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> | |
47 | #include <linux/refcount.h> | |
48 | #include <linux/uio.h> | |
6b47ee6e | 49 | #include <linux/bits.h> |
2b188cc1 JA |
50 | |
51 | #include <linux/sched/signal.h> | |
52 | #include <linux/fs.h> | |
53 | #include <linux/file.h> | |
54 | #include <linux/fdtable.h> | |
55 | #include <linux/mm.h> | |
56 | #include <linux/mman.h> | |
57 | #include <linux/mmu_context.h> | |
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> |
2b188cc1 | 78 | |
c826bd7a DD |
79 | #define CREATE_TRACE_POINTS |
80 | #include <trace/events/io_uring.h> | |
81 | ||
2b188cc1 JA |
82 | #include <uapi/linux/io_uring.h> |
83 | ||
84 | #include "internal.h" | |
561fb04a | 85 | #include "io-wq.h" |
2b188cc1 | 86 | |
5277deaa | 87 | #define IORING_MAX_ENTRIES 32768 |
33a107f0 | 88 | #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES) |
65e19f54 JA |
89 | |
90 | /* | |
91 | * Shift of 9 is 512 entries, or exactly one page on 64-bit archs | |
92 | */ | |
93 | #define IORING_FILE_TABLE_SHIFT 9 | |
94 | #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT) | |
95 | #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1) | |
96 | #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE) | |
2b188cc1 JA |
97 | |
98 | struct io_uring { | |
99 | u32 head ____cacheline_aligned_in_smp; | |
100 | u32 tail ____cacheline_aligned_in_smp; | |
101 | }; | |
102 | ||
1e84b97b | 103 | /* |
75b28aff HV |
104 | * This data is shared with the application through the mmap at offsets |
105 | * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING. | |
1e84b97b SB |
106 | * |
107 | * The offsets to the member fields are published through struct | |
108 | * io_sqring_offsets when calling io_uring_setup. | |
109 | */ | |
75b28aff | 110 | struct io_rings { |
1e84b97b SB |
111 | /* |
112 | * Head and tail offsets into the ring; the offsets need to be | |
113 | * masked to get valid indices. | |
114 | * | |
75b28aff HV |
115 | * The kernel controls head of the sq ring and the tail of the cq ring, |
116 | * and the application controls tail of the sq ring and the head of the | |
117 | * cq ring. | |
1e84b97b | 118 | */ |
75b28aff | 119 | struct io_uring sq, cq; |
1e84b97b | 120 | /* |
75b28aff | 121 | * Bitmasks to apply to head and tail offsets (constant, equals |
1e84b97b SB |
122 | * ring_entries - 1) |
123 | */ | |
75b28aff HV |
124 | u32 sq_ring_mask, cq_ring_mask; |
125 | /* Ring sizes (constant, power of 2) */ | |
126 | u32 sq_ring_entries, cq_ring_entries; | |
1e84b97b SB |
127 | /* |
128 | * Number of invalid entries dropped by the kernel due to | |
129 | * invalid index stored in array | |
130 | * | |
131 | * Written by the kernel, shouldn't be modified by the | |
132 | * application (i.e. get number of "new events" by comparing to | |
133 | * cached value). | |
134 | * | |
135 | * After a new SQ head value was read by the application this | |
136 | * counter includes all submissions that were dropped reaching | |
137 | * the new SQ head (and possibly more). | |
138 | */ | |
75b28aff | 139 | u32 sq_dropped; |
1e84b97b SB |
140 | /* |
141 | * Runtime flags | |
142 | * | |
143 | * Written by the kernel, shouldn't be modified by the | |
144 | * application. | |
145 | * | |
146 | * The application needs a full memory barrier before checking | |
147 | * for IORING_SQ_NEED_WAKEUP after updating the sq tail. | |
148 | */ | |
75b28aff | 149 | u32 sq_flags; |
1e84b97b SB |
150 | /* |
151 | * Number of completion events lost because the queue was full; | |
152 | * this should be avoided by the application by making sure | |
0b4295b5 | 153 | * there are not more requests pending than there is space in |
1e84b97b SB |
154 | * the completion queue. |
155 | * | |
156 | * Written by the kernel, shouldn't be modified by the | |
157 | * application (i.e. get number of "new events" by comparing to | |
158 | * cached value). | |
159 | * | |
160 | * As completion events come in out of order this counter is not | |
161 | * ordered with any other data. | |
162 | */ | |
75b28aff | 163 | u32 cq_overflow; |
1e84b97b SB |
164 | /* |
165 | * Ring buffer of completion events. | |
166 | * | |
167 | * The kernel writes completion events fresh every time they are | |
168 | * produced, so the application is allowed to modify pending | |
169 | * entries. | |
170 | */ | |
75b28aff | 171 | struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp; |
2b188cc1 JA |
172 | }; |
173 | ||
edafccee JA |
174 | struct io_mapped_ubuf { |
175 | u64 ubuf; | |
176 | size_t len; | |
177 | struct bio_vec *bvec; | |
178 | unsigned int nr_bvecs; | |
179 | }; | |
180 | ||
65e19f54 JA |
181 | struct fixed_file_table { |
182 | struct file **files; | |
31b51510 JA |
183 | }; |
184 | ||
05f3fb3c JA |
185 | enum { |
186 | FFD_F_ATOMIC, | |
187 | }; | |
188 | ||
189 | struct fixed_file_data { | |
190 | struct fixed_file_table *table; | |
191 | struct io_ring_ctx *ctx; | |
192 | ||
193 | struct percpu_ref refs; | |
194 | struct llist_head put_llist; | |
195 | unsigned long state; | |
196 | struct work_struct ref_work; | |
197 | struct completion done; | |
198 | }; | |
199 | ||
2b188cc1 JA |
200 | struct io_ring_ctx { |
201 | struct { | |
202 | struct percpu_ref refs; | |
203 | } ____cacheline_aligned_in_smp; | |
204 | ||
205 | struct { | |
206 | unsigned int flags; | |
69b3e546 JA |
207 | int compat: 1; |
208 | int account_mem: 1; | |
209 | int cq_overflow_flushed: 1; | |
210 | int drain_next: 1; | |
f2842ab5 | 211 | int eventfd_async: 1; |
2b188cc1 | 212 | |
75b28aff HV |
213 | /* |
214 | * Ring buffer of indices into array of io_uring_sqe, which is | |
215 | * mmapped by the application using the IORING_OFF_SQES offset. | |
216 | * | |
217 | * This indirection could e.g. be used to assign fixed | |
218 | * io_uring_sqe entries to operations and only submit them to | |
219 | * the queue when needed. | |
220 | * | |
221 | * The kernel modifies neither the indices array nor the entries | |
222 | * array. | |
223 | */ | |
224 | u32 *sq_array; | |
2b188cc1 JA |
225 | unsigned cached_sq_head; |
226 | unsigned sq_entries; | |
227 | unsigned sq_mask; | |
