
9a56a232 JA |
2858 | return state->file; |
2859 | } | |
2860 | ||
4503b767 JA |
2861 | static bool io_bdev_nowait(struct block_device *bdev) |
2862 | { | |
9ba0d0c8 | 2863 | return !bdev || blk_queue_nowait(bdev_get_queue(bdev)); |
4503b767 JA |
2864 | } |
2865 | ||
2b188cc1 JA |
2866 | /* |
2867 | * If we tracked the file through the SCM inflight mechanism, we could support | |
2868 | * any file. For now, just ensure that anything potentially problematic is done | |
2869 | * inline. | |
2870 | */ | |
af197f50 | 2871 | static bool io_file_supports_async(struct file *file, int rw) |
2b188cc1 JA |
2872 | { |
2873 | umode_t mode = file_inode(file)->i_mode; | |
2874 | ||
4503b767 | 2875 | if (S_ISBLK(mode)) { |
4e7b5671 CH |
2876 | if (IS_ENABLED(CONFIG_BLOCK) && |
2877 | io_bdev_nowait(I_BDEV(file->f_mapping->host))) | |
4503b767 JA |
2878 | return true; |
2879 | return false; | |
2880 | } | |
2881 | if (S_ISCHR(mode) || S_ISSOCK(mode)) | |
2b188cc1 | 2882 | return true; |
4503b767 | 2883 | if (S_ISREG(mode)) { |
4e7b5671 CH |
2884 | if (IS_ENABLED(CONFIG_BLOCK) && |
2885 | io_bdev_nowait(file->f_inode->i_sb->s_bdev) && | |
4503b767 JA |
2886 | file->f_op != &io_uring_fops) |
2887 | return true; | |
2888 | return false; | |
2889 | } | |
2b188cc1 | 2890 | |
c5b85625 JA |
2891 | /* any ->read/write should understand O_NONBLOCK */ |
2892 | if (file->f_flags & O_NONBLOCK) | |
2893 | return true; | |
2894 | ||
af197f50 JA |
2895 | if (!(file->f_mode & FMODE_NOWAIT)) |
2896 | return false; | |
2897 | ||
2898 | if (rw == READ) | |
2899 | return file->f_op->read_iter != NULL; | |
2900 | ||
2901 | return file->f_op->write_iter != NULL; | |
2b188cc1 JA |
2902 | } |
2903 | ||
a88fc400 | 2904 | static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2b188cc1 | 2905 | { |
def596e9 | 2906 | struct io_ring_ctx *ctx = req->ctx; |
9adbd45d | 2907 | struct kiocb *kiocb = &req->rw.kiocb; |
09bb8394 JA |
2908 | unsigned ioprio; |
2909 | int ret; | |
2b188cc1 | 2910 | |
491381ce JA |
2911 | if (S_ISREG(file_inode(req->file)->i_mode)) |
2912 | req->flags |= REQ_F_ISREG; | |
2913 | ||
2b188cc1 | 2914 | kiocb->ki_pos = READ_ONCE(sqe->off); |
ba04291e JA |
2915 | if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) { |
2916 | req->flags |= REQ_F_CUR_POS; | |
2917 | kiocb->ki_pos = req->file->f_pos; | |
2918 | } | |
2b188cc1 | 2919 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); |
3e577dcd PB |
2920 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); |
2921 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
2922 | if (unlikely(ret)) | |
2923 | return ret; | |
2b188cc1 JA |
2924 | |
2925 | ioprio = READ_ONCE(sqe->ioprio); | |
2926 | if (ioprio) { | |
2927 | ret = ioprio_check_cap(ioprio); | |
2928 | if (ret) | |
09bb8394 | 2929 | return ret; |
2b188cc1 JA |
2930 | |
2931 | kiocb->ki_ioprio = ioprio; | |
2932 | } else | |
2933 | kiocb->ki_ioprio = get_current_ioprio(); | |
2934 | ||
8449eeda | 2935 | /* don't allow async punt if RWF_NOWAIT was requested */ |
c5b85625 | 2936 | if (kiocb->ki_flags & IOCB_NOWAIT) |
8449eeda SB |
2937 | req->flags |= REQ_F_NOWAIT; |
2938 | ||
def596e9 | 2939 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
2940 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
2941 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 2942 | return -EOPNOTSUPP; |
2b188cc1 | 2943 | |
def596e9 JA |
2944 | kiocb->ki_flags |= IOCB_HIPRI; |
2945 | kiocb->ki_complete = io_complete_rw_iopoll; | |
65a6543d | 2946 | req->iopoll_completed = 0; |
def596e9 | 2947 | } else { |
09bb8394 JA |
2948 | if (kiocb->ki_flags & IOCB_HIPRI) |
2949 | return -EINVAL; | |
def596e9 JA |
2950 | kiocb->ki_complete = io_complete_rw; |
2951 | } | |
9adbd45d | 2952 | |
3529d8c2 JA |
2953 | req->rw.addr = READ_ONCE(sqe->addr); |
2954 | req->rw.len = READ_ONCE(sqe->len); | |
4f4eeba8 | 2955 | req->buf_index = READ_ONCE(sqe->buf_index); |
2b188cc1 | 2956 | return 0; |
2b188cc1 JA |
2957 | } |
2958 | ||
2959 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
2960 | { | |
2961 | switch (ret) { | |
2962 | case -EIOCBQUEUED: | |
2963 | break; | |
2964 | case -ERESTARTSYS: | |
2965 | case -ERESTARTNOINTR: | |
2966 | case -ERESTARTNOHAND: | |
2967 | case -ERESTART_RESTARTBLOCK: | |
2968 | /* | |
2969 | * We can't just restart the syscall, since previously | |
2970 | * submitted sqes may already be in progress. Just fail this | |
2971 | * IO with EINTR. | |
2972 | */ | |
2973 | ret = -EINTR; | |
df561f66 | 2974 | fallthrough; |
2b188cc1 JA |
2975 | default: |
2976 | kiocb->ki_complete(kiocb, ret, 0); | |
2977 | } | |
2978 | } | |
2979 | ||
a1d7c393 JA |
2980 | static void kiocb_done(struct kiocb *kiocb, ssize_t ret, |
2981 | struct io_comp_state *cs) | |
ba816ad6 | 2982 | { |
ba04291e | 2983 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
e8c2bc1f | 2984 | struct io_async_rw *io = req->async_data; |
ba04291e | 2985 | |
227c0c96 | 2986 | /* add previously done IO, if any */ |
e8c2bc1f | 2987 | if (io && io->bytes_done > 0) { |
227c0c96 | 2988 | if (ret < 0) |
e8c2bc1f | 2989 | ret = io->bytes_done; |
227c0c96 | 2990 | else |
e8c2bc1f | 2991 | ret += io->bytes_done; |
227c0c96 JA |
2992 | } |
2993 | ||
ba04291e JA |
2994 | if (req->flags & REQ_F_CUR_POS) |
2995 | req->file->f_pos = kiocb->ki_pos; | |
bcaec089 | 2996 | if (ret >= 0 && kiocb->ki_complete == io_complete_rw) |
a1d7c393 | 2997 | __io_complete_rw(req, ret, 0, cs); |
ba816ad6 JA |
2998 | else |
2999 | io_rw_done(kiocb, ret); | |
3000 | } | |
3001 | ||
9adbd45d | 3002 | static ssize_t io_import_fixed(struct io_kiocb *req, int rw, |
7d009165 | 3003 | struct iov_iter *iter) |
edafccee | 3004 | { |
9adbd45d JA |
3005 | struct io_ring_ctx *ctx = req->ctx; |
3006 | size_t len = req->rw.len; | |
edafccee | 3007 | struct io_mapped_ubuf *imu; |
4be1c615 | 3008 | u16 index, buf_index = req->buf_index; |
edafccee JA |
3009 | size_t offset; |
3010 | u64 buf_addr; | |
3011 | ||
edafccee JA |
3012 | if (unlikely(buf_index >= ctx->nr_user_bufs)) |
3013 | return -EFAULT; | |
edafccee JA |
3014 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); |
3015 | imu = &ctx->user_bufs[index]; | |
9adbd45d | 3016 | buf_addr = req->rw.addr; |
edafccee JA |
3017 | |
3018 | /* overflow */ | |
3019 | if (buf_addr + len < buf_addr) | |
3020 | return -EFAULT; | |
3021 | /* not inside the mapped region */ | |
3022 | if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len) | |
3023 | return -EFAULT; | |
3024 | ||
3025 | /* | |
3026 | * May not be a start of buffer, set size appropriately | |
3027 | * and advance us to the beginning. | |
3028 | */ | |
3029 | offset = buf_addr - imu->ubuf; | |
3030 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
bd11b3a3 JA |
3031 | |
3032 | if (offset) { | |
3033 | /* | |
3034 | * Don't use iov_iter_advance() here, as it's really slow for | |
3035 | * using the latter parts of a big fixed buffer - it iterates | |
3036 | * over each segment manually. We can cheat a bit here, because | |
3037 | * we know that: | |
3038 | * | |
3039 | * 1) it's a BVEC iter, we set it up | |
3040 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
3041 | * first and last bvec | |
3042 | * | |
3043 | * So just find our index, and adjust the iterator afterwards. | |
3044 | * If the offset is within the first bvec (or the whole first | |
3045 | * bvec, just use iov_iter_advance(). This makes it easier | |
3046 | * since we can just skip the first segment, which may not | |
3047 | * be PAGE_SIZE aligned. | |
3048 | */ | |
3049 | const struct bio_vec *bvec = imu->bvec; | |
3050 | ||
3051 | if (offset <= bvec->bv_len) { | |
3052 | iov_iter_advance(iter, offset); | |
3053 | } else { | |
3054 | unsigned long seg_skip; | |
3055 | ||
3056 | /* skip first vec */ | |
3057 | offset -= bvec->bv_len; | |
3058 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
3059 | ||
3060 | iter->bvec = bvec + seg_skip; | |
3061 | iter->nr_segs -= seg_skip; | |
99c79f66 | 3062 | iter->count -= bvec->bv_len + offset; |
bd11b3a3 | 3063 | iter->iov_offset = offset & ~PAGE_MASK; |
bd11b3a3 JA |
3064 | } |
3065 | } | |
3066 | ||
5e559561 | 3067 | return len; |
edafccee JA |
3068 | } |
3069 | ||
bcda7baa JA |
3070 | static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock) |
3071 | { | |
3072 | if (needs_lock) | |
3073 | mutex_unlock(&ctx->uring_lock); | |
3074 | } | |
3075 | ||
3076 | static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock) | |
3077 | { | |
3078 | /* | |
3079 | * "Normal" inline submissions always hold the uring_lock, since we | |
3080 | * grab it from the system call. Same is true for the SQPOLL offload. | |
3081 | * The only exception is when we've detached the request and issue it | |
3082 | * from an async worker thread, grab the lock for that case. | |
3083 | */ | |
3084 | if (needs_lock) | |
3085 | mutex_lock(&ctx->uring_lock); | |
3086 | } | |
3087 | ||
3088 | static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len, | |
3089 | int bgid, struct io_buffer *kbuf, | |
3090 | bool needs_lock) | |
3091 | { | |
3092 | struct io_buffer *head; | |
3093 | ||
3094 | if (req->flags & REQ_F_BUFFER_SELECTED) | |
3095 | return kbuf; | |
3096 | ||
3097 | io_ring_submit_lock(req->ctx, needs_lock); | |
3098 | ||
3099 | lockdep_assert_held(&req->ctx->uring_lock); | |
3100 | ||
3101 | head = idr_find(&req->ctx->io_buffer_idr, bgid); | |
3102 | if (head) { | |
3103 | if (!list_empty(&head->list)) { | |
3104 | kbuf = list_last_entry(&head->list, struct io_buffer, | |
3105 | list); | |
3106 | list_del(&kbuf->list); | |
3107 | } else { | |
3108 | kbuf = head; | |
3109 | idr_remove(&req->ctx->io_buffer_idr, bgid); | |
3110 | } | |
3111 | if (*len > kbuf->len) | |
3112 | *len = kbuf->len; | |
3113 | } else { | |
3114 | kbuf = ERR_PTR(-ENOBUFS); | |
3115 | } | |
3116 | ||
3117 | io_ring_submit_unlock(req->ctx, needs_lock); | |
3118 | ||
3119 | return kbuf; | |
3120 | } | |
3121 | ||
4d954c25 JA |
3122 | static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len, |
3123 | bool needs_lock) | |
3124 | { | |
3125 | struct io_buffer *kbuf; | |
4f4eeba8 | 3126 | u16 bgid; |
4d954c25 JA |
3127 | |
3128 | kbuf = (struct io_buffer *) (unsigned long) req->rw.addr; | |
4f4eeba8 | 3129 | bgid = req->buf_index; |
4d954c25 JA |
3130 | kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock); |
3131 | if (IS_ERR(kbuf)) | |
3132 | return kbuf; | |
3133 | req->rw.addr = (u64) (unsigned long) kbuf; | |
3134 | req->flags |= REQ_F_BUFFER_SELECTED; | |
3135 | return u64_to_user_ptr(kbuf->addr); | |
3136 | } | |
3137 | ||
3138 | #ifdef CONFIG_COMPAT | |
3139 | static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov, | |
3140 | bool needs_lock) | |
3141 | { | |
3142 | struct compat_iovec __user *uiov; | |
3143 | compat_ssize_t clen; | |
3144 | void __user *buf; | |
3145 | ssize_t len; | |
3146 | ||
3147 | uiov = u64_to_user_ptr(req->rw.addr); | |
3148 | if (!access_ok(uiov, sizeof(*uiov))) | |
3149 | return -EFAULT; | |
3150 | if (__get_user(clen, &uiov->iov_len)) | |
3151 | return -EFAULT; | |
3152 | if (clen < 0) | |
3153 | return -EINVAL; | |
3154 | ||
3155 | len = clen; | |
3156 | buf = io_rw_buffer_select(req, &len, needs_lock); | |
3157 | if (IS_ERR(buf)) | |
3158 | return PTR_ERR(buf); | |
3159 | iov[0].iov_base = buf; | |
3160 | iov[0].iov_len = (compat_size_t) len; | |
3161 | return 0; | |
3162 | } | |
3163 | #endif | |
3164 | ||
3165 | static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov, | |
3166 | bool needs_lock) | |
3167 | { | |
3168 | struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr); | |
3169 | void __user *buf; | |
3170 | ssize_t len; | |
3171 | ||
3172 | if (copy_from_user(iov, uiov, sizeof(*uiov))) | |
3173 | return -EFAULT; | |
3174 | ||
3175 | len = iov[0].iov_len; | |
3176 | if (len < 0) | |
3177 | return -EINVAL; | |
3178 | buf = io_rw_buffer_select(req, &len, needs_lock); | |
3179 | if (IS_ERR(buf)) | |
3180 | return PTR_ERR(buf); | |
3181 | iov[0].iov_base = buf; | |
3182 | iov[0].iov_len = len; | |
3183 | return 0; | |
3184 | } | |
3185 | ||
3186 | static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov, | |
3187 | bool needs_lock) | |
3188 | { | |
dddb3e26 JA |
3189 | if (req->flags & REQ_F_BUFFER_SELECTED) { |
3190 | struct io_buffer *kbuf; | |
3191 | ||
3192 | kbuf = (struct io_buffer *) (unsigned long) req->rw.addr; | |
3193 | iov[0].iov_base = u64_to_user_ptr(kbuf->addr); | |
3194 | iov[0].iov_len = kbuf->len; | |
4d954c25 | 3195 | return 0; |
dddb3e26 | 3196 | } |
dd201662 | 3197 | if (req->rw.len != 1) |
4d954c25 JA |
3198 | return -EINVAL; |
3199 | ||
3200 | #ifdef CONFIG_COMPAT | |
3201 | if (req->ctx->compat) | |
3202 | return io_compat_import(req, iov, needs_lock); | |
3203 | #endif | |
3204 | ||
3205 | return __io_iov_buffer_select(req, iov, needs_lock); | |
3206 | } | |
3207 | ||
2846c481 | 3208 | static ssize_t io_import_iovec(int rw, struct io_kiocb *req, |
8452fd0c JA |
3209 | struct iovec **iovec, struct iov_iter *iter, |
3210 | bool needs_lock) | |
2b188cc1 | 3211 | { |
9adbd45d JA |
3212 | void __user *buf = u64_to_user_ptr(req->rw.addr); |
3213 | size_t sqe_len = req->rw.len; | |
4d954c25 | 3214 | ssize_t ret; |
edafccee JA |
3215 | u8 opcode; |
3216 | ||
d625c6ee | 3217 | opcode = req->opcode; |
7d009165 | 3218 | if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) { |
edafccee | 3219 | *iovec = NULL; |
9adbd45d | 3220 | return io_import_fixed(req, rw, iter); |
edafccee | 3221 | } |
2b188cc1 | 3222 | |
bcda7baa | 3223 | /* buffer index only valid with fixed read/write, or buffer select */ |
4f4eeba8 | 3224 | if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT)) |
9adbd45d JA |
3225 | return -EINVAL; |
3226 | ||
3a6820f2 | 3227 | if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) { |
bcda7baa | 3228 | if (req->flags & REQ_F_BUFFER_SELECT) { |
4d954c25 | 3229 | buf = io_rw_buffer_select(req, &sqe_len, needs_lock); |
867a23ea | 3230 | if (IS_ERR(buf)) |
4d954c25 | 3231 | return PTR_ERR(buf); |
3f9d6441 | 3232 | req->rw.len = sqe_len; |
bcda7baa JA |
3233 | } |
3234 | ||
3a6820f2 JA |
3235 | ret = import_single_range(rw, buf, sqe_len, *iovec, iter); |
3236 | *iovec = NULL; | |
10fc72e4 | 3237 | return ret; |
3a6820f2 JA |
3238 | } |
3239 | ||
4d954c25 JA |
3240 | if (req->flags & REQ_F_BUFFER_SELECT) { |
3241 | ret = io_iov_buffer_select(req, *iovec, needs_lock); | |
3f9d6441 JA |
3242 | if (!ret) { |
3243 | ret = (*iovec)->iov_len; | |
3244 | iov_iter_init(iter, rw, *iovec, 1, ret); | |
3245 | } | |
4d954c25 JA |
3246 | *iovec = NULL; |
3247 | return ret; | |
3248 | } | |
3249 | ||
89cd35c5 CH |
3250 | return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter, |
3251 | req->ctx->compat); | |
2b188cc1 JA |
3252 | } |
3253 | ||
0fef9483 JA |
3254 | static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb) |
3255 | { | |
5b09e37e | 3256 | return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos; |
0fef9483 JA |
3257 | } |
3258 | ||
31b51510 | 3259 | /* |
32960613 JA |
3260 | * For files that don't have ->read_iter() and ->write_iter(), handle them |
3261 | * by looping over ->read() or ->write() manually. | |
31b51510 | 3262 | */ |
4017eb91 | 3263 | static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter) |
32960613 | 3264 | { |
4017eb91 JA |
3265 | struct kiocb *kiocb = &req->rw.kiocb; |
3266 | struct file *file = req->file; | |
32960613 JA |
3267 | ssize_t ret = 0; |
3268 | ||
3269 | /* | |
3270 | * Don't support polled IO through this interface, and we can't | |
3271 | * support non-blocking either. For the latter, this just causes | |
3272 | * the kiocb to be handled from an async context. | |
3273 | */ | |
3274 | if (kiocb->ki_flags & IOCB_HIPRI) | |
3275 | return -EOPNOTSUPP; | |
3276 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
3277 | return -EAGAIN; | |
3278 | ||
3279 | while (iov_iter_count(iter)) { | |
311ae9e1 | 3280 | struct iovec iovec; |
32960613 JA |
3281 | ssize_t nr; |
3282 | ||
311ae9e1 PB |
3283 | if (!iov_iter_is_bvec(iter)) { |
3284 | iovec = iov_iter_iovec(iter); | |
3285 | } else { | |
4017eb91 JA |
3286 | iovec.iov_base = u64_to_user_ptr(req->rw.addr); |
3287 | iovec.iov_len = req->rw.len; | |
311ae9e1 PB |
3288 | } |
3289 | ||
32960613 JA |
3290 | if (rw == READ) { |
3291 | nr = file->f_op->read(file, iovec.iov_base, | |
0fef9483 | 3292 | iovec.iov_len, io_kiocb_ppos(kiocb)); |
32960613 JA |
3293 | } else { |
3294 | nr = file->f_op->write(file, iovec.iov_base, | |
0fef9483 | 3295 | iovec.iov_len, io_kiocb_ppos(kiocb)); |
32960613 JA |
3296 | } |
3297 | ||
3298 | if (nr < 0) { | |
3299 | if (!ret) | |
3300 | ret = nr; | |
3301 | break; | |
3302 | } | |
3303 | ret += nr; | |
3304 | if (nr != iovec.iov_len) | |
3305 | break; | |
4017eb91 JA |
3306 | req->rw.len -= nr; |
3307 | req->rw.addr += nr; | |
32960613 JA |
3308 | iov_iter_advance(iter, nr); |
3309 | } | |
3310 | ||
3311 | return ret; | |
3312 | } | |
3313 | ||
ff6165b2 JA |
3314 | static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec, |
3315 | const struct iovec *fast_iov, struct iov_iter *iter) | |
f67676d1 | 3316 | { |
e8c2bc1f | 3317 | struct io_async_rw *rw = req->async_data; |
b64e3444 | 3318 | |
ff6165b2 | 3319 | memcpy(&rw->iter, iter, sizeof(*iter)); |
afb87658 | 3320 | rw->free_iovec = iovec; |
227c0c96 | 3321 | rw->bytes_done = 0; |
ff6165b2 | 3322 | /* can only be fixed buffers, no need to do anything */ |
9c3a205c | 3323 | if (iov_iter_is_bvec(iter)) |
ff6165b2 | 3324 | return; |
b64e3444 | 3325 | if (!iovec) { |
ff6165b2 JA |
3326 | unsigned iov_off = 0; |
3327 | ||
3328 | rw->iter.iov = rw->fast_iov; | |
3329 | if (iter->iov != fast_iov) { | |
3330 | iov_off = iter->iov - fast_iov; | |
3331 | rw->iter.iov += iov_off; | |
3332 | } | |
3333 | if (rw->fast_iov != fast_iov) | |
3334 | memcpy(rw->fast_iov + iov_off, fast_iov + iov_off, | |
45097dae | 3335 | sizeof(struct iovec) * iter->nr_segs); |
99bc4c38 PB |
3336 | } else { |
3337 | req->flags |= REQ_F_NEED_CLEANUP; | |
f67676d1 JA |
3338 | } |
3339 | } | |
3340 | ||
e8c2bc1f | 3341 | static inline int __io_alloc_async_data(struct io_kiocb *req) |
3d9932a8 | 3342 | { |
e8c2bc1f JA |
3343 | WARN_ON_ONCE(!io_op_defs[req->opcode].async_size); |
3344 | req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL); | |
3345 | return req->async_data == NULL; | |
3d9932a8 XW |
3346 | } |
3347 | ||
e8c2bc1f | 3348 | static int io_alloc_async_data(struct io_kiocb *req) |
f67676d1 | 3349 | { |
e8c2bc1f | 3350 | if (!io_op_defs[req->opcode].needs_async_data) |
d3656344 | 3351 | return 0; |
3d9932a8 | 3352 | |
e8c2bc1f | 3353 | return __io_alloc_async_data(req); |
b7bb4f7d JA |
3354 | } |
3355 | ||
ff6165b2 JA |
3356 | static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec, |
3357 | const struct iovec *fast_iov, | |
227c0c96 | 3358 | struct iov_iter *iter, bool force) |
b7bb4f7d | 3359 | { |
e8c2bc1f | 3360 | if (!force && !io_op_defs[req->opcode].needs_async_data) |
74566df3 | 3361 | return 0; |
e8c2bc1f JA |
3362 | if (!req->async_data) { |
3363 | if (__io_alloc_async_data(req)) | |
5d204bcf | 3364 | return -ENOMEM; |
b7bb4f7d | 3365 | |
ff6165b2 | 3366 | io_req_map_rw(req, iovec, fast_iov, iter); |
5d204bcf | 3367 | } |
b7bb4f7d | 3368 | return 0; |
f67676d1 JA |
3369 | } |
3370 | ||
73debe68 | 3371 | static inline int io_rw_prep_async(struct io_kiocb *req, int rw) |
c3e330a4 | 3372 | { |
e8c2bc1f | 3373 | struct io_async_rw *iorw = req->async_data; |
f4bff104 | 3374 | struct iovec *iov = iorw->fast_iov; |
c3e330a4 PB |
3375 | ssize_t ret; |
3376 | ||
2846c481 | 3377 | ret = io_import_iovec(rw, req, &iov, &iorw->iter, false); |
c3e330a4 PB |
3378 | if (unlikely(ret < 0)) |
3379 | return ret; | |
3380 | ||
ab0b196c PB |
3381 | iorw->bytes_done = 0; |
3382 | iorw->free_iovec = iov; | |
3383 | if (iov) | |
3384 | req->flags |= REQ_F_NEED_CLEANUP; | |
c3e330a4 PB |
3385 | return 0; |
3386 | } | |
3387 | ||
73debe68 | 3388 | static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f67676d1 JA |
3389 | { |
3390 | ssize_t ret; | |
3391 | ||
a88fc400 | 3392 | ret = io_prep_rw(req, sqe); |
3529d8c2 JA |
3393 | if (ret) |
3394 | return ret; | |
f67676d1 | 3395 | |
3529d8c2 JA |
3396 | if (unlikely(!(req->file->f_mode & FMODE_READ))) |
3397 | return -EBADF; | |
f67676d1 | 3398 | |
5f798bea | 3399 | /* either don't need iovec imported or already have it */ |
2d199895 | 3400 | if (!req->async_data) |
3529d8c2 | 3401 | return 0; |
73debe68 | 3402 | return io_rw_prep_async(req, READ); |
f67676d1 JA |
3403 | } |
3404 | ||
c1dd91d1 JA |
3405 | /* |
3406 | * This is our waitqueue callback handler, registered through lock_page_async() | |
3407 | * when we initially tried to do the IO with the iocb armed our waitqueue. | |
3408 | * This gets called when the page is unlocked, and we generally expect that to | |
3409 | * happen when the page IO is completed and the page is now uptodate. This will | |
3410 | * queue a task_work based retry of the operation, attempting to copy the data | |
3411 | * again. If the latter fails because the page was NOT uptodate, then we will | |
3412 | * do a thread based blocking retry of the operation. That's the unexpected | |
3413 | * slow path. | |
3414 | */ | |
bcf5a063 JA |
3415 | static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode, |
3416 | int sync, void *arg) | |
3417 | { | |
3418 | struct wait_page_queue *wpq; | |
3419 | struct io_kiocb *req = wait->private; | |
bcf5a063 | 3420 | struct wait_page_key *key = arg; |
bcf5a063 JA |
3421 | int ret; |
3422 | ||
3423 | wpq = container_of(wait, struct wait_page_queue, wait); | |
3424 | ||
cdc8fcb4 LT |
3425 | if (!wake_page_match(wpq, key)) |
3426 | return 0; | |
3427 | ||
c8d317aa | 3428 | req->rw.kiocb.ki_flags &= ~IOCB_WAITQ; |
bcf5a063 JA |
3429 | list_del_init(&wait->entry); |
3430 | ||
e7375122 | 3431 | init_task_work(&req->task_work, io_req_task_submit); |
6d816e08 JA |
3432 | percpu_ref_get(&req->ctx->refs); |
3433 | ||
bcf5a063 JA |
3434 | /* submit ref gets dropped, acquire a new one */ |
3435 | refcount_inc(&req->refs); | |
355fb9e2 | 3436 | ret = io_req_task_work_add(req); |
bcf5a063 | 3437 | if (unlikely(ret)) { |
c2c4c83c JA |
3438 | struct task_struct *tsk; |
3439 | ||
bcf5a063 | 3440 | /* queue just for cancelation */ |
e7375122 | 3441 | init_task_work(&req->task_work, io_req_task_cancel); |
bcf5a063 | 3442 | tsk = io_wq_get_task(req->ctx->io_wq); |
91989c70 | 3443 | task_work_add(tsk, &req->task_work, TWA_NONE); |
c2c4c83c | 3444 | wake_up_process(tsk); |
bcf5a063 | 3445 | } |
bcf5a063 JA |
3446 | return 1; |
3447 | } | |
3448 | ||
c1dd91d1 JA |
3449 | /* |
3450 | * This controls whether a given IO request should be armed for async page | |
3451 | * based retry. If we return false here, the request is handed to the async | |
3452 | * worker threads for retry. If we're doing buffered reads on a regular file, | |
3453 | * we prepare a private wait_page_queue entry and retry the operation. This | |
3454 | * will either succeed because the page is now uptodate and unlocked, or it | |
3455 | * will register a callback when the page is unlocked at IO completion. Through | |
3456 | * that callback, io_uring uses task_work to setup a retry of the operation. | |
3457 | * That retry will attempt the buffered read again. The retry will generally | |
3458 | * succeed, or in rare cases where it fails, we then fall back to using the | |
3459 | * async worker threads for a blocking retry. | |
3460 | */ | |
227c0c96 | 3461 | static bool io_rw_should_retry(struct io_kiocb *req) |
f67676d1 | 3462 | { |
e8c2bc1f JA |
3463 | struct io_async_rw *rw = req->async_data; |
3464 | struct wait_page_queue *wait = &rw->wpq; | |
bcf5a063 | 3465 | struct kiocb *kiocb = &req->rw.kiocb; |
f67676d1 | 3466 | |
bcf5a063 JA |
3467 | /* never retry for NOWAIT, we just complete with -EAGAIN */ |
3468 | if (req->flags & REQ_F_NOWAIT) | |
3469 | return false; | |
f67676d1 | 3470 | |
227c0c96 | 3471 | /* Only for buffered IO */ |
3b2a4439 | 3472 | if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI)) |
bcf5a063 | 3473 | return false; |
3b2a4439 | 3474 | |
bcf5a063 JA |
3475 | /* |
3476 | * just use poll if we can, and don't attempt if the fs doesn't | |
3477 | * support callback based unlocks | |
3478 | */ | |
3479 | if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC)) | |
3480 | return false; | |
f67676d1 | 3481 | |
3b2a4439 JA |
3482 | wait->wait.func = io_async_buf_func; |
3483 | wait->wait.private = req; | |
3484 | wait->wait.flags = 0; | |
3485 | INIT_LIST_HEAD(&wait->wait.entry); | |
3486 | kiocb->ki_flags |= IOCB_WAITQ; | |
c8d317aa | 3487 | kiocb->ki_flags &= ~IOCB_NOWAIT; |
3b2a4439 | 3488 | kiocb->ki_waitq = wait; |
3b2a4439 | 3489 | return true; |
bcf5a063 JA |
3490 | } |
3491 | ||
3492 | static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter) | |
3493 | { | |
3494 | if (req->file->f_op->read_iter) | |
3495 | return call_read_iter(req->file, &req->rw.kiocb, iter); | |
2dd2111d | 3496 | else if (req->file->f_op->read) |
4017eb91 | 3497 | return loop_rw_iter(READ, req, iter); |
2dd2111d GH |
3498 | else |
3499 | return -EINVAL; | |
f67676d1 JA |
3500 | } |
3501 | ||
a1d7c393 JA |
3502 | static int io_read(struct io_kiocb *req, bool force_nonblock, |
3503 | struct io_comp_state *cs) | |
2b188cc1 JA |
3504 | { |
3505 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
9adbd45d | 3506 | struct kiocb *kiocb = &req->rw.kiocb; |
ff6165b2 | 3507 | struct iov_iter __iter, *iter = &__iter; |
e8c2bc1f | 3508 | struct io_async_rw *rw = req->async_data; |
227c0c96 | 3509 | ssize_t io_size, ret, ret2; |
f5cac8b1 | 3510 | bool no_async; |
ff6165b2 | 3511 | |
2846c481 | 3512 | if (rw) { |
e8c2bc1f | 3513 | iter = &rw->iter; |
2846c481 PB |
3514 | iovec = NULL; |
3515 | } else { | |
3516 | ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock); | |
3517 | if (ret < 0) | |
3518 | return ret; | |
3519 | } | |
632546c4 | 3520 | io_size = iov_iter_count(iter); |
fa15bafb | 3521 | req->result = io_size; |
227c0c96 | 3522 | ret = 0; |
2b188cc1 | 3523 | |
fd6c2e4c JA |
3524 | /* Ensure we clear previously set non-block flag */ |
3525 | if (!force_nonblock) | |
29de5f6a | 3526 | kiocb->ki_flags &= ~IOCB_NOWAIT; |
a88fc400 PB |
3527 | else |
3528 | kiocb->ki_flags |= IOCB_NOWAIT; | |
3529 | ||
fd6c2e4c | 3530 | |
24c74678 | 3531 | /* If the file doesn't support async, just async punt */ |
f5cac8b1 JA |
3532 | no_async = force_nonblock && !io_file_supports_async(req->file, READ); |
3533 | if (no_async) | |
f67676d1 | 3534 | goto copy_iov; |
9e645e11 | 3535 | |
632546c4 | 3536 | ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size); |
fa15bafb PB |
3537 | if (unlikely(ret)) |
3538 | goto out_free; | |
2b188cc1 | 3539 | |
227c0c96 | 3540 | ret = io_iter_do_read(req, iter); |
32960613 | 3541 | |
227c0c96 JA |
3542 | if (!ret) { |
3543 | goto done; | |
3544 | } else if (ret == -EIOCBQUEUED) { | |
3545 | ret = 0; | |
3546 | goto out_free; | |
3547 | } else if (ret == -EAGAIN) { | |
eefdf30f JA |
3548 | /* IOPOLL retry should happen for io-wq threads */ |
3549 | if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
f91daf56 | 3550 | goto done; |
355afaeb JA |
3551 | /* no retry on NONBLOCK marked file */ |
3552 | if (req->file->f_flags & O_NONBLOCK) | |
3553 | goto done; | |
84216315 | 3554 | /* some cases will consume bytes even on error returns */ |
632546c4 | 3555 | iov_iter_revert(iter, io_size - iov_iter_count(iter)); |
f38c7e3a JA |
3556 | ret = 0; |
3557 | goto copy_iov; | |
227c0c96 | 3558 | } else if (ret < 0) { |
00d23d51 JA |
3559 | /* make sure -ERESTARTSYS -> -EINTR is done */ |
3560 | goto done; | |
227c0c96 JA |
3561 | } |
3562 | ||
3563 | /* read it all, or we did blocking attempt. no retry. */ | |
f91daf56 | 3564 | if (!iov_iter_count(iter) || !force_nonblock || |
9a173346 | 3565 | (req->file->f_flags & O_NONBLOCK) || !(req->flags & REQ_F_ISREG)) |
227c0c96 JA |
3566 | goto done; |
3567 | ||
3568 | io_size -= ret; | |
3569 | copy_iov: | |
3570 | ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true); | |
3571 | if (ret2) { | |
3572 | ret = ret2; | |
3573 | goto out_free; | |
3574 | } | |
f5cac8b1 JA |
3575 | if (no_async) |
3576 | return -EAGAIN; | |
e8c2bc1f | 3577 | rw = req->async_data; |
227c0c96 JA |
3578 | /* it's copied and will be cleaned with ->io */ |
3579 | iovec = NULL; | |
3580 | /* now use our persistent iterator, if we aren't already */ | |
e8c2bc1f | 3581 | iter = &rw->iter; |
227c0c96 | 3582 | retry: |
e8c2bc1f | 3583 | rw->bytes_done += ret; |
227c0c96 JA |
3584 | /* if we can retry, do so with the callbacks armed */ |
3585 | if (!io_rw_should_retry(req)) { | |
fa15bafb PB |
3586 | kiocb->ki_flags &= ~IOCB_WAITQ; |
3587 | return -EAGAIN; | |
2b188cc1 | 3588 | } |
227c0c96 JA |
3589 | |
3590 | /* | |
3591 | * Now retry read with the IOCB_WAITQ parts set in the iocb. If we | |
3592 | * get -EIOCBQUEUED, then we'll get a notification when the desired | |
3593 | * page gets unlocked. We can also get a partial read here, and if we | |
3594 | * do, then just retry at the new offset. | |
3595 | */ | |
3596 | ret = io_iter_do_read(req, iter); | |
3597 | if (ret == -EIOCBQUEUED) { | |
3598 | ret = 0; | |
3599 | goto out_free; | |
3600 | } else if (ret > 0 && ret < io_size) { | |
3601 | /* we got some bytes, but not all. retry. */ | |
3602 | goto retry; | |
3603 | } | |
3604 | done: | |
3605 | kiocb_done(kiocb, ret, cs); | |
3606 | ret = 0; | |
f67676d1 | 3607 | out_free: |
f261c168 | 3608 | /* it's reportedly faster than delegating the null check to kfree() */ |
252917c3 | 3609 | if (iovec) |
6f2cc166 | 3610 | kfree(iovec); |
2b188cc1 JA |
3611 | return ret; |
3612 | } | |
3613 | ||
73debe68 | 3614 | static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f67676d1 JA |
3615 | { |
3616 | ssize_t ret; | |
3617 | ||
a88fc400 | 3618 | ret = io_prep_rw(req, sqe); |
3529d8c2 JA |
3619 | if (ret) |
3620 | return ret; | |
f67676d1 | 3621 | |
3529d8c2 JA |
3622 | if (unlikely(!(req->file->f_mode & FMODE_WRITE))) |
3623 | return -EBADF; | |
f67676d1 | 3624 | |
5f798bea | 3625 | /* either don't need iovec imported or already have it */ |
2d199895 | 3626 | if (!req->async_data) |
3529d8c2 | 3627 | return 0; |
73debe68 | 3628 | return io_rw_prep_async(req, WRITE); |
f67676d1 JA |
3629 | } |
3630 | ||
a1d7c393 JA |
3631 | static int io_write(struct io_kiocb *req, bool force_nonblock, |
3632 | struct io_comp_state *cs) | |
2b188cc1 JA |
3633 | { |
3634 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
9adbd45d | 3635 | struct kiocb *kiocb = &req->rw.kiocb; |
ff6165b2 | 3636 | struct iov_iter __iter, *iter = &__iter; |
e8c2bc1f | 3637 | struct io_async_rw *rw = req->async_data; |
fa15bafb | 3638 | ssize_t ret, ret2, io_size; |
2b188cc1 | 3639 | |
2846c481 | 3640 | if (rw) { |
e8c2bc1f | 3641 | iter = &rw->iter; |
2846c481 PB |
3642 | iovec = NULL; |
3643 | } else { | |
3644 | ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock); | |
3645 | if (ret < 0) | |
3646 | return ret; | |
3647 | } | |
632546c4 | 3648 | io_size = iov_iter_count(iter); |
fa15bafb | 3649 | req->result = io_size; |
2b188cc1 | 3650 | |
fd6c2e4c JA |
3651 | /* Ensure we clear previously set non-block flag */ |
3652 | if (!force_nonblock) | |
a88fc400 PB |
3653 | kiocb->ki_flags &= ~IOCB_NOWAIT; |
3654 | else | |
3655 | kiocb->ki_flags |= IOCB_NOWAIT; | |
fd6c2e4c | 3656 | |
24c74678 | 3657 | /* If the file doesn't support async, just async punt */ |
af197f50 | 3658 | if (force_nonblock && !io_file_supports_async(req->file, WRITE)) |
f67676d1 | 3659 | goto copy_iov; |
31b51510 | 3660 | |
10d59345 JA |
3661 | /* file path doesn't support NOWAIT for non-direct_IO */ |
3662 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) && | |
3663 | (req->flags & REQ_F_ISREG)) | |
f67676d1 | 3664 | goto copy_iov; |
31b51510 | 3665 | |
632546c4 | 3666 | ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size); |
fa15bafb PB |
3667 | if (unlikely(ret)) |
3668 | goto out_free; | |
4ed734b0 | 3669 | |
fa15bafb PB |
3670 | /* |
3671 | * Open-code file_start_write here to grab freeze protection, | |
3672 | * which will be released by another thread in | |
3673 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
3674 | * released so that it doesn't complain about the held lock when | |
3675 | * we return to userspace. | |
3676 | */ | |
3677 | if (req->flags & REQ_F_ISREG) { | |
8a3c84b6 | 3678 | sb_start_write(file_inode(req->file)->i_sb); |
fa15bafb PB |
3679 | __sb_writers_release(file_inode(req->file)->i_sb, |
3680 | SB_FREEZE_WRITE); | |
3681 | } | |
3682 | kiocb->ki_flags |= IOCB_WRITE; | |
4ed734b0 | 3683 | |
fa15bafb | 3684 | if (req->file->f_op->write_iter) |
ff6165b2 | 3685 | ret2 = call_write_iter(req->file, kiocb, iter); |
2dd2111d | 3686 | else if (req->file->f_op->write) |
4017eb91 | 3687 | ret2 = loop_rw_iter(WRITE, req, iter); |
2dd2111d GH |
3688 | else |
3689 | ret2 = -EINVAL; | |
4ed734b0 | 3690 | |
fa15bafb PB |
3691 | /* |
3692 | * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just | |
3693 | * retry them without IOCB_NOWAIT. | |
3694 | */ | |
3695 | if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT)) | |
3696 | ret2 = -EAGAIN; | |
355afaeb JA |
3697 | /* no retry on NONBLOCK marked file */ |
3698 | if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK)) | |
3699 | goto done; | |
fa15bafb | 3700 | if (!force_nonblock || ret2 != -EAGAIN) { |
eefdf30f JA |
3701 | /* IOPOLL retry should happen for io-wq threads */ |
3702 | if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN) | |
3703 | goto copy_iov; | |
355afaeb | 3704 | done: |
fa15bafb PB |
3705 | kiocb_done(kiocb, ret2, cs); |
3706 | } else { | |
f67676d1 | 3707 | copy_iov: |
84216315 | 3708 | /* some cases will consume bytes even on error returns */ |
632546c4 | 3709 | iov_iter_revert(iter, io_size - iov_iter_count(iter)); |
227c0c96 | 3710 | ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false); |
ff6165b2 JA |
3711 | if (!ret) |
3712 | return -EAGAIN; | |
2b188cc1 | 3713 | } |
31b51510 | 3714 | out_free: |
f261c168 | 3715 | /* it's reportedly faster than delegating the null check to kfree() */ |
252917c3 | 3716 | if (iovec) |
6f2cc166 | 3717 | kfree(iovec); |
2b188cc1 JA |
3718 | return ret; |
3719 | } | |
3720 | ||
80a261fd JA |
3721 | static int io_renameat_prep(struct io_kiocb *req, |
3722 | const struct io_uring_sqe *sqe) | |
3723 | { | |
3724 | struct io_rename *ren = &req->rename; | |
3725 | const char __user *oldf, *newf; | |
3726 | ||
3727 | if (unlikely(req->flags & REQ_F_FIXED_FILE)) | |
3728 | return -EBADF; | |
3729 | ||
3730 | ren->old_dfd = READ_ONCE(sqe->fd); | |
3731 | oldf = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
3732 | newf = u64_to_user_ptr(READ_ONCE(sqe->addr2)); | |
3733 | ren->new_dfd = READ_ONCE(sqe->len); | |
3734 | ren->flags = READ_ONCE(sqe->rename_flags); | |
3735 | ||
3736 | ren->oldpath = getname(oldf); | |
3737 | if (IS_ERR(ren->oldpath)) | |
3738 | return PTR_ERR(ren->oldpath); | |
3739 | ||
3740 | ren->newpath = getname(newf); | |
3741 | if (IS_ERR(ren->newpath)) { | |
3742 | putname(ren->oldpath); | |
3743 | return PTR_ERR(ren->newpath); | |
3744 | } | |
3745 | ||
3746 | req->flags |= REQ_F_NEED_CLEANUP; | |
3747 | return 0; | |
3748 | } | |
3749 | ||
3750 | static int io_renameat(struct io_kiocb *req, bool force_nonblock) | |
3751 | { | |
3752 | struct io_rename *ren = &req->rename; | |
3753 | int ret; | |
3754 | ||
3755 | if (force_nonblock) | |
3756 | return -EAGAIN; | |
3757 | ||
3758 | ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd, | |
3759 | ren->newpath, ren->flags); | |
3760 | ||
3761 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
3762 | if (ret < 0) | |
3763 | req_set_fail_links(req); | |
3764 | io_req_complete(req, ret); | |
3765 | return 0; | |
3766 | } | |
3767 | ||
14a1143b JA |
3768 | static int io_unlinkat_prep(struct io_kiocb *req, |
3769 | const struct io_uring_sqe *sqe) | |
3770 | { | |
3771 | struct io_unlink *un = &req->unlink; | |
3772 | const char __user *fname; | |
3773 | ||
3774 | if (unlikely(req->flags & REQ_F_FIXED_FILE)) | |
3775 | return -EBADF; | |
3776 | ||
3777 | un->dfd = READ_ONCE(sqe->fd); | |
3778 | ||
3779 | un->flags = READ_ONCE(sqe->unlink_flags); | |
3780 | if (un->flags & ~AT_REMOVEDIR) | |
3781 | return -EINVAL; | |
3782 | ||
3783 | fname = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
3784 | un->filename = getname(fname); | |
3785 | if (IS_ERR(un->filename)) | |
3786 | return PTR_ERR(un->filename); | |
3787 | ||
3788 | req->flags |= REQ_F_NEED_CLEANUP; | |
3789 | return 0; | |
3790 | } | |
3791 | ||
3792 | static int io_unlinkat(struct io_kiocb *req, bool force_nonblock) | |
3793 | { | |
3794 | struct io_unlink *un = &req->unlink; | |
3795 | int ret; | |
3796 | ||
3797 | if (force_nonblock) | |
3798 | return -EAGAIN; | |
3799 | ||
3800 | if (un->flags & AT_REMOVEDIR) | |
3801 | ret = do_rmdir(un->dfd, un->filename); | |
3802 | else | |
3803 | ret = do_unlinkat(un->dfd, un->filename); | |
3804 | ||
3805 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
3806 | if (ret < 0) | |
3807 | req_set_fail_links(req); | |
3808 | io_req_complete(req, ret); | |
3809 | return 0; | |
3810 | } | |
3811 | ||
36f4fa68 JA |
3812 | static int io_shutdown_prep(struct io_kiocb *req, |
3813 | const struct io_uring_sqe *sqe) | |
3814 | { | |
3815 | #if defined(CONFIG_NET) | |
3816 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
3817 | return -EINVAL; | |
3818 | if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags || | |
3819 | sqe->buf_index) | |
3820 | return -EINVAL; | |
3821 | ||
3822 | req->shutdown.how = READ_ONCE(sqe->len); | |
3823 | return 0; | |
3824 | #else | |
3825 | return -EOPNOTSUPP; | |
3826 | #endif | |
3827 | } | |
3828 | ||
3829 | static int io_shutdown(struct io_kiocb *req, bool force_nonblock) | |
3830 | { | |
3831 | #if defined(CONFIG_NET) | |
3832 | struct socket *sock; | |
3833 | int ret; | |
3834 | ||
3835 | if (force_nonblock) | |
3836 | return -EAGAIN; | |
3837 | ||
48aba79b | 3838 | sock = sock_from_file(req->file); |
36f4fa68 | 3839 | if (unlikely(!sock)) |
48aba79b | 3840 | return -ENOTSOCK; |
36f4fa68 JA |
3841 | |
3842 | ret = __sys_shutdown_sock(sock, req->shutdown.how); | |
a146468d JA |
3843 | if (ret < 0) |
3844 | req_set_fail_links(req); | |
36f4fa68 JA |
3845 | io_req_complete(req, ret); |
3846 | return 0; | |
3847 | #else | |
3848 | return -EOPNOTSUPP; | |
3849 | #endif | |
3850 | } | |
3851 | ||
f2a8d5c7 PB |
3852 | static int __io_splice_prep(struct io_kiocb *req, |
3853 | const struct io_uring_sqe *sqe) | |
7d67af2c PB |
3854 | { |
3855 | struct io_splice* sp = &req->splice; | |
3856 | unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL; | |
7d67af2c | 3857 | |
3232dd02 PB |
3858 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3859 | return -EINVAL; | |
7d67af2c PB |
3860 | |
3861 | sp->file_in = NULL; | |
7d67af2c PB |
3862 | sp->len = READ_ONCE(sqe->len); |
3863 | sp->flags = READ_ONCE(sqe->splice_flags); | |
3864 | ||
3865 | if (unlikely(sp->flags & ~valid_flags)) | |
3866 | return -EINVAL; | |
3867 | ||
8371adf5 PB |
3868 | sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), |
3869 | (sp->flags & SPLICE_F_FD_IN_FIXED)); | |
3870 | if (!sp->file_in) | |
3871 | return -EBADF; | |
7d67af2c PB |
3872 | req->flags |= REQ_F_NEED_CLEANUP; |
3873 | ||
7cdaf587 XW |
3874 | if (!S_ISREG(file_inode(sp->file_in)->i_mode)) { |
3875 | /* | |
3876 | * Splice operation will be punted aync, and here need to | |
3877 | * modify io_wq_work.flags, so initialize io_wq_work firstly. | |
3878 | */ | |
3879 | io_req_init_async(req); | |
7d67af2c | 3880 | req->work.flags |= IO_WQ_WORK_UNBOUND; |
7cdaf587 | 3881 | } |
7d67af2c PB |
3882 | |
3883 | return 0; | |
3884 | } | |
3885 | ||
f2a8d5c7 PB |
3886 | static int io_tee_prep(struct io_kiocb *req, |
3887 | const struct io_uring_sqe *sqe) | |
3888 | { | |
3889 | if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off)) | |
3890 | return -EINVAL; | |
3891 | return __io_splice_prep(req, sqe); | |
3892 | } | |
3893 | ||
3894 | static int io_tee(struct io_kiocb *req, bool force_nonblock) | |
3895 | { | |
3896 | struct io_splice *sp = &req->splice; | |
3897 | struct file *in = sp->file_in; | |
3898 | struct file *out = sp->file_out; | |
3899 | unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED; | |
3900 | long ret = 0; | |
3901 | ||
3902 | if (force_nonblock) | |
3903 | return -EAGAIN; | |
3904 | if (sp->len) | |
3905 | ret = do_tee(in, out, sp->len, flags); | |
3906 | ||
3907 | io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED)); | |
3908 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
3909 | ||
f2a8d5c7 PB |
3910 | if (ret != sp->len) |
3911 | req_set_fail_links(req); | |
e1e16097 | 3912 | io_req_complete(req, ret); |
f2a8d5c7 PB |
3913 | return 0; |
3914 | } | |
3915 | ||
3916 | static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
3917 | { | |
3918 | struct io_splice* sp = &req->splice; | |
3919 | ||
3920 | sp->off_in = READ_ONCE(sqe->splice_off_in); | |
3921 | sp->off_out = READ_ONCE(sqe->off); | |
3922 | return __io_splice_prep(req, sqe); | |
3923 | } | |
3924 | ||
014db007 | 3925 | static int io_splice(struct io_kiocb *req, bool force_nonblock) |
7d67af2c PB |
3926 | { |
3927 | struct io_splice *sp = &req->splice; | |
3928 | struct file *in = sp->file_in; | |
3929 | struct file *out = sp->file_out; | |
3930 | unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED; | |
3931 | loff_t *poff_in, *poff_out; | |
c9687426 | 3932 | long ret = 0; |
7d67af2c | 3933 | |
2fb3e822 PB |
3934 | if (force_nonblock) |
3935 | return -EAGAIN; | |
7d67af2c PB |
3936 | |
3937 | poff_in = (sp->off_in == -1) ? NULL : &sp->off_in; | |
3938 | poff_out = (sp->off_out == -1) ? NULL : &sp->off_out; | |
c9687426 | 3939 | |
948a7749 | 3940 | if (sp->len) |
c9687426 | 3941 | ret = do_splice(in, poff_in, out, poff_out, sp->len, flags); |
7d67af2c PB |
3942 | |
3943 | io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED)); | |
3944 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
3945 | ||
7d67af2c PB |
3946 | if (ret != sp->len) |
3947 | req_set_fail_links(req); | |
e1e16097 | 3948 | io_req_complete(req, ret); |
7d67af2c PB |
3949 | return 0; |
3950 | } | |
3951 | ||
2b188cc1 JA |
3952 | /* |
3953 | * IORING_OP_NOP just posts a completion event, nothing else. | |
3954 | */ | |
229a7b63 | 3955 | static int io_nop(struct io_kiocb *req, struct io_comp_state *cs) |
2b188cc1 JA |
3956 | { |
3957 | struct io_ring_ctx *ctx = req->ctx; | |
2b188cc1 | 3958 | |
def596e9 JA |
3959 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
3960 | return -EINVAL; | |
3961 | ||
229a7b63 | 3962 | __io_req_complete(req, 0, 0, cs); |
2b188cc1 JA |
3963 | return 0; |
3964 | } | |
3965 | ||
3529d8c2 | 3966 | static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
c992fe29 | 3967 | { |
6b06314c | 3968 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 3969 | |
09bb8394 JA |
3970 | if (!req->file) |
3971 | return -EBADF; | |
c992fe29 | 3972 | |
6b06314c | 3973 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 3974 | return -EINVAL; |
edafccee | 3975 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
3976 | return -EINVAL; |
3977 | ||
8ed8d3c3 JA |
3978 | req->sync.flags = READ_ONCE(sqe->fsync_flags); |
3979 | if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC)) | |
3980 | return -EINVAL; | |
3981 | ||
3982 | req->sync.off = READ_ONCE(sqe->off); | |
3983 | req->sync.len = READ_ONCE(sqe->len); | |
c992fe29 CH |
3984 | return 0; |
3985 | } | |
3986 | ||
ac45abc0 | 3987 | static int io_fsync(struct io_kiocb *req, bool force_nonblock) |
8ed8d3c3 | 3988 | { |
8ed8d3c3 | 3989 | loff_t end = req->sync.off + req->sync.len; |
8ed8d3c3 JA |
3990 | int ret; |
3991 | ||
ac45abc0 PB |
3992 | /* fsync always requires a blocking context */ |
3993 | if (force_nonblock) | |
3994 | return -EAGAIN; | |
3995 | ||
9adbd45d | 3996 | ret = vfs_fsync_range(req->file, req->sync.off, |
8ed8d3c3 JA |
3997 | end > 0 ? end : LLONG_MAX, |
3998 | req->sync.flags & IORING_FSYNC_DATASYNC); | |
3999 | if (ret < 0) | |
4000 | req_set_fail_links(req); | |
e1e16097 | 4001 | io_req_complete(req, ret); |
c992fe29 CH |
4002 | return 0; |
4003 | } | |
4004 | ||
d63d1b5e JA |
4005 | static int io_fallocate_prep(struct io_kiocb *req, |
4006 | const struct io_uring_sqe *sqe) | |
4007 | { | |
4008 | if (sqe->ioprio || sqe->buf_index || sqe->rw_flags) | |
4009 | return -EINVAL; | |
3232dd02 PB |
4010 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4011 | return -EINVAL; | |
d63d1b5e JA |
4012 | |
4013 | req->sync.off = READ_ONCE(sqe->off); | |
4014 | req->sync.len = READ_ONCE(sqe->addr); | |
4015 | req->sync.mode = READ_ONCE(sqe->len); | |
4016 | return 0; | |
4017 | } | |
4018 | ||
014db007 | 4019 | static int io_fallocate(struct io_kiocb *req, bool force_nonblock) |
5d17b4a4 | 4020 | { |
ac45abc0 PB |
4021 | int ret; |
4022 | ||
d63d1b5e | 4023 | /* fallocate always requiring blocking context */ |
ac45abc0 | 4024 | if (force_nonblock) |
5d17b4a4 | 4025 | return -EAGAIN; |
ac45abc0 PB |
4026 | ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off, |
4027 | req->sync.len); | |
ac45abc0 PB |
4028 | if (ret < 0) |
4029 | req_set_fail_links(req); | |
e1e16097 | 4030 | io_req_complete(req, ret); |
5d17b4a4 JA |
4031 | return 0; |
4032 | } | |
4033 | ||
ec65fea5 | 4034 | static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
b7bb4f7d | 4035 | { |
f8748881 | 4036 | const char __user *fname; |
15b71abe | 4037 | int ret; |
b7bb4f7d | 4038 | |
ec65fea5 | 4039 | if (unlikely(sqe->ioprio || sqe->buf_index)) |
15b71abe | 4040 | return -EINVAL; |
ec65fea5 | 4041 | if (unlikely(req->flags & REQ_F_FIXED_FILE)) |
cf3040ca | 4042 | return -EBADF; |
03b1230c | 4043 | |
ec65fea5 PB |
4044 | /* open.how should be already initialised */ |
4045 | if (!(req->open.how.flags & O_PATH) && force_o_largefile()) | |
08a1d26e | 4046 | req->open.how.flags |= O_LARGEFILE; |
3529d8c2 | 4047 | |
25e72d10 PB |
4048 | req->open.dfd = READ_ONCE(sqe->fd); |
4049 | fname = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
f8748881 | 4050 | req->open.filename = getname(fname); |
15b71abe JA |
4051 | if (IS_ERR(req->open.filename)) { |
4052 | ret = PTR_ERR(req->open.filename); | |
4053 | req->open.filename = NULL; | |
4054 | return ret; | |
4055 | } | |
4022e7af | 4056 | req->open.nofile = rlimit(RLIMIT_NOFILE); |
944d1444 | 4057 | req->open.ignore_nonblock = false; |
8fef80bf | 4058 | req->flags |= REQ_F_NEED_CLEANUP; |
15b71abe | 4059 | return 0; |
03b1230c JA |
4060 | } |
4061 | ||
ec65fea5 PB |
4062 | static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4063 | { | |
4064 | u64 flags, mode; | |
4065 | ||
14587a46 | 4066 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4eb8dded | 4067 | return -EINVAL; |
ec65fea5 PB |
4068 | mode = READ_ONCE(sqe->len); |
4069 | flags = READ_ONCE(sqe->open_flags); | |
4070 | req->open.how = build_open_how(flags, mode); | |
4071 | return __io_openat_prep(req, sqe); | |
4072 | } | |
4073 | ||
cebdb986 | 4074 | static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
aa1fa28f | 4075 | { |
cebdb986 | 4076 | struct open_how __user *how; |
cebdb986 | 4077 | size_t len; |
0fa03c62 JA |
4078 | int ret; |
4079 | ||
14587a46 | 4080 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4eb8dded | 4081 | return -EINVAL; |
cebdb986 JA |
4082 | how = u64_to_user_ptr(READ_ONCE(sqe->addr2)); |
4083 | len = READ_ONCE(sqe->len); | |
cebdb986 JA |
4084 | if (len < OPEN_HOW_SIZE_VER0) |
4085 | return -EINVAL; | |
3529d8c2 | 4086 | |
cebdb986 JA |
4087 | ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how, |
4088 | len); | |
4089 | if (ret) | |
4090 | return ret; | |
3529d8c2 | 4091 | |
ec65fea5 | 4092 | return __io_openat_prep(req, sqe); |
cebdb986 JA |
4093 | } |
4094 | ||
014db007 | 4095 | static int io_openat2(struct io_kiocb *req, bool force_nonblock) |
15b71abe JA |
4096 | { |
4097 | struct open_flags op; | |
15b71abe JA |
4098 | struct file *file; |
4099 | int ret; | |
4100 | ||
944d1444 | 4101 | if (force_nonblock && !req->open.ignore_nonblock) |
15b71abe | 4102 | return -EAGAIN; |
15b71abe | 4103 | |
cebdb986 | 4104 | ret = build_open_flags(&req->open.how, &op); |
15b71abe JA |
4105 | if (ret) |
4106 | goto err; | |
4107 | ||
4022e7af | 4108 | ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile); |
15b71abe JA |
4109 | if (ret < 0) |
4110 | goto err; | |
4111 | ||
4112 | file = do_filp_open(req->open.dfd, req->open.filename, &op); | |
4113 | if (IS_ERR(file)) { | |
4114 | put_unused_fd(ret); | |
4115 | ret = PTR_ERR(file); | |
944d1444 JA |
4116 | /* |
4117 | * A work-around to ensure that /proc/self works that way | |
4118 | * that it should - if we get -EOPNOTSUPP back, then assume | |
4119 | * that proc_self_get_link() failed us because we're in async | |
4120 | * context. We should be safe to retry this from the task | |
4121 | * itself with force_nonblock == false set, as it should not | |
4122 | * block on lookup. Would be nice to know this upfront and | |
4123 | * avoid the async dance, but doesn't seem feasible. | |
4124 | */ | |
4125 | if (ret == -EOPNOTSUPP && io_wq_current_is_worker()) { | |
4126 | req->open.ignore_nonblock = true; | |
4127 | refcount_inc(&req->refs); | |
4128 | io_req_task_queue(req); | |
4129 | return 0; | |
4130 | } | |
15b71abe JA |
4131 | } else { |
4132 | fsnotify_open(file); | |
4133 | fd_install(ret, file); | |
4134 | } | |
4135 | err: | |
4136 | putname(req->open.filename); | |
8fef80bf | 4137 | req->flags &= ~REQ_F_NEED_CLEANUP; |
15b71abe JA |
4138 | if (ret < 0) |
4139 | req_set_fail_links(req); | |
e1e16097 | 4140 | io_req_complete(req, ret); |
15b71abe JA |
4141 | return 0; |
4142 | } | |
4143 | ||
014db007 | 4144 | static int io_openat(struct io_kiocb *req, bool force_nonblock) |
cebdb986 | 4145 | { |
014db007 | 4146 | return io_openat2(req, force_nonblock); |
cebdb986 JA |
4147 | } |
4148 | ||
067524e9 JA |
4149 | static int io_remove_buffers_prep(struct io_kiocb *req, |
4150 | const struct io_uring_sqe *sqe) | |
4151 | { | |
4152 | struct io_provide_buf *p = &req->pbuf; | |
4153 | u64 tmp; | |
4154 | ||
4155 | if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off) | |
4156 | return -EINVAL; | |
4157 | ||
4158 | tmp = READ_ONCE(sqe->fd); | |
4159 | if (!tmp || tmp > USHRT_MAX) | |
4160 | return -EINVAL; | |
4161 | ||
4162 | memset(p, 0, sizeof(*p)); | |
4163 | p->nbufs = tmp; | |
4164 | p->bgid = READ_ONCE(sqe->buf_group); | |
4165 | return 0; | |
4166 | } | |
4167 | ||
4168 | static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf, | |
4169 | int bgid, unsigned nbufs) | |
4170 | { | |
4171 | unsigned i = 0; | |
4172 | ||
4173 | /* shouldn't happen */ | |
4174 | if (!nbufs) | |
4175 | return 0; | |
4176 | ||
4177 | /* the head kbuf is the list itself */ | |
4178 | while (!list_empty(&buf->list)) { | |
4179 | struct io_buffer *nxt; | |
4180 | ||
4181 | nxt = list_first_entry(&buf->list, struct io_buffer, list); | |
4182 | list_del(&nxt->list); | |
4183 | kfree(nxt); | |
4184 | if (++i == nbufs) | |
4185 | return i; | |
4186 | } | |
4187 | i++; | |
4188 | kfree(buf); | |
4189 | idr_remove(&ctx->io_buffer_idr, bgid); | |
4190 | ||
4191 | return i; | |
4192 | } | |
4193 | ||
229a7b63 JA |
4194 | static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock, |
4195 | struct io_comp_state *cs) | |
067524e9 JA |
4196 | { |
4197 | struct io_provide_buf *p = &req->pbuf; | |
4198 | struct io_ring_ctx *ctx = req->ctx; | |
4199 | struct io_buffer *head; | |
4200 | int ret = 0; | |
4201 | ||
4202 | io_ring_submit_lock(ctx, !force_nonblock); | |
4203 | ||
4204 | lockdep_assert_held(&ctx->uring_lock); | |
4205 | ||
4206 | ret = -ENOENT; | |
4207 | head = idr_find(&ctx->io_buffer_idr, p->bgid); | |
4208 | if (head) | |
4209 | ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs); | |
067524e9 JA |
4210 | if (ret < 0) |
4211 | req_set_fail_links(req); | |
067524e9 | 4212 | |
31bff9a5 PB |
4213 | /* need to hold the lock to complete IOPOLL requests */ |
4214 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
4215 | __io_req_complete(req, ret, 0, cs); | |
4216 | io_ring_submit_unlock(ctx, !force_nonblock); | |
4217 | } else { | |
4218 | io_ring_submit_unlock(ctx, !force_nonblock); | |
4219 | __io_req_complete(req, ret, 0, cs); | |
4220 | } | |
067524e9 JA |
4221 | return 0; |
4222 | } | |
4223 | ||
ddf0322d JA |
4224 | static int io_provide_buffers_prep(struct io_kiocb *req, |
4225 | const struct io_uring_sqe *sqe) | |
4226 | { | |
4227 | struct io_provide_buf *p = &req->pbuf; | |
4228 | u64 tmp; | |
4229 | ||
4230 | if (sqe->ioprio || sqe->rw_flags) | |
4231 | return -EINVAL; | |
4232 | ||
4233 | tmp = READ_ONCE(sqe->fd); | |
4234 | if (!tmp || tmp > USHRT_MAX) | |
4235 | return -E2BIG; | |
4236 | p->nbufs = tmp; | |
4237 | p->addr = READ_ONCE(sqe->addr); | |
4238 | p->len = READ_ONCE(sqe->len); | |
4239 | ||
efe68c1c | 4240 | if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs))) |
ddf0322d JA |
4241 | return -EFAULT; |
4242 | ||
4243 | p->bgid = READ_ONCE(sqe->buf_group); | |
4244 | tmp = READ_ONCE(sqe->off); | |
4245 | if (tmp > USHRT_MAX) | |
4246 | return -E2BIG; | |
4247 | p->bid = tmp; | |
4248 | return 0; | |
4249 | } | |
4250 | ||
4251 | static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head) | |
4252 | { | |
4253 | struct io_buffer *buf; | |
4254 | u64 addr = pbuf->addr; | |
4255 | int i, bid = pbuf->bid; | |
4256 | ||
4257 | for (i = 0; i < pbuf->nbufs; i++) { | |
4258 | buf = kmalloc(sizeof(*buf), GFP_KERNEL); | |
4259 | if (!buf) | |
4260 | break; | |
4261 | ||
4262 | buf->addr = addr; | |
4263 | buf->len = pbuf->len; | |
4264 | buf->bid = bid; | |
4265 | addr += pbuf->len; | |
4266 | bid++; | |
4267 | if (!*head) { | |
4268 | INIT_LIST_HEAD(&buf->list); | |
4269 | *head = buf; | |
4270 | } else { | |
4271 | list_add_tail(&buf->list, &(*head)->list); | |
4272 | } | |
4273 | } | |
4274 | ||
4275 | return i ? i : -ENOMEM; | |
4276 | } | |
4277 | ||
229a7b63 JA |
4278 | static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock, |
4279 | struct io_comp_state *cs) | |
ddf0322d JA |
4280 | { |
4281 | struct io_provide_buf *p = &req->pbuf; | |
4282 | struct io_ring_ctx *ctx = req->ctx; | |
4283 | struct io_buffer *head, *list; | |
4284 | int ret = 0; | |
4285 | ||
4286 | io_ring_submit_lock(ctx, !force_nonblock); | |
4287 | ||
4288 | lockdep_assert_held(&ctx->uring_lock); | |
4289 | ||
4290 | list = head = idr_find(&ctx->io_buffer_idr, p->bgid); | |
4291 | ||
4292 | ret = io_add_buffers(p, &head); | |
4293 | if (ret < 0) | |
4294 | goto out; | |
4295 | ||
4296 | if (!list) { | |
4297 | ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1, | |
4298 | GFP_KERNEL); | |
4299 | if (ret < 0) { | |
067524e9 | 4300 | __io_remove_buffers(ctx, head, p->bgid, -1U); |
ddf0322d JA |
4301 | goto out; |
4302 | } | |
4303 | } | |
4304 | out: | |
ddf0322d JA |
4305 | if (ret < 0) |
4306 | req_set_fail_links(req); | |
31bff9a5 PB |
4307 | |
4308 | /* need to hold the lock to complete IOPOLL requests */ | |
4309 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
4310 | __io_req_complete(req, ret, 0, cs); | |
4311 | io_ring_submit_unlock(ctx, !force_nonblock); | |
4312 | } else { | |
4313 | io_ring_submit_unlock(ctx, !force_nonblock); | |
4314 | __io_req_complete(req, ret, 0, cs); | |
4315 | } | |
ddf0322d | 4316 | return 0; |
cebdb986 JA |
4317 | } |
4318 | ||
3e4827b0 JA |
4319 | static int io_epoll_ctl_prep(struct io_kiocb *req, |
4320 | const struct io_uring_sqe *sqe) | |
4321 | { | |
4322 | #if defined(CONFIG_EPOLL) | |
4323 | if (sqe->ioprio || sqe->buf_index) | |
4324 | return -EINVAL; | |
6ca56f84 | 4325 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL))) |
3232dd02 | 4326 | return -EINVAL; |
3e4827b0 JA |
4327 | |
4328 | req->epoll.epfd = READ_ONCE(sqe->fd); | |
4329 | req->epoll.op = READ_ONCE(sqe->len); | |
4330 | req->epoll.fd = READ_ONCE(sqe->off); | |
4331 | ||
4332 | if (ep_op_has_event(req->epoll.op)) { | |
4333 | struct epoll_event __user *ev; | |
4334 | ||
4335 | ev = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
4336 | if (copy_from_user(&req->epoll.event, ev, sizeof(*ev))) | |
4337 | return -EFAULT; | |
4338 | } | |
4339 | ||
4340 | return 0; | |
4341 | #else | |
4342 | return -EOPNOTSUPP; | |
4343 | #endif | |
4344 | } | |
4345 | ||
229a7b63 JA |
4346 | static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock, |
4347 | struct io_comp_state *cs) | |
3e4827b0 JA |
4348 | { |
4349 | #if defined(CONFIG_EPOLL) | |
4350 | struct io_epoll *ie = &req->epoll; | |
4351 | int ret; | |
4352 | ||
4353 | ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock); | |
4354 | if (force_nonblock && ret == -EAGAIN) | |
4355 | return -EAGAIN; | |
4356 | ||
4357 | if (ret < 0) | |
4358 | req_set_fail_links(req); | |
229a7b63 | 4359 | __io_req_complete(req, ret, 0, cs); |
3e4827b0 JA |
4360 | return 0; |
4361 | #else | |
4362 | return -EOPNOTSUPP; | |
4363 | #endif | |
4364 | } | |
4365 | ||
c1ca757b JA |
4366 | static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4367 | { | |
4368 | #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU) | |
4369 | if (sqe->ioprio || sqe->buf_index || sqe->off) | |
4370 | return -EINVAL; | |
3232dd02 PB |
4371 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4372 | return -EINVAL; | |
c1ca757b JA |
4373 | |
4374 | req->madvise.addr = READ_ONCE(sqe->addr); | |
4375 | req->madvise.len = READ_ONCE(sqe->len); | |
4376 | req->madvise.advice = READ_ONCE(sqe->fadvise_advice); | |
4377 | return 0; | |
4378 | #else | |
4379 | return -EOPNOTSUPP; | |
4380 | #endif | |
4381 | } | |
4382 | ||
014db007 | 4383 | static int io_madvise(struct io_kiocb *req, bool force_nonblock) |
c1ca757b JA |
4384 | { |
4385 | #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU) | |
4386 | struct io_madvise *ma = &req->madvise; | |
4387 | int ret; | |
4388 | ||
4389 | if (force_nonblock) | |
4390 | return -EAGAIN; | |
4391 | ||
0726b01e | 4392 | ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice); |
c1ca757b JA |
4393 | if (ret < 0) |
4394 | req_set_fail_links(req); | |
e1e16097 | 4395 | io_req_complete(req, ret); |
c1ca757b JA |
4396 | return 0; |
4397 | #else | |
4398 | return -EOPNOTSUPP; | |
4399 | #endif | |
4400 | } | |
4401 | ||
4840e418 JA |
4402 | static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4403 | { | |
4404 | if (sqe->ioprio || sqe->buf_index || sqe->addr) | |
4405 | return -EINVAL; | |
3232dd02 PB |
4406 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4407 | return -EINVAL; | |
4840e418 JA |
4408 | |
4409 | req->fadvise.offset = READ_ONCE(sqe->off); | |
4410 | req->fadvise.len = READ_ONCE(sqe->len); | |
4411 | req->fadvise.advice = READ_ONCE(sqe->fadvise_advice); | |
4412 | return 0; | |
4413 | } | |
4414 | ||
014db007 | 4415 | static int io_fadvise(struct io_kiocb *req, bool force_nonblock) |
4840e418 JA |
4416 | { |
4417 | struct io_fadvise *fa = &req->fadvise; | |
4418 | int ret; | |
4419 | ||
3e69426d JA |
4420 | if (force_nonblock) { |
4421 | switch (fa->advice) { | |
4422 | case POSIX_FADV_NORMAL: | |
4423 | case POSIX_FADV_RANDOM: | |
4424 | case POSIX_FADV_SEQUENTIAL: | |
4425 | break; | |
4426 | default: | |
4427 | return -EAGAIN; | |
4428 | } | |
4429 | } | |
4840e418 JA |
4430 | |
4431 | ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice); | |
4432 | if (ret < 0) | |
4433 | req_set_fail_links(req); | |
e1e16097 | 4434 | io_req_complete(req, ret); |
4840e418 JA |
4435 | return 0; |
4436 | } | |
4437 | ||
eddc7ef5 JA |
4438 | static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4439 | { | |
6ca56f84 | 4440 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL))) |
3232dd02 | 4441 | return -EINVAL; |
eddc7ef5 JA |
4442 | if (sqe->ioprio || sqe->buf_index) |
4443 | return -EINVAL; | |
9c280f90 | 4444 | if (req->flags & REQ_F_FIXED_FILE) |
cf3040ca | 4445 | return -EBADF; |
eddc7ef5 | 4446 | |
1d9e1288 BM |
4447 | req->statx.dfd = READ_ONCE(sqe->fd); |
4448 | req->statx.mask = READ_ONCE(sqe->len); | |
e62753e4 | 4449 | req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
1d9e1288 BM |
4450 | req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2)); |
4451 | req->statx.flags = READ_ONCE(sqe->statx_flags); | |
eddc7ef5 JA |
4452 | |
4453 | return 0; | |
4454 | } | |
4455 | ||
014db007 | 4456 | static int io_statx(struct io_kiocb *req, bool force_nonblock) |
eddc7ef5 | 4457 | { |
1d9e1288 | 4458 | struct io_statx *ctx = &req->statx; |
eddc7ef5 JA |
4459 | int ret; |
4460 | ||
5b0bbee4 JA |
4461 | if (force_nonblock) { |
4462 | /* only need file table for an actual valid fd */ | |
4463 | if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD) | |
4464 | req->flags |= REQ_F_NO_FILE_TABLE; | |
eddc7ef5 | 4465 | return -EAGAIN; |
5b0bbee4 | 4466 | } |
eddc7ef5 | 4467 | |
e62753e4 BM |
4468 | ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask, |
4469 | ctx->buffer); | |
eddc7ef5 | 4470 | |
eddc7ef5 JA |
4471 | if (ret < 0) |
4472 | req_set_fail_links(req); | |
e1e16097 | 4473 | io_req_complete(req, ret); |
eddc7ef5 JA |
4474 | return 0; |
4475 | } | |
4476 | ||
b5dba59e JA |
4477 | static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4478 | { | |
4479 | /* | |
4480 | * If we queue this for async, it must not be cancellable. That would | |
7cdaf587 XW |
4481 | * leave the 'file' in an undeterminate state, and here need to modify |
4482 | * io_wq_work.flags, so initialize io_wq_work firstly. | |
b5dba59e | 4483 | */ |
7cdaf587 | 4484 | io_req_init_async(req); |
b5dba59e | 4485 | |
14587a46 | 4486 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3232dd02 | 4487 | return -EINVAL; |
b5dba59e JA |
4488 | if (sqe->ioprio || sqe->off || sqe->addr || sqe->len || |
4489 | sqe->rw_flags || sqe->buf_index) | |
4490 | return -EINVAL; | |
9c280f90 | 4491 | if (req->flags & REQ_F_FIXED_FILE) |
cf3040ca | 4492 | return -EBADF; |
b5dba59e JA |
4493 | |
4494 | req->close.fd = READ_ONCE(sqe->fd); | |
0f212204 | 4495 | if ((req->file && req->file->f_op == &io_uring_fops)) |
fd2206e4 | 4496 | return -EBADF; |
b5dba59e | 4497 | |
3af73b28 | 4498 | req->close.put_file = NULL; |
b5dba59e | 4499 | return 0; |
b5dba59e JA |
4500 | } |
4501 | ||
229a7b63 JA |
4502 | static int io_close(struct io_kiocb *req, bool force_nonblock, |
4503 | struct io_comp_state *cs) | |
b5dba59e | 4504 | { |
3af73b28 | 4505 | struct io_close *close = &req->close; |
b5dba59e JA |
4506 | int ret; |
4507 | ||
3af73b28 PB |
4508 | /* might be already done during nonblock submission */ |
4509 | if (!close->put_file) { | |
9fe83c43 | 4510 | ret = close_fd_get_file(close->fd, &close->put_file); |
3af73b28 PB |
4511 | if (ret < 0) |
4512 | return (ret == -ENOENT) ? -EBADF : ret; | |
4513 | } | |
b5dba59e JA |
4514 | |
4515 | /* if the file has a flush method, be safe and punt to async */ | |
3af73b28 | 4516 | if (close->put_file->f_op->flush && force_nonblock) { |
607ec89e JA |
4517 | /* not safe to cancel at this point */ |
4518 | req->work.flags |= IO_WQ_WORK_NO_CANCEL; | |
24c74678 PB |
4519 | /* was never set, but play safe */ |
4520 | req->flags &= ~REQ_F_NOWAIT; | |
0bf0eefd | 4521 | /* avoid grabbing files - we don't need the files */ |
24c74678 | 4522 | req->flags |= REQ_F_NO_FILE_TABLE; |
0bf0eefd | 4523 | return -EAGAIN; |
a2100672 | 4524 | } |
b5dba59e | 4525 | |
3af73b28 | 4526 | /* No ->flush() or already async, safely close from here */ |
98447d65 | 4527 | ret = filp_close(close->put_file, req->work.identity->files); |
3af73b28 PB |
4528 | if (ret < 0) |
4529 | req_set_fail_links(req); | |
3af73b28 PB |
4530 | fput(close->put_file); |
4531 | close->put_file = NULL; | |
229a7b63 | 4532 | __io_req_complete(req, ret, 0, cs); |
1a417f4e | 4533 | return 0; |
b5dba59e JA |
4534 | } |
4535 | ||
3529d8c2 | 4536 | static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
5d17b4a4 JA |
4537 | { |
4538 | struct io_ring_ctx *ctx = req->ctx; | |
5d17b4a4 JA |
4539 | |
4540 | if (!req->file) | |
4541 | return -EBADF; | |
5d17b4a4 JA |
4542 | |
4543 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
4544 | return -EINVAL; | |
4545 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) | |
4546 | return -EINVAL; | |
4547 | ||
8ed8d3c3 JA |
4548 | req->sync.off = READ_ONCE(sqe->off); |
4549 | req->sync.len = READ_ONCE(sqe->len); | |
4550 | req->sync.flags = READ_ONCE(sqe->sync_range_flags); | |
8ed8d3c3 JA |
4551 | return 0; |
4552 | } | |
4553 | ||
ac45abc0 | 4554 | static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock) |
8ed8d3c3 | 4555 | { |
8ed8d3c3 JA |
4556 | int ret; |
4557 | ||
ac45abc0 PB |
4558 | /* sync_file_range always requires a blocking context */ |
4559 | if (force_nonblock) | |
4560 | return -EAGAIN; | |
4561 | ||
9adbd45d | 4562 | ret = sync_file_range(req->file, req->sync.off, req->sync.len, |
8ed8d3c3 JA |
4563 | req->sync.flags); |
4564 | if (ret < 0) | |
4565 | req_set_fail_links(req); | |
e1e16097 | 4566 | io_req_complete(req, ret); |
5d17b4a4 JA |
4567 | return 0; |
4568 | } | |
4569 | ||
469956e8 | 4570 | #if defined(CONFIG_NET) |
02d27d89 PB |
4571 | static int io_setup_async_msg(struct io_kiocb *req, |
4572 | struct io_async_msghdr *kmsg) | |
4573 | { | |
e8c2bc1f JA |
4574 | struct io_async_msghdr *async_msg = req->async_data; |
4575 | ||
4576 | if (async_msg) | |
02d27d89 | 4577 | return -EAGAIN; |
e8c2bc1f | 4578 | if (io_alloc_async_data(req)) { |
02d27d89 PB |
4579 | if (kmsg->iov != kmsg->fast_iov) |
4580 | kfree(kmsg->iov); | |
4581 | return -ENOMEM; | |
4582 | } | |
e8c2bc1f | 4583 | async_msg = req->async_data; |
02d27d89 | 4584 | req->flags |= REQ_F_NEED_CLEANUP; |
e8c2bc1f | 4585 | memcpy(async_msg, kmsg, sizeof(*kmsg)); |
02d27d89 PB |
4586 | return -EAGAIN; |
4587 | } | |
4588 | ||
2ae523ed PB |
4589 | static int io_sendmsg_copy_hdr(struct io_kiocb *req, |
4590 | struct io_async_msghdr *iomsg) | |
4591 | { | |
4592 | iomsg->iov = iomsg->fast_iov; | |
4593 | iomsg->msg.msg_name = &iomsg->addr; | |
4594 | return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg, | |
4595 | req->sr_msg.msg_flags, &iomsg->iov); | |
4596 | } | |
4597 | ||
3529d8c2 | 4598 | static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
03b1230c | 4599 | { |
e8c2bc1f | 4600 | struct io_async_msghdr *async_msg = req->async_data; |
e47293fd | 4601 | struct io_sr_msg *sr = &req->sr_msg; |
99bc4c38 | 4602 | int ret; |
03b1230c | 4603 | |
d2b6f48b PB |
4604 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4605 | return -EINVAL; | |
4606 | ||
e47293fd | 4607 | sr->msg_flags = READ_ONCE(sqe->msg_flags); |
270a5940 | 4608 | sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
fddaface | 4609 | sr->len = READ_ONCE(sqe->len); |
3529d8c2 | 4610 | |
d8768362 JA |
4611 | #ifdef CONFIG_COMPAT |
4612 | if (req->ctx->compat) | |
4613 | sr->msg_flags |= MSG_CMSG_COMPAT; | |
4614 | #endif | |
4615 | ||
e8c2bc1f | 4616 | if (!async_msg || !io_op_defs[req->opcode].needs_async_data) |
3529d8c2 | 4617 | return 0; |
e8c2bc1f | 4618 | ret = io_sendmsg_copy_hdr(req, async_msg); |
99bc4c38 PB |
4619 | if (!ret) |
4620 | req->flags |= REQ_F_NEED_CLEANUP; | |
4621 | return ret; | |
03b1230c JA |
4622 | } |
4623 | ||
229a7b63 JA |
4624 | static int io_sendmsg(struct io_kiocb *req, bool force_nonblock, |
4625 | struct io_comp_state *cs) | |
aa1fa28f | 4626 | { |
6b754c8b | 4627 | struct io_async_msghdr iomsg, *kmsg; |
0fa03c62 | 4628 | struct socket *sock; |
7a7cacba | 4629 | unsigned flags; |
0fa03c62 JA |
4630 | int ret; |
4631 | ||
dba4a925 | 4632 | sock = sock_from_file(req->file); |
7a7cacba | 4633 | if (unlikely(!sock)) |
dba4a925 | 4634 | return -ENOTSOCK; |
3529d8c2 | 4635 | |
e8c2bc1f JA |
4636 | if (req->async_data) { |
4637 | kmsg = req->async_data; | |
4638 | kmsg->msg.msg_name = &kmsg->addr; | |
7a7cacba PB |
4639 | /* if iov is set, it's allocated already */ |
4640 | if (!kmsg->iov) | |
4641 | kmsg->iov = kmsg->fast_iov; | |
4642 | kmsg->msg.msg_iter.iov = kmsg->iov; | |
4643 | } else { | |
4644 | ret = io_sendmsg_copy_hdr(req, &iomsg); | |
4645 | if (ret) | |
4646 | return ret; | |
4647 | kmsg = &iomsg; | |
0fa03c62 | 4648 | } |
0fa03c62 | 4649 | |
7a7cacba PB |
4650 | flags = req->sr_msg.msg_flags; |
4651 | if (flags & MSG_DONTWAIT) | |
4652 | req->flags |= REQ_F_NOWAIT; | |
4653 | else if (force_nonblock) | |
4654 | flags |= MSG_DONTWAIT; | |
e47293fd | 4655 | |
7a7cacba PB |
4656 | ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags); |
4657 | if (force_nonblock && ret == -EAGAIN) | |
4658 | return io_setup_async_msg(req, kmsg); | |
4659 | if (ret == -ERESTARTSYS) | |
4660 | ret = -EINTR; | |
0fa03c62 | 4661 | |
6b754c8b | 4662 | if (kmsg->iov != kmsg->fast_iov) |
0b416c3e | 4663 | kfree(kmsg->iov); |
99bc4c38 | 4664 | req->flags &= ~REQ_F_NEED_CLEANUP; |
4e88d6e7 JA |
4665 | if (ret < 0) |
4666 | req_set_fail_links(req); | |
229a7b63 | 4667 | __io_req_complete(req, ret, 0, cs); |
5d17b4a4 | 4668 | return 0; |
03b1230c | 4669 | } |
aa1fa28f | 4670 | |
229a7b63 JA |
4671 | static int io_send(struct io_kiocb *req, bool force_nonblock, |
4672 | struct io_comp_state *cs) | |
fddaface | 4673 | { |
7a7cacba PB |
4674 | struct io_sr_msg *sr = &req->sr_msg; |