6c271ce2 | 228 | unsigned sq_thread_idle; |
498ccd9e | 229 | unsigned cached_sq_dropped; |
206aefde | 230 | atomic_t cached_cq_overflow; |
ad3eb2c8 | 231 | unsigned long sq_check_overflow; |
de0617e4 JA |
232 | |
233 | struct list_head defer_list; | |
5262f567 | 234 | struct list_head timeout_list; |
1d7bb1d5 | 235 | struct list_head cq_overflow_list; |
fcb323cc JA |
236 | |
237 | wait_queue_head_t inflight_wait; | |
ad3eb2c8 | 238 | struct io_uring_sqe *sq_sqes; |
2b188cc1 JA |
239 | } ____cacheline_aligned_in_smp; |
240 | ||
206aefde JA |
241 | struct io_rings *rings; |
242 | ||
2b188cc1 | 243 | /* IO offload */ |
561fb04a | 244 | struct io_wq *io_wq; |
6c271ce2 | 245 | struct task_struct *sqo_thread; /* if using sq thread polling */ |
2b188cc1 | 246 | struct mm_struct *sqo_mm; |
6c271ce2 | 247 | wait_queue_head_t sqo_wait; |
75b28aff | 248 | |
6b06314c JA |
249 | /* |
250 | * If used, fixed file set. Writers must ensure that ->refs is dead, | |
251 | * readers must ensure that ->refs is alive as long as the file* is | |
252 | * used. Only updated through io_uring_register(2). | |
253 | */ | |
05f3fb3c | 254 | struct fixed_file_data *file_data; |
6b06314c | 255 | unsigned nr_user_files; |
b14cca0c PB |
256 | int ring_fd; |
257 | struct file *ring_file; | |
6b06314c | 258 | |
edafccee JA |
259 | /* if used, fixed mapped user buffers */ |
260 | unsigned nr_user_bufs; | |
261 | struct io_mapped_ubuf *user_bufs; | |
262 | ||
2b188cc1 JA |
263 | struct user_struct *user; |
264 | ||
0b8c0ec7 | 265 | const struct cred *creds; |
181e448d | 266 | |
206aefde JA |
267 | /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */ |
268 | struct completion *completions; | |
269 | ||
0ddf92e8 JA |
270 | /* if all else fails... */ |
271 | struct io_kiocb *fallback_req; | |
272 | ||
206aefde JA |
273 | #if defined(CONFIG_UNIX) |
274 | struct socket *ring_sock; | |
275 | #endif | |
276 | ||
071698e1 JA |
277 | struct idr personality_idr; |
278 | ||
206aefde JA |
279 | struct { |
280 | unsigned cached_cq_tail; | |
281 | unsigned cq_entries; | |
282 | unsigned cq_mask; | |
283 | atomic_t cq_timeouts; | |
ad3eb2c8 | 284 | unsigned long cq_check_overflow; |
206aefde JA |
285 | struct wait_queue_head cq_wait; |
286 | struct fasync_struct *cq_fasync; | |
287 | struct eventfd_ctx *cq_ev_fd; | |
288 | } ____cacheline_aligned_in_smp; | |
2b188cc1 JA |
289 | |
290 | struct { | |
291 | struct mutex uring_lock; | |
292 | wait_queue_head_t wait; | |
293 | } ____cacheline_aligned_in_smp; | |
294 | ||
295 | struct { | |
296 | spinlock_t completion_lock; | |
e94f141b JA |
297 | struct llist_head poll_llist; |
298 | ||
def596e9 JA |
299 | /* |
300 | * ->poll_list is protected by the ctx->uring_lock for | |
301 | * io_uring instances that don't use IORING_SETUP_SQPOLL. | |
302 | * For SQPOLL, only the single threaded io_sq_thread() will | |
303 | * manipulate the list, hence no extra locking is needed there. | |
304 | */ | |
305 | struct list_head poll_list; | |
78076bb6 JA |
306 | struct hlist_head *cancel_hash; |
307 | unsigned cancel_hash_bits; | |
e94f141b | 308 | bool poll_multi_file; |
31b51510 | 309 | |
fcb323cc JA |
310 | spinlock_t inflight_lock; |
311 | struct list_head inflight_list; | |
2b188cc1 | 312 | } ____cacheline_aligned_in_smp; |
2b188cc1 JA |
313 | }; |
314 | ||
09bb8394 JA |
315 | /* |
316 | * First field must be the file pointer in all the | |
317 | * iocb unions! See also 'struct kiocb' in <linux/fs.h> | |
318 | */ | |
221c5eb2 JA |
319 | struct io_poll_iocb { |
320 | struct file *file; | |
0969e783 JA |
321 | union { |
322 | struct wait_queue_head *head; | |
323 | u64 addr; | |
324 | }; | |
221c5eb2 | 325 | __poll_t events; |
8c838788 | 326 | bool done; |
221c5eb2 | 327 | bool canceled; |
392edb45 | 328 | struct wait_queue_entry wait; |
221c5eb2 JA |
329 | }; |
330 | ||
b5dba59e JA |
331 | struct io_close { |
332 | struct file *file; | |
333 | struct file *put_file; | |
334 | int fd; | |
335 | }; | |
336 | ||
ad8a48ac JA |
337 | struct io_timeout_data { |
338 | struct io_kiocb *req; | |
339 | struct hrtimer timer; | |
340 | struct timespec64 ts; | |
341 | enum hrtimer_mode mode; | |
cc42e0ac | 342 | u32 seq_offset; |
ad8a48ac JA |
343 | }; |
344 | ||
8ed8d3c3 JA |
345 | struct io_accept { |
346 | struct file *file; | |
347 | struct sockaddr __user *addr; | |
348 | int __user *addr_len; | |
349 | int flags; | |
350 | }; | |
351 | ||
352 | struct io_sync { | |
353 | struct file *file; | |
354 | loff_t len; | |
355 | loff_t off; | |
356 | int flags; | |
d63d1b5e | 357 | int mode; |
8ed8d3c3 JA |
358 | }; |
359 | ||
fbf23849 JA |
360 | struct io_cancel { |
361 | struct file *file; | |
362 | u64 addr; | |
363 | }; | |
364 | ||
b29472ee JA |
365 | struct io_timeout { |
366 | struct file *file; | |
367 | u64 addr; | |
368 | int flags; | |
26a61679 | 369 | unsigned count; |
b29472ee JA |
370 | }; |
371 | ||
9adbd45d JA |
372 | struct io_rw { |
373 | /* NOTE: kiocb has the file as the first member, so don't do it here */ | |
374 | struct kiocb kiocb; | |
375 | u64 addr; | |
376 | u64 len; | |
377 | }; | |
378 | ||
3fbb51c1 JA |
379 | struct io_connect { |
380 | struct file *file; | |
381 | struct sockaddr __user *addr; | |
382 | int addr_len; | |
383 | }; | |
384 | ||
e47293fd JA |
385 | struct io_sr_msg { |
386 | struct file *file; | |
fddaface JA |
387 | union { |
388 | struct user_msghdr __user *msg; | |
389 | void __user *buf; | |
390 | }; | |
e47293fd | 391 | int msg_flags; |
fddaface | 392 | size_t len; |
e47293fd JA |
393 | }; |
394 | ||
15b71abe JA |
395 | struct io_open { |
396 | struct file *file; | |
397 | int dfd; | |
eddc7ef5 | 398 | union { |
eddc7ef5 JA |
399 | unsigned mask; |
400 | }; | |
15b71abe | 401 | struct filename *filename; |
eddc7ef5 | 402 | struct statx __user *buffer; |
c12cedf2 | 403 | struct open_how how; |
15b71abe JA |
404 | }; |
405 | ||
05f3fb3c JA |
406 | struct io_files_update { |
407 | struct file *file; | |
408 | u64 arg; | |
409 | u32 nr_args; | |
410 | u32 offset; | |
411 | }; | |
412 | ||
4840e418 JA |
413 | struct io_fadvise { |
414 | struct file *file; | |
415 | u64 offset; | |
416 | u32 len; | |
417 | u32 advice; | |
418 | }; | |
419 | ||
c1ca757b JA |
420 | struct io_madvise { |
421 | struct file *file; | |
422 | u64 addr; | |
423 | u32 len; | |
424 | u32 advice; | |
425 | }; | |
426 | ||
3e4827b0 JA |
427 | struct io_epoll { |
428 | struct file *file; | |
429 | int epfd; | |
430 | int op; | |
431 | int fd; | |
432 | struct epoll_event event; | |
e47293fd JA |
433 | }; |
434 | ||
f499a021 JA |
435 | struct io_async_connect { |
436 | struct sockaddr_storage address; | |
437 | }; | |
438 | ||
03b1230c JA |
439 | struct io_async_msghdr { |
440 | struct iovec fast_iov[UIO_FASTIOV]; | |
441 | struct iovec *iov; | |
442 | struct sockaddr __user *uaddr; | |
443 | struct msghdr msg; | |
444 | }; | |
445 | ||
f67676d1 JA |
446 | struct io_async_rw { |
447 | struct iovec fast_iov[UIO_FASTIOV]; | |
448 | struct iovec *iov; | |
449 | ssize_t nr_segs; | |
450 | ssize_t size; | |
451 | }; | |
452 | ||
15b71abe JA |
453 | struct io_async_open { |
454 | struct filename *filename; | |
455 | }; | |
456 | ||
1a6b74fc | 457 | struct io_async_ctx { |
f67676d1 JA |
458 | union { |
459 | struct io_async_rw rw; | |
03b1230c | 460 | struct io_async_msghdr msg; |
f499a021 | 461 | struct io_async_connect connect; |
2d28390a | 462 | struct io_timeout_data timeout; |
15b71abe | 463 | struct io_async_open open; |
f67676d1 | 464 | }; |
1a6b74fc JA |
465 | }; |
466 | ||
6b47ee6e PB |
467 | enum { |
468 | REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT, | |
469 | REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT, | |
470 | REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT, | |
471 | REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT, | |
472 | REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT, | |
473 | ||
474 | REQ_F_LINK_NEXT_BIT, | |
475 | REQ_F_FAIL_LINK_BIT, | |
476 | REQ_F_INFLIGHT_BIT, | |
477 | REQ_F_CUR_POS_BIT, | |
478 | REQ_F_NOWAIT_BIT, | |
479 | REQ_F_IOPOLL_COMPLETED_BIT, | |
480 | REQ_F_LINK_TIMEOUT_BIT, | |
481 | REQ_F_TIMEOUT_BIT, | |
482 | REQ_F_ISREG_BIT, | |
483 | REQ_F_MUST_PUNT_BIT, | |
484 | REQ_F_TIMEOUT_NOSEQ_BIT, | |
485 | REQ_F_COMP_LOCKED_BIT, | |
486 | }; | |
487 | ||
488 | enum { | |
489 | /* ctx owns file */ | |
490 | REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT), | |
491 | /* drain existing IO first */ | |
492 | REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT), | |
493 | /* linked sqes */ | |
494 | REQ_F_LINK = BIT(REQ_F_LINK_BIT), | |
495 | /* doesn't sever on completion < 0 */ | |
496 | REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT), | |
497 | /* IOSQE_ASYNC */ | |
498 | REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT), | |
499 | ||
500 | /* already grabbed next link */ | |
501 | REQ_F_LINK_NEXT = BIT(REQ_F_LINK_NEXT_BIT), | |
502 | /* fail rest of links */ | |
503 | REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT), | |
504 | /* on inflight list */ | |
505 | REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT), | |
506 | /* read/write uses file position */ | |
507 | REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT), | |
508 | /* must not punt to workers */ | |
509 | REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT), | |
510 | /* polled IO has completed */ | |
511 | REQ_F_IOPOLL_COMPLETED = BIT(REQ_F_IOPOLL_COMPLETED_BIT), | |
512 | /* has linked timeout */ | |
513 | REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT), | |
514 | /* timeout request */ | |
515 | REQ_F_TIMEOUT = BIT(REQ_F_TIMEOUT_BIT), | |
516 | /* regular file */ | |
517 | REQ_F_ISREG = BIT(REQ_F_ISREG_BIT), | |
518 | /* must be punted even for NONBLOCK */ | |
519 | REQ_F_MUST_PUNT = BIT(REQ_F_MUST_PUNT_BIT), | |
520 | /* no timeout sequence */ | |
521 | REQ_F_TIMEOUT_NOSEQ = BIT(REQ_F_TIMEOUT_NOSEQ_BIT), | |
522 | /* completion under lock */ | |
523 | REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT), | |
524 | }; | |
525 | ||
09bb8394 JA |
526 | /* |
527 | * NOTE! Each of the iocb union members has the file pointer | |
528 | * as the first entry in their struct definition. So you can | |
529 | * access the file pointer through any of the sub-structs, | |
530 | * or directly as just 'ki_filp' in this struct. | |
531 | */ | |
2b188cc1 | 532 | struct io_kiocb { |
221c5eb2 | 533 | union { |
09bb8394 | 534 | struct file *file; |
9adbd45d | 535 | struct io_rw rw; |
221c5eb2 | 536 | struct io_poll_iocb poll; |
8ed8d3c3 JA |
537 | struct io_accept accept; |
538 | struct io_sync sync; | |
fbf23849 | 539 | struct io_cancel cancel; |
b29472ee | 540 | struct io_timeout timeout; |
3fbb51c1 | 541 | struct io_connect connect; |
e47293fd | 542 | struct io_sr_msg sr_msg; |
15b71abe | 543 | struct io_open open; |
b5dba59e | 544 | struct io_close close; |
05f3fb3c | 545 | struct io_files_update files_update; |
4840e418 | 546 | struct io_fadvise fadvise; |
c1ca757b | 547 | struct io_madvise madvise; |
3e4827b0 | 548 | struct io_epoll epoll; |
221c5eb2 | 549 | }; |
2b188cc1 | 550 | |
1a6b74fc | 551 | struct io_async_ctx *io; |
b14cca0c PB |
552 | /* |
553 | * llist_node is only used for poll deferred completions | |
554 | */ | |
555 | struct llist_node llist_node; | |
cf6fd4bd PB |
556 | bool has_user; |
557 | bool in_async; | |
558 | bool needs_fixed_file; | |
d625c6ee | 559 | u8 opcode; |
2b188cc1 JA |
560 | |
561 | struct io_ring_ctx *ctx; | |
eac406c6 JA |
562 | union { |
563 | struct list_head list; | |
78076bb6 | 564 | struct hlist_node hash_node; |
eac406c6 | 565 | }; |
9e645e11 | 566 | struct list_head link_list; |
2b188cc1 | 567 | unsigned int flags; |
c16361c1 | 568 | refcount_t refs; |
2b188cc1 | 569 | u64 user_data; |
9e645e11 | 570 | u32 result; |
de0617e4 | 571 | u32 sequence; |
2b188cc1 | 572 | |
fcb323cc JA |
573 | struct list_head inflight_entry; |
574 | ||
561fb04a | 575 | struct io_wq_work work; |
2b188cc1 JA |
576 | }; |
577 | ||
578 | #define IO_PLUG_THRESHOLD 2 | |
def596e9 | 579 | #define IO_IOPOLL_BATCH 8 |
2b188cc1 | 580 | |
9a56a232 JA |
581 | struct io_submit_state { |
582 | struct blk_plug plug; | |
583 | ||
2579f913 JA |
584 | /* |
585 | * io_kiocb alloc cache | |
586 | */ | |
587 | void *reqs[IO_IOPOLL_BATCH]; | |