4675 | struct msghdr msg; | |
4676 | struct iovec iov; | |
fddaface | 4677 | struct socket *sock; |
7a7cacba | 4678 | unsigned flags; |
fddaface JA |
4679 | int ret; |
4680 | ||
dba4a925 | 4681 | sock = sock_from_file(req->file); |
7a7cacba | 4682 | if (unlikely(!sock)) |
dba4a925 | 4683 | return -ENOTSOCK; |
fddaface | 4684 | |
7a7cacba PB |
4685 | ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter); |
4686 | if (unlikely(ret)) | |
14db8411 | 4687 | return ret; |
fddaface | 4688 | |
7a7cacba PB |
4689 | msg.msg_name = NULL; |
4690 | msg.msg_control = NULL; | |
4691 | msg.msg_controllen = 0; | |
4692 | msg.msg_namelen = 0; | |
fddaface | 4693 | |
7a7cacba PB |
4694 | flags = req->sr_msg.msg_flags; |
4695 | if (flags & MSG_DONTWAIT) | |
4696 | req->flags |= REQ_F_NOWAIT; | |
4697 | else if (force_nonblock) | |
4698 | flags |= MSG_DONTWAIT; | |
fddaface | 4699 | |
7a7cacba PB |
4700 | msg.msg_flags = flags; |
4701 | ret = sock_sendmsg(sock, &msg); | |
4702 | if (force_nonblock && ret == -EAGAIN) | |
4703 | return -EAGAIN; | |
4704 | if (ret == -ERESTARTSYS) | |
4705 | ret = -EINTR; | |
fddaface | 4706 | |
fddaface JA |
4707 | if (ret < 0) |
4708 | req_set_fail_links(req); | |
229a7b63 | 4709 | __io_req_complete(req, ret, 0, cs); |
fddaface | 4710 | return 0; |
fddaface JA |
4711 | } |
4712 | ||
1400e697 PB |
4713 | static int __io_recvmsg_copy_hdr(struct io_kiocb *req, |
4714 | struct io_async_msghdr *iomsg) | |
52de1fe1 JA |
4715 | { |
4716 | struct io_sr_msg *sr = &req->sr_msg; | |
4717 | struct iovec __user *uiov; | |
4718 | size_t iov_len; | |
4719 | int ret; | |
4720 | ||
1400e697 PB |
4721 | ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg, |
4722 | &iomsg->uaddr, &uiov, &iov_len); | |
52de1fe1 JA |
4723 | if (ret) |
4724 | return ret; | |
4725 | ||
4726 | if (req->flags & REQ_F_BUFFER_SELECT) { | |
4727 | if (iov_len > 1) | |
4728 | return -EINVAL; | |
1400e697 | 4729 | if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov))) |
52de1fe1 | 4730 | return -EFAULT; |
1400e697 PB |
4731 | sr->len = iomsg->iov[0].iov_len; |
4732 | iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1, | |
52de1fe1 | 4733 | sr->len); |
1400e697 | 4734 | iomsg->iov = NULL; |
52de1fe1 | 4735 | } else { |
89cd35c5 CH |
4736 | ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV, |
4737 | &iomsg->iov, &iomsg->msg.msg_iter, | |
4738 | false); | |
52de1fe1 JA |
4739 | if (ret > 0) |
4740 | ret = 0; | |
4741 | } | |
4742 | ||
4743 | return ret; | |
4744 | } | |
4745 | ||
4746 | #ifdef CONFIG_COMPAT | |
4747 | static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req, | |
1400e697 | 4748 | struct io_async_msghdr *iomsg) |
52de1fe1 JA |
4749 | { |
4750 | struct compat_msghdr __user *msg_compat; | |
4751 | struct io_sr_msg *sr = &req->sr_msg; | |
4752 | struct compat_iovec __user *uiov; | |
4753 | compat_uptr_t ptr; | |
4754 | compat_size_t len; | |
4755 | int ret; | |
4756 | ||
270a5940 | 4757 | msg_compat = (struct compat_msghdr __user *) sr->umsg; |
1400e697 | 4758 | ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr, |
52de1fe1 JA |
4759 | &ptr, &len); |
4760 | if (ret) | |
4761 | return ret; | |
4762 | ||
4763 | uiov = compat_ptr(ptr); | |
4764 | if (req->flags & REQ_F_BUFFER_SELECT) { | |
4765 | compat_ssize_t clen; | |
4766 | ||
4767 | if (len > 1) | |
4768 | return -EINVAL; | |
4769 | if (!access_ok(uiov, sizeof(*uiov))) | |
4770 | return -EFAULT; | |
4771 | if (__get_user(clen, &uiov->iov_len)) | |
4772 | return -EFAULT; | |
4773 | if (clen < 0) | |
4774 | return -EINVAL; | |
2d280bc8 PB |
4775 | sr->len = clen; |
4776 | iomsg->iov[0].iov_len = clen; | |
1400e697 | 4777 | iomsg->iov = NULL; |
52de1fe1 | 4778 | } else { |
89cd35c5 CH |
4779 | ret = __import_iovec(READ, (struct iovec __user *)uiov, len, |
4780 | UIO_FASTIOV, &iomsg->iov, | |
4781 | &iomsg->msg.msg_iter, true); | |
52de1fe1 JA |
4782 | if (ret < 0) |
4783 | return ret; | |
4784 | } | |
4785 | ||
4786 | return 0; | |
4787 | } | |
4788 | #endif | |
4789 | ||
1400e697 PB |
4790 | static int io_recvmsg_copy_hdr(struct io_kiocb *req, |
4791 | struct io_async_msghdr *iomsg) | |
52de1fe1 | 4792 | { |
1400e697 PB |
4793 | iomsg->msg.msg_name = &iomsg->addr; |
4794 | iomsg->iov = iomsg->fast_iov; | |
52de1fe1 JA |
4795 | |
4796 | #ifdef CONFIG_COMPAT | |
4797 | if (req->ctx->compat) | |
1400e697 | 4798 | return __io_compat_recvmsg_copy_hdr(req, iomsg); |
fddaface | 4799 | #endif |
52de1fe1 | 4800 | |
1400e697 | 4801 | return __io_recvmsg_copy_hdr(req, iomsg); |
52de1fe1 JA |
4802 | } |
4803 | ||
bcda7baa | 4804 | static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req, |
7fbb1b54 | 4805 | bool needs_lock) |
bcda7baa JA |
4806 | { |
4807 | struct io_sr_msg *sr = &req->sr_msg; | |
4808 | struct io_buffer *kbuf; | |
4809 | ||
bcda7baa JA |
4810 | kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock); |
4811 | if (IS_ERR(kbuf)) | |
4812 | return kbuf; | |
4813 | ||
4814 | sr->kbuf = kbuf; | |
4815 | req->flags |= REQ_F_BUFFER_SELECTED; | |
bcda7baa | 4816 | return kbuf; |
fddaface JA |
4817 | } |
4818 | ||
7fbb1b54 PB |
4819 | static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req) |
4820 | { | |
4821 | return io_put_kbuf(req, req->sr_msg.kbuf); | |
4822 | } | |
4823 | ||
3529d8c2 JA |
4824 | static int io_recvmsg_prep(struct io_kiocb *req, |
4825 | const struct io_uring_sqe *sqe) | |
aa1fa28f | 4826 | { |
e8c2bc1f | 4827 | struct io_async_msghdr *async_msg = req->async_data; |
e47293fd | 4828 | struct io_sr_msg *sr = &req->sr_msg; |
99bc4c38 | 4829 | int ret; |
3529d8c2 | 4830 | |
d2b6f48b PB |
4831 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4832 | return -EINVAL; | |
4833 | ||
3529d8c2 | 4834 | sr->msg_flags = READ_ONCE(sqe->msg_flags); |
270a5940 | 4835 | sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
0b7b21e4 | 4836 | sr->len = READ_ONCE(sqe->len); |
bcda7baa | 4837 | sr->bgid = READ_ONCE(sqe->buf_group); |
06b76d44 | 4838 | |
d8768362 JA |
4839 | #ifdef CONFIG_COMPAT |
4840 | if (req->ctx->compat) | |
4841 | sr->msg_flags |= MSG_CMSG_COMPAT; | |
4842 | #endif | |
4843 | ||
e8c2bc1f | 4844 | if (!async_msg || !io_op_defs[req->opcode].needs_async_data) |
5f798bea | 4845 | return 0; |
e8c2bc1f | 4846 | ret = io_recvmsg_copy_hdr(req, async_msg); |
99bc4c38 PB |
4847 | if (!ret) |
4848 | req->flags |= REQ_F_NEED_CLEANUP; | |
4849 | return ret; | |
aa1fa28f JA |
4850 | } |
4851 | ||
229a7b63 JA |
4852 | static int io_recvmsg(struct io_kiocb *req, bool force_nonblock, |
4853 | struct io_comp_state *cs) | |
aa1fa28f | 4854 | { |
6b754c8b | 4855 | struct io_async_msghdr iomsg, *kmsg; |
03b1230c | 4856 | struct socket *sock; |
7fbb1b54 | 4857 | struct io_buffer *kbuf; |
7a7cacba | 4858 | unsigned flags; |
52de1fe1 | 4859 | int ret, cflags = 0; |
03b1230c | 4860 | |
dba4a925 | 4861 | sock = sock_from_file(req->file); |
7a7cacba | 4862 | if (unlikely(!sock)) |
dba4a925 | 4863 | return -ENOTSOCK; |
3529d8c2 | 4864 | |
e8c2bc1f JA |
4865 | if (req->async_data) { |
4866 | kmsg = req->async_data; | |
4867 | kmsg->msg.msg_name = &kmsg->addr; | |
7a7cacba PB |
4868 | /* if iov is set, it's allocated already */ |
4869 | if (!kmsg->iov) | |
4870 | kmsg->iov = kmsg->fast_iov; | |
4871 | kmsg->msg.msg_iter.iov = kmsg->iov; | |
4872 | } else { | |
4873 | ret = io_recvmsg_copy_hdr(req, &iomsg); | |
4874 | if (ret) | |
681fda8d | 4875 | return ret; |
7a7cacba PB |
4876 | kmsg = &iomsg; |
4877 | } | |
03b1230c | 4878 | |
bc02ef33 | 4879 | if (req->flags & REQ_F_BUFFER_SELECT) { |
7fbb1b54 | 4880 | kbuf = io_recv_buffer_select(req, !force_nonblock); |
bc02ef33 | 4881 | if (IS_ERR(kbuf)) |
52de1fe1 | 4882 | return PTR_ERR(kbuf); |
7a7cacba PB |
4883 | kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr); |
4884 | iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov, | |
4885 | 1, req->sr_msg.len); | |
4886 | } | |
52de1fe1 | 4887 | |
7a7cacba PB |
4888 | flags = req->sr_msg.msg_flags; |
4889 | if (flags & MSG_DONTWAIT) | |
4890 | req->flags |= REQ_F_NOWAIT; | |
4891 | else if (force_nonblock) | |
4892 | flags |= MSG_DONTWAIT; | |
e47293fd | 4893 | |
7a7cacba PB |
4894 | ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg, |
4895 | kmsg->uaddr, flags); | |
0e1b6fe3 PB |
4896 | if (force_nonblock && ret == -EAGAIN) |
4897 | return io_setup_async_msg(req, kmsg); | |
7a7cacba PB |
4898 | if (ret == -ERESTARTSYS) |
4899 | ret = -EINTR; | |
03b1230c | 4900 | |
7fbb1b54 PB |
4901 | if (req->flags & REQ_F_BUFFER_SELECTED) |
4902 | cflags = io_put_recv_kbuf(req); | |
6b754c8b | 4903 | if (kmsg->iov != kmsg->fast_iov) |
0b416c3e | 4904 | kfree(kmsg->iov); |
99bc4c38 | 4905 | req->flags &= ~REQ_F_NEED_CLEANUP; |
4e88d6e7 JA |
4906 | if (ret < 0) |
4907 | req_set_fail_links(req); | |
229a7b63 | 4908 | __io_req_complete(req, ret, cflags, cs); |
03b1230c | 4909 | return 0; |
0fa03c62 | 4910 | } |
5d17b4a4 | 4911 | |
229a7b63 JA |
4912 | static int io_recv(struct io_kiocb *req, bool force_nonblock, |
4913 | struct io_comp_state *cs) | |
fddaface | 4914 | { |
6b754c8b | 4915 | struct io_buffer *kbuf; |
7a7cacba PB |
4916 | struct io_sr_msg *sr = &req->sr_msg; |
4917 | struct msghdr msg; | |
4918 | void __user *buf = sr->buf; | |
fddaface | 4919 | struct socket *sock; |
7a7cacba PB |
4920 | struct iovec iov; |
4921 | unsigned flags; | |
bcda7baa | 4922 | int ret, cflags = 0; |
fddaface | 4923 | |
dba4a925 | 4924 | sock = sock_from_file(req->file); |
7a7cacba | 4925 | if (unlikely(!sock)) |
dba4a925 | 4926 | return -ENOTSOCK; |
fddaface | 4927 | |
bc02ef33 | 4928 | if (req->flags & REQ_F_BUFFER_SELECT) { |
7fbb1b54 | 4929 | kbuf = io_recv_buffer_select(req, !force_nonblock); |
bcda7baa JA |
4930 | if (IS_ERR(kbuf)) |
4931 | return PTR_ERR(kbuf); | |
7a7cacba | 4932 | buf = u64_to_user_ptr(kbuf->addr); |
bc02ef33 | 4933 | } |
bcda7baa | 4934 | |
7a7cacba | 4935 | ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter); |
14c32eee PB |
4936 | if (unlikely(ret)) |
4937 | goto out_free; | |
fddaface | 4938 | |
7a7cacba PB |
4939 | msg.msg_name = NULL; |
4940 | msg.msg_control = NULL; | |
4941 | msg.msg_controllen = 0; | |
4942 | msg.msg_namelen = 0; | |
4943 | msg.msg_iocb = NULL; | |
4944 | msg.msg_flags = 0; | |
fddaface | 4945 | |
7a7cacba PB |
4946 | flags = req->sr_msg.msg_flags; |
4947 | if (flags & MSG_DONTWAIT) | |
4948 | req->flags |= REQ_F_NOWAIT; | |
4949 | else if (force_nonblock) | |
4950 | flags |= MSG_DONTWAIT; | |
4951 | ||
4952 | ret = sock_recvmsg(sock, &msg, flags); | |
4953 | if (force_nonblock && ret == -EAGAIN) | |
4954 | return -EAGAIN; | |
4955 | if (ret == -ERESTARTSYS) | |
4956 | ret = -EINTR; | |
14c32eee | 4957 | out_free: |
7fbb1b54 PB |
4958 | if (req->flags & REQ_F_BUFFER_SELECTED) |
4959 | cflags = io_put_recv_kbuf(req); | |
fddaface JA |
4960 | if (ret < 0) |
4961 | req_set_fail_links(req); | |
229a7b63 | 4962 | __io_req_complete(req, ret, cflags, cs); |
fddaface | 4963 | return 0; |
fddaface JA |
4964 | } |
4965 | ||
3529d8c2 | 4966 | static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
17f2fe35 | 4967 | { |
8ed8d3c3 JA |
4968 | struct io_accept *accept = &req->accept; |
4969 | ||
14587a46 | 4970 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
17f2fe35 | 4971 | return -EINVAL; |
8042d6ce | 4972 | if (sqe->ioprio || sqe->len || sqe->buf_index) |
17f2fe35 JA |
4973 | return -EINVAL; |
4974 | ||
d55e5f5b JA |
4975 | accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
4976 | accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2)); | |
8ed8d3c3 | 4977 | accept->flags = READ_ONCE(sqe->accept_flags); |
09952e3e | 4978 | accept->nofile = rlimit(RLIMIT_NOFILE); |
8ed8d3c3 | 4979 | return 0; |
8ed8d3c3 | 4980 | } |
17f2fe35 | 4981 | |
229a7b63 JA |
4982 | static int io_accept(struct io_kiocb *req, bool force_nonblock, |
4983 | struct io_comp_state *cs) | |
8ed8d3c3 JA |
4984 | { |
4985 | struct io_accept *accept = &req->accept; | |
ac45abc0 | 4986 | unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0; |
8ed8d3c3 JA |
4987 | int ret; |
4988 | ||
e697deed JX |
4989 | if (req->file->f_flags & O_NONBLOCK) |
4990 | req->flags |= REQ_F_NOWAIT; | |
4991 | ||
8ed8d3c3 | 4992 | ret = __sys_accept4_file(req->file, file_flags, accept->addr, |
09952e3e JA |
4993 | accept->addr_len, accept->flags, |
4994 | accept->nofile); | |
8ed8d3c3 | 4995 | if (ret == -EAGAIN && force_nonblock) |
17f2fe35 | 4996 | return -EAGAIN; |
ac45abc0 PB |
4997 | if (ret < 0) { |
4998 | if (ret == -ERESTARTSYS) | |
4999 | ret = -EINTR; | |
4e88d6e7 | 5000 | req_set_fail_links(req); |
ac45abc0 | 5001 | } |
229a7b63 | 5002 | __io_req_complete(req, ret, 0, cs); |
17f2fe35 | 5003 | return 0; |
8ed8d3c3 JA |
5004 | } |
5005 | ||
3529d8c2 | 5006 | static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f499a021 | 5007 | { |
3529d8c2 | 5008 | struct io_connect *conn = &req->connect; |
e8c2bc1f | 5009 | struct io_async_connect *io = req->async_data; |
f499a021 | 5010 | |
14587a46 | 5011 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3fbb51c1 JA |
5012 | return -EINVAL; |
5013 | if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags) | |
5014 | return -EINVAL; | |
5015 | ||
3529d8c2 JA |
5016 | conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
5017 | conn->addr_len = READ_ONCE(sqe->addr2); | |
5018 | ||
5019 | if (!io) | |
5020 | return 0; | |
5021 | ||
5022 | return move_addr_to_kernel(conn->addr, conn->addr_len, | |
e8c2bc1f | 5023 | &io->address); |
f499a021 JA |
5024 | } |
5025 | ||
229a7b63 JA |
5026 | static int io_connect(struct io_kiocb *req, bool force_nonblock, |
5027 | struct io_comp_state *cs) | |
f8e85cf2 | 5028 | { |
e8c2bc1f | 5029 | struct io_async_connect __io, *io; |
f8e85cf2 | 5030 | unsigned file_flags; |
3fbb51c1 | 5031 | int ret; |
f8e85cf2 | 5032 | |
e8c2bc1f JA |
5033 | if (req->async_data) { |
5034 | io = req->async_data; | |
f499a021 | 5035 | } else { |
3529d8c2 JA |
5036 | ret = move_addr_to_kernel(req->connect.addr, |
5037 | req->connect.addr_len, | |
e8c2bc1f | 5038 | &__io.address); |
f499a021 JA |
5039 | if (ret) |
5040 | goto out; | |
5041 | io = &__io; | |
5042 | } | |
5043 | ||
3fbb51c1 JA |
5044 | file_flags = force_nonblock ? O_NONBLOCK : 0; |
5045 | ||
e8c2bc1f | 5046 | ret = __sys_connect_file(req->file, &io->address, |
3fbb51c1 | 5047 | req->connect.addr_len, file_flags); |
87f80d62 | 5048 | if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) { |
e8c2bc1f | 5049 | if (req->async_data) |
b7bb4f7d | 5050 | return -EAGAIN; |
e8c2bc1f | 5051 | if (io_alloc_async_data(req)) { |
f499a021 JA |
5052 | ret = -ENOMEM; |
5053 | goto out; | |
5054 | } | |
e8c2bc1f JA |
5055 | io = req->async_data; |
5056 | memcpy(req->async_data, &__io, sizeof(__io)); | |
f8e85cf2 | 5057 | return -EAGAIN; |
f499a021 | 5058 | } |
f8e85cf2 JA |
5059 | if (ret == -ERESTARTSYS) |
5060 | ret = -EINTR; | |
f499a021 | 5061 | out: |
4e88d6e7 JA |
5062 | if (ret < 0) |
5063 | req_set_fail_links(req); | |
229a7b63 | 5064 | __io_req_complete(req, ret, 0, cs); |
f8e85cf2 | 5065 | return 0; |
469956e8 Y |
5066 | } |
5067 | #else /* !CONFIG_NET */ | |
5068 | static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
5069 | { | |
f8e85cf2 | 5070 | return -EOPNOTSUPP; |
f8e85cf2 JA |
5071 | } |
5072 | ||
1e16c2f9 RD |
5073 | static int io_sendmsg(struct io_kiocb *req, bool force_nonblock, |
5074 | struct io_comp_state *cs) | |
469956e8 Y |
5075 | { |
5076 | return -EOPNOTSUPP; | |
5077 | } | |
5078 | ||
1e16c2f9 RD |
5079 | static int io_send(struct io_kiocb *req, bool force_nonblock, |
5080 | struct io_comp_state *cs) | |
469956e8 Y |
5081 | { |
5082 | return -EOPNOTSUPP; | |
5083 | } | |
5084 | ||
5085 | static int io_recvmsg_prep(struct io_kiocb *req, | |
5086 | const struct io_uring_sqe *sqe) | |
5087 | { | |
5088 | return -EOPNOTSUPP; | |
5089 | } | |
5090 | ||
1e16c2f9 RD |
5091 | static int io_recvmsg(struct io_kiocb *req, bool force_nonblock, |
5092 | struct io_comp_state *cs) | |
469956e8 Y |
5093 | { |
5094 | return -EOPNOTSUPP; | |
5095 | } | |
5096 | ||
1e16c2f9 RD |
5097 | static int io_recv(struct io_kiocb *req, bool force_nonblock, |
5098 | struct io_comp_state *cs) | |
469956e8 Y |
5099 | { |
5100 | return -EOPNOTSUPP; | |
5101 | } | |
5102 | ||
5103 | static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
5104 | { | |
5105 | return -EOPNOTSUPP; | |
5106 | } | |
5107 | ||
1e16c2f9 RD |
5108 | static int io_accept(struct io_kiocb *req, bool force_nonblock, |
5109 | struct io_comp_state *cs) | |
469956e8 Y |
5110 | { |
5111 | return -EOPNOTSUPP; | |
5112 | } | |
ce593a6c | 5113 | |
469956e8 Y |
5114 | static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
5115 | { | |
5116 | return -EOPNOTSUPP; | |
5117 | } | |
5118 | ||
1e16c2f9 RD |
5119 | static int io_connect(struct io_kiocb *req, bool force_nonblock, |
5120 | struct io_comp_state *cs) | |
469956e8 | 5121 | { |
f8e85cf2 | 5122 | return -EOPNOTSUPP; |
ce593a6c | 5123 | } |
469956e8 | 5124 | #endif /* CONFIG_NET */ |
f8e85cf2 | 5125 | |
d7718a9d JA |
5126 | struct io_poll_table { |
5127 | struct poll_table_struct pt; | |
5128 | struct io_kiocb *req; | |
5129 | int error; | |
5130 | }; | |
ce593a6c | 5131 | |
d7718a9d JA |
5132 | static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll, |
5133 | __poll_t mask, task_work_func_t func) | |
5134 | { | |
aa96bf8a | 5135 | int ret; |
d7718a9d JA |
5136 | |
5137 | /* for instances that support it check for an event match first: */ | |
5138 | if (mask && !(mask & poll->events)) | |
5139 | return 0; | |
5140 | ||
5141 | trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask); | |
5142 | ||
5143 | list_del_init(&poll->wait.entry); | |
5144 | ||
d7718a9d JA |
5145 | req->result = mask; |
5146 | init_task_work(&req->task_work, func); | |
6d816e08 JA |
5147 | percpu_ref_get(&req->ctx->refs); |
5148 | ||
d7718a9d | 5149 | /* |
e3aabf95 JA |
5150 | * If this fails, then the task is exiting. When a task exits, the |
5151 | * work gets canceled, so just cancel this request as well instead | |
5152 | * of executing it. We can't safely execute it anyway, as we may not | |
5153 | * have the needed state needed for it anyway. | |
d7718a9d | 5154 | */ |
355fb9e2 | 5155 | ret = io_req_task_work_add(req); |
aa96bf8a | 5156 | if (unlikely(ret)) { |
c2c4c83c JA |
5157 | struct task_struct *tsk; |
5158 | ||
e3aabf95 | 5159 | WRITE_ONCE(poll->canceled, true); |
aa96bf8a | 5160 | tsk = io_wq_get_task(req->ctx->io_wq); |
91989c70 | 5161 | task_work_add(tsk, &req->task_work, TWA_NONE); |
ce593a6c | 5162 | wake_up_process(tsk); |
aa96bf8a | 5163 | } |
d7718a9d JA |
5164 | return 1; |
5165 | } | |
5166 | ||
74ce6ce4 JA |
5167 | static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll) |
5168 | __acquires(&req->ctx->completion_lock) | |
5169 | { | |
5170 | struct io_ring_ctx *ctx = req->ctx; | |
5171 | ||
5172 | if (!req->result && !READ_ONCE(poll->canceled)) { | |
5173 | struct poll_table_struct pt = { ._key = poll->events }; | |
5174 | ||
5175 | req->result = vfs_poll(req->file, &pt) & poll->events; | |
5176 | } | |
5177 | ||
5178 | spin_lock_irq(&ctx->completion_lock); | |
5179 | if (!req->result && !READ_ONCE(poll->canceled)) { | |
5180 | add_wait_queue(poll->head, &poll->wait); | |
5181 | return true; | |
5182 | } | |
5183 | ||
5184 | return false; | |
5185 | } | |
5186 | ||
d4e7cd36 | 5187 | static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req) |
18bceab1 | 5188 | { |
e8c2bc1f | 5189 | /* pure poll stashes this in ->async_data, poll driven retry elsewhere */ |
d4e7cd36 | 5190 | if (req->opcode == IORING_OP_POLL_ADD) |
e8c2bc1f | 5191 | return req->async_data; |
d4e7cd36 JA |
5192 | return req->apoll->double_poll; |
5193 | } | |
5194 | ||
5195 | static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req) | |
5196 | { | |
5197 | if (req->opcode == IORING_OP_POLL_ADD) | |
5198 | return &req->poll; | |
5199 | return &req->apoll->poll; | |
5200 | } | |
5201 | ||
5202 | static void io_poll_remove_double(struct io_kiocb *req) | |
5203 | { | |
5204 | struct io_poll_iocb *poll = io_poll_get_double(req); | |
18bceab1 JA |
5205 | |
5206 | lockdep_assert_held(&req->ctx->completion_lock); | |
5207 | ||
5208 | if (poll && poll->head) { | |
5209 | struct wait_queue_head *head = poll->head; | |
5210 | ||
5211 | spin_lock(&head->lock); | |
5212 | list_del_init(&poll->wait.entry); | |
5213 | if (poll->wait.private) | |
5214 | refcount_dec(&req->refs); | |
5215 | poll->head = NULL; | |
5216 | spin_unlock(&head->lock); | |
5217 | } | |
5218 | } | |
5219 | ||
5220 | static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error) | |
5221 | { | |
5222 | struct io_ring_ctx *ctx = req->ctx; | |
5223 | ||
d4e7cd36 | 5224 | io_poll_remove_double(req); |
18bceab1 JA |
5225 | req->poll.done = true; |
5226 | io_cqring_fill_event(req, error ? error : mangle_poll(mask)); | |
5227 | io_commit_cqring(ctx); | |
5228 | } | |
5229 | ||
dd221f46 | 5230 | static void io_poll_task_func(struct callback_head *cb) |
18bceab1 | 5231 | { |
dd221f46 | 5232 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); |
18bceab1 | 5233 | struct io_ring_ctx *ctx = req->ctx; |
dd221f46 | 5234 | struct io_kiocb *nxt; |
18bceab1 JA |
5235 | |
5236 | if (io_poll_rewait(req, &req->poll)) { | |
5237 | spin_unlock_irq(&ctx->completion_lock); | |
dd221f46 PB |
5238 | } else { |
5239 | hash_del(&req->hash_node); | |
5240 | io_poll_complete(req, req->result, 0); | |
5241 | spin_unlock_irq(&ctx->completion_lock); | |
18bceab1 | 5242 | |
dd221f46 PB |
5243 | nxt = io_put_req_find_next(req); |
5244 | io_cqring_ev_posted(ctx); | |
5245 | if (nxt) | |
5246 | __io_req_task_submit(nxt); | |
5247 | } | |
18bceab1 | 5248 | |
6d816e08 | 5249 | percpu_ref_put(&ctx->refs); |
18bceab1 JA |
5250 | } |
5251 | ||
5252 | static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode, | |
5253 | int sync, void *key) | |
5254 | { | |
5255 | struct io_kiocb *req = wait->private; | |
d4e7cd36 | 5256 | struct io_poll_iocb *poll = io_poll_get_single(req); |
18bceab1 JA |
5257 | __poll_t mask = key_to_poll(key); |
5258 | ||
5259 | /* for instances that support it check for an event match first: */ | |
5260 | if (mask && !(mask & poll->events)) | |
5261 | return 0; | |
5262 | ||
8706e04e JA |
5263 | list_del_init(&wait->entry); |
5264 | ||
807abcb0 | 5265 | if (poll && poll->head) { |
18bceab1 JA |
5266 | bool done; |
5267 | ||
807abcb0 JA |
5268 | spin_lock(&poll->head->lock); |
5269 | done = list_empty(&poll->wait.entry); | |
18bceab1 | 5270 | if (!done) |
807abcb0 | 5271 | list_del_init(&poll->wait.entry); |
d4e7cd36 JA |
5272 | /* make sure double remove sees this as being gone */ |
5273 | wait->private = NULL; | |
807abcb0 | 5274 | spin_unlock(&poll->head->lock); |
c8b5e260 JA |
5275 | if (!done) { |
5276 | /* use wait func handler, so it matches the rq type */ | |
5277 | poll->wait.func(&poll->wait, mode, sync, key); | |
5278 | } | |
18bceab1 JA |
5279 | } |
5280 | refcount_dec(&req->refs); | |
5281 | return 1; | |
5282 | } | |
5283 | ||
5284 | static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events, | |
5285 | wait_queue_func_t wake_func) | |
5286 | { | |
5287 | poll->head = NULL; | |
5288 | poll->done = false; | |
5289 | poll->canceled = false; | |
5290 | poll->events = events; | |
5291 | INIT_LIST_HEAD(&poll->wait.entry); | |
5292 | init_waitqueue_func_entry(&poll->wait, wake_func); | |
5293 | } | |
5294 | ||
5295 | static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt, | |
807abcb0 JA |
5296 | struct wait_queue_head *head, |
5297 | struct io_poll_iocb **poll_ptr) | |
18bceab1 JA |
5298 | { |
5299 | struct io_kiocb *req = pt->req; | |
5300 | ||
5301 | /* | |
5302 | * If poll->head is already set, it's because the file being polled | |
5303 | * uses multiple waitqueues for poll handling (eg one for read, one | |
5304 | * for write). Setup a separate io_poll_iocb if this happens. | |
5305 | */ | |
5306 | if (unlikely(poll->head)) { | |
58852d4d PB |
5307 | struct io_poll_iocb *poll_one = poll; |
5308 | ||
18bceab1 | 5309 | /* already have a 2nd entry, fail a third attempt */ |
807abcb0 | 5310 | if (*poll_ptr) { |
18bceab1 JA |
5311 | pt->error = -EINVAL; |
5312 | return; | |
5313 | } | |
5314 | poll = kmalloc(sizeof(*poll), GFP_ATOMIC); | |
5315 | if (!poll) { | |
5316 | pt->error = -ENOMEM; | |
5317 | return; | |
5318 | } | |
58852d4d | 5319 | io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake); |
18bceab1 JA |
5320 | refcount_inc(&req->refs); |
5321 | poll->wait.private = req; | |
807abcb0 | 5322 | *poll_ptr = poll; |
18bceab1 JA |
5323 | } |
5324 | ||
5325 | pt->error = 0; | |
5326 | poll->head = head; | |
a31eb4a2 JX |
5327 | |
5328 | if (poll->events & EPOLLEXCLUSIVE) | |
5329 | add_wait_queue_exclusive(head, &poll->wait); | |
5330 | else | |
5331 | add_wait_queue(head, &poll->wait); | |
18bceab1 JA |
5332 | } |
5333 | ||
5334 | static void io_async_queue_proc(struct file *file, struct wait_queue_head *head, | |
5335 | struct poll_table_struct *p) | |
5336 | { | |
5337 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
807abcb0 | 5338 | struct async_poll *apoll = pt->req->apoll; |
18bceab1 | 5339 | |
807abcb0 | 5340 | __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll); |
18bceab1 JA |
5341 | } |
5342 | ||
d7718a9d JA |
5343 | static void io_async_task_func(struct callback_head *cb) |
5344 | { | |
5345 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); | |
5346 | struct async_poll *apoll = req->apoll; | |
5347 | struct io_ring_ctx *ctx = req->ctx; | |
5348 | ||
5349 | trace_io_uring_task_run(req->ctx, req->opcode, req->user_data); | |
5350 | ||
74ce6ce4 | 5351 | if (io_poll_rewait(req, &apoll->poll)) { |
d7718a9d | 5352 | spin_unlock_irq(&ctx->completion_lock); |
6d816e08 | 5353 | percpu_ref_put(&ctx->refs); |
74ce6ce4 | 5354 | return; |
d7718a9d JA |
5355 | } |
5356 | ||
31067255 | 5357 | /* If req is still hashed, it cannot have been canceled. Don't check. */ |
0be0b0e3 | 5358 | if (hash_hashed(&req->hash_node)) |
74ce6ce4 | 5359 | hash_del(&req->hash_node); |
2bae047e | 5360 | |
d4e7cd36 | 5361 | io_poll_remove_double(req); |
74ce6ce4 JA |
5362 | spin_unlock_irq(&ctx->completion_lock); |
5363 | ||
0be0b0e3 PB |
5364 | if (!READ_ONCE(apoll->poll.canceled)) |
5365 | __io_req_task_submit(req); | |
5366 | else | |
5367 | __io_req_task_cancel(req, -ECANCELED); | |
aa340845 | 5368 | |
6d816e08 | 5369 | percpu_ref_put(&ctx->refs); |
807abcb0 | 5370 | kfree(apoll->double_poll); |
31067255 | 5371 | kfree(apoll); |
d7718a9d JA |
5372 | } |
5373 | ||
5374 | static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
5375 | void *key) | |
5376 | { | |
5377 | struct io_kiocb *req = wait->private; | |
5378 | struct io_poll_iocb *poll = &req->apoll->poll; | |
5379 | ||
5380 | trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data, | |
5381 | key_to_poll(key)); | |
5382 | ||
5383 | return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func); | |
5384 | } | |
5385 | ||
5386 | static void io_poll_req_insert(struct io_kiocb *req) | |
5387 | { | |
5388 | struct io_ring_ctx *ctx = req->ctx; | |
5389 | struct hlist_head *list; | |
5390 | ||
5391 | list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)]; | |
5392 | hlist_add_head(&req->hash_node, list); | |
5393 | } | |
5394 | ||
5395 | static __poll_t __io_arm_poll_handler(struct io_kiocb *req, | |
5396 | struct io_poll_iocb *poll, | |
5397 | struct io_poll_table *ipt, __poll_t mask, | |
5398 | wait_queue_func_t wake_func) | |
5399 | __acquires(&ctx->completion_lock) | |
5400 | { | |
5401 | struct io_ring_ctx *ctx = req->ctx; | |
5402 | bool cancel = false; | |
5403 | ||
4d52f338 | 5404 | INIT_HLIST_NODE(&req->hash_node); |
18bceab1 | 5405 | io_init_poll_iocb(poll, mask, wake_func); |
b90cd197 | 5406 | poll->file = req->file; |
18bceab1 | 5407 | poll->wait.private = req; |
d7718a9d JA |
5408 | |
5409 | ipt->pt._key = mask; | |
5410 | ipt->req = req; | |
5411 | ipt->error = -EINVAL; | |
5412 | ||
d7718a9d JA |
5413 | mask = vfs_poll(req->file, &ipt->pt) & poll->events; |
5414 | ||
5415 | spin_lock_irq(&ctx->completion_lock); | |
5416 | if (likely(poll->head)) { | |
5417 | spin_lock(&poll->head->lock); | |
5418 | if (unlikely(list_empty(&poll->wait.entry))) { | |
5419 | if (ipt->error) | |
5420 | cancel = true; | |
5421 | ipt->error = 0; | |
5422 | mask = 0; | |
5423 | } | |
5424 | if (mask || ipt->error) | |
5425 | list_del_init(&poll->wait.entry); | |
5426 | else if (cancel) | |
5427 | WRITE_ONCE(poll->canceled, true); | |
5428 | else if (!poll->done) /* actually waiting for an event */ | |
5429 | io_poll_req_insert(req); | |
5430 | spin_unlock(&poll->head->lock); | |
5431 | } | |
5432 | ||
5433 | return mask; | |
5434 | } | |
5435 | ||
5436 | static bool io_arm_poll_handler(struct io_kiocb *req) | |
5437 | { | |
5438 | const struct io_op_def *def = &io_op_defs[req->opcode]; | |
5439 | struct io_ring_ctx *ctx = req->ctx; | |
5440 | struct async_poll *apoll; | |
5441 | struct io_poll_table ipt; | |
5442 | __poll_t mask, ret; | |
9dab14b8 | 5443 | int rw; |
d7718a9d JA |
5444 | |
5445 | if (!req->file || !file_can_poll(req->file)) | |
5446 | return false; | |
24c74678 | 5447 | if (req->flags & REQ_F_POLLED) |
d7718a9d | 5448 | return false; |
9dab14b8 JA |
5449 | if (def->pollin) |
5450 | rw = READ; | |
5451 | else if (def->pollout) | |
5452 | rw = WRITE; | |
5453 | else | |
5454 | return false; | |
5455 | /* if we can't nonblock try, then no point in arming a poll handler */ | |
5456 | if (!io_file_supports_async(req->file, rw)) | |
d7718a9d JA |
5457 | return false; |
5458 | ||
5459 | apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC); | |
5460 | if (unlikely(!apoll)) | |
5461 | return false; | |
807abcb0 | 5462 | apoll->double_poll = NULL; |
d7718a9d JA |
5463 | |
5464 | req->flags |= REQ_F_POLLED; | |
d7718a9d | 5465 | req->apoll = apoll; |
d7718a9d | 5466 | |
8755d97a | 5467 | mask = 0; |
d7718a9d | 5468 | if (def->pollin) |
8755d97a | 5469 | mask |= POLLIN | POLLRDNORM; |
d7718a9d JA |
5470 | if (def->pollout) |
5471 | mask |= POLLOUT | POLLWRNORM; | |
901341bb LH |
5472 | |
5473 | /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */ | |
5474 | if ((req->opcode == IORING_OP_RECVMSG) && | |
5475 | (req->sr_msg.msg_flags & MSG_ERRQUEUE)) | |
5476 | mask &= ~POLLIN; | |
5477 | ||
d7718a9d JA |
5478 | mask |= POLLERR | POLLPRI; |
5479 | ||
5480 | ipt.pt._qproc = io_async_queue_proc; | |
5481 | ||
5482 | ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, | |
5483 | io_async_wake); | |
a36da65c | 5484 | if (ret || ipt.error) { |
d4e7cd36 | 5485 | io_poll_remove_double(req); |
d7718a9d | 5486 | spin_unlock_irq(&ctx->completion_lock); |
807abcb0 | 5487 | kfree(apoll->double_poll); |
d7718a9d JA |
5488 | kfree(apoll); |
5489 | return false; | |
5490 | } | |
5491 | spin_unlock_irq(&ctx->completion_lock); | |
5492 | trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask, | |
5493 | apoll->poll.events); | |
5494 | return true; | |
5495 | } | |
5496 | ||
5497 | static bool __io_poll_remove_one(struct io_kiocb *req, | |
5498 | struct io_poll_iocb *poll) | |
221c5eb2 | 5499 | { |
b41e9852 | 5500 | bool do_complete = false; |
221c5eb2 JA |
5501 | |
5502 | spin_lock(&poll->head->lock); | |
5503 | WRITE_ONCE(poll->canceled, true); | |
392edb45 JA |
5504 | if (!list_empty(&poll->wait.entry)) { |
5505 | list_del_init(&poll->wait.entry); | |
b41e9852 | 5506 | do_complete = true; |
221c5eb2 JA |
5507 | } |
5508 | spin_unlock(&poll->head->lock); | |
3bfa5bcb | 5509 | hash_del(&req->hash_node); |
d7718a9d JA |
5510 | return do_complete; |
5511 | } | |
5512 | ||
5513 | static bool io_poll_remove_one(struct io_kiocb *req) | |
5514 | { | |
5515 | bool do_complete; | |
5516 | ||
d4e7cd36 JA |
5517 | io_poll_remove_double(req); |
5518 | ||
d7718a9d JA |
5519 | if (req->opcode == IORING_OP_POLL_ADD) { |
5520 | do_complete = __io_poll_remove_one(req, &req->poll); | |
5521 | } else { | |
3bfa5bcb JA |
5522 | struct async_poll *apoll = req->apoll; |
5523 | ||
d7718a9d | 5524 | /* non-poll requests have submit ref still */ |
3bfa5bcb JA |
5525 | do_complete = __io_poll_remove_one(req, &apoll->poll); |
5526 | if (do_complete) { | |
d7718a9d | 5527 | io_put_req(req); |
807abcb0 | 5528 | kfree(apoll->double_poll); |
3bfa5bcb JA |
5529 | kfree(apoll); |
5530 | } | |
b1f573bd XW |
5531 | } |
5532 | ||
b41e9852 JA |
5533 | if (do_complete) { |
5534 | io_cqring_fill_event(req, -ECANCELED); | |
5535 | io_commit_cqring(req->ctx); | |
f254ac04 | 5536 | req_set_fail_links(req); |
216578e5 | 5537 | io_put_req_deferred(req, 1); |
b41e9852 JA |
5538 | } |
5539 | ||
5540 | return do_complete; | |
221c5eb2 JA |
5541 | } |
5542 | ||
76e1b642 JA |
5543 | /* |
5544 | * Returns true if we found and killed one or more poll requests | |
5545 | */ | |
6b81928d PB |
5546 | static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk, |
5547 | struct files_struct *files) | |
221c5eb2 | 5548 | { |
78076bb6 | 5549 | struct hlist_node *tmp; |
221c5eb2 | 5550 | struct io_kiocb *req; |
8e2e1faf | 5551 | int posted = 0, i; |
221c5eb2 JA |
5552 | |
5553 | spin_lock_irq(&ctx->completion_lock); | |
78076bb6 JA |
5554 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { |
5555 | struct hlist_head *list; | |
5556 | ||
5557 | list = &ctx->cancel_hash[i]; | |
f3606e3a | 5558 | hlist_for_each_entry_safe(req, tmp, list, hash_node) { |
6b81928d | 5559 | if (io_match_task(req, tsk, files)) |
f3606e3a JA |
5560 | posted += io_poll_remove_one(req); |
5561 | } | |
221c5eb2 JA |
5562 | } |
5563 | spin_unlock_irq(&ctx->completion_lock); | |
b41e9852 | 5564 | |
8e2e1faf JA |
5565 | if (posted) |
5566 | io_cqring_ev_posted(ctx); | |
76e1b642 JA |
5567 | |
5568 | return posted != 0; | |
221c5eb2 JA |
5569 | } |
5570 | ||
47f46768 JA |
5571 | static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr) |
5572 | { | |
78076bb6 | 5573 | struct hlist_head *list; |
47f46768 JA |
5574 | struct io_kiocb *req; |
5575 | ||
78076bb6 JA |
5576 | list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)]; |
5577 | hlist_for_each_entry(req, list, hash_node) { | |
b41e9852 JA |
5578 | if (sqe_addr != req->user_data) |
5579 | continue; | |
5580 | if (io_poll_remove_one(req)) | |
eac406c6 | 5581 | return 0; |
b41e9852 | 5582 | return -EALREADY; |
47f46768 JA |
5583 | } |
5584 | ||
5585 | return -ENOENT; | |
5586 | } | |
5587 | ||
3529d8c2 JA |
5588 | static int io_poll_remove_prep(struct io_kiocb *req, |
5589 | const struct io_uring_sqe *sqe) | |
0969e783 | 5590 | { |
0969e783 JA |
5591 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5592 | return -EINVAL; | |
5593 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index || | |
5594 | sqe->poll_events) | |
5595 | return -EINVAL; | |
5596 | ||
018043be | 5597 | req->poll_remove.addr = READ_ONCE(sqe->addr); |
0969e783 JA |
5598 | return 0; |
5599 | } | |
5600 | ||
221c5eb2 JA |
5601 | /* |
5602 | * Find a running poll command that matches one specified in sqe->addr, | |
5603 | * and remove it if found. | |
5604 | */ | |
fc4df999 | 5605 | static int io_poll_remove(struct io_kiocb *req) |
221c5eb2 JA |
5606 | { |
5607 | struct io_ring_ctx *ctx = req->ctx; | |
47f46768 | 5608 | int ret; |
221c5eb2 | 5609 | |
221c5eb2 | 5610 | spin_lock_irq(&ctx->completion_lock); |
018043be | 5611 | ret = io_poll_cancel(ctx, req->poll_remove.addr); |
221c5eb2 JA |
5612 | spin_unlock_irq(&ctx->completion_lock); |
5613 | ||
4e88d6e7 JA |
5614 | if (ret < 0) |
5615 | req_set_fail_links(req); | |
e1e16097 | 5616 | io_req_complete(req, ret); |
221c5eb2 JA |
5617 | return 0; |
5618 | } | |
5619 | ||
221c5eb2 JA |
5620 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, |
5621 | void *key) | |
5622 | { | |
c2f2eb7d JA |
5623 | struct io_kiocb *req = wait->private; |
5624 | struct io_poll_iocb *poll = &req->poll; | |
221c5eb2 | 5625 | |
d7718a9d | 5626 | return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func); |
221c5eb2 JA |
5627 | } |
5628 | ||
221c5eb2 JA |
5629 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, |
5630 | struct poll_table_struct *p) | |
5631 | { | |
5632 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
5633 | ||
e8c2bc1f | 5634 | __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data); |
eac406c6 JA |
5635 | } |
5636 | ||
3529d8c2 | 5637 | static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
221c5eb2 JA |
5638 | { |
5639 | struct io_poll_iocb *poll = &req->poll; | |
5769a351 | 5640 | u32 events; |
221c5eb2 JA |
5641 | |
5642 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
5643 | return -EINVAL; | |
5644 | if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
5645 | return -EINVAL; | |
5646 | ||
5769a351 JX |
5647 | events = READ_ONCE(sqe->poll32_events); |
5648 | #ifdef __BIG_ENDIAN | |
5649 | events = swahw32(events); | |
5650 | #endif | |
a31eb4a2 JX |
5651 | poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP | |
5652 | (events & EPOLLEXCLUSIVE); | |
0969e783 JA |
5653 | return 0; |
5654 | } | |
5655 | ||
014db007 | 5656 | static int io_poll_add(struct io_kiocb *req) |
0969e783 JA |
5657 | { |
5658 | struct io_poll_iocb *poll = &req->poll; | |
5659 | struct io_ring_ctx *ctx = req->ctx; | |
5660 | struct io_poll_table ipt; | |
0969e783 | 5661 | __poll_t mask; |
0969e783 | 5662 | |
d7718a9d | 5663 | ipt.pt._qproc = io_poll_queue_proc; |
36703247 | 5664 | |
d7718a9d JA |
5665 | mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events, |
5666 | io_poll_wake); | |
221c5eb2 | 5667 | |
8c838788 | 5668 | if (mask) { /* no async, we'd stolen it */ |
221c5eb2 | 5669 | ipt.error = 0; |
b0dd8a41 | 5670 | io_poll_complete(req, mask, 0); |
221c5eb2 | 5671 | } |
221c5eb2 JA |
5672 | spin_unlock_irq(&ctx->completion_lock); |
5673 | ||
8c838788 JA |
5674 | if (mask) { |
5675 | io_cqring_ev_posted(ctx); | |
014db007 | 5676 | io_put_req(req); |
221c5eb2 | 5677 | } |
8c838788 | 5678 | return ipt.error; |
221c5eb2 JA |
5679 | } |
5680 | ||
5262f567 JA |
5681 | static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer) |
5682 | { | |
ad8a48ac JA |
5683 | struct io_timeout_data *data = container_of(timer, |
5684 | struct io_timeout_data, timer); | |
5685 | struct io_kiocb *req = data->req; | |
5686 | struct io_ring_ctx *ctx = req->ctx; | |
5262f567 JA |
5687 | unsigned long flags; |
5688 | ||
5262f567 | 5689 | spin_lock_irqsave(&ctx->completion_lock, flags); |
a71976f3 | 5690 | list_del_init(&req->timeout.list); |
01cec8c1 PB |
5691 | atomic_set(&req->ctx->cq_timeouts, |
5692 | atomic_read(&req->ctx->cq_timeouts) + 1); | |
5693 | ||
78e19bbe | 5694 | io_cqring_fill_event(req, -ETIME); |
5262f567 JA |
5695 | io_commit_cqring(ctx); |
5696 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
5697 | ||
5698 | io_cqring_ev_posted(ctx); | |
4e88d6e7 | 5699 | req_set_fail_links(req); |
5262f567 JA |
5700 | io_put_req(req); |
5701 | return HRTIMER_NORESTART; | |
5702 | } | |
5703 | ||
fbd15848 PB |
5704 | static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx, |
5705 | __u64 user_data) | |
f254ac04 | 5706 | { |
fbd15848 | 5707 | struct io_timeout_data *io; |
47f46768 JA |
5708 | struct io_kiocb *req; |
5709 | int ret = -ENOENT; | |
f254ac04 | 5710 | |
135fcde8 | 5711 | list_for_each_entry(req, &ctx->timeout_list, timeout.list) { |
47f46768 | 5712 | if (user_data == req->user_data) { |
47f46768 JA |
5713 | ret = 0; |
5714 | break; | |
5715 | } | |
5716 | } | |
5717 | ||
5718 | if (ret == -ENOENT) | |
fbd15848 PB |
5719 | return ERR_PTR(ret); |
5720 | ||
5721 | io = req->async_data; | |
e8c2bc1f | 5722 | ret = hrtimer_try_to_cancel(&io->timer); |
f254ac04 | 5723 | if (ret == -1) |
fbd15848 | 5724 | return ERR_PTR(-EALREADY); |
a71976f3 | 5725 | list_del_init(&req->timeout.list); |
fbd15848 PB |
5726 | return req; |
5727 | } | |
47f46768 | 5728 | |
fbd15848 PB |
5729 | static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data) |
5730 | { | |
5731 | struct io_kiocb *req = io_timeout_extract(ctx, user_data); | |
5732 | ||
5733 | if (IS_ERR(req)) | |
5734 | return PTR_ERR(req); | |
f254ac04 JA |
5735 | |
5736 | req_set_fail_links(req); | |
f254ac04 | 5737 | io_cqring_fill_event(req, -ECANCELED); |
216578e5 | 5738 | io_put_req_deferred(req, 1); |
f254ac04 JA |
5739 | return 0; |
5740 | } | |
5741 | ||
9c8e11b3 PB |
5742 | static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data, |
5743 | struct timespec64 *ts, enum hrtimer_mode mode) | |
47f46768 | 5744 | { |
9c8e11b3 PB |
5745 | struct io_kiocb *req = io_timeout_extract(ctx, user_data); |
5746 | struct io_timeout_data *data; | |
47f46768 | 5747 | |
9c8e11b3 PB |
5748 | if (IS_ERR(req)) |
5749 | return PTR_ERR(req); | |
47f46768 | 5750 | |
9c8e11b3 PB |
5751 | req->timeout.off = 0; /* noseq */ |
5752 | data = req->async_data; | |
5753 | list_add_tail(&req->timeout.list, &ctx->timeout_list); | |
5754 | hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode); | |
5755 | data->timer.function = io_timeout_fn; | |
5756 | hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode); | |
5757 | return 0; | |
47f46768 JA |
5758 | } |
5759 | ||
3529d8c2 JA |
5760 | static int io_timeout_remove_prep(struct io_kiocb *req, |
5761 | const struct io_uring_sqe *sqe) | |
b29472ee | 5762 | { |
9c8e11b3 PB |
5763 | struct io_timeout_rem *tr = &req->timeout_rem; |
5764 | ||
b29472ee JA |
5765 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5766 | return -EINVAL; | |
61710e43 DA |
5767 | if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) |
5768 | return -EINVAL; | |
9c8e11b3 | 5769 | if (sqe->ioprio || sqe->buf_index || sqe->len) |
b29472ee JA |
5770 | return -EINVAL; |
5771 | ||
9c8e11b3 PB |
5772 | tr->addr = READ_ONCE(sqe->addr); |
5773 | tr->flags = READ_ONCE(sqe->timeout_flags); | |
5774 | if (tr->flags & IORING_TIMEOUT_UPDATE) { | |
5775 | if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS)) | |
5776 | return -EINVAL; | |
5777 | if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2))) | |
5778 | return -EFAULT; | |
5779 | } else if (tr->flags) { | |
5780 | /* timeout removal doesn't support flags */ | |
b29472ee | 5781 | return -EINVAL; |
9c8e11b3 | 5782 | } |
b29472ee | 5783 | |
b29472ee JA |
5784 | return 0; |
5785 | } | |
5786 | ||
11365043 JA |
5787 | /* |
5788 | * Remove or update an existing timeout command | |
5789 | */ | |
fc4df999 | 5790 | static int io_timeout_remove(struct io_kiocb *req) |
11365043 | 5791 | { |
9c8e11b3 | 5792 | struct io_timeout_rem *tr = &req->timeout_rem; |
11365043 | 5793 | struct io_ring_ctx *ctx = req->ctx; |
47f46768 | 5794 | int ret; |
11365043 | 5795 | |
11365043 | 5796 | spin_lock_irq(&ctx->completion_lock); |
9c8e11b3 PB |
5797 | if (req->timeout_rem.flags & IORING_TIMEOUT_UPDATE) { |
5798 | enum hrtimer_mode mode = (tr->flags & IORING_TIMEOUT_ABS) | |
5799 | ? HRTIMER_MODE_ABS : HRTIMER_MODE_REL; | |
5800 | ||
5801 | ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode); | |
5802 | } else { | |
5803 | ret = io_timeout_cancel(ctx, tr->addr); | |
5804 | } | |
11365043 | 5805 | |
47f46768 | 5806 | io_cqring_fill_event(req, ret); |
11365043 JA |
5807 | io_commit_cqring(ctx); |
5808 | spin_unlock_irq(&ctx->completion_lock); | |
5262f567 | 5809 | io_cqring_ev_posted(ctx); |
4e88d6e7 JA |
5810 | if (ret < 0) |
5811 | req_set_fail_links(req); | |
ec9c02ad | 5812 | io_put_req(req); |
11365043 | 5813 | return 0; |
5262f567 JA |
5814 | } |
5815 | ||
3529d8c2 | 5816 | static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2d28390a | 5817 | bool is_timeout_link) |
5262f567 | 5818 | { |
ad8a48ac | 5819 | struct io_timeout_data *data; |
a41525ab | 5820 | unsigned flags; |
56080b02 | 5821 | u32 off = READ_ONCE(sqe->off); |
5262f567 | 5822 | |
ad8a48ac | 5823 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5262f567 | 5824 | return -EINVAL; |
ad8a48ac | 5825 | if (sqe->ioprio || sqe->buf_index || sqe->len != 1) |
a41525ab | 5826 | return -EINVAL; |
56080b02 | 5827 | if (off && is_timeout_link) |
2d28390a | 5828 | return -EINVAL; |
a41525ab JA |
5829 | flags = READ_ONCE(sqe->timeout_flags); |
5830 | if (flags & ~IORING_TIMEOUT_ABS) | |
5262f567 | 5831 | return -EINVAL; |
bdf20073 | 5832 | |
bfe68a22 | 5833 | req->timeout.off = off; |
26a61679 | 5834 | |
e8c2bc1f | 5835 | if (!req->async_data && io_alloc_async_data(req)) |
26a61679 JA |
5836 | return -ENOMEM; |
5837 | ||
e8c2bc1f | 5838 | data = req->async_data; |
ad8a48ac | 5839 | data->req = req; |
ad8a48ac JA |
5840 | |
5841 | if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr))) | |
5262f567 JA |
5842 | return -EFAULT; |
5843 | ||
11365043 | 5844 | if (flags & IORING_TIMEOUT_ABS) |
ad8a48ac | 5845 | data->mode = HRTIMER_MODE_ABS; |
11365043 | 5846 | else |
ad8a48ac | 5847 | data->mode = HRTIMER_MODE_REL; |
11365043 | 5848 | |
ad8a48ac JA |
5849 | hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode); |
5850 | return 0; | |
5851 | } | |
5852 | ||
fc4df999 | 5853 | static int io_timeout(struct io_kiocb *req) |
ad8a48ac | 5854 | { |
ad8a48ac | 5855 | struct io_ring_ctx *ctx = req->ctx; |
e8c2bc1f | 5856 | struct io_timeout_data *data = req->async_data; |
ad8a48ac | 5857 | struct list_head *entry; |
bfe68a22 | 5858 | u32 tail, off = req->timeout.off; |
ad8a48ac | 5859 | |
733f5c95 | 5860 | spin_lock_irq(&ctx->completion_lock); |
93bd25bb | 5861 | |
5262f567 JA |
5862 | /* |
5863 | * sqe->off holds how many events that need to occur for this | |
93bd25bb JA |
5864 | * timeout event to be satisfied. If it isn't set, then this is |
5865 | * a pure timeout request, sequence isn't used. | |
5262f567 | 5866 | */ |
8eb7e2d0 | 5867 | if (io_is_timeout_noseq(req)) { |
93bd25bb JA |
5868 | entry = ctx->timeout_list.prev; |
5869 | goto add; | |
5870 | } | |
5262f567 | 5871 | |
bfe68a22 PB |
5872 | tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts); |
5873 | req->timeout.target_seq = tail + off; | |
5262f567 | 5874 | |
f010505b MDG |
5875 | /* Update the last seq here in case io_flush_timeouts() hasn't. |
5876 | * This is safe because ->completion_lock is held, and submissions | |
5877 | * and completions are never mixed in the same ->completion_lock section. | |
5878 | */ | |
5879 | ctx->cq_last_tm_flush = tail; | |
5880 | ||
5262f567 JA |
5881 | /* |
5882 | * Insertion sort, ensuring the first entry in the list is always | |
5883 | * the one we need first. | |
5884 | */ | |
5262f567 | 5885 | list_for_each_prev(entry, &ctx->timeout_list) { |
135fcde8 PB |
5886 | struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, |
5887 | timeout.list); | |
5262f567 | 5888 | |
8eb7e2d0 | 5889 | if (io_is_timeout_noseq(nxt)) |
93bd25bb | 5890 | continue; |
bfe68a22 PB |
5891 | /* nxt.seq is behind @tail, otherwise would've been completed */ |
5892 | if (off >= nxt->timeout.target_seq - tail) | |
5262f567 JA |
5893 | break; |
5894 | } | |
93bd25bb | 5895 | add: |
135fcde8 | 5896 | list_add(&req->timeout.list, entry); |
ad8a48ac JA |
5897 | data->timer.function = io_timeout_fn; |
5898 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode); | |
5262f567 | 5899 | spin_unlock_irq(&ctx->completion_lock); |
5262f567 JA |
5900 | return 0; |
5901 | } | |
5262f567 | 5902 | |
62755e35 JA |
5903 | static bool io_cancel_cb(struct io_wq_work *work, void *data) |
5904 | { | |
5905 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
5906 | ||
5907 | return req->user_data == (unsigned long) data; | |
5908 | } | |
5909 | ||
e977d6d3 | 5910 | static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr) |
62755e35 | 5911 | { |
62755e35 | 5912 | enum io_wq_cancel cancel_ret; |
62755e35 JA |
5913 | int ret = 0; |
5914 | ||
4f26bda1 | 5915 | cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false); |
62755e35 JA |
5916 | switch (cancel_ret) { |
5917 | case IO_WQ_CANCEL_OK: | |
5918 | ret = 0; | |
5919 | break; | |
5920 | case IO_WQ_CANCEL_RUNNING: | |
5921 | ret = -EALREADY; | |
5922 | break; | |
5923 | case IO_WQ_CANCEL_NOTFOUND: | |
5924 | ret = -ENOENT; | |
5925 | break; | |
5926 | } | |
5927 | ||
e977d6d3 JA |
5928 | return ret; |
5929 | } | |
5930 | ||
47f46768 JA |
5931 | static void io_async_find_and_cancel(struct io_ring_ctx *ctx, |
5932 | struct io_kiocb *req, __u64 sqe_addr, | |
014db007 | 5933 | int success_ret) |
47f46768 JA |
5934 | { |
5935 | unsigned long flags; | |
5936 | int ret; | |
5937 | ||
5938 | ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr); | |
5939 | if (ret != -ENOENT) { | |
5940 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
5941 | goto done; | |
5942 | } | |
5943 | ||
5944 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
5945 | ret = io_timeout_cancel(ctx, sqe_addr); | |
5946 | if (ret != -ENOENT) | |
5947 | goto done; | |
5948 | ret = io_poll_cancel(ctx, sqe_addr); | |
5949 | done: | |
b0dd8a41 JA |
5950 | if (!ret) |
5951 | ret = success_ret; | |
47f46768 JA |
5952 | io_cqring_fill_event(req, ret); |
5953 | io_commit_cqring(ctx); | |
5954 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
5955 | io_cqring_ev_posted(ctx); | |
5956 | ||
4e88d6e7 JA |
5957 | if (ret < 0) |
5958 | req_set_fail_links(req); | |
014db007 | 5959 | io_put_req(req); |
47f46768 JA |
5960 | } |
5961 | ||
3529d8c2 JA |
5962 | static int io_async_cancel_prep(struct io_kiocb *req, |
5963 | const struct io_uring_sqe *sqe) | |
e977d6d3 | 5964 | { |
fbf23849 | 5965 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
e977d6d3 | 5966 | return -EINVAL; |
61710e43 DA |
5967 | if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) |
5968 | return -EINVAL; | |
5969 | if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags) | |
e977d6d3 JA |
5970 | return -EINVAL; |
5971 | ||
fbf23849 JA |
5972 | req->cancel.addr = READ_ONCE(sqe->addr); |
5973 | return 0; | |
5974 | } | |
5975 | ||
014db007 | 5976 | static int io_async_cancel(struct io_kiocb *req) |
fbf23849 JA |
5977 | { |
5978 | struct io_ring_ctx *ctx = req->ctx; | |
fbf23849 | 5979 | |
014db007 | 5980 | io_async_find_and_cancel(ctx, req, req->cancel.addr, 0); |
5262f567 JA |
5981 | return 0; |
5982 | } | |
5983 | ||
05f3fb3c JA |
5984 | static int io_files_update_prep(struct io_kiocb *req, |
5985 | const struct io_uring_sqe *sqe) | |
5986 | { | |
6ca56f84 JA |
5987 | if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL)) |
5988 | return -EINVAL; | |
61710e43 DA |
5989 | if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) |
5990 | return -EINVAL; | |
5991 | if (sqe->ioprio || sqe->rw_flags) | |
05f3fb3c JA |
5992 | return -EINVAL; |
5993 | ||
5994 | req->files_update.offset = READ_ONCE(sqe->off); | |
5995 | req->files_update.nr_args = READ_ONCE(sqe->len); | |
5996 | if (!req->files_update.nr_args) | |
5997 | return -EINVAL; | |
5998 | req->files_update.arg = READ_ONCE(sqe->addr); | |
5999 | return 0; | |
6000 | } | |
6001 | ||
229a7b63 JA |
6002 | static int io_files_update(struct io_kiocb *req, bool force_nonblock, |
6003 | struct io_comp_state *cs) | |
fbf23849 JA |
6004 | { |
6005 | struct io_ring_ctx *ctx = req->ctx; | |
05f3fb3c JA |
6006 | struct io_uring_files_update up; |
6007 | int ret; | |
fbf23849 | 6008 | |
f86cd20c | 6009 | if (force_nonblock) |
05f3fb3c | 6010 | return -EAGAIN; |
05f3fb3c JA |
6011 | |
6012 | up.offset = req->files_update.offset; | |
6013 | up.fds = req->files_update.arg; | |
6014 | ||
6015 | mutex_lock(&ctx->uring_lock); | |
6016 | ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args); | |
6017 | mutex_unlock(&ctx->uring_lock); | |
6018 | ||
6019 | if (ret < 0) | |
6020 | req_set_fail_links(req); | |
229a7b63 | 6021 | __io_req_complete(req, ret, 0, cs); |
5262f567 JA |
6022 | return 0; |
6023 | } | |
6024 | ||
bfe76559 | 6025 | static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f67676d1 | 6026 | { |
d625c6ee | 6027 | switch (req->opcode) { |
e781573e | 6028 | case IORING_OP_NOP: |
bfe76559 | 6029 | return 0; |
f67676d1 JA |
6030 | case IORING_OP_READV: |
6031 | case IORING_OP_READ_FIXED: | |
3a6820f2 | 6032 | case IORING_OP_READ: |
bfe76559 | 6033 | return io_read_prep(req, sqe); |
f67676d1 JA |
6034 | case IORING_OP_WRITEV: |
6035 | case IORING_OP_WRITE_FIXED: | |
3a6820f2 | 6036 | case IORING_OP_WRITE: |
bfe76559 | 6037 | return io_write_prep(req, sqe); |
0969e783 | 6038 | case IORING_OP_POLL_ADD: |
bfe76559 | 6039 | return io_poll_add_prep(req, sqe); |
0969e783 | 6040 | case IORING_OP_POLL_REMOVE: |
bfe76559 | 6041 | return io_poll_remove_prep(req, sqe); |
8ed8d3c3 | 6042 | case IORING_OP_FSYNC: |
bfe76559 | 6043 | return io_prep_fsync(req, sqe); |
8ed8d3c3 | 6044 | case IORING_OP_SYNC_FILE_RANGE: |
bfe76559 | 6045 | return io_prep_sfr(req, sqe); |
03b1230c | 6046 | case IORING_OP_SENDMSG: |
fddaface | 6047 | case IORING_OP_SEND: |
bfe76559 | 6048 | return io_sendmsg_prep(req, sqe); |
03b1230c | 6049 | case IORING_OP_RECVMSG: |
fddaface | 6050 | case IORING_OP_RECV: |
bfe76559 | 6051 | return io_recvmsg_prep(req, sqe); |
f499a021 | 6052 | case IORING_OP_CONNECT: |
bfe76559 | 6053 | return io_connect_prep(req, sqe); |
2d28390a | 6054 | case IORING_OP_TIMEOUT: |
bfe76559 | 6055 | return io_timeout_prep(req, sqe, false); |
b29472ee | 6056 | case IORING_OP_TIMEOUT_REMOVE: |
bfe76559 | 6057 | return io_timeout_remove_prep(req, sqe); |
fbf23849 | 6058 | case IORING_OP_ASYNC_CANCEL: |
bfe76559 | 6059 | return io_async_cancel_prep(req, sqe); |
2d28390a | 6060 | case IORING_OP_LINK_TIMEOUT: |
bfe76559 | 6061 | return io_timeout_prep(req, sqe, true); |
8ed8d3c3 | 6062 | case IORING_OP_ACCEPT: |
bfe76559 | 6063 | return io_accept_prep(req, sqe); |
d63d1b5e | 6064 | case IORING_OP_FALLOCATE: |
bfe76559 | 6065 | return io_fallocate_prep(req, sqe); |
15b71abe | 6066 | case IORING_OP_OPENAT: |
bfe76559 | 6067 | return io_openat_prep(req, sqe); |
b5dba59e | 6068 | case IORING_OP_CLOSE: |
bfe76559 | 6069 | return io_close_prep(req, sqe); |
05f3fb3c | 6070 | case IORING_OP_FILES_UPDATE: |
bfe76559 | 6071 | return io_files_update_prep(req, sqe); |
eddc7ef5 | 6072 | case IORING_OP_STATX: |
bfe76559 | 6073 | return io_statx_prep(req, sqe); |
4840e418 | 6074 | case IORING_OP_FADVISE: |
bfe76559 | 6075 | return io_fadvise_prep(req, sqe); |
c1ca757b | 6076 | case IORING_OP_MADVISE: |
bfe76559 | 6077 | return io_madvise_prep(req, sqe); |
cebdb986 | 6078 | case IORING_OP_OPENAT2: |
bfe76559 | 6079 | return io_openat2_prep(req, sqe); |
3e4827b0 | 6080 | case IORING_OP_EPOLL_CTL: |
bfe76559 | 6081 | return io_epoll_ctl_prep(req, sqe); |
7d67af2c | 6082 | case IORING_OP_SPLICE: |
bfe76559 | 6083 | return io_splice_prep(req, sqe); |
ddf0322d | 6084 | case IORING_OP_PROVIDE_BUFFERS: |
bfe76559 | 6085 | return io_provide_buffers_prep(req, sqe); |
067524e9 | 6086 | case IORING_OP_REMOVE_BUFFERS: |
bfe76559 | 6087 | return io_remove_buffers_prep(req, sqe); |
f2a8d5c7 | 6088 | case IORING_OP_TEE: |
bfe76559 | 6089 | return io_tee_prep(req, sqe); |
36f4fa68 JA |
6090 | case IORING_OP_SHUTDOWN: |
6091 | return io_shutdown_prep(req, sqe); | |
80a261fd JA |
6092 | case IORING_OP_RENAMEAT: |
6093 | return io_renameat_prep(req, sqe); | |
14a1143b JA |
6094 | case IORING_OP_UNLINKAT: |
6095 | return io_unlinkat_prep(req, sqe); | |
f67676d1 JA |
6096 | } |
6097 | ||
bfe76559 PB |
6098 | printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n", |
6099 | req->opcode); | |
6100 | return-EINVAL; | |
6101 | } | |
6102 | ||
6103 | static int io_req_defer_prep(struct io_kiocb *req, | |
6104 | const struct io_uring_sqe *sqe) | |
6105 | { | |
bfe76559 PB |
6106 | if (!sqe) |
6107 | return 0; | |
6108 | if (io_alloc_async_data(req)) | |
6109 | return -EAGAIN; | |
bfe76559 | 6110 | return io_req_prep(req, sqe); |
f67676d1 JA |
6111 | } |
6112 | ||
9cf7c104 PB |
6113 | static u32 io_get_sequence(struct io_kiocb *req) |
6114 | { | |
6115 | struct io_kiocb *pos; | |
6116 | struct io_ring_ctx *ctx = req->ctx; | |
f2f87370 | 6117 | u32 total_submitted, nr_reqs = 0; |
9cf7c104 | 6118 | |
f2f87370 PB |
6119 | io_for_each_link(pos, req) |
6120 | nr_reqs++; | |
9cf7c104 PB |
6121 | |
6122 | total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped; | |
6123 | return total_submitted - nr_reqs; | |
6124 | } | |
6125 | ||
3529d8c2 | 6126 | static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
de0617e4 | 6127 | { |
a197f664 | 6128 | struct io_ring_ctx *ctx = req->ctx; |
27dc8338 | 6129 | struct io_defer_entry *de; |
f67676d1 | 6130 | int ret; |
9cf7c104 | 6131 | u32 seq; |
de0617e4 | 6132 | |
9d858b21 | 6133 | /* Still need defer if there is pending req in defer list. */ |
9cf7c104 PB |
6134 | if (likely(list_empty_careful(&ctx->defer_list) && |
6135 | !(req->flags & REQ_F_IO_DRAIN))) | |
6136 | return 0; | |
6137 | ||
6138 | seq = io_get_sequence(req); | |
6139 | /* Still a chance to pass the sequence check */ | |
6140 | if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) | |
de0617e4 JA |
6141 | return 0; |
6142 | ||
e8c2bc1f | 6143 | if (!req->async_data) { |
650b5481 | 6144 | ret = io_req_defer_prep(req, sqe); |
327d6d96 | 6145 | if (ret) |
650b5481 PB |
6146 | return ret; |
6147 | } | |
cbdcb435 | 6148 | io_prep_async_link(req); |
27dc8338 PB |
6149 | de = kmalloc(sizeof(*de), GFP_KERNEL); |
6150 | if (!de) | |
6151 | return -ENOMEM; | |
2d28390a | 6152 | |
de0617e4 | 6153 | spin_lock_irq(&ctx->completion_lock); |
9cf7c104 | 6154 | if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { |
de0617e4 | 6155 | spin_unlock_irq(&ctx->completion_lock); |
27dc8338 | 6156 | kfree(de); |
ae34817b PB |
6157 | io_queue_async_work(req); |
6158 | return -EIOCBQUEUED; | |
de0617e4 JA |
6159 | } |
6160 | ||
915967f6 | 6161 | trace_io_uring_defer(ctx, req, req->user_data); |
27dc8338 | 6162 | de->req = req; |
9cf7c104 | 6163 | de->seq = seq; |
27dc8338 | 6164 | list_add_tail(&de->list, &ctx->defer_list); |
de0617e4 JA |
6165 | spin_unlock_irq(&ctx->completion_lock); |
6166 | return -EIOCBQUEUED; | |
6167 | } | |
6168 | ||
f573d384 | 6169 | static void io_req_drop_files(struct io_kiocb *req) |
99bc4c38 | 6170 | { |
f573d384 | 6171 | struct io_ring_ctx *ctx = req->ctx; |
c98de08c | 6172 | struct io_uring_task *tctx = req->task->io_uring; |
f573d384 JA |
6173 | unsigned long flags; |
6174 | ||
02a13674 JA |
6175 | if (req->work.flags & IO_WQ_WORK_FILES) { |
6176 | put_files_struct(req->work.identity->files); | |
6177 | put_nsproxy(req->work.identity->nsproxy); | |
6178 | } | |
f573d384 JA |
6179 | spin_lock_irqsave(&ctx->inflight_lock, flags); |
6180 | list_del(&req->inflight_entry); | |
f573d384 JA |
6181 | spin_unlock_irqrestore(&ctx->inflight_lock, flags); |
6182 | req->flags &= ~REQ_F_INFLIGHT; | |
dfead8a8 | 6183 | req->work.flags &= ~IO_WQ_WORK_FILES; |
dfea9fce PB |
6184 | if (atomic_read(&tctx->in_idle)) |
6185 | wake_up(&tctx->wait); | |
f573d384 | 6186 | } |
99bc4c38 | 6187 | |
3ca405eb | 6188 | static void __io_clean_op(struct io_kiocb *req) |
99bc4c38 | 6189 | { |
0e1b6fe3 PB |
6190 | if (req->flags & REQ_F_BUFFER_SELECTED) { |
6191 | switch (req->opcode) { | |
6192 | case IORING_OP_READV: | |
6193 | case IORING_OP_READ_FIXED: | |
6194 | case IORING_OP_READ: | |
bcda7baa | 6195 | kfree((void *)(unsigned long)req->rw.addr); |
0e1b6fe3 PB |
6196 | break; |
6197 | case IORING_OP_RECVMSG: | |
6198 | case IORING_OP_RECV: | |
bcda7baa | 6199 | kfree(req->sr_msg.kbuf); |
0e1b6fe3 PB |
6200 | break; |
6201 | } | |
6202 | req->flags &= ~REQ_F_BUFFER_SELECTED; | |
99bc4c38 PB |
6203 | } |
6204 | ||
0e1b6fe3 PB |
6205 | if (req->flags & REQ_F_NEED_CLEANUP) { |
6206 | switch (req->opcode) { | |
6207 | case IORING_OP_READV: | |
6208 | case IORING_OP_READ_FIXED: | |
6209 | case IORING_OP_READ: | |
6210 | case IORING_OP_WRITEV: | |
6211 | case IORING_OP_WRITE_FIXED: | |
e8c2bc1f JA |
6212 | case IORING_OP_WRITE: { |
6213 | struct io_async_rw *io = req->async_data; | |
6214 | if (io->free_iovec) | |
6215 | kfree(io->free_iovec); | |
0e1b6fe3 | 6216 | break; |
e8c2bc1f | 6217 | } |
0e1b6fe3 | 6218 | case IORING_OP_RECVMSG: |
e8c2bc1f JA |
6219 | case IORING_OP_SENDMSG: { |
6220 | struct io_async_msghdr *io = req->async_data; | |
6221 | if (io->iov != io->fast_iov) | |
6222 | kfree(io->iov); | |
0e1b6fe3 | 6223 | break; |
e8c2bc1f | 6224 | } |
0e1b6fe3 PB |
6225 | case IORING_OP_SPLICE: |
6226 | case IORING_OP_TEE: | |
6227 | io_put_file(req, req->splice.file_in, | |
6228 | (req->splice.flags & SPLICE_F_FD_IN_FIXED)); | |
6229 | break; | |
f3cd4850 JA |
6230 | case IORING_OP_OPENAT: |
6231 | case IORING_OP_OPENAT2: | |
6232 | if (req->open.filename) | |
6233 | putname(req->open.filename); | |
6234 | break; | |
80a261fd JA |
6235 | case IORING_OP_RENAMEAT: |
6236 | putname(req->rename.oldpath); | |
6237 | putname(req->rename.newpath); | |
6238 | break; | |
14a1143b JA |
6239 | case IORING_OP_UNLINKAT: |
6240 | putname(req->unlink.filename); | |
6241 | break; | |
0e1b6fe3 PB |
6242 | } |
6243 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
99bc4c38 | 6244 | } |
99bc4c38 PB |
6245 | } |
6246 | ||
c1379e24 PB |
6247 | static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock, |
6248 | struct io_comp_state *cs) | |
2b188cc1 | 6249 | { |
a197f664 | 6250 | struct io_ring_ctx *ctx = req->ctx; |
d625c6ee | 6251 | int ret; |
2b188cc1 | 6252 | |
d625c6ee | 6253 | switch (req->opcode) { |
2b188cc1 | 6254 | case IORING_OP_NOP: |
229a7b63 | 6255 | ret = io_nop(req, cs); |
2b188cc1 JA |
6256 | break; |
6257 | case IORING_OP_READV: | |
edafccee | 6258 | case IORING_OP_READ_FIXED: |
3a6820f2 | 6259 | case IORING_OP_READ: |
a1d7c393 | 6260 | ret = io_read(req, force_nonblock, cs); |
edafccee | 6261 | break; |
3529d8c2 | 6262 | case IORING_OP_WRITEV: |
edafccee | 6263 | case IORING_OP_WRITE_FIXED: |
3a6820f2 | 6264 | case IORING_OP_WRITE: |
a1d7c393 | 6265 | ret = io_write(req, force_nonblock, cs); |
2b188cc1 | 6266 | break; |
c992fe29 | 6267 | case IORING_OP_FSYNC: |
014db007 | 6268 | ret = io_fsync(req, force_nonblock); |
c992fe29 | 6269 | break; |
221c5eb2 | 6270 | case IORING_OP_POLL_ADD: |
014db007 | 6271 | ret = io_poll_add(req); |
221c5eb2 JA |
6272 | break; |
6273 | case IORING_OP_POLL_REMOVE: | |
fc4df999 | 6274 | ret = io_poll_remove(req); |
221c5eb2 | 6275 | break; |
5d17b4a4 | 6276 | case IORING_OP_SYNC_FILE_RANGE: |
014db007 | 6277 | ret = io_sync_file_range(req, force_nonblock); |
5d17b4a4 | 6278 | break; |
0fa03c62 | 6279 | case IORING_OP_SENDMSG: |
062d04d7 PB |
6280 | ret = io_sendmsg(req, force_nonblock, cs); |
6281 | break; | |
fddaface | 6282 | case IORING_OP_SEND: |
062d04d7 | 6283 | ret = io_send(req, force_nonblock, cs); |
0fa03c62 | 6284 | break; |
aa1fa28f | 6285 | case IORING_OP_RECVMSG: |
062d04d7 PB |
6286 | ret = io_recvmsg(req, force_nonblock, cs); |
6287 | break; | |
fddaface | 6288 | case IORING_OP_RECV: |
062d04d7 | 6289 | ret = io_recv(req, force_nonblock, cs); |
aa1fa28f | 6290 | break; |
5262f567 | 6291 | case IORING_OP_TIMEOUT: |
fc4df999 | 6292 | ret = io_timeout(req); |
5262f567 | 6293 | break; |
11365043 | 6294 | case IORING_OP_TIMEOUT_REMOVE: |
fc4df999 | 6295 | ret = io_timeout_remove(req); |
11365043 | 6296 | break; |
17f2fe35 | 6297 | case IORING_OP_ACCEPT: |
229a7b63 | 6298 | ret = io_accept(req, force_nonblock, cs); |
17f2fe35 | 6299 | break; |
f8e85cf2 | 6300 | case IORING_OP_CONNECT: |
229a7b63 | 6301 | ret = io_connect(req, force_nonblock, cs); |
f8e85cf2 | 6302 | break; |
62755e35 | 6303 | case IORING_OP_ASYNC_CANCEL: |
014db007 | 6304 | ret = io_async_cancel(req); |
62755e35 | 6305 | break; |
d63d1b5e | 6306 | case IORING_OP_FALLOCATE: |
014db007 | 6307 | ret = io_fallocate(req, force_nonblock); |
d63d1b5e | 6308 | break; |
15b71abe | 6309 | case IORING_OP_OPENAT: |
014db007 | 6310 | ret = io_openat(req, force_nonblock); |
15b71abe | 6311 | break; |
b5dba59e | 6312 | case IORING_OP_CLOSE: |
229a7b63 | 6313 | ret = io_close(req, force_nonblock, cs); |
b5dba59e | 6314 | break; |
05f3fb3c | 6315 | case IORING_OP_FILES_UPDATE: |
229a7b63 | 6316 | ret = io_files_update(req, force_nonblock, cs); |
05f3fb3c | 6317 | break; |
eddc7ef5 | 6318 | case IORING_OP_STATX: |
014db007 | 6319 | ret = io_statx(req, force_nonblock); |
eddc7ef5 | 6320 | break; |
4840e418 | 6321 | case IORING_OP_FADVISE: |
014db007 | 6322 | ret = io_fadvise(req, force_nonblock); |
4840e418 | 6323 | break; |
c1ca757b | 6324 | case IORING_OP_MADVISE: |
014db007 | 6325 | ret = io_madvise(req, force_nonblock); |
c1ca757b | 6326 | break; |
cebdb986 | 6327 | case IORING_OP_OPENAT2: |
014db007 | 6328 | ret = io_openat2(req, force_nonblock); |
cebdb986 | 6329 | break; |
3e4827b0 | 6330 | case IORING_OP_EPOLL_CTL: |
229a7b63 | 6331 | ret = io_epoll_ctl(req, force_nonblock, cs); |
3e4827b0 | 6332 | break; |
7d67af2c | 6333 | case IORING_OP_SPLICE: |
014db007 | 6334 | ret = io_splice(req, force_nonblock); |
7d67af2c | 6335 | break; |
ddf0322d | 6336 | case IORING_OP_PROVIDE_BUFFERS: |
229a7b63 | 6337 | ret = io_provide_buffers(req, force_nonblock, cs); |
ddf0322d | 6338 | break; |
067524e9 | 6339 | case IORING_OP_REMOVE_BUFFERS: |
229a7b63 | 6340 | ret = io_remove_buffers(req, force_nonblock, cs); |
3e4827b0 | 6341 | break; |
f2a8d5c7 | 6342 | case IORING_OP_TEE: |
f2a8d5c7 PB |
6343 | ret = io_tee(req, force_nonblock); |
6344 | break; | |
36f4fa68 JA |
6345 | case IORING_OP_SHUTDOWN: |
6346 | ret = io_shutdown(req, force_nonblock); | |
6347 | break; | |
80a261fd JA |
6348 | case IORING_OP_RENAMEAT: |
6349 | ret = io_renameat(req, force_nonblock); | |
6350 | break; | |
14a1143b JA |
6351 | case IORING_OP_UNLINKAT: |
6352 | ret = io_unlinkat(req, force_nonblock); | |
6353 | break; | |
2b188cc1 JA |
6354 | default: |
6355 | ret = -EINVAL; | |
6356 | break; | |
6357 | } | |
6358 | ||
def596e9 JA |
6359 | if (ret) |
6360 | return ret; | |
6361 | ||
b532576e JA |
6362 | /* If the op doesn't have a file, we're not polling for it */ |
6363 | if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) { | |
11ba820b JA |
6364 | const bool in_async = io_wq_current_is_worker(); |
6365 | ||
11ba820b JA |
6366 | /* workqueue context doesn't hold uring_lock, grab it now */ |
6367 | if (in_async) | |
6368 | mutex_lock(&ctx->uring_lock); | |
6369 | ||
2e9dbe90 | 6370 | io_iopoll_req_issued(req, in_async); |
11ba820b JA |
6371 | |
6372 | if (in_async) | |
6373 | mutex_unlock(&ctx->uring_lock); | |
def596e9 JA |
6374 | } |
6375 | ||
6376 | return 0; | |
2b188cc1 JA |
6377 | } |
6378 | ||
f4db7182 | 6379 | static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work) |
2b188cc1 JA |
6380 | { |
6381 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
6df1db6b | 6382 | struct io_kiocb *timeout; |
561fb04a | 6383 | int ret = 0; |
2b188cc1 | 6384 | |
6df1db6b PB |
6385 | timeout = io_prep_linked_timeout(req); |
6386 | if (timeout) | |
6387 | io_queue_linked_timeout(timeout); | |
d4c81f38 | 6388 | |
0c9d5ccd JA |
6389 | /* if NO_CANCEL is set, we must still run the work */ |
6390 | if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) == | |
6391 | IO_WQ_WORK_CANCEL) { | |
561fb04a | 6392 | ret = -ECANCELED; |
0c9d5ccd | 6393 | } |
31b51510 | 6394 | |
561fb04a | 6395 | if (!ret) { |
561fb04a | 6396 | do { |
c1379e24 | 6397 | ret = io_issue_sqe(req, false, NULL); |
561fb04a JA |
6398 | /* |
6399 | * We can get EAGAIN for polled IO even though we're | |
6400 | * forcing a sync submission from here, since we can't | |
6401 | * wait for request slots on the block side. | |
6402 | */ | |
6403 | if (ret != -EAGAIN) | |
6404 | break; | |
6405 | cond_resched(); | |
6406 | } while (1); | |
6407 | } | |
31b51510 | 6408 | |
561fb04a | 6409 | if (ret) { |
c07e6719 XW |
6410 | struct io_ring_ctx *lock_ctx = NULL; |
6411 | ||
6412 | if (req->ctx->flags & IORING_SETUP_IOPOLL) | |
6413 | lock_ctx = req->ctx; | |
6414 | ||
dad1b124 | 6415 | /* |
c07e6719 XW |
6416 | * io_iopoll_complete() does not hold completion_lock to |
6417 | * complete polled io, so here for polled io, we can not call | |
6418 | * io_req_complete() directly, otherwise there maybe concurrent | |
6419 | * access to cqring, defer_list, etc, which is not safe. Given | |
6420 | * that io_iopoll_complete() is always called under uring_lock, | |
6421 | * so here for polled io, we also get uring_lock to complete | |
6422 | * it. | |
dad1b124 | 6423 | */ |
c07e6719 XW |
6424 | if (lock_ctx) |
6425 | mutex_lock(&lock_ctx->uring_lock); | |
dad1b124 | 6426 | |
c07e6719 XW |
6427 | req_set_fail_links(req); |
6428 | io_req_complete(req, ret); | |
6429 | ||
6430 | if (lock_ctx) | |
6431 | mutex_unlock(&lock_ctx->uring_lock); | |
edafccee | 6432 | } |
2b188cc1 | 6433 | |
f4db7182 | 6434 | return io_steal_work(req); |
2b188cc1 JA |
6435 | } |
6436 | ||
65e19f54 JA |
6437 | static inline struct file *io_file_from_index(struct io_ring_ctx *ctx, |
6438 | int index) | |
6439 | { | |
6440 | struct fixed_file_table *table; | |
6441 | ||
05f3fb3c | 6442 | table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT]; |
84695089 | 6443 | return table->files[index & IORING_FILE_TABLE_MASK]; |
65e19f54 JA |
6444 | } |
6445 | ||
8371adf5 PB |
6446 | static struct file *io_file_get(struct io_submit_state *state, |
6447 | struct io_kiocb *req, int fd, bool fixed) | |
09bb8394 | 6448 | { |
a197f664 | 6449 | struct io_ring_ctx *ctx = req->ctx; |
8da11c19 | 6450 | struct file *file; |
09bb8394 | 6451 | |
8da11c19 | 6452 | if (fixed) { |
479f517b | 6453 | if (unlikely((unsigned int)fd >= ctx->nr_user_files)) |
8371adf5 | 6454 | return NULL; |
b7620121 | 6455 | fd = array_index_nospec(fd, ctx->nr_user_files); |