588 | unsigned int free_reqs; | |
589 | unsigned int cur_req; | |
590 | ||
9a56a232 JA |
591 | /* |
592 | * File reference cache | |
593 | */ | |
594 | struct file *file; | |
595 | unsigned int fd; | |
596 | unsigned int has_refs; | |
597 | unsigned int used_refs; | |
598 | unsigned int ios_left; | |
599 | }; | |
600 | ||
d3656344 JA |
601 | struct io_op_def { |
602 | /* needs req->io allocated for deferral/async */ | |
603 | unsigned async_ctx : 1; | |
604 | /* needs current->mm setup, does mm access */ | |
605 | unsigned needs_mm : 1; | |
606 | /* needs req->file assigned */ | |
607 | unsigned needs_file : 1; | |
608 | /* needs req->file assigned IFF fd is >= 0 */ | |
609 | unsigned fd_non_neg : 1; | |
610 | /* hash wq insertion if file is a regular file */ | |
611 | unsigned hash_reg_file : 1; | |
612 | /* unbound wq insertion if file is a non-regular file */ | |
613 | unsigned unbound_nonreg_file : 1; | |
66f4af93 JA |
614 | /* opcode is not supported by this kernel */ |
615 | unsigned not_supported : 1; | |
f86cd20c JA |
616 | /* needs file table */ |
617 | unsigned file_table : 1; | |
d3656344 JA |
618 | }; |
619 | ||
620 | static const struct io_op_def io_op_defs[] = { | |
0463b6c5 PB |
621 | [IORING_OP_NOP] = {}, |
622 | [IORING_OP_READV] = { | |
d3656344 JA |
623 | .async_ctx = 1, |
624 | .needs_mm = 1, | |
625 | .needs_file = 1, | |
626 | .unbound_nonreg_file = 1, | |
627 | }, | |
0463b6c5 | 628 | [IORING_OP_WRITEV] = { |
d3656344 JA |
629 | .async_ctx = 1, |
630 | .needs_mm = 1, | |
631 | .needs_file = 1, | |
632 | .hash_reg_file = 1, | |
633 | .unbound_nonreg_file = 1, | |
634 | }, | |
0463b6c5 | 635 | [IORING_OP_FSYNC] = { |
d3656344 JA |
636 | .needs_file = 1, |
637 | }, | |
0463b6c5 | 638 | [IORING_OP_READ_FIXED] = { |
d3656344 JA |
639 | .needs_file = 1, |
640 | .unbound_nonreg_file = 1, | |
641 | }, | |
0463b6c5 | 642 | [IORING_OP_WRITE_FIXED] = { |
d3656344 JA |
643 | .needs_file = 1, |
644 | .hash_reg_file = 1, | |
645 | .unbound_nonreg_file = 1, | |
646 | }, | |
0463b6c5 | 647 | [IORING_OP_POLL_ADD] = { |
d3656344 JA |
648 | .needs_file = 1, |
649 | .unbound_nonreg_file = 1, | |
650 | }, | |
0463b6c5 PB |
651 | [IORING_OP_POLL_REMOVE] = {}, |
652 | [IORING_OP_SYNC_FILE_RANGE] = { | |
d3656344 JA |
653 | .needs_file = 1, |
654 | }, | |
0463b6c5 | 655 | [IORING_OP_SENDMSG] = { |
d3656344 JA |
656 | .async_ctx = 1, |
657 | .needs_mm = 1, | |
658 | .needs_file = 1, | |
659 | .unbound_nonreg_file = 1, | |
660 | }, | |
0463b6c5 | 661 | [IORING_OP_RECVMSG] = { |
d3656344 JA |
662 | .async_ctx = 1, |
663 | .needs_mm = 1, | |
664 | .needs_file = 1, | |
665 | .unbound_nonreg_file = 1, | |
666 | }, | |
0463b6c5 | 667 | [IORING_OP_TIMEOUT] = { |
d3656344 JA |
668 | .async_ctx = 1, |
669 | .needs_mm = 1, | |
670 | }, | |
0463b6c5 PB |
671 | [IORING_OP_TIMEOUT_REMOVE] = {}, |
672 | [IORING_OP_ACCEPT] = { | |
d3656344 JA |
673 | .needs_mm = 1, |
674 | .needs_file = 1, | |
675 | .unbound_nonreg_file = 1, | |
f86cd20c | 676 | .file_table = 1, |
d3656344 | 677 | }, |
0463b6c5 PB |
678 | [IORING_OP_ASYNC_CANCEL] = {}, |
679 | [IORING_OP_LINK_TIMEOUT] = { | |
d3656344 JA |
680 | .async_ctx = 1, |
681 | .needs_mm = 1, | |
682 | }, | |
0463b6c5 | 683 | [IORING_OP_CONNECT] = { |
d3656344 JA |
684 | .async_ctx = 1, |
685 | .needs_mm = 1, | |
686 | .needs_file = 1, | |
687 | .unbound_nonreg_file = 1, | |
688 | }, | |
0463b6c5 | 689 | [IORING_OP_FALLOCATE] = { |
d3656344 JA |
690 | .needs_file = 1, |
691 | }, | |
0463b6c5 | 692 | [IORING_OP_OPENAT] = { |
d3656344 JA |
693 | .needs_file = 1, |
694 | .fd_non_neg = 1, | |
f86cd20c | 695 | .file_table = 1, |
d3656344 | 696 | }, |
0463b6c5 | 697 | [IORING_OP_CLOSE] = { |
d3656344 | 698 | .needs_file = 1, |
f86cd20c | 699 | .file_table = 1, |
d3656344 | 700 | }, |
0463b6c5 | 701 | [IORING_OP_FILES_UPDATE] = { |
d3656344 | 702 | .needs_mm = 1, |
f86cd20c | 703 | .file_table = 1, |
d3656344 | 704 | }, |
0463b6c5 | 705 | [IORING_OP_STATX] = { |
d3656344 JA |
706 | .needs_mm = 1, |
707 | .needs_file = 1, | |
708 | .fd_non_neg = 1, | |
709 | }, | |
0463b6c5 | 710 | [IORING_OP_READ] = { |
3a6820f2 JA |
711 | .needs_mm = 1, |
712 | .needs_file = 1, | |
713 | .unbound_nonreg_file = 1, | |
714 | }, | |
0463b6c5 | 715 | [IORING_OP_WRITE] = { |
3a6820f2 JA |
716 | .needs_mm = 1, |
717 | .needs_file = 1, | |
718 | .unbound_nonreg_file = 1, | |
719 | }, | |
0463b6c5 | 720 | [IORING_OP_FADVISE] = { |
4840e418 JA |
721 | .needs_file = 1, |
722 | }, | |
0463b6c5 | 723 | [IORING_OP_MADVISE] = { |
c1ca757b JA |
724 | .needs_mm = 1, |
725 | }, | |
0463b6c5 | 726 | [IORING_OP_SEND] = { |
fddaface JA |
727 | .needs_mm = 1, |
728 | .needs_file = 1, | |
729 | .unbound_nonreg_file = 1, | |
730 | }, | |
0463b6c5 | 731 | [IORING_OP_RECV] = { |
fddaface JA |
732 | .needs_mm = 1, |
733 | .needs_file = 1, | |
734 | .unbound_nonreg_file = 1, | |
735 | }, | |
0463b6c5 | 736 | [IORING_OP_OPENAT2] = { |
cebdb986 JA |
737 | .needs_file = 1, |
738 | .fd_non_neg = 1, | |
f86cd20c | 739 | .file_table = 1, |
cebdb986 | 740 | }, |
3e4827b0 JA |
741 | [IORING_OP_EPOLL_CTL] = { |
742 | .unbound_nonreg_file = 1, | |
743 | .file_table = 1, | |
744 | }, | |
d3656344 JA |
745 | }; |
746 | ||
561fb04a | 747 | static void io_wq_submit_work(struct io_wq_work **workptr); |
78e19bbe | 748 | static void io_cqring_fill_event(struct io_kiocb *req, long res); |
ec9c02ad | 749 | static void io_put_req(struct io_kiocb *req); |
978db57e | 750 | static void __io_double_put_req(struct io_kiocb *req); |
94ae5e77 JA |
751 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req); |
752 | static void io_queue_linked_timeout(struct io_kiocb *req); | |
05f3fb3c JA |
753 | static int __io_sqe_files_update(struct io_ring_ctx *ctx, |
754 | struct io_uring_files_update *ip, | |
755 | unsigned nr_args); | |
f86cd20c | 756 | static int io_grab_files(struct io_kiocb *req); |
de0617e4 | 757 | |
2b188cc1 JA |
758 | static struct kmem_cache *req_cachep; |
759 | ||
760 | static const struct file_operations io_uring_fops; | |
761 | ||
762 | struct sock *io_uring_get_socket(struct file *file) | |
763 | { | |
764 | #if defined(CONFIG_UNIX) | |