8da11c19 | 6456 | file = io_file_from_index(ctx, fd); |
36f72fe2 | 6457 | io_set_resource_node(req); |
09bb8394 | 6458 | } else { |
c826bd7a | 6459 | trace_io_uring_file_get(ctx, fd); |
8da11c19 | 6460 | file = __io_file_get(state, fd); |
09bb8394 JA |
6461 | } |
6462 | ||
f609cbb8 PB |
6463 | if (file && file->f_op == &io_uring_fops && |
6464 | !(req->flags & REQ_F_INFLIGHT)) { | |
02a13674 JA |
6465 | io_req_init_async(req); |
6466 | req->flags |= REQ_F_INFLIGHT; | |
6467 | ||
6468 | spin_lock_irq(&ctx->inflight_lock); | |
6469 | list_add(&req->inflight_entry, &ctx->inflight_list); | |
6470 | spin_unlock_irq(&ctx->inflight_lock); | |
6471 | } | |
6472 | ||
8371adf5 | 6473 | return file; |
09bb8394 JA |
6474 | } |
6475 | ||
2665abfd | 6476 | static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer) |
2b188cc1 | 6477 | { |
ad8a48ac JA |
6478 | struct io_timeout_data *data = container_of(timer, |
6479 | struct io_timeout_data, timer); | |
90cd7e42 | 6480 | struct io_kiocb *prev, *req = data->req; |
2665abfd | 6481 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd | 6482 | unsigned long flags; |
2665abfd JA |
6483 | |
6484 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
90cd7e42 PB |
6485 | prev = req->timeout.head; |
6486 | req->timeout.head = NULL; | |
2665abfd JA |
6487 | |
6488 | /* | |
6489 | * We don't expect the list to be empty, that will only happen if we | |
6490 | * race with the completion of the linked work. | |
6491 | */ | |
90cd7e42 | 6492 | if (prev && refcount_inc_not_zero(&prev->refs)) |
f2f87370 | 6493 | io_remove_next_linked(prev); |
90cd7e42 PB |
6494 | else |
6495 | prev = NULL; | |
2665abfd JA |
6496 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
6497 | ||
6498 | if (prev) { | |
4e88d6e7 | 6499 | req_set_fail_links(prev); |
014db007 | 6500 | io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME); |
76a46e06 | 6501 | io_put_req(prev); |
47f46768 | 6502 | } else { |
e1e16097 | 6503 | io_req_complete(req, -ETIME); |
2665abfd | 6504 | } |
2665abfd JA |
6505 | return HRTIMER_NORESTART; |
6506 | } | |
6507 | ||
7271ef3a | 6508 | static void __io_queue_linked_timeout(struct io_kiocb *req) |
2665abfd | 6509 | { |
76a46e06 | 6510 | /* |
f2f87370 PB |
6511 | * If the back reference is NULL, then our linked request finished |
6512 | * before we got a chance to setup the timer | |
76a46e06 | 6513 | */ |
90cd7e42 | 6514 | if (req->timeout.head) { |
e8c2bc1f | 6515 | struct io_timeout_data *data = req->async_data; |
94ae5e77 | 6516 | |
ad8a48ac JA |
6517 | data->timer.function = io_link_timeout_fn; |
6518 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), | |
6519 | data->mode); | |
2665abfd | 6520 | } |
7271ef3a JA |
6521 | } |
6522 | ||
6523 | static void io_queue_linked_timeout(struct io_kiocb *req) | |
6524 | { | |
6525 | struct io_ring_ctx *ctx = req->ctx; | |
6526 | ||
6527 | spin_lock_irq(&ctx->completion_lock); | |
6528 | __io_queue_linked_timeout(req); | |
76a46e06 | 6529 | spin_unlock_irq(&ctx->completion_lock); |
2665abfd | 6530 | |
2665abfd | 6531 | /* drop submission reference */ |
76a46e06 JA |
6532 | io_put_req(req); |
6533 | } | |
2665abfd | 6534 | |
ad8a48ac | 6535 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) |
2665abfd | 6536 | { |
f2f87370 | 6537 | struct io_kiocb *nxt = req->link; |
2665abfd | 6538 | |
f2f87370 PB |
6539 | if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) || |
6540 | nxt->opcode != IORING_OP_LINK_TIMEOUT) | |
76a46e06 | 6541 | return NULL; |
2665abfd | 6542 | |
90cd7e42 | 6543 | nxt->timeout.head = req; |
900fad45 | 6544 | nxt->flags |= REQ_F_LTIMEOUT_ACTIVE; |
76a46e06 | 6545 | req->flags |= REQ_F_LINK_TIMEOUT; |
76a46e06 | 6546 | return nxt; |
2665abfd JA |
6547 | } |
6548 | ||
c1379e24 | 6549 | static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs) |
2b188cc1 | 6550 | { |
4a0a7a18 | 6551 | struct io_kiocb *linked_timeout; |
193155c8 | 6552 | const struct cred *old_creds = NULL; |
e0c5c576 | 6553 | int ret; |
2b188cc1 | 6554 | |
4a0a7a18 JA |
6555 | again: |
6556 | linked_timeout = io_prep_linked_timeout(req); | |
6557 | ||
2e5aa6cb PB |
6558 | if ((req->flags & REQ_F_WORK_INITIALIZED) && |
6559 | (req->work.flags & IO_WQ_WORK_CREDS) && | |
98447d65 | 6560 | req->work.identity->creds != current_cred()) { |
193155c8 JA |
6561 | if (old_creds) |
6562 | revert_creds(old_creds); | |
98447d65 | 6563 | if (old_creds == req->work.identity->creds) |
193155c8 JA |
6564 | old_creds = NULL; /* restored original creds */ |
6565 | else | |
98447d65 | 6566 | old_creds = override_creds(req->work.identity->creds); |
193155c8 JA |
6567 | } |
6568 | ||
c1379e24 | 6569 | ret = io_issue_sqe(req, true, cs); |
491381ce JA |
6570 | |
6571 | /* | |
6572 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
6573 | * doesn't support non-blocking read/write attempts | |
6574 | */ | |
24c74678 | 6575 | if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) { |
f063c547 | 6576 | if (!io_arm_poll_handler(req)) { |
f063c547 PB |
6577 | /* |
6578 | * Queued up for async execution, worker will release | |
6579 | * submit reference when the iocb is actually submitted. | |
6580 | */ | |
6581 | io_queue_async_work(req); | |
2b188cc1 | 6582 | } |
bbad27b2 | 6583 | |
f063c547 PB |
6584 | if (linked_timeout) |
6585 | io_queue_linked_timeout(linked_timeout); | |
0d63c148 PB |
6586 | } else if (likely(!ret)) { |
6587 | /* drop submission reference */ | |
6588 | req = io_put_req_find_next(req); | |
6589 | if (linked_timeout) | |
6590 | io_queue_linked_timeout(linked_timeout); | |
e65ef56d | 6591 | |
0d63c148 PB |
6592 | if (req) { |
6593 | if (!(req->flags & REQ_F_FORCE_ASYNC)) | |
6594 | goto again; | |
6595 | io_queue_async_work(req); | |
6596 | } | |
6597 | } else { | |
652532ad PB |
6598 | /* un-prep timeout, so it'll be killed as any other linked */ |
6599 | req->flags &= ~REQ_F_LINK_TIMEOUT; | |
4e88d6e7 | 6600 | req_set_fail_links(req); |
e65ef56d | 6601 | io_put_req(req); |
e1e16097 | 6602 | io_req_complete(req, ret); |
9e645e11 | 6603 | } |
652532ad | 6604 | |
193155c8 JA |
6605 | if (old_creds) |
6606 | revert_creds(old_creds); | |
2b188cc1 JA |
6607 | } |
6608 | ||
f13fad7b JA |
6609 | static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
6610 | struct io_comp_state *cs) | |
4fe2c963 JL |
6611 | { |
6612 | int ret; | |
6613 | ||
3529d8c2 | 6614 | ret = io_req_defer(req, sqe); |
4fe2c963 JL |
6615 | if (ret) { |
6616 | if (ret != -EIOCBQUEUED) { | |
1118591a | 6617 | fail_req: |
4e88d6e7 | 6618 | req_set_fail_links(req); |
e1e16097 JA |
6619 | io_put_req(req); |
6620 | io_req_complete(req, ret); | |
4fe2c963 | 6621 | } |
2550878f | 6622 | } else if (req->flags & REQ_F_FORCE_ASYNC) { |
e8c2bc1f | 6623 | if (!req->async_data) { |
bd2ab18a | 6624 | ret = io_req_defer_prep(req, sqe); |
327d6d96 | 6625 | if (unlikely(ret)) |
bd2ab18a PB |
6626 | goto fail_req; |
6627 | } | |
ce35a47a JA |
6628 | io_queue_async_work(req); |
6629 | } else { | |
c1379e24 PB |
6630 | if (sqe) { |
6631 | ret = io_req_prep(req, sqe); | |
6632 | if (unlikely(ret)) | |
6633 | goto fail_req; | |
6634 | } | |
6635 | __io_queue_sqe(req, cs); | |
ce35a47a | 6636 | } |
4fe2c963 JL |
6637 | } |
6638 | ||
f13fad7b JA |
6639 | static inline void io_queue_link_head(struct io_kiocb *req, |
6640 | struct io_comp_state *cs) | |
4fe2c963 | 6641 | { |
94ae5e77 | 6642 | if (unlikely(req->flags & REQ_F_FAIL_LINK)) { |
e1e16097 JA |
6643 | io_put_req(req); |
6644 | io_req_complete(req, -ECANCELED); | |
1b4a51b6 | 6645 | } else |
f13fad7b | 6646 | io_queue_sqe(req, NULL, cs); |
4fe2c963 JL |
6647 | } |
6648 | ||
863e0560 PB |
6649 | struct io_submit_link { |
6650 | struct io_kiocb *head; | |
6651 | struct io_kiocb *last; | |
6652 | }; | |
6653 | ||
1d4240cc | 6654 | static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
863e0560 | 6655 | struct io_submit_link *link, struct io_comp_state *cs) |
9e645e11 | 6656 | { |
a197f664 | 6657 | struct io_ring_ctx *ctx = req->ctx; |
ef4ff581 | 6658 | int ret; |
9e645e11 | 6659 | |
9e645e11 JA |
6660 | /* |
6661 | * If we already have a head request, queue this one for async | |
6662 | * submittal once the head completes. If we don't have a head but | |
6663 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
6664 | * submitted sync once the chain is complete. If none of those | |
6665 | * conditions are true (normal request), then just queue it. | |
6666 | */ | |
863e0560 PB |
6667 | if (link->head) { |
6668 | struct io_kiocb *head = link->head; | |
4e88d6e7 | 6669 | |
8cdf2193 PB |
6670 | /* |
6671 | * Taking sequential execution of a link, draining both sides | |
6672 | * of the link also fullfils IOSQE_IO_DRAIN semantics for all | |
6673 | * requests in the link. So, it drains the head and the | |
6674 | * next after the link request. The last one is done via | |
6675 | * drain_next flag to persist the effect across calls. | |
6676 | */ | |
ef4ff581 | 6677 | if (req->flags & REQ_F_IO_DRAIN) { |
711be031 PB |
6678 | head->flags |= REQ_F_IO_DRAIN; |
6679 | ctx->drain_next = 1; | |
6680 | } | |
3529d8c2 | 6681 | ret = io_req_defer_prep(req, sqe); |
327d6d96 | 6682 | if (unlikely(ret)) { |
4e88d6e7 | 6683 | /* fail even hard links since we don't submit */ |
9d76377f | 6684 | head->flags |= REQ_F_FAIL_LINK; |
1d4240cc | 6685 | return ret; |
2d28390a | 6686 | } |
9d76377f | 6687 | trace_io_uring_link(ctx, req, head); |
f2f87370 | 6688 | link->last->link = req; |
863e0560 | 6689 | link->last = req; |
32fe525b PB |
6690 | |
6691 | /* last request of a link, enqueue the link */ | |
ef4ff581 | 6692 | if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) { |
f13fad7b | 6693 | io_queue_link_head(head, cs); |
863e0560 | 6694 | link->head = NULL; |
32fe525b | 6695 | } |
9e645e11 | 6696 | } else { |
711be031 PB |
6697 | if (unlikely(ctx->drain_next)) { |
6698 | req->flags |= REQ_F_IO_DRAIN; | |
ef4ff581 | 6699 | ctx->drain_next = 0; |
711be031 | 6700 | } |
ef4ff581 | 6701 | if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) { |
711be031 | 6702 | ret = io_req_defer_prep(req, sqe); |
327d6d96 | 6703 | if (unlikely(ret)) |
711be031 | 6704 | req->flags |= REQ_F_FAIL_LINK; |
863e0560 PB |
6705 | link->head = req; |
6706 | link->last = req; | |
711be031 | 6707 | } else { |
f13fad7b | 6708 | io_queue_sqe(req, sqe, cs); |
711be031 | 6709 | } |
9e645e11 | 6710 | } |
2e6e1fde | 6711 | |
1d4240cc | 6712 | return 0; |
9e645e11 JA |
6713 | } |
6714 | ||
9a56a232 JA |
6715 | /* |
6716 | * Batched submission is done, ensure local IO is flushed out. | |
6717 | */ | |
6718 | static void io_submit_state_end(struct io_submit_state *state) | |
6719 | { | |
f13fad7b JA |
6720 | if (!list_empty(&state->comp.list)) |
6721 | io_submit_flush_completions(&state->comp); | |
27926b68 JA |
6722 | if (state->plug_started) |
6723 | blk_finish_plug(&state->plug); | |
9f13c35b | 6724 | io_state_file_put(state); |
2579f913 | 6725 | if (state->free_reqs) |
6c8a3134 | 6726 | kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs); |
9a56a232 JA |
6727 | } |
6728 | ||
6729 | /* | |
6730 | * Start submission side cache. | |
6731 | */ | |
6732 | static void io_submit_state_start(struct io_submit_state *state, | |
013538bd | 6733 | struct io_ring_ctx *ctx, unsigned int max_ios) |
9a56a232 | 6734 | { |
27926b68 | 6735 | state->plug_started = false; |
013538bd JA |
6736 | state->comp.nr = 0; |
6737 | INIT_LIST_HEAD(&state->comp.list); | |
6738 | state->comp.ctx = ctx; | |
2579f913 | 6739 | state->free_reqs = 0; |
6e1271e6 | 6740 | state->file_refs = 0; |
9a56a232 JA |
6741 | state->ios_left = max_ios; |
6742 | } | |
6743 | ||
2b188cc1 JA |
6744 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
6745 | { | |
75b28aff | 6746 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 6747 | |
caf582c6 PB |
6748 | /* |
6749 | * Ensure any loads from the SQEs are done at this point, | |
6750 | * since once we write the new head, the application could | |
6751 | * write new data to them. | |
6752 | */ | |
6753 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
6754 | } |
6755 | ||
2b188cc1 | 6756 | /* |
3529d8c2 | 6757 | * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory |
2b188cc1 JA |
6758 | * that is mapped by userspace. This means that care needs to be taken to |
6759 | * ensure that reads are stable, as we cannot rely on userspace always | |
6760 | * being a good citizen. If members of the sqe are validated and then later | |
6761 | * used, it's important that those reads are done through READ_ONCE() to | |
6762 | * prevent a re-load down the line. | |
6763 | */ | |
709b302f | 6764 | static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx) |
2b188cc1 | 6765 | { |
75b28aff | 6766 | u32 *sq_array = ctx->sq_array; |
2b188cc1 JA |
6767 | unsigned head; |
6768 | ||
6769 | /* | |
6770 | * The cached sq head (or cq tail) serves two purposes: | |
6771 | * | |
6772 | * 1) allows us to batch the cost of updating the user visible | |
6773 | * head updates. | |
6774 | * 2) allows the kernel side to track the head on its own, even | |
6775 | * though the application is the one updating it. | |
6776 | */ | |
ee7d46d9 | 6777 | head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]); |
709b302f PB |
6778 | if (likely(head < ctx->sq_entries)) |
6779 | return &ctx->sq_sqes[head]; | |
2b188cc1 JA |
6780 | |
6781 | /* drop invalid entries */ | |
498ccd9e | 6782 | ctx->cached_sq_dropped++; |
ee7d46d9 | 6783 | WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped); |
709b302f PB |
6784 | return NULL; |
6785 | } | |
6786 | ||
6787 | static inline void io_consume_sqe(struct io_ring_ctx *ctx) | |
6788 | { | |
6789 | ctx->cached_sq_head++; | |
2b188cc1 JA |
6790 | } |
6791 | ||
21b55dbc SG |
6792 | /* |
6793 | * Check SQE restrictions (opcode and flags). | |
6794 | * | |
6795 | * Returns 'true' if SQE is allowed, 'false' otherwise. | |
6796 | */ | |
6797 | static inline bool io_check_restriction(struct io_ring_ctx *ctx, | |
6798 | struct io_kiocb *req, | |
6799 | unsigned int sqe_flags) | |
6800 | { | |
6801 | if (!ctx->restricted) | |
6802 | return true; | |
6803 | ||
6804 | if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) | |
6805 | return false; | |
6806 | ||
6807 | if ((sqe_flags & ctx->restrictions.sqe_flags_required) != | |
6808 | ctx->restrictions.sqe_flags_required) | |
6809 | return false; | |
6810 | ||
6811 | if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | | |
6812 | ctx->restrictions.sqe_flags_required)) | |
6813 | return false; | |
6814 | ||
6815 | return true; | |
6816 | } | |
6817 | ||
ef4ff581 PB |
6818 | #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \ |
6819 | IOSQE_IO_HARDLINK | IOSQE_ASYNC | \ | |
6820 | IOSQE_BUFFER_SELECT) | |
6821 | ||
6822 | static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, | |
6823 | const struct io_uring_sqe *sqe, | |
0cdaf760 | 6824 | struct io_submit_state *state) |
0553b8bd | 6825 | { |
ef4ff581 | 6826 | unsigned int sqe_flags; |
71b547c0 | 6827 | int id, ret; |
ef4ff581 | 6828 | |
0553b8bd PB |
6829 | req->opcode = READ_ONCE(sqe->opcode); |
6830 | req->user_data = READ_ONCE(sqe->user_data); | |
e8c2bc1f | 6831 | req->async_data = NULL; |
0553b8bd PB |
6832 | req->file = NULL; |
6833 | req->ctx = ctx; | |
6834 | req->flags = 0; | |
f2f87370 | 6835 | req->link = NULL; |
36f72fe2 | 6836 | req->fixed_file_refs = NULL; |
0553b8bd PB |
6837 | /* one is dropped after submission, the other at completion */ |
6838 | refcount_set(&req->refs, 2); | |
4dd2824d | 6839 | req->task = current; |
0553b8bd | 6840 | req->result = 0; |
ef4ff581 PB |
6841 | |
6842 | if (unlikely(req->opcode >= IORING_OP_LAST)) | |
6843 | return -EINVAL; | |
6844 | ||
28cea78a | 6845 | if (unlikely(io_sq_thread_acquire_mm_files(ctx, req))) |
9d8426a0 | 6846 | return -EFAULT; |
ef4ff581 PB |
6847 | |
6848 | sqe_flags = READ_ONCE(sqe->flags); | |
6849 | /* enforce forwards compatibility on users */ | |
6850 | if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) | |
6851 | return -EINVAL; | |
6852 | ||
21b55dbc SG |
6853 | if (unlikely(!io_check_restriction(ctx, req, sqe_flags))) |
6854 | return -EACCES; | |
6855 | ||
ef4ff581 PB |
6856 | if ((sqe_flags & IOSQE_BUFFER_SELECT) && |
6857 | !io_op_defs[req->opcode].buffer_select) | |
6858 | return -EOPNOTSUPP; | |
6859 | ||
6860 | id = READ_ONCE(sqe->personality); | |
6861 | if (id) { | |
1e6fa521 JA |
6862 | struct io_identity *iod; |
6863 | ||
1e6fa521 JA |
6864 | iod = idr_find(&ctx->personality_idr, id); |
6865 | if (unlikely(!iod)) | |
ef4ff581 | 6866 | return -EINVAL; |
1e6fa521 | 6867 | refcount_inc(&iod->count); |
ec99ca6c PB |
6868 | |
6869 | __io_req_init_async(req); | |
1e6fa521 JA |
6870 | get_cred(iod->creds); |
6871 | req->work.identity = iod; | |
dfead8a8 | 6872 | req->work.flags |= IO_WQ_WORK_CREDS; |
ef4ff581 PB |
6873 | } |
6874 | ||
6875 | /* same numerical values with corresponding REQ_F_*, safe to copy */ | |
c11368a5 | 6876 | req->flags |= sqe_flags; |
ef4ff581 | 6877 | |
27926b68 JA |
6878 | /* |
6879 | * Plug now if we have more than 1 IO left after this, and the target | |
6880 | * is potentially a read/write to block based storage. | |
6881 | */ | |
6882 | if (!state->plug_started && state->ios_left > 1 && | |
6883 | io_op_defs[req->opcode].plug) { | |
6884 | blk_start_plug(&state->plug); | |
6885 | state->plug_started = true; | |
6886 | } | |
6887 | ||
bd5bbda7 PB |
6888 | ret = 0; |
6889 | if (io_op_defs[req->opcode].needs_file) { | |
6890 | bool fixed = req->flags & REQ_F_FIXED_FILE; | |
6891 | ||
6892 | req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed); | |
6893 | if (unlikely(!req->file && | |
6894 | !io_op_defs[req->opcode].needs_file_no_error)) | |
6895 | ret = -EBADF; | |
6896 | } | |
63ff8223 | 6897 | |
71b547c0 PB |
6898 | state->ios_left--; |
6899 | return ret; | |
0553b8bd PB |
6900 | } |
6901 | ||
0f212204 | 6902 | static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) |
6c271ce2 | 6903 | { |
ac8691c4 | 6904 | struct io_submit_state state; |
863e0560 | 6905 | struct io_submit_link link; |
9e645e11 | 6906 | int i, submitted = 0; |
6c271ce2 | 6907 | |
c4a2ed72 | 6908 | /* if we have a backlog and couldn't flush it all, return BUSY */ |
ad3eb2c8 | 6909 | if (test_bit(0, &ctx->sq_check_overflow)) { |
6c503150 | 6910 | if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL)) |
ad3eb2c8 JA |
6911 | return -EBUSY; |
6912 | } | |
6c271ce2 | 6913 | |
ee7d46d9 PB |
6914 | /* make sure SQ entry isn't read before tail */ |
6915 | nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx)); | |
9ef4f124 | 6916 | |
2b85edfc PB |
6917 | if (!percpu_ref_tryget_many(&ctx->refs, nr)) |
6918 | return -EAGAIN; | |
6c271ce2 | 6919 | |
d8a6df10 | 6920 | percpu_counter_add(¤t->io_uring->inflight, nr); |
faf7b51c | 6921 | refcount_add(nr, ¤t->usage); |
6c271ce2 | 6922 | |
013538bd | 6923 | io_submit_state_start(&state, ctx, nr); |
863e0560 | 6924 | link.head = NULL; |
b14cca0c | 6925 | |
6c271ce2 | 6926 | for (i = 0; i < nr; i++) { |
3529d8c2 | 6927 | const struct io_uring_sqe *sqe; |
196be95c | 6928 | struct io_kiocb *req; |
1cb1edb2 | 6929 | int err; |
fb5ccc98 | 6930 | |
b1e50e54 PB |
6931 | sqe = io_get_sqe(ctx); |
6932 | if (unlikely(!sqe)) { | |
6933 | io_consume_sqe(ctx); | |
6934 | break; | |
6935 | } | |
ac8691c4 | 6936 | req = io_alloc_req(ctx, &state); |
196be95c PB |
6937 | if (unlikely(!req)) { |
6938 | if (!submitted) | |
6939 | submitted = -EAGAIN; | |
fb5ccc98 | 6940 | break; |
196be95c | 6941 | } |
709b302f | 6942 | io_consume_sqe(ctx); |
d3656344 JA |
6943 | /* will complete beyond this point, count as submitted */ |
6944 | submitted++; | |
6945 | ||
692d8363 | 6946 | err = io_init_req(ctx, req, sqe, &state); |
ef4ff581 | 6947 | if (unlikely(err)) { |
1cb1edb2 | 6948 | fail_req: |
e1e16097 JA |
6949 | io_put_req(req); |
6950 | io_req_complete(req, err); | |
196be95c PB |
6951 | break; |
6952 | } | |
fb5ccc98 | 6953 | |
354420f7 | 6954 | trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data, |
0cdaf760 | 6955 | true, io_async_submit(ctx)); |
f13fad7b | 6956 | err = io_submit_sqe(req, sqe, &link, &state.comp); |
1d4240cc PB |
6957 | if (err) |
6958 | goto fail_req; | |
6c271ce2 JA |
6959 | } |
6960 | ||
9466f437 PB |
6961 | if (unlikely(submitted != nr)) { |
6962 | int ref_used = (submitted == -EAGAIN) ? 0 : submitted; | |
d8a6df10 JA |
6963 | struct io_uring_task *tctx = current->io_uring; |
6964 | int unused = nr - ref_used; | |
9466f437 | 6965 | |
d8a6df10 JA |
6966 | percpu_ref_put_many(&ctx->refs, unused); |
6967 | percpu_counter_sub(&tctx->inflight, unused); | |
6968 | put_task_struct_many(current, unused); | |
9466f437 | 6969 | } |
863e0560 PB |
6970 | if (link.head) |
6971 | io_queue_link_head(link.head, &state.comp); | |
ac8691c4 | 6972 | io_submit_state_end(&state); |
6c271ce2 | 6973 | |
ae9428ca PB |
6974 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
6975 | io_commit_sqring(ctx); | |
6976 | ||
6c271ce2 JA |
6977 | return submitted; |
6978 | } | |
6979 | ||
23b3628e XW |
6980 | static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx) |
6981 | { | |
6982 | /* Tell userspace we may need a wakeup call */ | |
6983 | spin_lock_irq(&ctx->completion_lock); | |
6984 | ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP; | |
6985 | spin_unlock_irq(&ctx->completion_lock); | |
6986 | } | |
6987 | ||
6988 | static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx) | |
6989 | { | |
6990 | spin_lock_irq(&ctx->completion_lock); | |
6991 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; | |
6992 | spin_unlock_irq(&ctx->completion_lock); | |
6993 | } | |
6994 | ||
08369246 | 6995 | static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries) |
6c271ce2 | 6996 | { |
c8d1ba58 | 6997 | unsigned int to_submit; |
bdcd3eab | 6998 | int ret = 0; |
6c271ce2 | 6999 | |
c8d1ba58 | 7000 | to_submit = io_sqring_entries(ctx); |
e95eee2d JA |
7001 | /* if we're handling multiple rings, cap submit size for fairness */ |
7002 | if (cap_entries && to_submit > 8) | |
7003 | to_submit = 8; | |
7004 | ||
906a3c6f | 7005 | if (!list_empty(&ctx->iopoll_list) || to_submit) { |
c8d1ba58 | 7006 | unsigned nr_events = 0; |
a4c0b3de | 7007 | |
c8d1ba58 | 7008 | mutex_lock(&ctx->uring_lock); |
906a3c6f | 7009 | if (!list_empty(&ctx->iopoll_list)) |
c8d1ba58 | 7010 | io_do_iopoll(ctx, &nr_events, 0); |
906a3c6f | 7011 | |
d9d05217 PB |
7012 | if (to_submit && !ctx->sqo_dead && |
7013 | likely(!percpu_ref_is_dying(&ctx->refs))) | |
08369246 | 7014 | ret = io_submit_sqes(ctx, to_submit); |
c8d1ba58 JA |
7015 | mutex_unlock(&ctx->uring_lock); |
7016 | } | |
6c271ce2 | 7017 | |
90554200 JA |
7018 | if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait)) |
7019 | wake_up(&ctx->sqo_sq_wait); | |
6c271ce2 | 7020 | |
08369246 XW |
7021 | return ret; |
7022 | } | |
6c271ce2 | 7023 | |
08369246 XW |
7024 | static void io_sqd_update_thread_idle(struct io_sq_data *sqd) |
7025 | { | |
7026 | struct io_ring_ctx *ctx; | |
7027 | unsigned sq_thread_idle = 0; | |
6c271ce2 | 7028 | |
08369246 XW |
7029 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) { |
7030 | if (sq_thread_idle < ctx->sq_thread_idle) | |
7031 | sq_thread_idle = ctx->sq_thread_idle; | |
c8d1ba58 | 7032 | } |
c1edbf5f | 7033 | |
08369246 | 7034 | sqd->sq_thread_idle = sq_thread_idle; |
c8d1ba58 | 7035 | } |
6c271ce2 | 7036 | |
69fb2131 JA |
7037 | static void io_sqd_init_new(struct io_sq_data *sqd) |
7038 | { | |
7039 | struct io_ring_ctx *ctx; | |
7040 | ||
7041 | while (!list_empty(&sqd->ctx_new_list)) { | |
7042 | ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list); | |
69fb2131 JA |
7043 | list_move_tail(&ctx->sqd_list, &sqd->ctx_list); |
7044 | complete(&ctx->sq_thread_comp); | |
7045 | } | |
08369246 XW |
7046 | |
7047 | io_sqd_update_thread_idle(sqd); | |
69fb2131 JA |
7048 | } |
7049 | ||
c8d1ba58 JA |
7050 | static int io_sq_thread(void *data) |
7051 | { | |
91d8f519 | 7052 | struct cgroup_subsys_state *cur_css = NULL; |
28cea78a JA |
7053 | struct files_struct *old_files = current->files; |
7054 | struct nsproxy *old_nsproxy = current->nsproxy; | |
69fb2131 JA |
7055 | const struct cred *old_cred = NULL; |
7056 | struct io_sq_data *sqd = data; | |
7057 | struct io_ring_ctx *ctx; | |
a0d9205f | 7058 | unsigned long timeout = 0; |
08369246 | 7059 | DEFINE_WAIT(wait); |
6c271ce2 | 7060 | |
28cea78a JA |
7061 | task_lock(current); |
7062 | current->files = NULL; | |
7063 | current->nsproxy = NULL; | |
7064 | task_unlock(current); | |
6c271ce2 | 7065 | |
69fb2131 | 7066 | while (!kthread_should_stop()) { |
08369246 XW |
7067 | int ret; |
7068 | bool cap_entries, sqt_spin, needs_sched; | |
c1edbf5f JA |
7069 | |
7070 | /* | |
69fb2131 JA |
7071 | * Any changes to the sqd lists are synchronized through the |
7072 | * kthread parking. This synchronizes the thread vs users, | |
7073 | * the users are synchronized on the sqd->ctx_lock. | |
c1edbf5f | 7074 | */ |
65b2b213 | 7075 | if (kthread_should_park()) { |
69fb2131 | 7076 | kthread_parkme(); |
65b2b213 XW |
7077 | /* |
7078 | * When sq thread is unparked, in case the previous park operation | |
7079 | * comes from io_put_sq_data(), which means that sq thread is going | |
7080 | * to be stopped, so here needs to have a check. | |
7081 | */ | |
7082 | if (kthread_should_stop()) | |
7083 | break; | |
7084 | } | |
7143b5ac | 7085 | |
08369246 | 7086 | if (unlikely(!list_empty(&sqd->ctx_new_list))) { |
69fb2131 | 7087 | io_sqd_init_new(sqd); |
08369246 XW |
7088 | timeout = jiffies + sqd->sq_thread_idle; |
7089 | } | |
6c271ce2 | 7090 | |
08369246 | 7091 | sqt_spin = false; |
e95eee2d | 7092 | cap_entries = !list_is_singular(&sqd->ctx_list); |
69fb2131 JA |
7093 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) { |
7094 | if (current->cred != ctx->creds) { | |
7095 | if (old_cred) | |
7096 | revert_creds(old_cred); | |
7097 | old_cred = override_creds(ctx->creds); | |
bdcd3eab | 7098 | } |
91d8f519 | 7099 | io_sq_thread_associate_blkcg(ctx, &cur_css); |
4ea33a97 JA |
7100 | #ifdef CONFIG_AUDIT |
7101 | current->loginuid = ctx->loginuid; | |
7102 | current->sessionid = ctx->sessionid; | |
7103 | #endif | |
bdcd3eab | 7104 | |
08369246 XW |
7105 | ret = __io_sq_thread(ctx, cap_entries); |
7106 | if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list))) | |
7107 | sqt_spin = true; | |
6c271ce2 | 7108 | |
28cea78a | 7109 | io_sq_thread_drop_mm_files(); |
69fb2131 | 7110 | } |
6c271ce2 | 7111 | |
08369246 | 7112 | if (sqt_spin || !time_after(jiffies, timeout)) { |
c8d1ba58 | 7113 | io_run_task_work(); |
d434ab6d | 7114 | io_sq_thread_drop_mm_files(); |
c8d1ba58 | 7115 | cond_resched(); |
08369246 XW |
7116 | if (sqt_spin) |
7117 | timeout = jiffies + sqd->sq_thread_idle; | |
7118 | continue; | |
7119 | } | |
7120 | ||
7121 | if (kthread_should_park()) | |
7122 | continue; | |
7123 | ||
7124 | needs_sched = true; | |
7125 | prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE); | |
7126 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) { | |
7127 | if ((ctx->flags & IORING_SETUP_IOPOLL) && | |
7128 | !list_empty_careful(&ctx->iopoll_list)) { | |
7129 | needs_sched = false; | |
7130 | break; | |
7131 | } | |
7132 | if (io_sqring_entries(ctx)) { | |
7133 | needs_sched = false; | |
7134 | break; | |
7135 | } | |
7136 | } | |
7137 | ||
7138 | if (needs_sched) { | |
69fb2131 JA |
7139 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) |
7140 | io_ring_set_wakeup_flag(ctx); | |
08369246 | 7141 | |
69fb2131 | 7142 | schedule(); |
69fb2131 JA |
7143 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) |
7144 | io_ring_clear_wakeup_flag(ctx); | |
6c271ce2 | 7145 | } |
08369246 XW |
7146 | |
7147 | finish_wait(&sqd->wait, &wait); | |
7148 | timeout = jiffies + sqd->sq_thread_idle; | |
6c271ce2 JA |
7149 | } |
7150 | ||
4c6e277c | 7151 | io_run_task_work(); |
d434ab6d | 7152 | io_sq_thread_drop_mm_files(); |
b41e9852 | 7153 | |
91d8f519 DZ |
7154 | if (cur_css) |
7155 | io_sq_thread_unassociate_blkcg(); | |
69fb2131 JA |
7156 | if (old_cred) |
7157 | revert_creds(old_cred); | |
06058632 | 7158 | |
28cea78a JA |
7159 | task_lock(current); |
7160 | current->files = old_files; | |
7161 | current->nsproxy = old_nsproxy; | |
7162 | task_unlock(current); | |
7163 | ||
2bbcd6d3 | 7164 | kthread_parkme(); |
06058632 | 7165 | |
6c271ce2 JA |
7166 | return 0; |
7167 | } | |
7168 | ||
bda52162 JA |
7169 | struct io_wait_queue { |
7170 | struct wait_queue_entry wq; | |
7171 | struct io_ring_ctx *ctx; | |
7172 | unsigned to_wait; | |
7173 | unsigned nr_timeouts; | |
7174 | }; | |
7175 | ||
6c503150 | 7176 | static inline bool io_should_wake(struct io_wait_queue *iowq) |
bda52162 JA |
7177 | { |
7178 | struct io_ring_ctx *ctx = iowq->ctx; | |
7179 | ||
7180 | /* | |
d195a66e | 7181 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
7182 | * started waiting. For timeouts, we always want to return to userspace, |
7183 | * regardless of event count. | |
7184 | */ | |
6c503150 | 7185 | return io_cqring_events(ctx) >= iowq->to_wait || |
bda52162 JA |
7186 | atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
7187 | } | |
7188 | ||
7189 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
7190 | int wake_flags, void *key) | |
7191 | { | |
7192 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
7193 | wq); | |
7194 | ||
6c503150 PB |
7195 | /* |
7196 | * Cannot safely flush overflowed CQEs from here, ensure we wake up | |
7197 | * the task, and the next invocation will do it. | |
7198 | */ | |
7199 | if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow)) | |
7200 | return autoremove_wake_function(curr, mode, wake_flags, key); | |
7201 | return -1; | |
bda52162 JA |
7202 | } |
7203 | ||
af9c1a44 JA |
7204 | static int io_run_task_work_sig(void) |
7205 | { | |
7206 | if (io_run_task_work()) | |
7207 | return 1; | |
7208 | if (!signal_pending(current)) | |
7209 | return 0; | |
792ee0f6 JA |
7210 | if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL)) |
7211 | return -ERESTARTSYS; | |
af9c1a44 JA |
7212 | return -EINTR; |
7213 | } | |
7214 | ||
2b188cc1 JA |
7215 | /* |
7216 | * Wait until events become available, if we don't already have some. The | |
7217 | * application must reap them itself, as they reside on the shared cq ring. | |
7218 | */ | |
7219 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
c73ebb68 HX |
7220 | const sigset_t __user *sig, size_t sigsz, |
7221 | struct __kernel_timespec __user *uts) | |
2b188cc1 | 7222 | { |
bda52162 JA |
7223 | struct io_wait_queue iowq = { |
7224 | .wq = { | |
7225 | .private = current, | |
7226 | .func = io_wake_function, | |
7227 | .entry = LIST_HEAD_INIT(iowq.wq.entry), | |
7228 | }, | |
7229 | .ctx = ctx, | |
7230 | .to_wait = min_events, | |
7231 | }; | |
75b28aff | 7232 | struct io_rings *rings = ctx->rings; |
c73ebb68 HX |
7233 | struct timespec64 ts; |
7234 | signed long timeout = 0; | |
e9ffa5c2 | 7235 | int ret = 0; |
2b188cc1 | 7236 | |
b41e9852 | 7237 | do { |
6c503150 PB |
7238 | io_cqring_overflow_flush(ctx, false, NULL, NULL); |
7239 | if (io_cqring_events(ctx) >= min_events) | |
b41e9852 | 7240 | return 0; |
4c6e277c | 7241 | if (!io_run_task_work()) |
b41e9852 | 7242 | break; |
b41e9852 | 7243 | } while (1); |
2b188cc1 JA |
7244 | |
7245 | if (sig) { | |
9e75ad5d AB |
7246 | #ifdef CONFIG_COMPAT |
7247 | if (in_compat_syscall()) | |
7248 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 7249 | sigsz); |
9e75ad5d AB |
7250 | else |
7251 | #endif | |
b772434b | 7252 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 7253 | |
2b188cc1 JA |
7254 | if (ret) |
7255 | return ret; | |
7256 | } | |
7257 | ||
c73ebb68 HX |
7258 | if (uts) { |
7259 | if (get_timespec64(&ts, uts)) | |
7260 | return -EFAULT; | |
7261 | timeout = timespec64_to_jiffies(&ts); | |
7262 | } | |
7263 | ||
bda52162 | 7264 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
c826bd7a | 7265 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 | 7266 | do { |
6c503150 | 7267 | io_cqring_overflow_flush(ctx, false, NULL, NULL); |
bda52162 JA |
7268 | prepare_to_wait_exclusive(&ctx->wait, &iowq.wq, |
7269 | TASK_INTERRUPTIBLE); | |
ce593a6c | 7270 | /* make sure we run task_work before checking for signals */ |
af9c1a44 | 7271 | ret = io_run_task_work_sig(); |
6195ba09 HX |
7272 | if (ret > 0) { |
7273 | finish_wait(&ctx->wait, &iowq.wq); | |
4c6e277c | 7274 | continue; |
6195ba09 | 7275 | } |
af9c1a44 | 7276 | else if (ret < 0) |
bda52162 | 7277 | break; |
6c503150 | 7278 | if (io_should_wake(&iowq)) |
ce593a6c | 7279 | break; |
6195ba09 HX |
7280 | if (test_bit(0, &ctx->cq_check_overflow)) { |
7281 | finish_wait(&ctx->wait, &iowq.wq); | |
6c503150 | 7282 | continue; |
6195ba09 | 7283 | } |
c73ebb68 HX |
7284 | if (uts) { |
7285 | timeout = schedule_timeout(timeout); | |
7286 | if (timeout == 0) { | |
7287 | ret = -ETIME; | |
7288 | break; | |
7289 | } | |
7290 | } else { | |
7291 | schedule(); | |
7292 | } | |
bda52162 JA |
7293 | } while (1); |
7294 | finish_wait(&ctx->wait, &iowq.wq); | |
7295 | ||
b7db41c9 | 7296 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 7297 | |
75b28aff | 7298 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
7299 | } |
7300 | ||
6b06314c JA |
7301 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
7302 | { | |
7303 | #if defined(CONFIG_UNIX) | |
7304 | if (ctx->ring_sock) { | |
7305 | struct sock *sock = ctx->ring_sock->sk; | |
7306 | struct sk_buff *skb; | |
7307 | ||
7308 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
7309 | kfree_skb(skb); | |
7310 | } | |
7311 | #else | |
7312 | int i; | |
7313 | ||
65e19f54 JA |
7314 | for (i = 0; i < ctx->nr_user_files; i++) { |
7315 | struct file *file; | |
7316 | ||
7317 | file = io_file_from_index(ctx, i); | |
7318 | if (file) | |
7319 | fput(file); | |
7320 | } | |
6b06314c JA |
7321 | #endif |
7322 | } | |
7323 | ||
05f3fb3c JA |
7324 | static void io_file_ref_kill(struct percpu_ref *ref) |
7325 | { | |
7326 | struct fixed_file_data *data; | |
7327 | ||
7328 | data = container_of(ref, struct fixed_file_data, refs); | |
7329 | complete(&data->done); | |
7330 | } | |
7331 | ||
1642b445 PB |
7332 | static void io_sqe_files_set_node(struct fixed_file_data *file_data, |
7333 | struct fixed_file_ref_node *ref_node) | |
7334 | { | |
7335 | spin_lock_bh(&file_data->lock); | |
7336 | file_data->node = ref_node; | |
7337 | list_add_tail(&ref_node->node, &file_data->ref_list); | |
7338 | spin_unlock_bh(&file_data->lock); | |
7339 | percpu_ref_get(&file_data->refs); | |
7340 | } | |
7341 | ||
6b06314c JA |
7342 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) |
7343 | { | |
05f3fb3c | 7344 | struct fixed_file_data *data = ctx->file_data; |
1ffc5422 | 7345 | struct fixed_file_ref_node *backup_node, *ref_node = NULL; |
65e19f54 | 7346 | unsigned nr_tables, i; |
1ffc5422 | 7347 | int ret; |
65e19f54 | 7348 | |
05f3fb3c | 7349 | if (!data) |
6b06314c | 7350 | return -ENXIO; |
1ffc5422 PB |
7351 | backup_node = alloc_fixed_file_ref_node(ctx); |
7352 | if (!backup_node) | |
7353 | return -ENOMEM; | |
6b06314c | 7354 | |
ac0648a5 | 7355 | spin_lock_bh(&data->lock); |
1e5d770b | 7356 | ref_node = data->node; |
ac0648a5 | 7357 | spin_unlock_bh(&data->lock); |
05589553 XW |
7358 | if (ref_node) |
7359 | percpu_ref_kill(&ref_node->refs); | |
7360 | ||
7361 | percpu_ref_kill(&data->refs); | |
7362 | ||
7363 | /* wait for all refs nodes to complete */ | |
4a38aed2 | 7364 | flush_delayed_work(&ctx->file_put_work); |
1ffc5422 PB |
7365 | do { |
7366 | ret = wait_for_completion_interruptible(&data->done); | |
7367 | if (!ret) | |
7368 | break; | |
7369 | ret = io_run_task_work_sig(); | |
7370 | if (ret < 0) { | |
7371 | percpu_ref_resurrect(&data->refs); | |
7372 | reinit_completion(&data->done); | |
7373 | io_sqe_files_set_node(data, backup_node); | |
7374 | return ret; | |
7375 | } | |
7376 | } while (1); | |
05f3fb3c | 7377 | |
6b06314c | 7378 | __io_sqe_files_unregister(ctx); |
65e19f54 JA |
7379 | nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE); |
7380 | for (i = 0; i < nr_tables; i++) | |
05f3fb3c JA |
7381 | kfree(data->table[i].files); |
7382 | kfree(data->table); | |
05589553 XW |
7383 | percpu_ref_exit(&data->refs); |
7384 | kfree(data); | |
05f3fb3c | 7385 | ctx->file_data = NULL; |
6b06314c | 7386 | ctx->nr_user_files = 0; |
1ffc5422 | 7387 | destroy_fixed_file_ref_node(backup_node); |
6b06314c JA |
7388 | return 0; |
7389 | } | |
7390 | ||
534ca6d6 | 7391 | static void io_put_sq_data(struct io_sq_data *sqd) |
6c271ce2 | 7392 | { |
534ca6d6 | 7393 | if (refcount_dec_and_test(&sqd->refs)) { |
2bbcd6d3 RP |
7394 | /* |
7395 | * The park is a bit of a work-around, without it we get | |
7396 | * warning spews on shutdown with SQPOLL set and affinity | |
7397 | * set to a single CPU. | |
7398 | */ | |
534ca6d6 JA |
7399 | if (sqd->thread) { |
7400 | kthread_park(sqd->thread); | |
7401 | kthread_stop(sqd->thread); | |
7402 | } | |
7403 | ||
7404 | kfree(sqd); | |
7405 | } | |
7406 | } | |
7407 | ||
aa06165d JA |
7408 | static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p) |
7409 | { | |
7410 | struct io_ring_ctx *ctx_attach; | |
7411 | struct io_sq_data *sqd; | |
7412 | struct fd f; | |
7413 | ||
7414 | f = fdget(p->wq_fd); | |
7415 | if (!f.file) | |
7416 | return ERR_PTR(-ENXIO); | |
7417 | if (f.file->f_op != &io_uring_fops) { | |
7418 | fdput(f); | |
7419 | return ERR_PTR(-EINVAL); | |
7420 | } | |
7421 | ||
7422 | ctx_attach = f.file->private_data; | |
7423 | sqd = ctx_attach->sq_data; | |
7424 | if (!sqd) { | |
7425 | fdput(f); | |
7426 | return ERR_PTR(-EINVAL); | |
7427 | } | |
7428 | ||
7429 | refcount_inc(&sqd->refs); | |
7430 | fdput(f); | |
7431 | return sqd; | |
7432 | } | |
7433 | ||
534ca6d6 JA |
7434 | static struct io_sq_data *io_get_sq_data(struct io_uring_params *p) |
7435 | { | |
7436 | struct io_sq_data *sqd; | |
7437 | ||
aa06165d JA |
7438 | if (p->flags & IORING_SETUP_ATTACH_WQ) |
7439 | return io_attach_sq_data(p); | |
7440 | ||
534ca6d6 JA |
7441 | sqd = kzalloc(sizeof(*sqd), GFP_KERNEL); |
7442 | if (!sqd) | |
7443 | return ERR_PTR(-ENOMEM); | |
7444 | ||
7445 | refcount_set(&sqd->refs, 1); | |
69fb2131 JA |
7446 | INIT_LIST_HEAD(&sqd->ctx_list); |
7447 | INIT_LIST_HEAD(&sqd->ctx_new_list); | |
7448 | mutex_init(&sqd->ctx_lock); | |
7449 | mutex_init(&sqd->lock); | |
534ca6d6 JA |
7450 | init_waitqueue_head(&sqd->wait); |
7451 | return sqd; | |
7452 | } | |
7453 | ||
69fb2131 JA |
7454 | static void io_sq_thread_unpark(struct io_sq_data *sqd) |
7455 | __releases(&sqd->lock) | |
7456 | { | |
7457 | if (!sqd->thread) | |
7458 | return; | |
7459 | kthread_unpark(sqd->thread); | |
7460 | mutex_unlock(&sqd->lock); | |
7461 | } | |
7462 | ||
7463 | static void io_sq_thread_park(struct io_sq_data *sqd) | |
7464 | __acquires(&sqd->lock) | |
7465 | { | |
7466 | if (!sqd->thread) | |
7467 | return; | |
7468 | mutex_lock(&sqd->lock); | |
7469 | kthread_park(sqd->thread); | |
7470 | } | |
7471 | ||
534ca6d6 JA |
7472 | static void io_sq_thread_stop(struct io_ring_ctx *ctx) |
7473 | { | |
7474 | struct io_sq_data *sqd = ctx->sq_data; | |
7475 | ||
7476 | if (sqd) { | |
7477 | if (sqd->thread) { | |
7478 | /* | |
7479 | * We may arrive here from the error branch in | |
7480 | * io_sq_offload_create() where the kthread is created | |
7481 | * without being waked up, thus wake it up now to make | |
7482 | * sure the wait will complete. | |
7483 | */ | |
7484 | wake_up_process(sqd->thread); | |
7485 | wait_for_completion(&ctx->sq_thread_comp); | |
69fb2131 JA |
7486 | |
7487 | io_sq_thread_park(sqd); | |
7488 | } | |
7489 | ||
7490 | mutex_lock(&sqd->ctx_lock); | |
7491 | list_del(&ctx->sqd_list); | |
08369246 | 7492 | io_sqd_update_thread_idle(sqd); |
69fb2131 JA |
7493 | mutex_unlock(&sqd->ctx_lock); |
7494 | ||
08369246 | 7495 | if (sqd->thread) |
69fb2131 | 7496 | io_sq_thread_unpark(sqd); |
534ca6d6 JA |
7497 | |
7498 | io_put_sq_data(sqd); | |
7499 | ctx->sq_data = NULL; | |
6c271ce2 JA |
7500 | } |
7501 | } | |
7502 | ||
6b06314c JA |
7503 | static void io_finish_async(struct io_ring_ctx *ctx) |
7504 | { | |
6c271ce2 JA |
7505 | io_sq_thread_stop(ctx); |
7506 | ||
561fb04a JA |
7507 | if (ctx->io_wq) { |
7508 | io_wq_destroy(ctx->io_wq); | |
7509 | ctx->io_wq = NULL; | |
6b06314c JA |
7510 | } |
7511 | } | |
7512 | ||
7513 | #if defined(CONFIG_UNIX) | |
6b06314c JA |
7514 | /* |
7515 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
7516 | * the io_uring can be safely unregistered on process exit, even if we have | |
7517 | * loops in the file referencing. | |
7518 | */ | |
7519 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
7520 | { | |
7521 | struct sock *sk = ctx->ring_sock->sk; | |
7522 | struct scm_fp_list *fpl; | |
7523 | struct sk_buff *skb; | |
08a45173 | 7524 | int i, nr_files; |
6b06314c | 7525 | |
6b06314c JA |
7526 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); |
7527 | if (!fpl) | |
7528 | return -ENOMEM; | |
7529 | ||
7530 | skb = alloc_skb(0, GFP_KERNEL); | |
7531 | if (!skb) { | |
7532 | kfree(fpl); | |
7533 | return -ENOMEM; | |
7534 | } | |
7535 | ||
7536 | skb->sk = sk; | |
6b06314c | 7537 | |
08a45173 | 7538 | nr_files = 0; |
6b06314c JA |
7539 | fpl->user = get_uid(ctx->user); |
7540 | for (i = 0; i < nr; i++) { | |
65e19f54 JA |
7541 | struct file *file = io_file_from_index(ctx, i + offset); |
7542 | ||
7543 | if (!file) | |
08a45173 | 7544 | continue; |
65e19f54 | 7545 | fpl->fp[nr_files] = get_file(file); |
08a45173 JA |
7546 | unix_inflight(fpl->user, fpl->fp[nr_files]); |
7547 | nr_files++; | |
6b06314c JA |
7548 | } |
7549 | ||
08a45173 JA |
7550 | if (nr_files) { |
7551 | fpl->max = SCM_MAX_FD; | |
7552 | fpl->count = nr_files; | |
7553 | UNIXCB(skb).fp = fpl; | |
05f3fb3c | 7554 | skb->destructor = unix_destruct_scm; |
08a45173 JA |
7555 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); |
7556 | skb_queue_head(&sk->sk_receive_queue, skb); | |
6b06314c | 7557 | |
08a45173 JA |
7558 | for (i = 0; i < nr_files; i++) |
7559 | fput(fpl->fp[i]); | |
7560 | } else { | |
7561 | kfree_skb(skb); | |
7562 | kfree(fpl); | |
7563 | } | |
6b06314c JA |
7564 | |
7565 | return 0; | |
7566 | } | |
7567 | ||
7568 | /* | |
7569 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
7570 | * causes regular reference counting to break down. We rely on the UNIX | |
7571 | * garbage collection to take care of this problem for us. | |
7572 | */ | |
7573 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
7574 | { | |
7575 | unsigned left, total; | |
7576 | int ret = 0; | |
7577 | ||
7578 | total = 0; | |
7579 | left = ctx->nr_user_files; | |
7580 | while (left) { | |
7581 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
7582 | |
7583 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
7584 | if (ret) | |
7585 | break; | |
7586 | left -= this_files; | |
7587 | total += this_files; | |
7588 | } | |
7589 | ||
7590 | if (!ret) | |
7591 | return 0; | |
7592 | ||
7593 | while (total < ctx->nr_user_files) { | |
65e19f54 JA |
7594 | struct file *file = io_file_from_index(ctx, total); |
7595 | ||
7596 | if (file) | |
7597 | fput(file); | |
6b06314c JA |
7598 | total++; |
7599 | } | |
7600 | ||
7601 | return ret; | |
7602 | } | |
7603 | #else | |
7604 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
7605 | { | |
7606 | return 0; | |
7607 | } | |
7608 | #endif | |
7609 | ||
5398ae69 PB |
7610 | static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data, |
7611 | unsigned nr_tables, unsigned nr_files) | |
65e19f54 JA |
7612 | { |
7613 | int i; | |
7614 | ||
7615 | for (i = 0; i < nr_tables; i++) { | |
5398ae69 | 7616 | struct fixed_file_table *table = &file_data->table[i]; |
65e19f54 JA |
7617 | unsigned this_files; |
7618 | ||
7619 | this_files = min(nr_files, IORING_MAX_FILES_TABLE); | |
7620 | table->files = kcalloc(this_files, sizeof(struct file *), | |
7621 | GFP_KERNEL); | |
7622 | if (!table->files) | |
7623 | break; | |
7624 | nr_files -= this_files; | |
7625 | } | |
7626 | ||
7627 | if (i == nr_tables) | |
7628 | return 0; | |
7629 | ||
7630 | for (i = 0; i < nr_tables; i++) { | |
5398ae69 | 7631 | struct fixed_file_table *table = &file_data->table[i]; |
65e19f54 JA |
7632 | kfree(table->files); |
7633 | } | |
7634 | return 1; | |
7635 | } | |
7636 | ||
05f3fb3c JA |
7637 | static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file) |
7638 | { | |
7639 | #if defined(CONFIG_UNIX) | |
7640 | struct sock *sock = ctx->ring_sock->sk; | |
7641 | struct sk_buff_head list, *head = &sock->sk_receive_queue; | |
7642 | struct sk_buff *skb; | |
7643 | int i; | |
7644 | ||
7645 | __skb_queue_head_init(&list); | |
7646 | ||
7647 | /* | |
7648 | * Find the skb that holds this file in its SCM_RIGHTS. When found, | |
7649 | * remove this entry and rearrange the file array. | |
7650 | */ | |
7651 | skb = skb_dequeue(head); | |
7652 | while (skb) { | |
7653 | struct scm_fp_list *fp; | |
7654 | ||
7655 | fp = UNIXCB(skb).fp; | |
7656 | for (i = 0; i < fp->count; i++) { | |
7657 | int left; | |
7658 | ||
7659 | if (fp->fp[i] != file) | |
7660 | continue; | |
7661 | ||
7662 | unix_notinflight(fp->user, fp->fp[i]); | |
7663 | left = fp->count - 1 - i; | |
7664 | if (left) { | |
7665 | memmove(&fp->fp[i], &fp->fp[i + 1], | |
7666 | left * sizeof(struct file *)); | |
7667 | } | |
7668 | fp->count--; | |
7669 | if (!fp->count) { | |
7670 | kfree_skb(skb); | |
7671 | skb = NULL; | |
7672 | } else { | |
7673 | __skb_queue_tail(&list, skb); | |
7674 | } | |
7675 | fput(file); | |
7676 | file = NULL; | |
7677 | break; | |
7678 | } | |
7679 | ||
7680 | if (!file) | |
7681 | break; | |
7682 | ||
7683 | __skb_queue_tail(&list, skb); | |
7684 | ||
7685 | skb = skb_dequeue(head); | |
7686 | } | |
7687 | ||
7688 | if (skb_peek(&list)) { | |
7689 | spin_lock_irq(&head->lock); | |
7690 | while ((skb = __skb_dequeue(&list)) != NULL) | |
7691 | __skb_queue_tail(head, skb); | |
7692 | spin_unlock_irq(&head->lock); | |
7693 | } | |
7694 | #else | |
7695 | fput(file); | |
7696 | #endif | |
7697 | } | |
7698 | ||
7699 | struct io_file_put { | |
05589553 | 7700 | struct list_head list; |
05f3fb3c | 7701 | struct file *file; |
05f3fb3c JA |
7702 | }; |
7703 | ||
4a38aed2 | 7704 | static void __io_file_put_work(struct fixed_file_ref_node *ref_node) |
65e19f54 | 7705 | { |
4a38aed2 JA |
7706 | struct fixed_file_data *file_data = ref_node->file_data; |
7707 | struct io_ring_ctx *ctx = file_data->ctx; | |
05f3fb3c | 7708 | struct io_file_put *pfile, *tmp; |
05589553 XW |
7709 | |
7710 | list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) { | |
6a4d07cd | 7711 | list_del(&pfile->list); |
05589553 XW |
7712 | io_ring_file_put(ctx, pfile->file); |
7713 | kfree(pfile); | |
65e19f54 | 7714 | } |
05589553 | 7715 | |
05589553 XW |
7716 | percpu_ref_exit(&ref_node->refs); |
7717 | kfree(ref_node); | |
7718 | percpu_ref_put(&file_data->refs); | |
2faf852d | 7719 | } |
65e19f54 | 7720 | |
4a38aed2 JA |
7721 | static void io_file_put_work(struct work_struct *work) |
7722 | { | |
7723 | struct io_ring_ctx *ctx; | |
7724 | struct llist_node *node; | |
7725 | ||
7726 | ctx = container_of(work, struct io_ring_ctx, file_put_work.work); | |
7727 | node = llist_del_all(&ctx->file_put_llist); | |
7728 | ||
7729 | while (node) { | |
7730 | struct fixed_file_ref_node *ref_node; | |
7731 | struct llist_node *next = node->next; | |
7732 | ||
7733 | ref_node = llist_entry(node, struct fixed_file_ref_node, llist); | |
7734 | __io_file_put_work(ref_node); | |
7735 | node = next; | |
7736 | } | |
7737 | } | |
7738 | ||
05589553 | 7739 | static void io_file_data_ref_zero(struct percpu_ref *ref) |
2faf852d | 7740 | { |
05589553 | 7741 | struct fixed_file_ref_node *ref_node; |
e297822b | 7742 | struct fixed_file_data *data; |
4a38aed2 | 7743 | struct io_ring_ctx *ctx; |
e297822b | 7744 | bool first_add = false; |
4a38aed2 | 7745 | int delay = HZ; |
65e19f54 | 7746 | |
05589553 | 7747 | ref_node = container_of(ref, struct fixed_file_ref_node, refs); |
e297822b PB |
7748 | data = ref_node->file_data; |
7749 | ctx = data->ctx; | |
7750 | ||
ac0648a5 | 7751 | spin_lock_bh(&data->lock); |
e297822b PB |
7752 | ref_node->done = true; |
7753 | ||
7754 | while (!list_empty(&data->ref_list)) { | |
7755 | ref_node = list_first_entry(&data->ref_list, | |
7756 | struct fixed_file_ref_node, node); | |
7757 | /* recycle ref nodes in order */ | |
7758 | if (!ref_node->done) | |
7759 | break; | |
7760 | list_del(&ref_node->node); | |
7761 | first_add |= llist_add(&ref_node->llist, &ctx->file_put_llist); | |
7762 | } | |
ac0648a5 | 7763 | spin_unlock_bh(&data->lock); |
05589553 | 7764 | |
e297822b | 7765 | if (percpu_ref_is_dying(&data->refs)) |
4a38aed2 | 7766 | delay = 0; |
05589553 | 7767 | |
4a38aed2 JA |
7768 | if (!delay) |
7769 | mod_delayed_work(system_wq, &ctx->file_put_work, 0); | |
7770 | else if (first_add) | |
7771 | queue_delayed_work(system_wq, &ctx->file_put_work, delay); | |
05f3fb3c | 7772 | } |
65e19f54 | 7773 | |
05589553 XW |
7774 | static struct fixed_file_ref_node *alloc_fixed_file_ref_node( |
7775 | struct io_ring_ctx *ctx) | |
05f3fb3c | 7776 | { |
05589553 | 7777 | struct fixed_file_ref_node *ref_node; |
05f3fb3c | 7778 | |
05589553 XW |
7779 | ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL); |
7780 | if (!ref_node) | |
3e2224c5 | 7781 | return NULL; |
05f3fb3c | 7782 | |
05589553 XW |
7783 | if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero, |
7784 | 0, GFP_KERNEL)) { | |
7785 | kfree(ref_node); | |
3e2224c5 | 7786 | return NULL; |
05589553 XW |
7787 | } |
7788 | INIT_LIST_HEAD(&ref_node->node); | |
7789 | INIT_LIST_HEAD(&ref_node->file_list); | |
05589553 | 7790 | ref_node->file_data = ctx->file_data; |
e297822b | 7791 | ref_node->done = false; |
05589553 | 7792 | return ref_node; |
05589553 XW |
7793 | } |
7794 | ||
7795 | static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node) | |
7796 | { | |
7797 | percpu_ref_exit(&ref_node->refs); | |
7798 | kfree(ref_node); | |
65e19f54 JA |
7799 | } |
7800 | ||
6b06314c JA |
7801 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, |
7802 | unsigned nr_args) | |
7803 | { | |
7804 | __s32 __user *fds = (__s32 __user *) arg; | |
600cf3f8 | 7805 | unsigned nr_tables, i; |
05f3fb3c | 7806 | struct file *file; |
600cf3f8 | 7807 | int fd, ret = -ENOMEM; |
05589553 | 7808 | struct fixed_file_ref_node *ref_node; |
5398ae69 | 7809 | struct fixed_file_data *file_data; |
6b06314c | 7810 | |
05f3fb3c | 7811 | if (ctx->file_data) |
6b06314c JA |
7812 | return -EBUSY; |
7813 | if (!nr_args) | |
7814 | return -EINVAL; | |
7815 | if (nr_args > IORING_MAX_FIXED_FILES) | |
7816 | return -EMFILE; | |
7817 | ||
5398ae69 PB |
7818 | file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL); |
7819 | if (!file_data) | |
05f3fb3c | 7820 | return -ENOMEM; |
5398ae69 PB |
7821 | file_data->ctx = ctx; |
7822 | init_completion(&file_data->done); | |
7823 | INIT_LIST_HEAD(&file_data->ref_list); | |
7824 | spin_lock_init(&file_data->lock); | |
05f3fb3c | 7825 | |
65e19f54 | 7826 | nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE); |
035fbafc | 7827 | file_data->table = kcalloc(nr_tables, sizeof(*file_data->table), |
5398ae69 | 7828 | GFP_KERNEL); |
600cf3f8 PB |
7829 | if (!file_data->table) |
7830 | goto out_free; | |
05f3fb3c | 7831 | |
5398ae69 | 7832 | if (percpu_ref_init(&file_data->refs, io_file_ref_kill, |
600cf3f8 PB |
7833 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) |
7834 | goto out_free; | |
6b06314c | 7835 | |
600cf3f8 PB |
7836 | if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args)) |
7837 | goto out_ref; | |
55cbc256 | 7838 | ctx->file_data = file_data; |
65e19f54 | 7839 | |
08a45173 | 7840 | for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { |
65e19f54 JA |
7841 | struct fixed_file_table *table; |
7842 | unsigned index; | |
7843 | ||
600cf3f8 PB |
7844 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) { |
7845 | ret = -EFAULT; | |
7846 | goto out_fput; | |
7847 | } | |
08a45173 | 7848 | /* allow sparse sets */ |
600cf3f8 | 7849 | if (fd == -1) |
08a45173 | 7850 | continue; |
6b06314c | 7851 | |
05f3fb3c | 7852 | file = fget(fd); |
6b06314c | 7853 | ret = -EBADF; |
05f3fb3c | 7854 | if (!file) |
600cf3f8 | 7855 | goto out_fput; |
05f3fb3c | 7856 | |
6b06314c JA |
7857 | /* |
7858 | * Don't allow io_uring instances to be registered. If UNIX | |
7859 | * isn't enabled, then this causes a reference cycle and this | |
7860 | * instance can never get freed. If UNIX is enabled we'll | |
7861 | * handle it just fine, but there's still no point in allowing | |
7862 | * a ring fd as it doesn't support regular read/write anyway. | |
7863 | */ | |
05f3fb3c JA |
7864 | if (file->f_op == &io_uring_fops) { |
7865 | fput(file); | |
600cf3f8 | 7866 | goto out_fput; |
6b06314c | 7867 | } |
600cf3f8 PB |
7868 | table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT]; |
7869 | index = i & IORING_FILE_TABLE_MASK; | |
05f3fb3c | 7870 | table->files[index] = file; |
6b06314c JA |
7871 | } |
7872 | ||
6b06314c | 7873 | ret = io_sqe_files_scm(ctx); |
05589553 | 7874 | if (ret) { |
6b06314c | 7875 | io_sqe_files_unregister(ctx); |
05589553 XW |
7876 | return ret; |
7877 | } | |
6b06314c | 7878 | |
05589553 | 7879 | ref_node = alloc_fixed_file_ref_node(ctx); |
3e2224c5 | 7880 | if (!ref_node) { |
05589553 | 7881 | io_sqe_files_unregister(ctx); |
3e2224c5 | 7882 | return -ENOMEM; |
05589553 XW |
7883 | } |
7884 | ||
1642b445 | 7885 | io_sqe_files_set_node(file_data, ref_node); |
6b06314c | 7886 | return ret; |
600cf3f8 PB |
7887 | out_fput: |
7888 | for (i = 0; i < ctx->nr_user_files; i++) { | |
7889 | file = io_file_from_index(ctx, i); | |
7890 | if (file) | |
7891 | fput(file); | |
7892 | } | |
7893 | for (i = 0; i < nr_tables; i++) | |
7894 | kfree(file_data->table[i].files); | |
7895 | ctx->nr_user_files = 0; | |
7896 | out_ref: | |
7897 | percpu_ref_exit(&file_data->refs); | |
7898 | out_free: | |
7899 | kfree(file_data->table); | |
7900 | kfree(file_data); | |
55cbc256 | 7901 | ctx->file_data = NULL; |
6b06314c JA |
7902 | return ret; |
7903 | } | |
7904 | ||
c3a31e60 JA |
7905 | static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file, |
7906 | int index) | |
7907 | { | |
7908 | #if defined(CONFIG_UNIX) | |
7909 | struct sock *sock = ctx->ring_sock->sk; | |
7910 | struct sk_buff_head *head = &sock->sk_receive_queue; | |
7911 | struct sk_buff *skb; | |
7912 | ||
7913 | /* | |
7914 | * See if we can merge this file into an existing skb SCM_RIGHTS | |
7915 | * file set. If there's no room, fall back to allocating a new skb | |
7916 | * and filling it in. | |
7917 | */ | |
7918 | spin_lock_irq(&head->lock); | |
7919 | skb = skb_peek(head); | |
7920 | if (skb) { | |
7921 | struct scm_fp_list *fpl = UNIXCB(skb).fp; | |
7922 | ||
7923 | if (fpl->count < SCM_MAX_FD) { | |
7924 | __skb_unlink(skb, head); | |
7925 | spin_unlock_irq(&head->lock); | |
7926 | fpl->fp[fpl->count] = get_file(file); | |
7927 | unix_inflight(fpl->user, fpl->fp[fpl->count]); | |
7928 | fpl->count++; | |
7929 | spin_lock_irq(&head->lock); | |
7930 | __skb_queue_head(head, skb); | |
7931 | } else { | |
7932 | skb = NULL; | |
7933 | } | |
7934 | } | |
7935 | spin_unlock_irq(&head->lock); | |
7936 | ||
7937 | if (skb) { | |
7938 | fput(file); | |
7939 | return 0; | |
7940 | } | |
7941 | ||
7942 | return __io_sqe_files_scm(ctx, 1, index); | |
7943 | #else | |
7944 | return 0; | |
7945 | #endif | |
7946 | } | |
7947 | ||
a5318d3c | 7948 | static int io_queue_file_removal(struct fixed_file_data *data, |
05589553 | 7949 | struct file *file) |
05f3fb3c | 7950 | { |
a5318d3c | 7951 | struct io_file_put *pfile; |
b2e96852 | 7952 | struct fixed_file_ref_node *ref_node = data->node; |
05f3fb3c | 7953 | |
05f3fb3c | 7954 | pfile = kzalloc(sizeof(*pfile), GFP_KERNEL); |
a5318d3c HD |
7955 | if (!pfile) |
7956 | return -ENOMEM; | |
05f3fb3c JA |
7957 | |
7958 | pfile->file = file; | |
05589553 XW |
7959 | list_add(&pfile->list, &ref_node->file_list); |
7960 | ||
a5318d3c | 7961 | return 0; |
05f3fb3c JA |
7962 | } |
7963 | ||
7964 | static int __io_sqe_files_update(struct io_ring_ctx *ctx, | |
7965 | struct io_uring_files_update *up, | |
7966 | unsigned nr_args) | |
7967 | { | |
7968 | struct fixed_file_data *data = ctx->file_data; | |
05589553 | 7969 | struct fixed_file_ref_node *ref_node; |
05f3fb3c | 7970 | struct file *file; |
c3a31e60 JA |
7971 | __s32 __user *fds; |
7972 | int fd, i, err; | |
7973 | __u32 done; | |
05589553 | 7974 | bool needs_switch = false; |
c3a31e60 | 7975 | |
05f3fb3c | 7976 | if (check_add_overflow(up->offset, nr_args, &done)) |
c3a31e60 JA |
7977 | return -EOVERFLOW; |
7978 | if (done > ctx->nr_user_files) | |
7979 | return -EINVAL; | |
7980 | ||
05589553 | 7981 | ref_node = alloc_fixed_file_ref_node(ctx); |
3e2224c5 MWO |
7982 | if (!ref_node) |
7983 | return -ENOMEM; | |
05589553 | 7984 | |
c3a31e60 | 7985 | done = 0; |
05f3fb3c | 7986 | fds = u64_to_user_ptr(up->fds); |
c3a31e60 | 7987 | while (nr_args) { |
65e19f54 JA |
7988 | struct fixed_file_table *table; |
7989 | unsigned index; | |
7990 | ||
c3a31e60 JA |
7991 | err = 0; |
7992 | if (copy_from_user(&fd, &fds[done], sizeof(fd))) { | |
7993 | err = -EFAULT; | |
7994 | break; | |
7995 | } | |
05f3fb3c JA |
7996 | i = array_index_nospec(up->offset, ctx->nr_user_files); |
7997 | table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT]; | |
65e19f54 JA |
7998 | index = i & IORING_FILE_TABLE_MASK; |
7999 | if (table->files[index]) { | |
98dfd502 | 8000 | file = table->files[index]; |
a5318d3c HD |
8001 | err = io_queue_file_removal(data, file); |
8002 | if (err) | |
8003 | break; | |
65e19f54 | 8004 | table->files[index] = NULL; |
05589553 | 8005 | needs_switch = true; |
c3a31e60 JA |
8006 | } |
8007 | if (fd != -1) { | |
c3a31e60 JA |
8008 | file = fget(fd); |
8009 | if (!file) { | |
8010 | err = -EBADF; | |
8011 | break; | |
8012 | } | |
8013 | /* | |
8014 | * Don't allow io_uring instances to be registered. If | |
8015 | * UNIX isn't enabled, then this causes a reference | |
8016 | * cycle and this instance can never get freed. If UNIX | |
8017 | * is enabled we'll handle it just fine, but there's | |
8018 | * still no point in allowing a ring fd as it doesn't | |
8019 | * support regular read/write anyway. | |
8020 | */ | |
8021 | if (file->f_op == &io_uring_fops) { | |
8022 | fput(file); | |
8023 | err = -EBADF; | |
8024 | break; | |
8025 | } | |
65e19f54 | 8026 | table->files[index] = file; |
c3a31e60 | 8027 | err = io_sqe_file_register(ctx, file, i); |
f3bd9dae | 8028 | if (err) { |
95d1c8e5 | 8029 | table->files[index] = NULL; |
f3bd9dae | 8030 | fput(file); |
c3a31e60 | 8031 | break; |
f3bd9dae | 8032 | } |
c3a31e60 JA |
8033 | } |
8034 | nr_args--; | |
8035 | done++; | |
05f3fb3c JA |
8036 | up->offset++; |
8037 | } | |
8038 | ||
05589553 | 8039 | if (needs_switch) { |
b2e96852 | 8040 | percpu_ref_kill(&data->node->refs); |
1642b445 | 8041 | io_sqe_files_set_node(data, ref_node); |
05589553 XW |
8042 | } else |
8043 | destroy_fixed_file_ref_node(ref_node); | |
c3a31e60 JA |
8044 | |
8045 | return done ? done : err; | |
8046 | } | |
05589553 | 8047 | |
05f3fb3c JA |
8048 | static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg, |
8049 | unsigned nr_args) | |
8050 | { | |
8051 | struct io_uring_files_update up; | |
8052 | ||
8053 | if (!ctx->file_data) | |
8054 | return -ENXIO; | |
8055 | if (!nr_args) | |
8056 | return -EINVAL; | |
8057 | if (copy_from_user(&up, arg, sizeof(up))) | |
8058 | return -EFAULT; | |
8059 | if (up.resv) | |
8060 | return -EINVAL; | |
8061 | ||
8062 | return __io_sqe_files_update(ctx, &up, nr_args); | |
8063 | } | |
c3a31e60 | 8064 | |
e9fd9396 | 8065 | static void io_free_work(struct io_wq_work *work) |
7d723065 JA |
8066 | { |
8067 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
8068 | ||
e9fd9396 | 8069 | /* Consider that io_steal_work() relies on this ref */ |
7d723065 JA |
8070 | io_put_req(req); |
8071 | } | |
8072 | ||
24369c2e PB |
8073 | static int io_init_wq_offload(struct io_ring_ctx *ctx, |
8074 | struct io_uring_params *p) | |
8075 | { | |
8076 | struct io_wq_data data; | |
8077 | struct fd f; | |
8078 | struct io_ring_ctx *ctx_attach; | |
8079 | unsigned int concurrency; | |
8080 | int ret = 0; | |
8081 | ||
8082 | data.user = ctx->user; | |
e9fd9396 | 8083 | data.free_work = io_free_work; |
f5fa38c5 | 8084 | data.do_work = io_wq_submit_work; |
24369c2e PB |
8085 | |
8086 | if (!(p->flags & IORING_SETUP_ATTACH_WQ)) { | |
8087 | /* Do QD, or 4 * CPUS, whatever is smallest */ | |
8088 | concurrency = min(ctx->sq_entries, 4 * num_online_cpus()); | |
8089 | ||
8090 | ctx->io_wq = io_wq_create(concurrency, &data); | |
8091 | if (IS_ERR(ctx->io_wq)) { | |
8092 | ret = PTR_ERR(ctx->io_wq); | |
8093 | ctx->io_wq = NULL; | |
8094 | } | |
8095 | return ret; | |
8096 | } | |
8097 | ||
8098 | f = fdget(p->wq_fd); | |
8099 | if (!f.file) | |
8100 | return -EBADF; | |
8101 | ||
8102 | if (f.file->f_op != &io_uring_fops) { | |
8103 | ret = -EINVAL; | |
8104 | goto out_fput; | |
8105 | } | |
8106 | ||
8107 | ctx_attach = f.file->private_data; | |
8108 | /* @io_wq is protected by holding the fd */ | |
8109 | if (!io_wq_get(ctx_attach->io_wq, &data)) { | |
8110 | ret = -EINVAL; | |
8111 | goto out_fput; | |
8112 | } | |
8113 | ||
8114 | ctx->io_wq = ctx_attach->io_wq; | |
8115 | out_fput: | |
8116 | fdput(f); | |
8117 | return ret; | |
8118 | } | |
8119 | ||
0f212204 JA |
8120 | static int io_uring_alloc_task_context(struct task_struct *task) |
8121 | { | |
8122 | struct io_uring_task *tctx; | |
d8a6df10 | 8123 | int ret; |
0f212204 JA |
8124 | |
8125 | tctx = kmalloc(sizeof(*tctx), GFP_KERNEL); | |
8126 | if (unlikely(!tctx)) | |
8127 | return -ENOMEM; | |
8128 | ||
d8a6df10 JA |
8129 | ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL); |
8130 | if (unlikely(ret)) { | |
8131 | kfree(tctx); | |
8132 | return ret; | |
8133 | } | |
8134 | ||
0f212204 JA |
8135 | xa_init(&tctx->xa); |
8136 | init_waitqueue_head(&tctx->wait); | |
8137 | tctx->last = NULL; | |
fdaf083c JA |
8138 | atomic_set(&tctx->in_idle, 0); |
8139 | tctx->sqpoll = false; | |
500a373d JA |
8140 | io_init_identity(&tctx->__identity); |
8141 | tctx->identity = &tctx->__identity; | |
0f212204 JA |
8142 | task->io_uring = tctx; |
8143 | return 0; | |
8144 | } | |
8145 | ||
8146 | void __io_uring_free(struct task_struct *tsk) | |
8147 | { | |
8148 | struct io_uring_task *tctx = tsk->io_uring; | |
8149 | ||
8150 | WARN_ON_ONCE(!xa_empty(&tctx->xa)); | |
500a373d JA |
8151 | WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1); |
8152 | if (tctx->identity != &tctx->__identity) | |
8153 | kfree(tctx->identity); | |
d8a6df10 | 8154 | percpu_counter_destroy(&tctx->inflight); |
0f212204 JA |
8155 | kfree(tctx); |
8156 | tsk->io_uring = NULL; | |
8157 | } | |
8158 | ||
7e84e1c7 SG |
8159 | static int io_sq_offload_create(struct io_ring_ctx *ctx, |
8160 | struct io_uring_params *p) | |
2b188cc1 JA |
8161 | { |
8162 | int ret; | |
8163 | ||
6c271ce2 | 8164 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
534ca6d6 JA |
8165 | struct io_sq_data *sqd; |
8166 | ||
3ec482d1 | 8167 | ret = -EPERM; |
ce59fc69 | 8168 | if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE)) |
3ec482d1 JA |
8169 | goto err; |
8170 | ||
534ca6d6 JA |
8171 | sqd = io_get_sq_data(p); |
8172 | if (IS_ERR(sqd)) { | |
8173 | ret = PTR_ERR(sqd); | |
8174 | goto err; | |
8175 | } | |
69fb2131 | 8176 | |
534ca6d6 | 8177 | ctx->sq_data = sqd; |
69fb2131 JA |
8178 | io_sq_thread_park(sqd); |
8179 | mutex_lock(&sqd->ctx_lock); | |
8180 | list_add(&ctx->sqd_list, &sqd->ctx_new_list); | |
8181 | mutex_unlock(&sqd->ctx_lock); | |
8182 | io_sq_thread_unpark(sqd); | |
534ca6d6 | 8183 | |
917257da JA |
8184 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
8185 | if (!ctx->sq_thread_idle) | |
8186 | ctx->sq_thread_idle = HZ; | |
8187 | ||
aa06165d JA |
8188 | if (sqd->thread) |
8189 | goto done; | |
8190 | ||
6c271ce2 | 8191 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 8192 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 8193 | |
917257da | 8194 | ret = -EINVAL; |
44a9bd18 JA |
8195 | if (cpu >= nr_cpu_ids) |
8196 | goto err; | |
7889f44d | 8197 | if (!cpu_online(cpu)) |
917257da JA |
8198 | goto err; |
8199 | ||
69fb2131 | 8200 | sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd, |
534ca6d6 | 8201 | cpu, "io_uring-sq"); |
6c271ce2 | 8202 | } else { |
69fb2131 | 8203 | sqd->thread = kthread_create(io_sq_thread, sqd, |
6c271ce2 JA |
8204 | "io_uring-sq"); |
8205 | } | |
534ca6d6 JA |
8206 | if (IS_ERR(sqd->thread)) { |
8207 | ret = PTR_ERR(sqd->thread); | |
8208 | sqd->thread = NULL; | |
6c271ce2 JA |
8209 | goto err; |
8210 | } | |
534ca6d6 | 8211 | ret = io_uring_alloc_task_context(sqd->thread); |
0f212204 JA |
8212 | if (ret) |
8213 | goto err; | |
6c271ce2 JA |
8214 | } else if (p->flags & IORING_SETUP_SQ_AFF) { |
8215 | /* Can't have SQ_AFF without SQPOLL */ | |
8216 | ret = -EINVAL; | |
8217 | goto err; | |
8218 | } | |
8219 | ||
aa06165d | 8220 | done: |
24369c2e PB |
8221 | ret = io_init_wq_offload(ctx, p); |
8222 | if (ret) | |
2b188cc1 | 8223 | goto err; |
2b188cc1 JA |
8224 | |
8225 | return 0; | |
8226 | err: | |
54a91f3b | 8227 | io_finish_async(ctx); |
2b188cc1 JA |
8228 | return ret; |
8229 | } | |
8230 | ||
7e84e1c7 SG |
8231 | static void io_sq_offload_start(struct io_ring_ctx *ctx) |
8232 | { | |
534ca6d6 JA |
8233 | struct io_sq_data *sqd = ctx->sq_data; |
8234 | ||
8235 | if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread) | |
8236 | wake_up_process(sqd->thread); | |
7e84e1c7 SG |
8237 | } |
8238 | ||
a087e2b5 BM |
8239 | static inline void __io_unaccount_mem(struct user_struct *user, |
8240 | unsigned long nr_pages) | |
2b188cc1 JA |
8241 | { |
8242 | atomic_long_sub(nr_pages, &user->locked_vm); | |
8243 | } | |
8244 | ||
a087e2b5 BM |
8245 | static inline int __io_account_mem(struct user_struct *user, |
8246 | unsigned long nr_pages) | |
2b188cc1 JA |
8247 | { |
8248 | unsigned long page_limit, cur_pages, new_pages; | |
8249 | ||
8250 | /* Don't allow more pages than we can safely lock */ | |
8251 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
8252 | ||
8253 | do { | |
8254 | cur_pages = atomic_long_read(&user->locked_vm); | |
8255 | new_pages = cur_pages + nr_pages; | |
8256 | if (new_pages > page_limit) | |
8257 | return -ENOMEM; | |
8258 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
8259 | new_pages) != cur_pages); | |
8260 | ||
8261 | return 0; | |
8262 | } | |
8263 | ||
2e0464d4 BM |
8264 | static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages, |
8265 | enum io_mem_account acct) | |
a087e2b5 | 8266 | { |
aad5d8da | 8267 | if (ctx->limit_mem) |
a087e2b5 | 8268 | __io_unaccount_mem(ctx->user, nr_pages); |
30975825 | 8269 | |
2aede0e4 | 8270 | if (ctx->mm_account) { |
4bc4a912 JA |
8271 | if (acct == ACCT_LOCKED) { |
8272 | mmap_write_lock(ctx->mm_account); | |
2aede0e4 | 8273 | ctx->mm_account->locked_vm -= nr_pages; |
4bc4a912 JA |
8274 | mmap_write_unlock(ctx->mm_account); |
8275 | }else if (acct == ACCT_PINNED) { | |
2aede0e4 | 8276 | atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm); |
4bc4a912 | 8277 | } |
2e0464d4 | 8278 | } |
a087e2b5 BM |
8279 | } |
8280 | ||
2e0464d4 BM |
8281 | static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages, |
8282 | enum io_mem_account acct) | |
a087e2b5 | 8283 | { |
30975825 BM |
8284 | int ret; |
8285 | ||
8286 | if (ctx->limit_mem) { | |
8287 | ret = __io_account_mem(ctx->user, nr_pages); | |
8288 | if (ret) | |
8289 | return ret; | |
8290 | } | |
8291 | ||
2aede0e4 | 8292 | if (ctx->mm_account) { |
4bc4a912 JA |
8293 | if (acct == ACCT_LOCKED) { |
8294 | mmap_write_lock(ctx->mm_account); | |
2aede0e4 | 8295 | ctx->mm_account->locked_vm += nr_pages; |
4bc4a912 JA |
8296 | mmap_write_unlock(ctx->mm_account); |
8297 | } else if (acct == ACCT_PINNED) { | |
2aede0e4 | 8298 | atomic64_add(nr_pages, &ctx->mm_account->pinned_vm); |
4bc4a912 | 8299 | } |
2e0464d4 | 8300 | } |
a087e2b5 BM |
8301 | |
8302 | return 0; | |
8303 | } | |
8304 | ||
2b188cc1 JA |
8305 | static void io_mem_free(void *ptr) |
8306 | { | |
52e04ef4 MR |
8307 | struct page *page; |
8308 | ||
8309 | if (!ptr) | |
8310 | return; | |
2b188cc1 | 8311 | |
52e04ef4 | 8312 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
8313 | if (put_page_testzero(page)) |
8314 | free_compound_page(page); | |
8315 | } | |
8316 | ||
8317 | static void *io_mem_alloc(size_t size) | |
8318 | { | |
8319 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
8320 | __GFP_NORETRY; | |
8321 | ||
8322 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
8323 | } | |
8324 | ||
75b28aff HV |
8325 | static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries, |
8326 | size_t *sq_offset) | |
8327 | { | |
8328 | struct io_rings *rings; | |
8329 | size_t off, sq_array_size; | |
8330 | ||
8331 | off = struct_size(rings, cqes, cq_entries); | |
8332 | if (off == SIZE_MAX) | |
8333 | return SIZE_MAX; | |
8334 | ||
8335 | #ifdef CONFIG_SMP | |
8336 | off = ALIGN(off, SMP_CACHE_BYTES); | |
8337 | if (off == 0) | |
8338 | return SIZE_MAX; | |
8339 | #endif | |
8340 | ||