765 | if (file->f_op == &io_uring_fops) { | |
766 | struct io_ring_ctx *ctx = file->private_data; | |
767 | ||
768 | return ctx->ring_sock->sk; | |
769 | } | |
770 | #endif | |
771 | return NULL; | |
772 | } | |
773 | EXPORT_SYMBOL(io_uring_get_socket); | |
774 | ||
775 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) | |
776 | { | |
777 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
778 | ||
206aefde | 779 | complete(&ctx->completions[0]); |
2b188cc1 JA |
780 | } |
781 | ||
782 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) | |
783 | { | |
784 | struct io_ring_ctx *ctx; | |
78076bb6 | 785 | int hash_bits; |
2b188cc1 JA |
786 | |
787 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
788 | if (!ctx) | |
789 | return NULL; | |
790 | ||
0ddf92e8 JA |
791 | ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL); |
792 | if (!ctx->fallback_req) | |
793 | goto err; | |
794 | ||
206aefde JA |
795 | ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL); |
796 | if (!ctx->completions) | |
797 | goto err; | |
798 | ||
78076bb6 JA |
799 | /* |
800 | * Use 5 bits less than the max cq entries, that should give us around | |
801 | * 32 entries per hash list if totally full and uniformly spread. | |
802 | */ | |
803 | hash_bits = ilog2(p->cq_entries); | |
804 | hash_bits -= 5; | |
805 | if (hash_bits <= 0) | |
806 | hash_bits = 1; | |
807 | ctx->cancel_hash_bits = hash_bits; | |
808 | ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head), | |
809 | GFP_KERNEL); | |
810 | if (!ctx->cancel_hash) | |
811 | goto err; | |
812 | __hash_init(ctx->cancel_hash, 1U << hash_bits); | |
813 | ||
21482896 | 814 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
206aefde JA |
815 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) |
816 | goto err; | |
2b188cc1 JA |
817 | |
818 | ctx->flags = p->flags; | |
819 | init_waitqueue_head(&ctx->cq_wait); | |
1d7bb1d5 | 820 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
206aefde JA |
821 | init_completion(&ctx->completions[0]); |
822 | init_completion(&ctx->completions[1]); | |
071698e1 | 823 | idr_init(&ctx->personality_idr); |
2b188cc1 JA |
824 | mutex_init(&ctx->uring_lock); |
825 | init_waitqueue_head(&ctx->wait); | |
826 | spin_lock_init(&ctx->completion_lock); | |
e94f141b | 827 | init_llist_head(&ctx->poll_llist); |
def596e9 | 828 | INIT_LIST_HEAD(&ctx->poll_list); |
de0617e4 | 829 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 830 | INIT_LIST_HEAD(&ctx->timeout_list); |
fcb323cc JA |
831 | init_waitqueue_head(&ctx->inflight_wait); |
832 | spin_lock_init(&ctx->inflight_lock); | |
833 | INIT_LIST_HEAD(&ctx->inflight_list); | |
2b188cc1 | 834 | return ctx; |
206aefde | 835 | err: |
0ddf92e8 JA |
836 | if (ctx->fallback_req) |
837 | kmem_cache_free(req_cachep, ctx->fallback_req); | |
206aefde | 838 | kfree(ctx->completions); |
78076bb6 | 839 | kfree(ctx->cancel_hash); |
206aefde JA |
840 | kfree(ctx); |
841 | return NULL; | |
2b188cc1 JA |
842 | } |
843 | ||
9d858b21 | 844 | static inline bool __req_need_defer(struct io_kiocb *req) |
7adf4eaf | 845 | { |
a197f664 JL |
846 | struct io_ring_ctx *ctx = req->ctx; |
847 | ||
498ccd9e JA |
848 | return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped |
849 | + atomic_read(&ctx->cached_cq_overflow); | |
7adf4eaf JA |
850 | } |
851 | ||
9d858b21 | 852 | static inline bool req_need_defer(struct io_kiocb *req) |
de0617e4 | 853 | { |
87987898 | 854 | if (unlikely(req->flags & REQ_F_IO_DRAIN)) |
9d858b21 | 855 | return __req_need_defer(req); |
de0617e4 | 856 | |
9d858b21 | 857 | return false; |
de0617e4 JA |
858 | } |
859 | ||
7adf4eaf | 860 | static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx) |
de0617e4 JA |
861 | { |
862 | struct io_kiocb *req; | |
863 | ||
7adf4eaf | 864 | req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list); |
9d858b21 | 865 | if (req && !req_need_defer(req)) { |
de0617e4 JA |
866 | list_del_init(&req->list); |
867 | return req; | |
868 | } | |
869 | ||
870 | return NULL; | |
871 | } | |
872 | ||
5262f567 JA |
873 | static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx) |
874 | { | |
7adf4eaf JA |
875 | struct io_kiocb *req; |
876 | ||
877 | req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list); | |
93bd25bb JA |
878 | if (req) { |
879 | if (req->flags & REQ_F_TIMEOUT_NOSEQ) | |
880 | return NULL; | |
fb4b3d3f | 881 | if (!__req_need_defer(req)) { |
93bd25bb JA |
882 | list_del_init(&req->list); |
883 | return req; | |
884 | } | |
7adf4eaf JA |
885 | } |
886 | ||
887 | return NULL; | |
5262f567 JA |
888 | } |
889 | ||
de0617e4 | 890 | static void __io_commit_cqring(struct io_ring_ctx *ctx) |
2b188cc1 | 891 | { |
75b28aff | 892 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 893 | |
07910158 PB |
894 | /* order cqe stores with ring update */ |
895 | smp_store_release(&rings->cq.tail, ctx->cached_cq_tail); | |
2b188cc1 | 896 | |
07910158 PB |
897 | if (wq_has_sleeper(&ctx->cq_wait)) { |
898 | wake_up_interruptible(&ctx->cq_wait); | |
899 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
2b188cc1 JA |
900 | } |
901 | } | |
902 | ||
cccf0ee8 JA |
903 | static inline void io_req_work_grab_env(struct io_kiocb *req, |
904 | const struct io_op_def *def) | |
905 | { | |
906 | if (!req->work.mm && def->needs_mm) { | |
907 | mmgrab(current->mm); | |
908 | req->work.mm = current->mm; | |
2b188cc1 | 909 | } |
cccf0ee8 JA |
910 | if (!req->work.creds) |
911 | req->work.creds = get_current_cred(); | |
2b188cc1 JA |
912 | } |
913 | ||
cccf0ee8 | 914 | static inline void io_req_work_drop_env(struct io_kiocb *req) |
18d9be1a | 915 | { |
cccf0ee8 JA |
916 | if (req->work.mm) { |
917 | mmdrop(req->work.mm); | |
918 | req->work.mm = NULL; | |
919 | } | |
920 | if (req->work.creds) { | |
921 | put_cred(req->work.creds); | |
922 | req->work.creds = NULL; | |
923 | } | |
561fb04a JA |
924 | } |
925 | ||
94ae5e77 JA |
926 | static inline bool io_prep_async_work(struct io_kiocb *req, |
927 | struct io_kiocb **link) | |
18d9be1a | 928 | { |
d3656344 | 929 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
561fb04a | 930 | bool do_hashed = false; |
54a91f3b | 931 | |