b36200f5 DV |
8341 | if (sq_offset) |
8342 | *sq_offset = off; | |
8343 | ||
75b28aff HV |
8344 | sq_array_size = array_size(sizeof(u32), sq_entries); |
8345 | if (sq_array_size == SIZE_MAX) | |
8346 | return SIZE_MAX; | |
8347 | ||
8348 | if (check_add_overflow(off, sq_array_size, &off)) | |
8349 | return SIZE_MAX; | |
8350 | ||
75b28aff HV |
8351 | return off; |
8352 | } | |
8353 | ||
2b188cc1 JA |
8354 | static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries) |
8355 | { | |
75b28aff | 8356 | size_t pages; |
2b188cc1 | 8357 | |
75b28aff HV |
8358 | pages = (size_t)1 << get_order( |
8359 | rings_size(sq_entries, cq_entries, NULL)); | |
8360 | pages += (size_t)1 << get_order( | |
8361 | array_size(sizeof(struct io_uring_sqe), sq_entries)); | |
2b188cc1 | 8362 | |
75b28aff | 8363 | return pages; |
2b188cc1 JA |
8364 | } |
8365 | ||
edafccee JA |
8366 | static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx) |
8367 | { | |
8368 | int i, j; | |
8369 | ||
8370 | if (!ctx->user_bufs) | |
8371 | return -ENXIO; | |
8372 | ||
8373 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
8374 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
8375 | ||
8376 | for (j = 0; j < imu->nr_bvecs; j++) | |
f1f6a7dd | 8377 | unpin_user_page(imu->bvec[j].bv_page); |
edafccee | 8378 | |
de293938 JA |
8379 | if (imu->acct_pages) |
8380 | io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED); | |
d4ef6475 | 8381 | kvfree(imu->bvec); |
edafccee JA |
8382 | imu->nr_bvecs = 0; |
8383 | } | |
8384 | ||
8385 | kfree(ctx->user_bufs); | |
8386 | ctx->user_bufs = NULL; | |
8387 | ctx->nr_user_bufs = 0; | |
8388 | return 0; | |
8389 | } | |
8390 | ||
8391 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
8392 | void __user *arg, unsigned index) | |
8393 | { | |
8394 | struct iovec __user *src; | |
8395 | ||
8396 | #ifdef CONFIG_COMPAT | |
8397 | if (ctx->compat) { | |
8398 | struct compat_iovec __user *ciovs; | |
8399 | struct compat_iovec ciov; | |
8400 | ||
8401 | ciovs = (struct compat_iovec __user *) arg; | |
8402 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
8403 | return -EFAULT; | |
8404 | ||
d55e5f5b | 8405 | dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base); |
edafccee JA |
8406 | dst->iov_len = ciov.iov_len; |
8407 | return 0; | |
8408 | } | |
8409 | #endif | |
8410 | src = (struct iovec __user *) arg; | |
8411 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
8412 | return -EFAULT; | |
8413 | return 0; | |
8414 | } | |
8415 | ||
de293938 JA |
8416 | /* |
8417 | * Not super efficient, but this is just a registration time. And we do cache | |
8418 | * the last compound head, so generally we'll only do a full search if we don't | |
8419 | * match that one. | |
8420 | * | |
8421 | * We check if the given compound head page has already been accounted, to | |
8422 | * avoid double accounting it. This allows us to account the full size of the | |
8423 | * page, not just the constituent pages of a huge page. | |
8424 | */ | |
8425 | static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages, | |
8426 | int nr_pages, struct page *hpage) | |
8427 | { | |
8428 | int i, j; | |
8429 | ||
8430 | /* check current page array */ | |
8431 | for (i = 0; i < nr_pages; i++) { | |
8432 | if (!PageCompound(pages[i])) | |
8433 | continue; | |
8434 | if (compound_head(pages[i]) == hpage) | |
8435 | return true; | |
8436 | } | |
8437 | ||
8438 | /* check previously registered pages */ | |
8439 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
8440 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
8441 | ||
8442 | for (j = 0; j < imu->nr_bvecs; j++) { | |
8443 | if (!PageCompound(imu->bvec[j].bv_page)) | |
8444 | continue; | |
8445 | if (compound_head(imu->bvec[j].bv_page) == hpage) | |
8446 | return true; | |
8447 | } | |
8448 | } | |
8449 | ||
8450 | return false; | |
8451 | } | |
8452 | ||
8453 | static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages, | |
8454 | int nr_pages, struct io_mapped_ubuf *imu, | |
8455 | struct page **last_hpage) | |
8456 | { | |
8457 | int i, ret; | |
8458 | ||
8459 | for (i = 0; i < nr_pages; i++) { | |
8460 | if (!PageCompound(pages[i])) { | |
8461 | imu->acct_pages++; | |
8462 | } else { | |
8463 | struct page *hpage; | |
8464 | ||
8465 | hpage = compound_head(pages[i]); | |
8466 | if (hpage == *last_hpage) | |
8467 | continue; | |
8468 | *last_hpage = hpage; | |
8469 | if (headpage_already_acct(ctx, pages, i, hpage)) | |
8470 | continue; | |
8471 | imu->acct_pages += page_size(hpage) >> PAGE_SHIFT; | |
8472 | } | |
8473 | } | |
8474 | ||
8475 | if (!imu->acct_pages) | |
8476 | return 0; | |
8477 | ||
8478 | ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED); | |
8479 | if (ret) | |
8480 | imu->acct_pages = 0; | |
8481 | return ret; | |
8482 | } | |
8483 | ||
edafccee JA |
8484 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, |
8485 | unsigned nr_args) | |
8486 | { | |
8487 | struct vm_area_struct **vmas = NULL; | |
8488 | struct page **pages = NULL; | |
de293938 | 8489 | struct page *last_hpage = NULL; |
edafccee JA |
8490 | int i, j, got_pages = 0; |
8491 | int ret = -EINVAL; | |
8492 | ||
8493 | if (ctx->user_bufs) | |
8494 | return -EBUSY; | |
8495 | if (!nr_args || nr_args > UIO_MAXIOV) | |
8496 | return -EINVAL; | |
8497 | ||
8498 | ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf), | |
8499 | GFP_KERNEL); | |
8500 | if (!ctx->user_bufs) | |
8501 | return -ENOMEM; | |
8502 | ||
8503 | for (i = 0; i < nr_args; i++) { | |
8504 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
8505 | unsigned long off, start, end, ubuf; | |
8506 | int pret, nr_pages; | |
8507 | struct iovec iov; | |
8508 | size_t size; | |
8509 | ||
8510 | ret = io_copy_iov(ctx, &iov, arg, i); | |
8511 | if (ret) | |
a278682d | 8512 | goto err; |
edafccee JA |
8513 | |
8514 | /* | |
8515 | * Don't impose further limits on the size and buffer | |
8516 | * constraints here, we'll -EINVAL later when IO is | |
8517 | * submitted if they are wrong. | |
8518 | */ | |
8519 | ret = -EFAULT; | |
8520 | if (!iov.iov_base || !iov.iov_len) | |
8521 | goto err; | |
8522 | ||
8523 | /* arbitrary limit, but we need something */ | |
8524 | if (iov.iov_len > SZ_1G) | |
8525 | goto err; | |
8526 | ||
8527 | ubuf = (unsigned long) iov.iov_base; | |
8528 | end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
8529 | start = ubuf >> PAGE_SHIFT; | |
8530 | nr_pages = end - start; | |
8531 | ||
edafccee JA |
8532 | ret = 0; |
8533 | if (!pages || nr_pages > got_pages) { | |
a8c73c1a DE |
8534 | kvfree(vmas); |
8535 | kvfree(pages); | |
d4ef6475 | 8536 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), |
edafccee | 8537 | GFP_KERNEL); |
d4ef6475 | 8538 | vmas = kvmalloc_array(nr_pages, |
edafccee JA |
8539 | sizeof(struct vm_area_struct *), |
8540 | GFP_KERNEL); | |
8541 | if (!pages || !vmas) { | |
8542 | ret = -ENOMEM; | |
edafccee JA |
8543 | goto err; |
8544 | } | |
8545 | got_pages = nr_pages; | |
8546 | } | |
8547 | ||
d4ef6475 | 8548 | imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec), |
edafccee JA |
8549 | GFP_KERNEL); |
8550 | ret = -ENOMEM; | |
de293938 | 8551 | if (!imu->bvec) |
edafccee | 8552 | goto err; |
edafccee JA |
8553 | |
8554 | ret = 0; | |
d8ed45c5 | 8555 | mmap_read_lock(current->mm); |
2113b05d | 8556 | pret = pin_user_pages(ubuf, nr_pages, |
932f4a63 IW |
8557 | FOLL_WRITE | FOLL_LONGTERM, |
8558 | pages, vmas); | |
edafccee JA |
8559 | if (pret == nr_pages) { |
8560 | /* don't support file backed memory */ | |
8561 | for (j = 0; j < nr_pages; j++) { | |
8562 | struct vm_area_struct *vma = vmas[j]; | |
8563 | ||
8564 | if (vma->vm_file && | |
8565 | !is_file_hugepages(vma->vm_file)) { | |
8566 | ret = -EOPNOTSUPP; | |
8567 | break; | |
8568 | } | |
8569 | } | |
8570 | } else { | |
8571 | ret = pret < 0 ? pret : -EFAULT; | |
8572 | } | |
d8ed45c5 | 8573 | mmap_read_unlock(current->mm); |
edafccee JA |
8574 | if (ret) { |
8575 | /* | |
8576 | * if we did partial map, or found file backed vmas, | |
8577 | * release any pages we did get | |
8578 | */ | |
27c4d3a3 | 8579 | if (pret > 0) |
f1f6a7dd | 8580 | unpin_user_pages(pages, pret); |
de293938 JA |
8581 | kvfree(imu->bvec); |
8582 | goto err; | |
8583 | } | |
8584 | ||
8585 | ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage); | |
8586 | if (ret) { | |
8587 | unpin_user_pages(pages, pret); | |
d4ef6475 | 8588 | kvfree(imu->bvec); |
edafccee JA |
8589 | goto err; |
8590 | } | |
8591 | ||
8592 | off = ubuf & ~PAGE_MASK; | |
8593 | size = iov.iov_len; | |
8594 | for (j = 0; j < nr_pages; j++) { | |
8595 | size_t vec_len; | |
8596 | ||
8597 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
8598 | imu->bvec[j].bv_page = pages[j]; | |
8599 | imu->bvec[j].bv_len = vec_len; | |
8600 | imu->bvec[j].bv_offset = off; | |
8601 | off = 0; | |
8602 | size -= vec_len; | |
8603 | } | |
8604 | /* store original address for later verification */ | |
8605 | imu->ubuf = ubuf; | |
8606 | imu->len = iov.iov_len; | |
8607 | imu->nr_bvecs = nr_pages; | |
8608 | ||
8609 | ctx->nr_user_bufs++; | |
8610 | } | |
d4ef6475 MR |
8611 | kvfree(pages); |
8612 | kvfree(vmas); | |
edafccee JA |
8613 | return 0; |
8614 | err: | |
d4ef6475 MR |
8615 | kvfree(pages); |
8616 | kvfree(vmas); | |
edafccee JA |
8617 | io_sqe_buffer_unregister(ctx); |
8618 | return ret; | |
8619 | } | |
8620 | ||
9b402849 JA |
8621 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
8622 | { | |
8623 | __s32 __user *fds = arg; | |
8624 | int fd; | |
8625 | ||
8626 | if (ctx->cq_ev_fd) | |
8627 | return -EBUSY; | |
8628 | ||
8629 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
8630 | return -EFAULT; | |
8631 | ||
8632 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
8633 | if (IS_ERR(ctx->cq_ev_fd)) { | |
8634 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
8635 | ctx->cq_ev_fd = NULL; | |
8636 | return ret; | |
8637 | } | |
8638 | ||
8639 | return 0; | |
8640 | } | |
8641 | ||
8642 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
8643 | { | |
8644 | if (ctx->cq_ev_fd) { | |
8645 | eventfd_ctx_put(ctx->cq_ev_fd); | |
8646 | ctx->cq_ev_fd = NULL; | |
8647 | return 0; | |
8648 | } | |
8649 | ||
8650 | return -ENXIO; | |
8651 | } | |
8652 | ||
5a2e745d JA |
8653 | static int __io_destroy_buffers(int id, void *p, void *data) |
8654 | { | |
8655 | struct io_ring_ctx *ctx = data; | |
8656 | struct io_buffer *buf = p; | |
8657 | ||
067524e9 | 8658 | __io_remove_buffers(ctx, buf, id, -1U); |
5a2e745d JA |
8659 | return 0; |
8660 | } | |
8661 | ||
8662 | static void io_destroy_buffers(struct io_ring_ctx *ctx) | |
8663 | { | |
8664 | idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx); | |
8665 | idr_destroy(&ctx->io_buffer_idr); | |
8666 | } | |
8667 | ||
2b188cc1 JA |
8668 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
8669 | { | |
6b06314c | 8670 | io_finish_async(ctx); |
5dbcad51 | 8671 | io_sqe_buffer_unregister(ctx); |
2aede0e4 JA |
8672 | |
8673 | if (ctx->sqo_task) { | |
8674 | put_task_struct(ctx->sqo_task); | |
8675 | ctx->sqo_task = NULL; | |
8676 | mmdrop(ctx->mm_account); | |
8677 | ctx->mm_account = NULL; | |
30975825 | 8678 | } |
def596e9 | 8679 | |
91d8f519 DZ |
8680 | #ifdef CONFIG_BLK_CGROUP |
8681 | if (ctx->sqo_blkcg_css) | |
8682 | css_put(ctx->sqo_blkcg_css); | |
8683 | #endif | |
8684 | ||
6b06314c | 8685 | io_sqe_files_unregister(ctx); |
9b402849 | 8686 | io_eventfd_unregister(ctx); |
5a2e745d | 8687 | io_destroy_buffers(ctx); |
41726c9a | 8688 | idr_destroy(&ctx->personality_idr); |
def596e9 | 8689 | |
2b188cc1 | 8690 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
8691 | if (ctx->ring_sock) { |
8692 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 8693 | sock_release(ctx->ring_sock); |
355e8d26 | 8694 | } |
2b188cc1 JA |
8695 | #endif |
8696 | ||
75b28aff | 8697 | io_mem_free(ctx->rings); |
2b188cc1 | 8698 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
8699 | |
8700 | percpu_ref_exit(&ctx->refs); | |
2b188cc1 | 8701 | free_uid(ctx->user); |
181e448d | 8702 | put_cred(ctx->creds); |
78076bb6 | 8703 | kfree(ctx->cancel_hash); |
0ddf92e8 | 8704 | kmem_cache_free(req_cachep, ctx->fallback_req); |
2b188cc1 JA |
8705 | kfree(ctx); |
8706 | } | |
8707 | ||
8708 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
8709 | { | |
8710 | struct io_ring_ctx *ctx = file->private_data; | |
8711 | __poll_t mask = 0; | |
8712 | ||
8713 | poll_wait(file, &ctx->cq_wait, wait); | |
4f7067c3 SB |
8714 | /* |
8715 | * synchronizes with barrier from wq_has_sleeper call in | |
8716 | * io_commit_cqring | |
8717 | */ | |
2b188cc1 | 8718 | smp_rmb(); |
90554200 | 8719 | if (!io_sqring_full(ctx)) |
2b188cc1 | 8720 | mask |= EPOLLOUT | EPOLLWRNORM; |
6c503150 PB |
8721 | io_cqring_overflow_flush(ctx, false, NULL, NULL); |
8722 | if (io_cqring_events(ctx)) | |
2b188cc1 JA |
8723 | mask |= EPOLLIN | EPOLLRDNORM; |
8724 | ||
8725 | return mask; | |
8726 | } | |
8727 | ||
8728 | static int io_uring_fasync(int fd, struct file *file, int on) | |
8729 | { | |
8730 | struct io_ring_ctx *ctx = file->private_data; | |
8731 | ||
8732 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
8733 | } | |
8734 | ||
071698e1 JA |
8735 | static int io_remove_personalities(int id, void *p, void *data) |
8736 | { | |
8737 | struct io_ring_ctx *ctx = data; | |
1e6fa521 | 8738 | struct io_identity *iod; |
071698e1 | 8739 | |
1e6fa521 JA |
8740 | iod = idr_remove(&ctx->personality_idr, id); |
8741 | if (iod) { | |
8742 | put_cred(iod->creds); | |
8743 | if (refcount_dec_and_test(&iod->count)) | |
8744 | kfree(iod); | |
8745 | } | |
071698e1 JA |
8746 | return 0; |
8747 | } | |
8748 | ||
85faa7b8 JA |
8749 | static void io_ring_exit_work(struct work_struct *work) |
8750 | { | |
b2edc0a7 PB |
8751 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, |
8752 | exit_work); | |
85faa7b8 | 8753 | |
56952e91 JA |
8754 | /* |
8755 | * If we're doing polled IO and end up having requests being | |
8756 | * submitted async (out-of-line), then completions can come in while | |
8757 | * we're waiting for refs to drop. We need to reap these manually, | |
8758 | * as nobody else will be looking for them. | |
8759 | */ | |
b2edc0a7 | 8760 | do { |
90df0853 | 8761 | __io_uring_cancel_task_requests(ctx, NULL); |
b2edc0a7 | 8762 | } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20)); |
85faa7b8 JA |
8763 | io_ring_ctx_free(ctx); |
8764 | } | |
8765 | ||
00c18640 JA |
8766 | static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data) |
8767 | { | |
8768 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
8769 | ||
8770 | return req->ctx == data; | |
8771 | } | |
8772 | ||
2b188cc1 JA |
8773 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
8774 | { | |
8775 | mutex_lock(&ctx->uring_lock); | |
8776 | percpu_ref_kill(&ctx->refs); | |
d9d05217 PB |
8777 | |
8778 | if (WARN_ON_ONCE((ctx->flags & IORING_SETUP_SQPOLL) && !ctx->sqo_dead)) | |
8779 | ctx->sqo_dead = 1; | |
8780 | ||
cda286f0 PB |
8781 | /* if force is set, the ring is going away. always drop after that */ |
8782 | ctx->cq_overflow_flushed = 1; | |
634578f8 | 8783 | if (ctx->rings) |
6c503150 | 8784 | __io_cqring_overflow_flush(ctx, true, NULL, NULL); |
2b188cc1 JA |
8785 | mutex_unlock(&ctx->uring_lock); |
8786 | ||
6b81928d PB |
8787 | io_kill_timeouts(ctx, NULL, NULL); |
8788 | io_poll_remove_all(ctx, NULL, NULL); | |
561fb04a JA |
8789 | |
8790 | if (ctx->io_wq) | |
00c18640 | 8791 | io_wq_cancel_cb(ctx->io_wq, io_cancel_ctx_cb, ctx, true); |
561fb04a | 8792 | |
15dff286 | 8793 | /* if we failed setting up the ctx, we might not have any rings */ |
b2edc0a7 | 8794 | io_iopoll_try_reap_events(ctx); |
071698e1 | 8795 | idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx); |
309fc03a JA |
8796 | |
8797 | /* | |
8798 | * Do this upfront, so we won't have a grace period where the ring | |
8799 | * is closed but resources aren't reaped yet. This can cause | |
8800 | * spurious failure in setting up a new ring. | |
8801 | */ | |
760618f7 JA |
8802 | io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries), |
8803 | ACCT_LOCKED); | |
309fc03a | 8804 | |
85faa7b8 | 8805 | INIT_WORK(&ctx->exit_work, io_ring_exit_work); |
fc666777 JA |
8806 | /* |
8807 | * Use system_unbound_wq to avoid spawning tons of event kworkers | |
8808 | * if we're exiting a ton of rings at the same time. It just adds | |
8809 | * noise and overhead, there's no discernable change in runtime | |
8810 | * over using system_wq. | |
8811 | */ | |
8812 | queue_work(system_unbound_wq, &ctx->exit_work); | |
2b188cc1 JA |
8813 | } |
8814 | ||
8815 | static int io_uring_release(struct inode *inode, struct file *file) | |
8816 | { | |
8817 | struct io_ring_ctx *ctx = file->private_data; | |
8818 | ||
8819 | file->private_data = NULL; | |
8820 | io_ring_ctx_wait_and_kill(ctx); | |
8821 | return 0; | |
8822 | } | |
8823 | ||
f6edbabb PB |
8824 | struct io_task_cancel { |
8825 | struct task_struct *task; | |
8826 | struct files_struct *files; | |
8827 | }; | |
f254ac04 | 8828 | |
f6edbabb | 8829 | static bool io_cancel_task_cb(struct io_wq_work *work, void *data) |
b711d4ea | 8830 | { |
9a472ef7 | 8831 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
f6edbabb | 8832 | struct io_task_cancel *cancel = data; |
9a472ef7 PB |
8833 | bool ret; |
8834 | ||
f6edbabb | 8835 | if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) { |
9a472ef7 PB |
8836 | unsigned long flags; |
8837 | struct io_ring_ctx *ctx = req->ctx; | |
8838 | ||
8839 | /* protect against races with linked timeouts */ | |
8840 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
f6edbabb | 8841 | ret = io_match_task(req, cancel->task, cancel->files); |
9a472ef7 PB |
8842 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
8843 | } else { | |
f6edbabb | 8844 | ret = io_match_task(req, cancel->task, cancel->files); |
9a472ef7 PB |
8845 | } |
8846 | return ret; | |
b711d4ea JA |
8847 | } |
8848 | ||
b7ddce3c | 8849 | static void io_cancel_defer_files(struct io_ring_ctx *ctx, |
ef9865a4 | 8850 | struct task_struct *task, |
b7ddce3c PB |
8851 | struct files_struct *files) |
8852 | { | |
8853 | struct io_defer_entry *de = NULL; | |
8854 | LIST_HEAD(list); | |
8855 | ||
8856 | spin_lock_irq(&ctx->completion_lock); | |
8857 | list_for_each_entry_reverse(de, &ctx->defer_list, list) { | |
08d23634 | 8858 | if (io_match_task(de->req, task, files)) { |
b7ddce3c PB |
8859 | list_cut_position(&list, &ctx->defer_list, &de->list); |
8860 | break; | |
8861 | } | |
8862 | } | |
8863 | spin_unlock_irq(&ctx->completion_lock); | |
8864 | ||
8865 | while (!list_empty(&list)) { | |
8866 | de = list_first_entry(&list, struct io_defer_entry, list); | |
8867 | list_del_init(&de->list); | |
8868 | req_set_fail_links(de->req); | |
8869 | io_put_req(de->req); | |
8870 | io_req_complete(de->req, -ECANCELED); | |
8871 | kfree(de); | |
8872 | } | |
8873 | } | |
8874 | ||
ca70f00b PB |
8875 | static int io_uring_count_inflight(struct io_ring_ctx *ctx, |
8876 | struct task_struct *task, | |
8877 | struct files_struct *files) | |
8878 | { | |
8879 | struct io_kiocb *req; | |
8880 | int cnt = 0; | |
8881 | ||
8882 | spin_lock_irq(&ctx->inflight_lock); | |
8883 | list_for_each_entry(req, &ctx->inflight_list, inflight_entry) | |
8884 | cnt += io_match_task(req, task, files); | |
8885 | spin_unlock_irq(&ctx->inflight_lock); | |
8886 | return cnt; | |
8887 | } | |
8888 | ||
b52fda00 | 8889 | static void io_uring_cancel_files(struct io_ring_ctx *ctx, |
df9923f9 | 8890 | struct task_struct *task, |
fcb323cc JA |
8891 | struct files_struct *files) |
8892 | { | |
fcb323cc | 8893 | while (!list_empty_careful(&ctx->inflight_list)) { |
bee749b1 | 8894 | struct io_task_cancel cancel = { .task = task, .files = files }; |
d8f1b971 | 8895 | DEFINE_WAIT(wait); |
ca70f00b | 8896 | int inflight; |
fcb323cc | 8897 | |
ca70f00b PB |
8898 | inflight = io_uring_count_inflight(ctx, task, files); |
8899 | if (!inflight) | |
fcb323cc | 8900 | break; |
f6edbabb PB |
8901 | |
8902 | io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, &cancel, true); | |
8903 | io_poll_remove_all(ctx, task, files); | |
8904 | io_kill_timeouts(ctx, task, files); | |
9d5c8190 | 8905 | io_cqring_overflow_flush(ctx, true, task, files); |
6200b0ae JA |
8906 | /* cancellations _may_ trigger task work */ |
8907 | io_run_task_work(); | |
ca70f00b PB |
8908 | |
8909 | prepare_to_wait(&task->io_uring->wait, &wait, | |
8910 | TASK_UNINTERRUPTIBLE); | |
8911 | if (inflight == io_uring_count_inflight(ctx, task, files)) | |
8912 | schedule(); | |
c98de08c | 8913 | finish_wait(&task->io_uring->wait, &wait); |
fcb323cc | 8914 | } |
44e728b8 PB |
8915 | } |
8916 | ||
b52fda00 PB |
8917 | static void __io_uring_cancel_task_requests(struct io_ring_ctx *ctx, |
8918 | struct task_struct *task) | |
0f212204 | 8919 | { |
b52fda00 | 8920 | while (1) { |
f6edbabb | 8921 | struct io_task_cancel cancel = { .task = task, .files = NULL, }; |
0f212204 | 8922 | enum io_wq_cancel cret; |
b52fda00 | 8923 | bool ret = false; |
0f212204 | 8924 | |
90df0853 PB |
8925 | if (ctx->io_wq) { |
8926 | cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, | |
8927 | &cancel, true); | |
8928 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
8929 | } | |
0f212204 JA |
8930 | |
8931 | /* SQPOLL thread does its own polling */ | |
8932 | if (!(ctx->flags & IORING_SETUP_SQPOLL)) { | |
8933 | while (!list_empty_careful(&ctx->iopoll_list)) { | |
8934 | io_iopoll_try_reap_events(ctx); | |
8935 | ret = true; | |
8936 | } | |
8937 | } | |
8938 | ||
6b81928d PB |
8939 | ret |= io_poll_remove_all(ctx, task, NULL); |
8940 | ret |= io_kill_timeouts(ctx, task, NULL); | |
55583d72 | 8941 | ret |= io_run_task_work(); |
b52fda00 PB |
8942 | if (!ret) |
8943 | break; | |
b52fda00 | 8944 | cond_resched(); |
0f212204 | 8945 | } |
0f212204 JA |
8946 | } |
8947 | ||
d9d05217 PB |
8948 | static void io_disable_sqo_submit(struct io_ring_ctx *ctx) |
8949 | { | |
d9d05217 PB |
8950 | mutex_lock(&ctx->uring_lock); |
8951 | ctx->sqo_dead = 1; | |
8952 | mutex_unlock(&ctx->uring_lock); | |
8953 | ||
8954 | /* make sure callers enter the ring to get error */ | |
b4411616 PB |
8955 | if (ctx->rings) |
8956 | io_ring_set_wakeup_flag(ctx); | |
d9d05217 PB |
8957 | } |
8958 | ||
0f212204 JA |
8959 | /* |
8960 | * We need to iteratively cancel requests, in case a request has dependent | |
8961 | * hard links. These persist even for failure of cancelations, hence keep | |
8962 | * looping until none are found. | |
8963 | */ | |
8964 | static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx, | |
8965 | struct files_struct *files) | |
8966 | { | |
8967 | struct task_struct *task = current; | |
8968 | ||
fdaf083c | 8969 | if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) { |
d9d05217 | 8970 | io_disable_sqo_submit(ctx); |
534ca6d6 | 8971 | task = ctx->sq_data->thread; |
fdaf083c JA |
8972 | atomic_inc(&task->io_uring->in_idle); |
8973 | io_sq_thread_park(ctx->sq_data); | |
8974 | } | |
0f212204 | 8975 | |
df9923f9 | 8976 | io_cancel_defer_files(ctx, task, files); |
0f212204 JA |
8977 | io_cqring_overflow_flush(ctx, true, task, files); |
8978 | ||
b52fda00 PB |
8979 | if (!files) |
8980 | __io_uring_cancel_task_requests(ctx, task); | |
bee749b1 PB |
8981 | else |
8982 | io_uring_cancel_files(ctx, task, files); | |
fdaf083c JA |
8983 | |
8984 | if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) { | |
8985 | atomic_dec(&task->io_uring->in_idle); | |
8986 | /* | |
8987 | * If the files that are going away are the ones in the thread | |
8988 | * identity, clear them out. | |
8989 | */ | |
8990 | if (task->io_uring->identity->files == files) | |
8991 | task->io_uring->identity->files = NULL; | |
8992 | io_sq_thread_unpark(ctx->sq_data); | |
8993 | } | |
0f212204 JA |
8994 | } |
8995 | ||
8996 | /* | |
8997 | * Note that this task has used io_uring. We use it for cancelation purposes. | |
8998 | */ | |
fdaf083c | 8999 | static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file) |
0f212204 | 9000 | { |
236434c3 | 9001 | struct io_uring_task *tctx = current->io_uring; |
a528b04e | 9002 | int ret; |
236434c3 MWO |
9003 | |
9004 | if (unlikely(!tctx)) { | |
0f212204 JA |
9005 | ret = io_uring_alloc_task_context(current); |
9006 | if (unlikely(ret)) | |
9007 | return ret; | |
236434c3 | 9008 | tctx = current->io_uring; |
0f212204 | 9009 | } |
236434c3 MWO |
9010 | if (tctx->last != file) { |
9011 | void *old = xa_load(&tctx->xa, (unsigned long)file); | |
0f212204 | 9012 | |
236434c3 | 9013 | if (!old) { |
0f212204 | 9014 | get_file(file); |
a528b04e PB |
9015 | ret = xa_err(xa_store(&tctx->xa, (unsigned long)file, |
9016 | file, GFP_KERNEL)); | |
9017 | if (ret) { | |
9018 | fput(file); | |
9019 | return ret; | |
9020 | } | |
0f212204 | 9021 | } |
236434c3 | 9022 | tctx->last = file; |
0f212204 JA |
9023 | } |
9024 | ||
fdaf083c JA |
9025 | /* |
9026 | * This is race safe in that the task itself is doing this, hence it | |
9027 | * cannot be going through the exit/cancel paths at the same time. | |
9028 | * This cannot be modified while exit/cancel is running. | |
9029 | */ | |
9030 | if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL)) | |
9031 | tctx->sqpoll = true; | |
9032 | ||
0f212204 JA |
9033 | return 0; |
9034 | } | |
9035 | ||
9036 | /* | |
9037 | * Remove this io_uring_file -> task mapping. | |
9038 | */ | |
9039 | static void io_uring_del_task_file(struct file *file) | |
9040 | { | |
9041 | struct io_uring_task *tctx = current->io_uring; | |
0f212204 JA |
9042 | |
9043 | if (tctx->last == file) | |
9044 | tctx->last = NULL; | |
5e2ed8c4 | 9045 | file = xa_erase(&tctx->xa, (unsigned long)file); |
0f212204 JA |
9046 | if (file) |
9047 | fput(file); | |
9048 | } | |
9049 | ||
de7f1d9e PB |
9050 | static void io_uring_remove_task_files(struct io_uring_task *tctx) |
9051 | { | |
9052 | struct file *file; | |
9053 | unsigned long index; | |
9054 | ||
9055 | xa_for_each(&tctx->xa, index, file) | |
9056 | io_uring_del_task_file(file); | |
9057 | } | |
9058 | ||
0f212204 JA |
9059 | void __io_uring_files_cancel(struct files_struct *files) |
9060 | { | |
9061 | struct io_uring_task *tctx = current->io_uring; | |
ce765372 MWO |
9062 | struct file *file; |
9063 | unsigned long index; | |
0f212204 JA |
9064 | |
9065 | /* make sure overflow events are dropped */ | |
fdaf083c | 9066 | atomic_inc(&tctx->in_idle); |
de7f1d9e PB |
9067 | xa_for_each(&tctx->xa, index, file) |
9068 | io_uring_cancel_task_requests(file->private_data, files); | |
fdaf083c | 9069 | atomic_dec(&tctx->in_idle); |
de7f1d9e PB |
9070 | |
9071 | if (files) | |
9072 | io_uring_remove_task_files(tctx); | |
fdaf083c JA |
9073 | } |
9074 | ||
9075 | static s64 tctx_inflight(struct io_uring_task *tctx) | |
9076 | { | |
9077 | unsigned long index; | |
9078 | struct file *file; | |
9079 | s64 inflight; | |
9080 | ||
9081 | inflight = percpu_counter_sum(&tctx->inflight); | |
9082 | if (!tctx->sqpoll) | |
9083 | return inflight; | |
9084 | ||
9085 | /* | |
9086 | * If we have SQPOLL rings, then we need to iterate and find them, and | |
9087 | * add the pending count for those. | |
9088 | */ | |
9089 | xa_for_each(&tctx->xa, index, file) { | |
9090 | struct io_ring_ctx *ctx = file->private_data; | |
9091 | ||
9092 | if (ctx->flags & IORING_SETUP_SQPOLL) { | |
9093 | struct io_uring_task *__tctx = ctx->sqo_task->io_uring; | |
9094 | ||
9095 | inflight += percpu_counter_sum(&__tctx->inflight); | |
9096 | } | |
9097 | } | |
9098 | ||
9099 | return inflight; | |
0f212204 JA |
9100 | } |
9101 | ||
0f212204 JA |
9102 | /* |
9103 | * Find any io_uring fd that this task has registered or done IO on, and cancel | |
9104 | * requests. | |
9105 | */ | |
9106 | void __io_uring_task_cancel(void) | |
9107 | { | |
9108 | struct io_uring_task *tctx = current->io_uring; | |
9109 | DEFINE_WAIT(wait); | |
d8a6df10 | 9110 | s64 inflight; |
0f212204 JA |
9111 | |
9112 | /* make sure overflow events are dropped */ | |
fdaf083c | 9113 | atomic_inc(&tctx->in_idle); |
0f212204 | 9114 | |
0b5cd6c3 PB |
9115 | /* trigger io_disable_sqo_submit() */ |
9116 | if (tctx->sqpoll) | |
9117 | __io_uring_files_cancel(NULL); | |
9118 | ||
d8a6df10 | 9119 | do { |
0f212204 | 9120 | /* read completions before cancelations */ |
fdaf083c | 9121 | inflight = tctx_inflight(tctx); |
d8a6df10 JA |
9122 | if (!inflight) |
9123 | break; | |
0f212204 JA |
9124 | __io_uring_files_cancel(NULL); |
9125 | ||
9126 | prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE); | |
9127 | ||
9128 | /* | |
a1bb3cd5 PB |
9129 | * If we've seen completions, retry without waiting. This |
9130 | * avoids a race where a completion comes in before we did | |
9131 | * prepare_to_wait(). | |
0f212204 | 9132 | */ |
a1bb3cd5 PB |
9133 | if (inflight == tctx_inflight(tctx)) |
9134 | schedule(); | |
f57555ed | 9135 | finish_wait(&tctx->wait, &wait); |
d8a6df10 | 9136 | } while (1); |
0f212204 | 9137 | |
fdaf083c | 9138 | atomic_dec(&tctx->in_idle); |
de7f1d9e PB |
9139 | |
9140 | io_uring_remove_task_files(tctx); | |
44e728b8 PB |
9141 | } |
9142 | ||
fcb323cc JA |
9143 | static int io_uring_flush(struct file *file, void *data) |
9144 | { | |
6b5733eb | 9145 | struct io_uring_task *tctx = current->io_uring; |
d9d05217 | 9146 | struct io_ring_ctx *ctx = file->private_data; |
6b5733eb | 9147 | |
84965ff8 JA |
9148 | if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) |
9149 | io_uring_cancel_task_requests(ctx, NULL); | |
9150 | ||
6b5733eb | 9151 | if (!tctx) |
4f793dc4 PB |
9152 | return 0; |
9153 | ||
6b5733eb PB |
9154 | /* we should have cancelled and erased it before PF_EXITING */ |
9155 | WARN_ON_ONCE((current->flags & PF_EXITING) && | |
9156 | xa_load(&tctx->xa, (unsigned long)file)); | |
9157 | ||
4f793dc4 PB |
9158 | /* |
9159 | * fput() is pending, will be 2 if the only other ref is our potential | |
9160 | * task file note. If the task is exiting, drop regardless of count. | |
9161 | */ | |
6b5733eb PB |
9162 | if (atomic_long_read(&file->f_count) != 2) |
9163 | return 0; | |
4f793dc4 | 9164 | |
d9d05217 PB |
9165 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
9166 | /* there is only one file note, which is owned by sqo_task */ | |
4325cb49 PB |
9167 | WARN_ON_ONCE(ctx->sqo_task != current && |
9168 | xa_load(&tctx->xa, (unsigned long)file)); | |
9169 | /* sqo_dead check is for when this happens after cancellation */ | |
9170 | WARN_ON_ONCE(ctx->sqo_task == current && !ctx->sqo_dead && | |
d9d05217 PB |
9171 | !xa_load(&tctx->xa, (unsigned long)file)); |
9172 | ||
9173 | io_disable_sqo_submit(ctx); | |
9174 | } | |
9175 | ||
9176 | if (!(ctx->flags & IORING_SETUP_SQPOLL) || ctx->sqo_task == current) | |
9177 | io_uring_del_task_file(file); | |
fcb323cc JA |
9178 | return 0; |
9179 | } | |
9180 | ||
6c5c240e RP |
9181 | static void *io_uring_validate_mmap_request(struct file *file, |
9182 | loff_t pgoff, size_t sz) | |
2b188cc1 | 9183 | { |
2b188cc1 | 9184 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 9185 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
9186 | struct page *page; |
9187 | void *ptr; | |
9188 | ||
9189 | switch (offset) { | |
9190 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
9191 | case IORING_OFF_CQ_RING: |
9192 | ptr = ctx->rings; | |
2b188cc1 JA |
9193 | break; |
9194 | case IORING_OFF_SQES: | |
9195 | ptr = ctx->sq_sqes; | |
9196 | break; | |
2b188cc1 | 9197 | default: |
6c5c240e | 9198 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
9199 | } |
9200 | ||
9201 | page = virt_to_head_page(ptr); | |
a50b854e | 9202 | if (sz > page_size(page)) |
6c5c240e RP |
9203 | return ERR_PTR(-EINVAL); |
9204 | ||
9205 | return ptr; | |
9206 | } | |
9207 | ||
9208 | #ifdef CONFIG_MMU | |
9209 | ||
9210 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
9211 | { | |
9212 | size_t sz = vma->vm_end - vma->vm_start; | |
9213 | unsigned long pfn; | |
9214 | void *ptr; | |
9215 | ||
9216 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
9217 | if (IS_ERR(ptr)) | |
9218 | return PTR_ERR(ptr); | |
2b188cc1 JA |
9219 | |
9220 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
9221 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
9222 | } | |
9223 | ||
6c5c240e RP |
9224 | #else /* !CONFIG_MMU */ |
9225 | ||
9226 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
9227 | { | |
9228 | return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL; | |
9229 | } | |
9230 | ||
9231 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
9232 | { | |
9233 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
9234 | } | |
9235 | ||
9236 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
9237 | unsigned long addr, unsigned long len, | |
9238 | unsigned long pgoff, unsigned long flags) | |
9239 | { | |
9240 | void *ptr; | |
9241 | ||
9242 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
9243 | if (IS_ERR(ptr)) | |
9244 | return PTR_ERR(ptr); | |
9245 | ||
9246 | return (unsigned long) ptr; | |
9247 | } | |
9248 | ||
9249 | #endif /* !CONFIG_MMU */ | |
9250 | ||
d9d05217 | 9251 | static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx) |
90554200 | 9252 | { |
d9d05217 | 9253 | int ret = 0; |
90554200 JA |
9254 | DEFINE_WAIT(wait); |
9255 | ||
9256 | do { | |
9257 | if (!io_sqring_full(ctx)) | |
9258 | break; | |
9259 | ||
9260 | prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE); | |
9261 | ||
d9d05217 PB |
9262 | if (unlikely(ctx->sqo_dead)) { |
9263 | ret = -EOWNERDEAD; | |
9264 | goto out; | |
9265 | } | |
9266 | ||
90554200 JA |
9267 | if (!io_sqring_full(ctx)) |
9268 | break; | |
9269 | ||
9270 | schedule(); | |
9271 | } while (!signal_pending(current)); | |
9272 | ||
9273 | finish_wait(&ctx->sqo_sq_wait, &wait); | |
d9d05217 PB |
9274 | out: |
9275 | return ret; | |
90554200 JA |
9276 | } |
9277 | ||
c73ebb68 HX |
9278 | static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz, |
9279 | struct __kernel_timespec __user **ts, | |
9280 | const sigset_t __user **sig) | |
9281 | { | |
9282 | struct io_uring_getevents_arg arg; | |
9283 | ||
9284 | /* | |
9285 | * If EXT_ARG isn't set, then we have no timespec and the argp pointer | |
9286 | * is just a pointer to the sigset_t. | |
9287 | */ | |
9288 | if (!(flags & IORING_ENTER_EXT_ARG)) { | |
9289 | *sig = (const sigset_t __user *) argp; | |
9290 | *ts = NULL; | |
9291 | return 0; | |
9292 | } | |
9293 | ||
9294 | /* | |
9295 | * EXT_ARG is set - ensure we agree on the size of it and copy in our | |
9296 | * timespec and sigset_t pointers if good. | |
9297 | */ | |
9298 | if (*argsz != sizeof(arg)) | |
9299 | return -EINVAL; | |