d3656344 JA |
932 | if (req->flags & REQ_F_ISREG) { |
933 | if (def->hash_reg_file) | |
3529d8c2 | 934 | do_hashed = true; |
d3656344 JA |
935 | } else { |
936 | if (def->unbound_nonreg_file) | |
3529d8c2 | 937 | req->work.flags |= IO_WQ_WORK_UNBOUND; |
54a91f3b | 938 | } |
cccf0ee8 JA |
939 | |
940 | io_req_work_grab_env(req, def); | |
54a91f3b | 941 | |
94ae5e77 | 942 | *link = io_prep_linked_timeout(req); |
561fb04a JA |
943 | return do_hashed; |
944 | } | |
945 | ||
a197f664 | 946 | static inline void io_queue_async_work(struct io_kiocb *req) |
561fb04a | 947 | { |
a197f664 | 948 | struct io_ring_ctx *ctx = req->ctx; |
94ae5e77 JA |
949 | struct io_kiocb *link; |
950 | bool do_hashed; | |
951 | ||
952 | do_hashed = io_prep_async_work(req, &link); | |
561fb04a JA |
953 | |
954 | trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work, | |
955 | req->flags); | |
956 | if (!do_hashed) { | |
957 | io_wq_enqueue(ctx->io_wq, &req->work); | |
958 | } else { | |
959 | io_wq_enqueue_hashed(ctx->io_wq, &req->work, | |
960 | file_inode(req->file)); | |
961 | } | |
94ae5e77 JA |
962 | |
963 | if (link) | |
964 | io_queue_linked_timeout(link); | |
18d9be1a JA |
965 | } |
966 | ||
5262f567 JA |
967 | static void io_kill_timeout(struct io_kiocb *req) |
968 | { | |
969 | int ret; | |
970 | ||
2d28390a | 971 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); |
5262f567 JA |
972 | if (ret != -1) { |
973 | atomic_inc(&req->ctx->cq_timeouts); | |
842f9612 | 974 | list_del_init(&req->list); |
78e19bbe | 975 | io_cqring_fill_event(req, 0); |
ec9c02ad | 976 | io_put_req(req); |
5262f567 JA |
977 | } |
978 | } | |
979 | ||
980 | static void io_kill_timeouts(struct io_ring_ctx *ctx) | |
981 | { | |
982 | struct io_kiocb *req, *tmp; | |
983 | ||
984 | spin_lock_irq(&ctx->completion_lock); | |
985 | list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list) | |
986 | io_kill_timeout(req); | |
987 | spin_unlock_irq(&ctx->completion_lock); | |
988 | } | |
989 | ||
de0617e4 JA |
990 | static void io_commit_cqring(struct io_ring_ctx *ctx) |
991 | { | |
992 | struct io_kiocb *req; | |
993 | ||
5262f567 JA |
994 | while ((req = io_get_timeout_req(ctx)) != NULL) |
995 | io_kill_timeout(req); | |
996 | ||
de0617e4 JA |
997 | __io_commit_cqring(ctx); |
998 | ||
87987898 | 999 | while ((req = io_get_deferred_req(ctx)) != NULL) |
a197f664 | 1000 | io_queue_async_work(req); |
de0617e4 JA |
1001 | } |
1002 | ||
2b188cc1 JA |
1003 | static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx) |
1004 | { | |
75b28aff | 1005 | struct io_rings *rings = ctx->rings; |
2b188cc1 JA |
1006 | unsigned tail; |
1007 | ||
1008 | tail = ctx->cached_cq_tail; | |
115e12e5 SB |
1009 | /* |
1010 | * writes to the cq entry need to come after reading head; the | |
1011 | * control dependency is enough as we're using WRITE_ONCE to | |
1012 | * fill the cq entry | |
1013 | */ | |
75b28aff | 1014 | if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries) |
2b188cc1 JA |
1015 | return NULL; |
1016 | ||
1017 | ctx->cached_cq_tail++; | |
75b28aff | 1018 | return &rings->cqes[tail & ctx->cq_mask]; |
2b188cc1 JA |
1019 | } |
1020 | ||
f2842ab5 JA |
1021 | static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx) |
1022 | { | |
1023 | if (!ctx->eventfd_async) | |
1024 | return true; | |
1025 | return io_wq_current_is_worker() || in_interrupt(); | |
1026 | } | |
1027 | ||
1d7bb1d5 JA |
1028 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
1029 | { | |
1030 | if (waitqueue_active(&ctx->wait)) | |
1031 | wake_up(&ctx->wait); | |
1032 | if (waitqueue_active(&ctx->sqo_wait)) | |
1033 | wake_up(&ctx->sqo_wait); | |
f2842ab5 | 1034 | if (ctx->cq_ev_fd && io_should_trigger_evfd(ctx)) |
1d7bb1d5 JA |
1035 | eventfd_signal(ctx->cq_ev_fd, 1); |
1036 | } | |
1037 | ||
c4a2ed72 JA |
1038 | /* Returns true if there are no backlogged entries after the flush */ |
1039 | static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force) | |
1d7bb1d5 JA |
1040 | { |
1041 | struct io_rings *rings = ctx->rings; | |
1042 | struct io_uring_cqe *cqe; | |
1043 | struct io_kiocb *req; | |
1044 | unsigned long flags; | |
1045 | LIST_HEAD(list); | |
1046 | ||
1047 | if (!force) { | |
1048 | if (list_empty_careful(&ctx->cq_overflow_list)) | |
c4a2ed72 | 1049 | return true; |
1d7bb1d5 JA |
1050 | if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) == |
1051 | rings->cq_ring_entries)) | |
c4a2ed72 | 1052 | return false; |
1d7bb1d5 JA |
1053 | } |
1054 | ||
1055 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
1056 | ||
1057 | /* if force is set, the ring is going away. always drop after that */ | |
1058 | if (force) | |
69b3e546 | 1059 | ctx->cq_overflow_flushed = 1; |
1d7bb1d5 | 1060 | |
c4a2ed72 | 1061 | cqe = NULL; |
1d7bb1d5 JA |
1062 | while (!list_empty(&ctx->cq_overflow_list)) { |
1063 | cqe = io_get_cqring(ctx); | |
1064 | if (!cqe && !force) | |
1065 | break; | |
1066 | ||
1067 | req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb, | |
1068 | list); | |
1069 | list_move(&req->list, &list); | |
1070 | if (cqe) { | |
1071 | WRITE_ONCE(cqe->user_data, req->user_data); | |
1072 | WRITE_ONCE(cqe->res, req->result); | |
1073 | WRITE_ONCE(cqe->flags, 0); | |
1074 | } else { | |
1075 | WRITE_ONCE(ctx->rings->cq_overflow, | |
1076 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
1077 | } | |
1078 | } | |
1079 | ||
1080 | io_commit_cqring(ctx); | |
ad3eb2c8 JA |
1081 | if (cqe) { |
1082 | clear_bit(0, &ctx->sq_check_overflow); | |
1083 | clear_bit(0, &ctx->cq_check_overflow); | |
1084 | } | |
1d7bb1d5 JA |
1085 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
1086 | io_cqring_ev_posted(ctx); | |
1087 | ||
1088 | while (!list_empty(&list)) { | |
1089 | req = list_first_entry(&list, struct io_kiocb, list); | |
1090 | list_del(&req->list); | |
ec9c02ad | 1091 | io_put_req(req); |
1d7bb1d5 | 1092 | } |
c4a2ed72 JA |
1093 | |
1094 | return cqe != NULL; | |
1d7bb1d5 JA |
1095 | } |
1096 | ||
78e19bbe | 1097 | static void io_cqring_fill_event(struct io_kiocb *req, long res) |
2b188cc1 | 1098 | { |
78e19bbe | 1099 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
1100 | struct io_uring_cqe *cqe; |
1101 | ||
78e19bbe | 1102 | trace_io_uring_complete(ctx, req->user_data, res); |
51c3ff62 | 1103 | |
2b188cc1 JA |
1104 | /* |
1105 | * If we can't get a cq entry, userspace overflowed the | |
1106 | * submission (by quite a lot). Increment the overflow count in | |
1107 | * the ring. | |
1108 | */ | |
1109 | cqe = io_get_cqring(ctx); | |
1d7bb1d5 | 1110 | if (likely(cqe)) { |
78e19bbe | 1111 | WRITE_ONCE(cqe->user_data, req->user_data); |
2b188cc1 | 1112 | WRITE_ONCE(cqe->res, res); |
c71ffb67 | 1113 | WRITE_ONCE(cqe->flags, 0); |
1d7bb1d5 | 1114 | } else if (ctx->cq_overflow_flushed) { |
498ccd9e JA |
1115 | WRITE_ONCE(ctx->rings->cq_overflow, |
1116 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
1d7bb1d5 | 1117 | } else { |
ad3eb2c8 JA |
1118 | if (list_empty(&ctx->cq_overflow_list)) { |
1119 | set_bit(0, &ctx->sq_check_overflow); | |
1120 | set_bit(0, &ctx->cq_check_overflow); | |
1121 | } | |
1d7bb1d5 JA |
1122 | refcount_inc(&req->refs); |
1123 | req->result = res; | |
1124 | list_add_tail(&req->list, &ctx->cq_overflow_list); | |
2b188cc1 JA |
1125 | } |
1126 | } | |
1127 | ||
78e19bbe | 1128 | static void io_cqring_add_event(struct io_kiocb *req, long res) |
2b188cc1 | 1129 | { |
78e19bbe | 1130 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
1131 | unsigned long flags; |
1132 | ||
1133 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
78e19bbe | 1134 | io_cqring_fill_event(req, res); |
2b188cc1 JA |
1135 | io_commit_cqring(ctx); |
1136 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1137 | ||
8c838788 | 1138 | io_cqring_ev_posted(ctx); |
2b188cc1 JA |
1139 | } |
1140 | ||
0ddf92e8 JA |
1141 | static inline bool io_is_fallback_req(struct io_kiocb *req) |
1142 | { | |
1143 | return req == (struct io_kiocb *) | |
1144 | ((unsigned long) req->ctx->fallback_req & ~1UL); | |
1145 | } | |
1146 | ||
1147 | static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx) | |
1148 | { | |
1149 | struct io_kiocb *req; | |
1150 | ||
1151 | req = ctx->fallback_req; | |
1152 | if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req)) | |
1153 | return req; | |
1154 | ||
1155 | return NULL; | |
1156 | } | |
1157 | ||
2579f913 JA |
1158 | static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx, |
1159 | struct io_submit_state *state) | |
2b188cc1 | 1160 | { |
fd6fab2c | 1161 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
2b188cc1 JA |
1162 | struct io_kiocb *req; |
1163 | ||
2579f913 | 1164 | if (!state) { |
fd6fab2c | 1165 | req = kmem_cache_alloc(req_cachep, gfp); |
2579f913 | 1166 | if (unlikely(!req)) |
0ddf92e8 | 1167 | goto fallback; |
2579f913 JA |
1168 | } else if (!state->free_reqs) { |
1169 | size_t sz; | |
1170 | int ret; | |
1171 | ||
1172 | sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs)); | |
fd6fab2c JA |
1173 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs); |
1174 | ||
1175 | /* | |
1176 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
1177 | * retry single alloc to be on the safe side. | |
1178 | */ | |
1179 | if (unlikely(ret <= 0)) { | |
1180 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
1181 | if (!state->reqs[0]) | |
0ddf92e8 | 1182 | goto fallback; |
fd6fab2c JA |
1183 | ret = 1; |
1184 | } | |
2579f913 JA |
1185 | state->free_reqs = ret - 1; |
1186 | state->cur_req = 1; | |
1187 | req = state->reqs[0]; | |
1188 | } else { | |
1189 | req = state->reqs[state->cur_req]; | |
1190 | state->free_reqs--; | |
1191 | state->cur_req++; | |
2b188cc1 JA |
1192 | } |
1193 | ||
0ddf92e8 | 1194 | got_it: |
1a6b74fc | 1195 | req->io = NULL; |
60c112b0 | 1196 | req->file = NULL; |
2579f913 JA |
1197 | req->ctx = ctx; |
1198 | req->flags = 0; | |
e65ef56d JA |
1199 | /* one is dropped after submission, the other at completion */ |
1200 | refcount_set(&req->refs, 2); | |
9e645e11 | 1201 | req->result = 0; |
561fb04a | 1202 | INIT_IO_WORK(&req->work, io_wq_submit_work); |
2579f913 | 1203 | return req; |
0ddf92e8 JA |
1204 | fallback: |
1205 | req = io_get_fallback_req(ctx); | |
1206 | if (req) | |
1207 | goto got_it; | |
6805b32e | 1208 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
1209 | return NULL; |
1210 | } | |
1211 | ||
2b85edfc | 1212 | static void __io_req_do_free(struct io_kiocb *req) |
def596e9 | 1213 | { |
2b85edfc PB |
1214 | if (likely(!io_is_fallback_req(req))) |
1215 | kmem_cache_free(req_cachep, req); | |
1216 | else | |
1217 | clear_bit_unlock(0, (unsigned long *) req->ctx->fallback_req); | |
1218 | } | |
1219 | ||
c6ca97b3 | 1220 | static void __io_req_aux_free(struct io_kiocb *req) |
2b188cc1 | 1221 | { |
fcb323cc JA |
1222 | struct io_ring_ctx *ctx = req->ctx; |
1223 | ||
96fd84d8 | 1224 | kfree(req->io); |
05f3fb3c JA |
1225 | if (req->file) { |
1226 | if (req->flags & REQ_F_FIXED_FILE) | |
1227 | percpu_ref_put(&ctx->file_data->refs); | |
1228 | else | |
1229 | fput(req->file); | |
def596e9 | 1230 | } |
cccf0ee8 JA |
1231 | |
1232 | io_req_work_drop_env(req); | |
def596e9 JA |
1233 | } |
1234 | ||
9e645e11 | 1235 | static void __io_free_req(struct io_kiocb *req) |
2b188cc1 | 1236 | { |
c6ca97b3 | 1237 | __io_req_aux_free(req); |
fcb323cc | 1238 | |
fcb323cc | 1239 | if (req->flags & REQ_F_INFLIGHT) { |
c6ca97b3 | 1240 | struct io_ring_ctx *ctx = req->ctx; |
fcb323cc JA |
1241 | unsigned long flags; |
1242 | ||
1243 | spin_lock_irqsave(&ctx->inflight_lock, flags); | |
1244 | list_del(&req->inflight_entry); | |
1245 | if (waitqueue_active(&ctx->inflight_wait)) | |
1246 | wake_up(&ctx->inflight_wait); | |
1247 | spin_unlock_irqrestore(&ctx->inflight_lock, flags); | |
1248 | } | |
2b85edfc PB |
1249 | |
1250 | percpu_ref_put(&req->ctx->refs); | |
1251 | __io_req_do_free(req); | |
e65ef56d JA |
1252 | } |
1253 | ||
c6ca97b3 JA |
1254 | struct req_batch { |
1255 | void *reqs[IO_IOPOLL_BATCH]; | |
1256 | int to_free; | |
1257 | int need_iter; | |
1258 | }; | |
1259 | ||
1260 | static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb) | |
1261 | { | |
10fef4be JA |
1262 | int fixed_refs = rb->to_free; |
1263 | ||
c6ca97b3 JA |
1264 | if (!rb->to_free) |
1265 | return; | |
1266 | if (rb->need_iter) { | |