9300 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
9301 | return -EFAULT; | |
9302 | *sig = u64_to_user_ptr(arg.sigmask); | |
9303 | *argsz = arg.sigmask_sz; | |
9304 | *ts = u64_to_user_ptr(arg.ts); | |
9305 | return 0; | |
9306 | } | |
9307 | ||
2b188cc1 | 9308 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
c73ebb68 HX |
9309 | u32, min_complete, u32, flags, const void __user *, argp, |
9310 | size_t, argsz) | |
2b188cc1 JA |
9311 | { |
9312 | struct io_ring_ctx *ctx; | |
9313 | long ret = -EBADF; | |
9314 | int submitted = 0; | |
9315 | struct fd f; | |
9316 | ||
4c6e277c | 9317 | io_run_task_work(); |
b41e9852 | 9318 | |
90554200 | 9319 | if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | |
c73ebb68 | 9320 | IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)) |
2b188cc1 JA |
9321 | return -EINVAL; |
9322 | ||
9323 | f = fdget(fd); | |
9324 | if (!f.file) | |
9325 | return -EBADF; | |
9326 | ||
9327 | ret = -EOPNOTSUPP; | |
9328 | if (f.file->f_op != &io_uring_fops) | |
9329 | goto out_fput; | |
9330 | ||
9331 | ret = -ENXIO; | |
9332 | ctx = f.file->private_data; | |
9333 | if (!percpu_ref_tryget(&ctx->refs)) | |
9334 | goto out_fput; | |
9335 | ||
7e84e1c7 SG |
9336 | ret = -EBADFD; |
9337 | if (ctx->flags & IORING_SETUP_R_DISABLED) | |
9338 | goto out; | |
9339 | ||
6c271ce2 JA |
9340 | /* |
9341 | * For SQ polling, the thread will do all submissions and completions. | |
9342 | * Just return the requested submit count, and wake the thread if | |
9343 | * we were asked to. | |
9344 | */ | |
b2a9eada | 9345 | ret = 0; |
6c271ce2 | 9346 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
6c503150 | 9347 | io_cqring_overflow_flush(ctx, false, NULL, NULL); |
89448c47 | 9348 | |
d9d05217 PB |
9349 | ret = -EOWNERDEAD; |
9350 | if (unlikely(ctx->sqo_dead)) | |
9351 | goto out; | |
6c271ce2 | 9352 | if (flags & IORING_ENTER_SQ_WAKEUP) |
534ca6d6 | 9353 | wake_up(&ctx->sq_data->wait); |
d9d05217 PB |
9354 | if (flags & IORING_ENTER_SQ_WAIT) { |
9355 | ret = io_sqpoll_wait_sq(ctx); | |
9356 | if (ret) | |
9357 | goto out; | |
9358 | } | |
6c271ce2 | 9359 | submitted = to_submit; |
b2a9eada | 9360 | } else if (to_submit) { |
fdaf083c | 9361 | ret = io_uring_add_task_file(ctx, f.file); |
0f212204 JA |
9362 | if (unlikely(ret)) |
9363 | goto out; | |
2b188cc1 | 9364 | mutex_lock(&ctx->uring_lock); |
0f212204 | 9365 | submitted = io_submit_sqes(ctx, to_submit); |
2b188cc1 | 9366 | mutex_unlock(&ctx->uring_lock); |
7c504e65 PB |
9367 | |
9368 | if (submitted != to_submit) | |
9369 | goto out; | |
2b188cc1 JA |
9370 | } |
9371 | if (flags & IORING_ENTER_GETEVENTS) { | |
c73ebb68 HX |
9372 | const sigset_t __user *sig; |
9373 | struct __kernel_timespec __user *ts; | |
9374 | ||
9375 | ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig); | |
9376 | if (unlikely(ret)) | |
9377 | goto out; | |
9378 | ||
2b188cc1 JA |
9379 | min_complete = min(min_complete, ctx->cq_entries); |
9380 | ||
32b2244a XW |
9381 | /* |
9382 | * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user | |
9383 | * space applications don't need to do io completion events | |
9384 | * polling again, they can rely on io_sq_thread to do polling | |
9385 | * work, which can reduce cpu usage and uring_lock contention. | |
9386 | */ | |
9387 | if (ctx->flags & IORING_SETUP_IOPOLL && | |
9388 | !(ctx->flags & IORING_SETUP_SQPOLL)) { | |
7668b92a | 9389 | ret = io_iopoll_check(ctx, min_complete); |
def596e9 | 9390 | } else { |
c73ebb68 | 9391 | ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts); |
def596e9 | 9392 | } |
2b188cc1 JA |
9393 | } |
9394 | ||
7c504e65 | 9395 | out: |
6805b32e | 9396 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
9397 | out_fput: |
9398 | fdput(f); | |
9399 | return submitted ? submitted : ret; | |
9400 | } | |
9401 | ||
bebdb65e | 9402 | #ifdef CONFIG_PROC_FS |
87ce955b JA |
9403 | static int io_uring_show_cred(int id, void *p, void *data) |
9404 | { | |
6b47ab81 JA |
9405 | struct io_identity *iod = p; |
9406 | const struct cred *cred = iod->creds; | |
87ce955b JA |
9407 | struct seq_file *m = data; |
9408 | struct user_namespace *uns = seq_user_ns(m); | |
9409 | struct group_info *gi; | |
9410 | kernel_cap_t cap; | |
9411 | unsigned __capi; | |
9412 | int g; | |
9413 | ||
9414 | seq_printf(m, "%5d\n", id); | |
9415 | seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid)); | |
9416 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid)); | |
9417 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid)); | |
9418 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid)); | |
9419 | seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid)); | |
9420 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid)); | |
9421 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid)); | |
9422 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid)); | |
9423 | seq_puts(m, "\n\tGroups:\t"); | |
9424 | gi = cred->group_info; | |
9425 | for (g = 0; g < gi->ngroups; g++) { | |
9426 | seq_put_decimal_ull(m, g ? " " : "", | |
9427 | from_kgid_munged(uns, gi->gid[g])); | |
9428 | } | |
9429 | seq_puts(m, "\n\tCapEff:\t"); | |
9430 | cap = cred->cap_effective; | |
9431 | CAP_FOR_EACH_U32(__capi) | |
9432 | seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8); | |
9433 | seq_putc(m, '\n'); | |
9434 | return 0; | |
9435 | } | |
9436 | ||
9437 | static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m) | |
9438 | { | |
dbbe9c64 | 9439 | struct io_sq_data *sq = NULL; |
fad8e0de | 9440 | bool has_lock; |
87ce955b JA |
9441 | int i; |
9442 | ||
fad8e0de JA |
9443 | /* |
9444 | * Avoid ABBA deadlock between the seq lock and the io_uring mutex, | |
9445 | * since fdinfo case grabs it in the opposite direction of normal use | |
9446 | * cases. If we fail to get the lock, we just don't iterate any | |
9447 | * structures that could be going away outside the io_uring mutex. | |
9448 | */ | |
9449 | has_lock = mutex_trylock(&ctx->uring_lock); | |
9450 | ||
dbbe9c64 JQ |
9451 | if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) |
9452 | sq = ctx->sq_data; | |
9453 | ||
9454 | seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1); | |
9455 | seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1); | |
87ce955b | 9456 | seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files); |
fad8e0de | 9457 | for (i = 0; has_lock && i < ctx->nr_user_files; i++) { |
87ce955b JA |
9458 | struct fixed_file_table *table; |
9459 | struct file *f; | |
9460 | ||
9461 | table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT]; | |
9462 | f = table->files[i & IORING_FILE_TABLE_MASK]; | |
9463 | if (f) | |
9464 | seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname); | |
9465 | else | |
9466 | seq_printf(m, "%5u: <none>\n", i); | |
9467 | } | |
9468 | seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs); | |
fad8e0de | 9469 | for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) { |
87ce955b JA |
9470 | struct io_mapped_ubuf *buf = &ctx->user_bufs[i]; |
9471 | ||
9472 | seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, | |
9473 | (unsigned int) buf->len); | |
9474 | } | |
fad8e0de | 9475 | if (has_lock && !idr_is_empty(&ctx->personality_idr)) { |
87ce955b JA |
9476 | seq_printf(m, "Personalities:\n"); |
9477 | idr_for_each(&ctx->personality_idr, io_uring_show_cred, m); | |
9478 | } | |
d7718a9d JA |
9479 | seq_printf(m, "PollList:\n"); |
9480 | spin_lock_irq(&ctx->completion_lock); | |
9481 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { | |
9482 | struct hlist_head *list = &ctx->cancel_hash[i]; | |
9483 | struct io_kiocb *req; | |
9484 | ||
9485 | hlist_for_each_entry(req, list, hash_node) | |
9486 | seq_printf(m, " op=%d, task_works=%d\n", req->opcode, | |
9487 | req->task->task_works != NULL); | |
9488 | } | |
9489 | spin_unlock_irq(&ctx->completion_lock); | |
fad8e0de JA |
9490 | if (has_lock) |
9491 | mutex_unlock(&ctx->uring_lock); | |
87ce955b JA |
9492 | } |
9493 | ||
9494 | static void io_uring_show_fdinfo(struct seq_file *m, struct file *f) | |
9495 | { | |
9496 | struct io_ring_ctx *ctx = f->private_data; | |
9497 | ||
9498 | if (percpu_ref_tryget(&ctx->refs)) { | |
9499 | __io_uring_show_fdinfo(ctx, m); | |
9500 | percpu_ref_put(&ctx->refs); | |
9501 | } | |
9502 | } | |
bebdb65e | 9503 | #endif |
87ce955b | 9504 | |
2b188cc1 JA |
9505 | static const struct file_operations io_uring_fops = { |
9506 | .release = io_uring_release, | |
fcb323cc | 9507 | .flush = io_uring_flush, |
2b188cc1 | 9508 | .mmap = io_uring_mmap, |
6c5c240e RP |
9509 | #ifndef CONFIG_MMU |
9510 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
9511 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
9512 | #endif | |
2b188cc1 JA |
9513 | .poll = io_uring_poll, |
9514 | .fasync = io_uring_fasync, | |
bebdb65e | 9515 | #ifdef CONFIG_PROC_FS |
87ce955b | 9516 | .show_fdinfo = io_uring_show_fdinfo, |
bebdb65e | 9517 | #endif |
2b188cc1 JA |
9518 | }; |
9519 | ||
9520 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
9521 | struct io_uring_params *p) | |
9522 | { | |
75b28aff HV |
9523 | struct io_rings *rings; |
9524 | size_t size, sq_array_offset; | |
2b188cc1 | 9525 | |
bd740481 JA |
9526 | /* make sure these are sane, as we already accounted them */ |
9527 | ctx->sq_entries = p->sq_entries; | |
9528 | ctx->cq_entries = p->cq_entries; | |
9529 | ||
75b28aff HV |
9530 | size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset); |
9531 | if (size == SIZE_MAX) | |
9532 | return -EOVERFLOW; | |
9533 | ||
9534 | rings = io_mem_alloc(size); | |
9535 | if (!rings) | |
2b188cc1 JA |
9536 | return -ENOMEM; |
9537 | ||
75b28aff HV |
9538 | ctx->rings = rings; |
9539 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
9540 | rings->sq_ring_mask = p->sq_entries - 1; | |
9541 | rings->cq_ring_mask = p->cq_entries - 1; | |
9542 | rings->sq_ring_entries = p->sq_entries; | |
9543 | rings->cq_ring_entries = p->cq_entries; | |
9544 | ctx->sq_mask = rings->sq_ring_mask; | |
9545 | ctx->cq_mask = rings->cq_ring_mask; | |
2b188cc1 JA |
9546 | |
9547 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 JA |
9548 | if (size == SIZE_MAX) { |
9549 | io_mem_free(ctx->rings); | |
9550 | ctx->rings = NULL; | |
2b188cc1 | 9551 | return -EOVERFLOW; |
eb065d30 | 9552 | } |
2b188cc1 JA |
9553 | |
9554 | ctx->sq_sqes = io_mem_alloc(size); | |
eb065d30 JA |
9555 | if (!ctx->sq_sqes) { |
9556 | io_mem_free(ctx->rings); | |
9557 | ctx->rings = NULL; | |
2b188cc1 | 9558 | return -ENOMEM; |
eb065d30 | 9559 | } |
2b188cc1 | 9560 | |
2b188cc1 JA |
9561 | return 0; |
9562 | } | |
9563 | ||
9faadcc8 PB |
9564 | static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file) |
9565 | { | |
9566 | int ret, fd; | |
9567 | ||
9568 | fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
9569 | if (fd < 0) | |
9570 | return fd; | |
9571 | ||
9572 | ret = io_uring_add_task_file(ctx, file); | |
9573 | if (ret) { | |
9574 | put_unused_fd(fd); | |
9575 | return ret; | |
9576 | } | |
9577 | fd_install(fd, file); | |
9578 | return fd; | |
9579 | } | |
9580 | ||
2b188cc1 JA |
9581 | /* |
9582 | * Allocate an anonymous fd, this is what constitutes the application | |
9583 | * visible backing of an io_uring instance. The application mmaps this | |
9584 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
9585 | * we have to tie this fd to a socket for file garbage collection purposes. | |
9586 | */ | |
9faadcc8 | 9587 | static struct file *io_uring_get_file(struct io_ring_ctx *ctx) |
2b188cc1 JA |
9588 | { |
9589 | struct file *file; | |
9faadcc8 | 9590 | #if defined(CONFIG_UNIX) |
2b188cc1 JA |
9591 | int ret; |
9592 | ||
2b188cc1 JA |
9593 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, |
9594 | &ctx->ring_sock); | |
9595 | if (ret) | |
9faadcc8 | 9596 | return ERR_PTR(ret); |
2b188cc1 JA |
9597 | #endif |
9598 | ||
2b188cc1 JA |
9599 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, |
9600 | O_RDWR | O_CLOEXEC); | |
2b188cc1 | 9601 | #if defined(CONFIG_UNIX) |
9faadcc8 PB |
9602 | if (IS_ERR(file)) { |
9603 | sock_release(ctx->ring_sock); | |
9604 | ctx->ring_sock = NULL; | |
9605 | } else { | |
9606 | ctx->ring_sock->file = file; | |
0f212204 | 9607 | } |
2b188cc1 | 9608 | #endif |
9faadcc8 | 9609 | return file; |
2b188cc1 JA |
9610 | } |
9611 | ||
7f13657d XW |
9612 | static int io_uring_create(unsigned entries, struct io_uring_params *p, |
9613 | struct io_uring_params __user *params) | |
2b188cc1 JA |
9614 | { |
9615 | struct user_struct *user = NULL; | |
9616 | struct io_ring_ctx *ctx; | |
9faadcc8 | 9617 | struct file *file; |
aad5d8da | 9618 | bool limit_mem; |
2b188cc1 JA |
9619 | int ret; |
9620 | ||
8110c1a6 | 9621 | if (!entries) |
2b188cc1 | 9622 | return -EINVAL; |
8110c1a6 JA |
9623 | if (entries > IORING_MAX_ENTRIES) { |
9624 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
9625 | return -EINVAL; | |
9626 | entries = IORING_MAX_ENTRIES; | |
9627 | } | |
2b188cc1 JA |
9628 | |
9629 | /* | |
9630 | * Use twice as many entries for the CQ ring. It's possible for the | |
9631 | * application to drive a higher depth than the size of the SQ ring, | |
9632 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
9633 | * some flexibility in overcommitting a bit. If the application has |
9634 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
9635 | * of CQ ring entries manually. | |
2b188cc1 JA |
9636 | */ |
9637 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
9638 | if (p->flags & IORING_SETUP_CQSIZE) { |
9639 | /* | |
9640 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
9641 | * to a power-of-two, if it isn't already. We do NOT impose | |
9642 | * any cq vs sq ring sizing. | |
9643 | */ | |
eb2667b3 | 9644 | if (!p->cq_entries) |
33a107f0 | 9645 | return -EINVAL; |
8110c1a6 JA |
9646 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
9647 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
9648 | return -EINVAL; | |
9649 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
9650 | } | |
eb2667b3 JQ |
9651 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
9652 | if (p->cq_entries < p->sq_entries) | |
9653 | return -EINVAL; | |
33a107f0 JA |
9654 | } else { |
9655 | p->cq_entries = 2 * p->sq_entries; | |
9656 | } | |
2b188cc1 JA |
9657 | |
9658 | user = get_uid(current_user()); | |
aad5d8da | 9659 | limit_mem = !capable(CAP_IPC_LOCK); |
2b188cc1 | 9660 | |
aad5d8da | 9661 | if (limit_mem) { |
a087e2b5 | 9662 | ret = __io_account_mem(user, |
2b188cc1 JA |
9663 | ring_pages(p->sq_entries, p->cq_entries)); |
9664 | if (ret) { | |
9665 | free_uid(user); | |
9666 | return ret; | |
9667 | } | |
9668 | } | |
9669 | ||
9670 | ctx = io_ring_ctx_alloc(p); | |
9671 | if (!ctx) { | |
aad5d8da | 9672 | if (limit_mem) |
a087e2b5 | 9673 | __io_unaccount_mem(user, ring_pages(p->sq_entries, |
2b188cc1 JA |
9674 | p->cq_entries)); |
9675 | free_uid(user); | |
9676 | return -ENOMEM; | |
9677 | } | |
9678 | ctx->compat = in_compat_syscall(); | |
2b188cc1 | 9679 | ctx->user = user; |
0b8c0ec7 | 9680 | ctx->creds = get_current_cred(); |
4ea33a97 JA |
9681 | #ifdef CONFIG_AUDIT |
9682 | ctx->loginuid = current->loginuid; | |
9683 | ctx->sessionid = current->sessionid; | |
9684 | #endif | |
2aede0e4 JA |
9685 | ctx->sqo_task = get_task_struct(current); |
9686 | ||
9687 | /* | |
9688 | * This is just grabbed for accounting purposes. When a process exits, | |
9689 | * the mm is exited and dropped before the files, hence we need to hang | |
9690 | * on to this mm purely for the purposes of being able to unaccount | |
9691 | * memory (locked/pinned vm). It's not used for anything else. | |
9692 | */ | |
6b7898eb | 9693 | mmgrab(current->mm); |
2aede0e4 | 9694 | ctx->mm_account = current->mm; |
6b7898eb | 9695 | |
91d8f519 DZ |
9696 | #ifdef CONFIG_BLK_CGROUP |
9697 | /* | |
9698 | * The sq thread will belong to the original cgroup it was inited in. | |
9699 | * If the cgroup goes offline (e.g. disabling the io controller), then | |
9700 | * issued bios will be associated with the closest cgroup later in the | |
9701 | * block layer. | |
9702 | */ | |
9703 | rcu_read_lock(); | |
9704 | ctx->sqo_blkcg_css = blkcg_css(); | |
9705 | ret = css_tryget_online(ctx->sqo_blkcg_css); | |
9706 | rcu_read_unlock(); | |
9707 | if (!ret) { | |
9708 | /* don't init against a dying cgroup, have the user try again */ | |
9709 | ctx->sqo_blkcg_css = NULL; | |
9710 | ret = -ENODEV; | |
9711 | goto err; | |
9712 | } | |
9713 | #endif | |
6b7898eb | 9714 | |
f74441e6 JA |
9715 | /* |
9716 | * Account memory _before_ installing the file descriptor. Once | |
9717 | * the descriptor is installed, it can get closed at any time. Also | |
9718 | * do this before hitting the general error path, as ring freeing | |
9719 | * will un-account as well. | |
9720 | */ | |
9721 | io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries), | |
9722 | ACCT_LOCKED); | |
9723 | ctx->limit_mem = limit_mem; | |
9724 | ||
2b188cc1 JA |
9725 | ret = io_allocate_scq_urings(ctx, p); |
9726 | if (ret) | |
9727 | goto err; | |
9728 | ||
7e84e1c7 | 9729 | ret = io_sq_offload_create(ctx, p); |
2b188cc1 JA |
9730 | if (ret) |
9731 | goto err; | |
9732 | ||
7e84e1c7 SG |
9733 | if (!(p->flags & IORING_SETUP_R_DISABLED)) |
9734 | io_sq_offload_start(ctx); | |
9735 | ||
2b188cc1 | 9736 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
9737 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
9738 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
9739 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
9740 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
9741 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
9742 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
9743 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
9744 | |
9745 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
9746 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
9747 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
9748 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
9749 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
9750 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
9751 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
0d9b5b3a | 9752 | p->cq_off.flags = offsetof(struct io_rings, cq_flags); |
ac90f249 | 9753 | |
7f13657d XW |
9754 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
9755 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | | |
5769a351 | 9756 | IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | |
c73ebb68 HX |
9757 | IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED | |
9758 | IORING_FEAT_EXT_ARG; | |
7f13657d XW |
9759 | |
9760 | if (copy_to_user(params, p, sizeof(*p))) { | |
9761 | ret = -EFAULT; | |
9762 | goto err; | |
9763 | } | |
d1719f70 | 9764 | |
9faadcc8 PB |
9765 | file = io_uring_get_file(ctx); |
9766 | if (IS_ERR(file)) { | |
9767 | ret = PTR_ERR(file); | |
9768 | goto err; | |
9769 | } | |
9770 | ||
044c1ab3 JA |
9771 | /* |
9772 | * Install ring fd as the very last thing, so we don't risk someone | |
9773 | * having closed it before we finish setup | |
9774 | */ | |
9faadcc8 PB |
9775 | ret = io_uring_install_fd(ctx, file); |
9776 | if (ret < 0) { | |
06585c49 | 9777 | io_disable_sqo_submit(ctx); |
9faadcc8 PB |
9778 | /* fput will clean it up */ |
9779 | fput(file); | |
9780 | return ret; | |
9781 | } | |
044c1ab3 | 9782 | |
c826bd7a | 9783 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
9784 | return ret; |
9785 | err: | |
d9d05217 | 9786 | io_disable_sqo_submit(ctx); |
2b188cc1 JA |
9787 | io_ring_ctx_wait_and_kill(ctx); |
9788 | return ret; | |
9789 | } | |
9790 | ||
9791 | /* | |
9792 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
9793 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
9794 | * params structure passed in. | |
9795 | */ | |
9796 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
9797 | { | |
9798 | struct io_uring_params p; | |
2b188cc1 JA |
9799 | int i; |
9800 | ||
9801 | if (copy_from_user(&p, params, sizeof(p))) | |
9802 | return -EFAULT; | |
9803 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
9804 | if (p.resv[i]) | |
9805 | return -EINVAL; | |
9806 | } | |
9807 | ||
6c271ce2 | 9808 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 9809 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
7e84e1c7 SG |
9810 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | |
9811 | IORING_SETUP_R_DISABLED)) | |
2b188cc1 JA |
9812 | return -EINVAL; |
9813 | ||
7f13657d | 9814 | return io_uring_create(entries, &p, params); |
2b188cc1 JA |
9815 | } |
9816 | ||
9817 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
9818 | struct io_uring_params __user *, params) | |
9819 | { | |
9820 | return io_uring_setup(entries, params); | |
9821 | } | |
9822 | ||
66f4af93 JA |
9823 | static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args) |
9824 | { | |
9825 | struct io_uring_probe *p; | |
9826 | size_t size; | |
9827 | int i, ret; | |
9828 | ||
9829 | size = struct_size(p, ops, nr_args); | |
9830 | if (size == SIZE_MAX) | |
9831 | return -EOVERFLOW; | |
9832 | p = kzalloc(size, GFP_KERNEL); | |
9833 | if (!p) | |
9834 | return -ENOMEM; | |
9835 | ||
9836 | ret = -EFAULT; | |
9837 | if (copy_from_user(p, arg, size)) | |
9838 | goto out; | |
9839 | ret = -EINVAL; | |
9840 | if (memchr_inv(p, 0, size)) | |
9841 | goto out; | |
9842 | ||
9843 | p->last_op = IORING_OP_LAST - 1; | |
9844 | if (nr_args > IORING_OP_LAST) | |
9845 | nr_args = IORING_OP_LAST; | |
9846 | ||
9847 | for (i = 0; i < nr_args; i++) { | |
9848 | p->ops[i].op = i; | |
9849 | if (!io_op_defs[i].not_supported) | |
9850 | p->ops[i].flags = IO_URING_OP_SUPPORTED; | |
9851 | } | |
9852 | p->ops_len = i; | |
9853 | ||
9854 | ret = 0; | |
9855 | if (copy_to_user(arg, p, size)) | |
9856 | ret = -EFAULT; | |
9857 | out: | |
9858 | kfree(p); | |
9859 | return ret; | |
9860 | } | |
9861 | ||
071698e1 JA |
9862 | static int io_register_personality(struct io_ring_ctx *ctx) |
9863 | { | |
1e6fa521 JA |
9864 | struct io_identity *id; |
9865 | int ret; | |
071698e1 | 9866 | |
1e6fa521 JA |
9867 | id = kmalloc(sizeof(*id), GFP_KERNEL); |
9868 | if (unlikely(!id)) | |
9869 | return -ENOMEM; | |
9870 | ||
9871 | io_init_identity(id); | |
9872 | id->creds = get_current_cred(); | |
9873 | ||
9874 | ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL); | |
9875 | if (ret < 0) { | |
9876 | put_cred(id->creds); | |
9877 | kfree(id); | |
9878 | } | |
9879 | return ret; | |
071698e1 JA |
9880 | } |
9881 | ||
9882 | static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) | |
9883 | { | |
1e6fa521 | 9884 | struct io_identity *iod; |
071698e1 | 9885 | |
1e6fa521 JA |
9886 | iod = idr_remove(&ctx->personality_idr, id); |
9887 | if (iod) { | |
9888 | put_cred(iod->creds); | |
9889 | if (refcount_dec_and_test(&iod->count)) | |
9890 | kfree(iod); | |
071698e1 JA |
9891 | return 0; |
9892 | } | |
9893 | ||
9894 | return -EINVAL; | |
9895 | } | |
9896 | ||
21b55dbc SG |
9897 | static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg, |
9898 | unsigned int nr_args) | |
9899 | { | |
9900 | struct io_uring_restriction *res; | |
9901 | size_t size; | |
9902 | int i, ret; | |
9903 | ||
7e84e1c7 SG |
9904 | /* Restrictions allowed only if rings started disabled */ |
9905 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
9906 | return -EBADFD; | |
9907 | ||
21b55dbc | 9908 | /* We allow only a single restrictions registration */ |
7e84e1c7 | 9909 | if (ctx->restrictions.registered) |
21b55dbc SG |
9910 | return -EBUSY; |
9911 | ||
9912 | if (!arg || nr_args > IORING_MAX_RESTRICTIONS) | |
9913 | return -EINVAL; | |
9914 | ||
9915 | size = array_size(nr_args, sizeof(*res)); | |
9916 | if (size == SIZE_MAX) | |
9917 | return -EOVERFLOW; | |
9918 | ||
9919 | res = memdup_user(arg, size); | |
9920 | if (IS_ERR(res)) | |
9921 | return PTR_ERR(res); | |
9922 | ||
9923 | ret = 0; | |
9924 | ||
9925 | for (i = 0; i < nr_args; i++) { | |
9926 | switch (res[i].opcode) { | |
9927 | case IORING_RESTRICTION_REGISTER_OP: | |
9928 | if (res[i].register_op >= IORING_REGISTER_LAST) { | |
9929 | ret = -EINVAL; | |
9930 | goto out; | |
9931 | } | |
9932 | ||
9933 | __set_bit(res[i].register_op, | |
9934 | ctx->restrictions.register_op); | |
9935 | break; | |
9936 | case IORING_RESTRICTION_SQE_OP: | |
9937 | if (res[i].sqe_op >= IORING_OP_LAST) { | |
9938 | ret = -EINVAL; | |
9939 | goto out; | |
9940 | } | |
9941 | ||
9942 | __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op); | |
9943 | break; | |
9944 | case IORING_RESTRICTION_SQE_FLAGS_ALLOWED: | |
9945 | ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags; | |
9946 | break; | |
9947 | case IORING_RESTRICTION_SQE_FLAGS_REQUIRED: | |
9948 | ctx->restrictions.sqe_flags_required = res[i].sqe_flags; | |
9949 | break; | |
9950 | default: | |
9951 | ret = -EINVAL; | |
9952 | goto out; | |
9953 | } | |
9954 | } | |
9955 | ||
9956 | out: | |
9957 | /* Reset all restrictions if an error happened */ | |
9958 | if (ret != 0) | |
9959 | memset(&ctx->restrictions, 0, sizeof(ctx->restrictions)); | |
9960 | else | |
7e84e1c7 | 9961 | ctx->restrictions.registered = true; |
21b55dbc SG |
9962 | |
9963 | kfree(res); | |
9964 | return ret; | |
9965 | } | |
9966 | ||
7e84e1c7 SG |
9967 | static int io_register_enable_rings(struct io_ring_ctx *ctx) |
9968 | { | |
9969 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
9970 | return -EBADFD; | |
9971 | ||
9972 | if (ctx->restrictions.registered) | |
9973 | ctx->restricted = 1; | |
9974 | ||
9975 | ctx->flags &= ~IORING_SETUP_R_DISABLED; | |
9976 | ||
9977 | io_sq_offload_start(ctx); | |
9978 | ||
9979 | return 0; | |
9980 | } | |
9981 | ||
071698e1 JA |
9982 | static bool io_register_op_must_quiesce(int op) |
9983 | { | |
9984 | switch (op) { | |
9985 | case IORING_UNREGISTER_FILES: | |
9986 | case IORING_REGISTER_FILES_UPDATE: | |
9987 | case IORING_REGISTER_PROBE: | |
9988 | case IORING_REGISTER_PERSONALITY: | |
9989 | case IORING_UNREGISTER_PERSONALITY: | |
9990 | return false; | |
9991 | default: | |
9992 | return true; | |
9993 | } | |
9994 | } | |
9995 | ||
edafccee JA |
9996 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
9997 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
9998 | __releases(ctx->uring_lock) |
9999 | __acquires(ctx->uring_lock) | |
edafccee JA |
10000 | { |
10001 | int ret; | |
10002 | ||
35fa71a0 JA |
10003 | /* |
10004 | * We're inside the ring mutex, if the ref is already dying, then | |
10005 | * someone else killed the ctx or is already going through | |
10006 | * io_uring_register(). | |
10007 | */ | |
10008 | if (percpu_ref_is_dying(&ctx->refs)) | |
10009 | return -ENXIO; | |
10010 | ||
071698e1 | 10011 | if (io_register_op_must_quiesce(opcode)) { |
05f3fb3c | 10012 | percpu_ref_kill(&ctx->refs); |
b19062a5 | 10013 | |
05f3fb3c JA |
10014 | /* |
10015 | * Drop uring mutex before waiting for references to exit. If | |
10016 | * another thread is currently inside io_uring_enter() it might | |
10017 | * need to grab the uring_lock to make progress. If we hold it | |
10018 | * here across the drain wait, then we can deadlock. It's safe | |
10019 | * to drop the mutex here, since no new references will come in | |
10020 | * after we've killed the percpu ref. | |
10021 | */ | |
10022 | mutex_unlock(&ctx->uring_lock); | |
af9c1a44 JA |
10023 | do { |
10024 | ret = wait_for_completion_interruptible(&ctx->ref_comp); | |
10025 | if (!ret) | |
10026 | break; | |
ed6930c9 JA |
10027 | ret = io_run_task_work_sig(); |
10028 | if (ret < 0) | |
10029 | break; | |
af9c1a44 JA |
10030 | } while (1); |
10031 | ||
05f3fb3c | 10032 | mutex_lock(&ctx->uring_lock); |
af9c1a44 | 10033 | |
c150368b JA |
10034 | if (ret) { |
10035 | percpu_ref_resurrect(&ctx->refs); | |
21b55dbc SG |
10036 | goto out_quiesce; |
10037 | } | |
10038 | } | |
10039 | ||
10040 | if (ctx->restricted) { | |
10041 | if (opcode >= IORING_REGISTER_LAST) { | |
10042 | ret = -EINVAL; | |
10043 | goto out; | |
10044 | } | |
10045 | ||
10046 | if (!test_bit(opcode, ctx->restrictions.register_op)) { | |
10047 | ret = -EACCES; | |
c150368b JA |
10048 | goto out; |
10049 | } | |
05f3fb3c | 10050 | } |
edafccee JA |
10051 | |
10052 | switch (opcode) { | |
10053 | case IORING_REGISTER_BUFFERS: | |
10054 | ret = io_sqe_buffer_register(ctx, arg, nr_args); | |
10055 | break; | |
10056 | case IORING_UNREGISTER_BUFFERS: | |
10057 | ret = -EINVAL; | |
10058 | if (arg || nr_args) | |
10059 | break; | |
10060 | ret = io_sqe_buffer_unregister(ctx); | |
10061 | break; | |
6b06314c JA |
10062 | case IORING_REGISTER_FILES: |
10063 | ret = io_sqe_files_register(ctx, arg, nr_args); | |
10064 | break; | |
10065 | case IORING_UNREGISTER_FILES: | |
10066 | ret = -EINVAL; | |
10067 | if (arg || nr_args) | |
10068 | break; | |
10069 | ret = io_sqe_files_unregister(ctx); | |
10070 | break; | |
c3a31e60 JA |
10071 | case IORING_REGISTER_FILES_UPDATE: |
10072 | ret = io_sqe_files_update(ctx, arg, nr_args); | |
10073 | break; | |
9b402849 | 10074 | case IORING_REGISTER_EVENTFD: |
f2842ab5 | 10075 | case IORING_REGISTER_EVENTFD_ASYNC: |
9b402849 JA |
10076 | ret = -EINVAL; |
10077 | if (nr_args != 1) | |
10078 | break; | |
10079 | ret = io_eventfd_register(ctx, arg); | |
f2842ab5 JA |
10080 | if (ret) |
10081 | break; | |
10082 | if (opcode == IORING_REGISTER_EVENTFD_ASYNC) | |
10083 | ctx->eventfd_async = 1; | |
10084 | else | |
10085 | ctx->eventfd_async = 0; | |
9b402849 JA |
10086 | break; |
10087 | case IORING_UNREGISTER_EVENTFD: | |
10088 | ret = -EINVAL; | |
10089 | if (arg || nr_args) | |
10090 | break; | |
10091 | ret = io_eventfd_unregister(ctx); | |
10092 | break; | |
66f4af93 JA |
10093 | case IORING_REGISTER_PROBE: |
10094 | ret = -EINVAL; | |
10095 | if (!arg || nr_args > 256) | |
10096 | break; | |
10097 | ret = io_probe(ctx, arg, nr_args); | |
10098 | break; | |
071698e1 JA |
10099 | case IORING_REGISTER_PERSONALITY: |
10100 | ret = -EINVAL; | |
10101 | if (arg || nr_args) | |
10102 | break; | |
10103 | ret = io_register_personality(ctx); | |
10104 | break; | |
10105 | case IORING_UNREGISTER_PERSONALITY: | |
10106 | ret = -EINVAL; | |
10107 | if (arg) | |
10108 | break; | |
10109 | ret = io_unregister_personality(ctx, nr_args); | |
10110 | break; | |
7e84e1c7 SG |
10111 | case IORING_REGISTER_ENABLE_RINGS: |
10112 | ret = -EINVAL; | |
10113 | if (arg || nr_args) | |
10114 | break; | |
10115 | ret = io_register_enable_rings(ctx); | |
10116 | break; | |
21b55dbc SG |
10117 | case IORING_REGISTER_RESTRICTIONS: |
10118 | ret = io_register_restrictions(ctx, arg, nr_args); | |
10119 | break; | |
edafccee JA |
10120 | default: |
10121 | ret = -EINVAL; | |
10122 | break; | |
10123 | } | |
10124 | ||
21b55dbc | 10125 | out: |
071698e1 | 10126 | if (io_register_op_must_quiesce(opcode)) { |
05f3fb3c | 10127 | /* bring the ctx back to life */ |
05f3fb3c | 10128 | percpu_ref_reinit(&ctx->refs); |
21b55dbc | 10129 | out_quiesce: |
0f158b4c | 10130 | reinit_completion(&ctx->ref_comp); |
05f3fb3c | 10131 | } |
edafccee JA |
10132 | return ret; |
10133 | } | |
10134 | ||
10135 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
10136 | void __user *, arg, unsigned int, nr_args) | |
10137 | { | |
10138 | struct io_ring_ctx *ctx; | |
10139 | long ret = -EBADF; | |
10140 | struct fd f; | |
10141 | ||
10142 | f = fdget(fd); | |
10143 | if (!f.file) | |
10144 | return -EBADF; | |
10145 | ||
10146 | ret = -EOPNOTSUPP; | |
10147 | if (f.file->f_op != &io_uring_fops) | |
10148 | goto out_fput; | |
10149 | ||
10150 | ctx = f.file->private_data; | |
10151 | ||
10152 | mutex_lock(&ctx->uring_lock); | |
10153 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
10154 | mutex_unlock(&ctx->uring_lock); | |
c826bd7a DD |
10155 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, |
10156 | ctx->cq_ev_fd != NULL, ret); | |
edafccee JA |
10157 | out_fput: |
10158 | fdput(f); | |
10159 | return ret; | |
10160 | } | |
10161 | ||
2b188cc1 JA |
10162 | static int __init io_uring_init(void) |
10163 | { | |
d7f62e82 SM |
10164 | #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \ |
10165 | BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ | |
10166 | BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \ | |
10167 | } while (0) | |
10168 | ||
10169 | #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ | |
10170 | __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename) | |
10171 | BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); | |
10172 | BUILD_BUG_SQE_ELEM(0, __u8, opcode); | |
10173 | BUILD_BUG_SQE_ELEM(1, __u8, flags); | |
10174 | BUILD_BUG_SQE_ELEM(2, __u16, ioprio); | |
10175 | BUILD_BUG_SQE_ELEM(4, __s32, fd); | |
10176 | BUILD_BUG_SQE_ELEM(8, __u64, off); | |
10177 | BUILD_BUG_SQE_ELEM(8, __u64, addr2); | |
10178 | BUILD_BUG_SQE_ELEM(16, __u64, addr); | |
7d67af2c | 10179 | BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); |
d7f62e82 SM |
10180 | BUILD_BUG_SQE_ELEM(24, __u32, len); |
10181 | BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); | |
10182 | BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); | |
10183 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); | |
10184 | BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); | |
5769a351 JX |
10185 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); |
10186 | BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); | |
d7f62e82 SM |
10187 | BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); |
10188 | BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); | |
10189 | BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); | |
10190 | BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); | |
10191 | BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); | |
10192 | BUILD_BUG_SQE_ELEM(28, __u32, open_flags); | |
10193 | BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); | |
10194 | BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); | |
7d67af2c | 10195 | BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); |
d7f62e82 SM |
10196 | BUILD_BUG_SQE_ELEM(32, __u64, user_data); |
10197 | BUILD_BUG_SQE_ELEM(40, __u16, buf_index); | |
10198 | BUILD_BUG_SQE_ELEM(42, __u16, personality); | |
7d67af2c | 10199 | BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); |
d7f62e82 | 10200 | |
d3656344 | 10201 | BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST); |
84557871 | 10202 | BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int)); |
2b188cc1 JA |
10203 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC); |
10204 | return 0; | |
10205 | }; | |
10206 | __initcall(io_uring_init); |