1267 | int i, inflight = 0; | |
1268 | unsigned long flags; | |
1269 | ||
10fef4be | 1270 | fixed_refs = 0; |
c6ca97b3 JA |
1271 | for (i = 0; i < rb->to_free; i++) { |
1272 | struct io_kiocb *req = rb->reqs[i]; | |
1273 | ||
10fef4be | 1274 | if (req->flags & REQ_F_FIXED_FILE) { |
c6ca97b3 | 1275 | req->file = NULL; |
10fef4be JA |
1276 | fixed_refs++; |
1277 | } | |
c6ca97b3 JA |
1278 | if (req->flags & REQ_F_INFLIGHT) |
1279 | inflight++; | |
c6ca97b3 JA |
1280 | __io_req_aux_free(req); |
1281 | } | |
1282 | if (!inflight) | |
1283 | goto do_free; | |
1284 | ||
1285 | spin_lock_irqsave(&ctx->inflight_lock, flags); | |
1286 | for (i = 0; i < rb->to_free; i++) { | |
1287 | struct io_kiocb *req = rb->reqs[i]; | |
1288 | ||
10fef4be | 1289 | if (req->flags & REQ_F_INFLIGHT) { |
c6ca97b3 JA |
1290 | list_del(&req->inflight_entry); |
1291 | if (!--inflight) | |
1292 | break; | |
1293 | } | |
1294 | } | |
1295 | spin_unlock_irqrestore(&ctx->inflight_lock, flags); | |
1296 | ||
1297 | if (waitqueue_active(&ctx->inflight_wait)) | |
1298 | wake_up(&ctx->inflight_wait); | |
1299 | } | |
1300 | do_free: | |
1301 | kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs); | |
10fef4be JA |
1302 | if (fixed_refs) |
1303 | percpu_ref_put_many(&ctx->file_data->refs, fixed_refs); | |
c6ca97b3 | 1304 | percpu_ref_put_many(&ctx->refs, rb->to_free); |
c6ca97b3 | 1305 | rb->to_free = rb->need_iter = 0; |
e65ef56d JA |
1306 | } |
1307 | ||
a197f664 | 1308 | static bool io_link_cancel_timeout(struct io_kiocb *req) |
2665abfd | 1309 | { |
a197f664 | 1310 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd JA |
1311 | int ret; |
1312 | ||
2d28390a | 1313 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); |
2665abfd | 1314 | if (ret != -1) { |
78e19bbe | 1315 | io_cqring_fill_event(req, -ECANCELED); |
2665abfd JA |
1316 | io_commit_cqring(ctx); |
1317 | req->flags &= ~REQ_F_LINK; | |
ec9c02ad | 1318 | io_put_req(req); |
2665abfd JA |
1319 | return true; |
1320 | } | |
1321 | ||
1322 | return false; | |
e65ef56d JA |
1323 | } |
1324 | ||
ba816ad6 | 1325 | static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr) |
9e645e11 | 1326 | { |
2665abfd | 1327 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd | 1328 | bool wake_ev = false; |
9e645e11 | 1329 | |
4d7dd462 JA |
1330 | /* Already got next link */ |
1331 | if (req->flags & REQ_F_LINK_NEXT) | |
1332 | return; | |
1333 | ||
9e645e11 JA |
1334 | /* |
1335 | * The list should never be empty when we are called here. But could | |
1336 | * potentially happen if the chain is messed up, check to be on the | |
1337 | * safe side. | |
1338 | */ | |
4493233e PB |
1339 | while (!list_empty(&req->link_list)) { |
1340 | struct io_kiocb *nxt = list_first_entry(&req->link_list, | |
1341 | struct io_kiocb, link_list); | |
94ae5e77 | 1342 | |
4493233e PB |
1343 | if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) && |
1344 | (nxt->flags & REQ_F_TIMEOUT))) { | |
1345 | list_del_init(&nxt->link_list); | |
94ae5e77 | 1346 | wake_ev |= io_link_cancel_timeout(nxt); |
94ae5e77 JA |
1347 | req->flags &= ~REQ_F_LINK_TIMEOUT; |
1348 | continue; | |
1349 | } | |
9e645e11 | 1350 | |
4493233e PB |
1351 | list_del_init(&req->link_list); |
1352 | if (!list_empty(&nxt->link_list)) | |
1353 | nxt->flags |= REQ_F_LINK; | |
b18fdf71 | 1354 | *nxtptr = nxt; |
94ae5e77 | 1355 | break; |
9e645e11 | 1356 | } |
2665abfd | 1357 | |
4d7dd462 | 1358 | req->flags |= REQ_F_LINK_NEXT; |
2665abfd JA |
1359 | if (wake_ev) |
1360 | io_cqring_ev_posted(ctx); | |
9e645e11 JA |
1361 | } |
1362 | ||
1363 | /* | |
1364 | * Called if REQ_F_LINK is set, and we fail the head request | |
1365 | */ | |
1366 | static void io_fail_links(struct io_kiocb *req) | |
1367 | { | |
2665abfd | 1368 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd JA |
1369 | unsigned long flags; |
1370 | ||
1371 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
9e645e11 JA |
1372 | |
1373 | while (!list_empty(&req->link_list)) { | |
4493233e PB |
1374 | struct io_kiocb *link = list_first_entry(&req->link_list, |
1375 | struct io_kiocb, link_list); | |
9e645e11 | 1376 | |
4493233e | 1377 | list_del_init(&link->link_list); |
c826bd7a | 1378 | trace_io_uring_fail_link(req, link); |
2665abfd JA |
1379 | |
1380 | if ((req->flags & REQ_F_LINK_TIMEOUT) && | |
d625c6ee | 1381 | link->opcode == IORING_OP_LINK_TIMEOUT) { |
a197f664 | 1382 | io_link_cancel_timeout(link); |
2665abfd | 1383 | } else { |
78e19bbe | 1384 | io_cqring_fill_event(link, -ECANCELED); |
978db57e | 1385 | __io_double_put_req(link); |
2665abfd | 1386 | } |
5d960724 | 1387 | req->flags &= ~REQ_F_LINK_TIMEOUT; |
9e645e11 | 1388 | } |
2665abfd JA |
1389 | |
1390 | io_commit_cqring(ctx); | |
1391 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1392 | io_cqring_ev_posted(ctx); | |
9e645e11 JA |
1393 | } |
1394 | ||
4d7dd462 | 1395 | static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt) |
9e645e11 | 1396 | { |
4d7dd462 | 1397 | if (likely(!(req->flags & REQ_F_LINK))) |
2665abfd | 1398 | return; |
2665abfd | 1399 | |
9e645e11 JA |
1400 | /* |
1401 | * If LINK is set, we have dependent requests in this chain. If we | |
1402 | * didn't fail this request, queue the first one up, moving any other | |
1403 | * dependencies to the next request. In case of failure, fail the rest | |
1404 | * of the chain. | |
1405 | */ | |
2665abfd JA |
1406 | if (req->flags & REQ_F_FAIL_LINK) { |
1407 | io_fail_links(req); | |
7c9e7f0f JA |
1408 | } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) == |
1409 | REQ_F_LINK_TIMEOUT) { | |
2665abfd JA |
1410 | struct io_ring_ctx *ctx = req->ctx; |
1411 | unsigned long flags; | |
1412 | ||
1413 | /* | |
1414 | * If this is a timeout link, we could be racing with the | |
1415 | * timeout timer. Grab the completion lock for this case to | |
7c9e7f0f | 1416 | * protect against that. |
2665abfd JA |
1417 | */ |
1418 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
1419 | io_req_link_next(req, nxt); | |
1420 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1421 | } else { | |
1422 | io_req_link_next(req, nxt); | |
9e645e11 | 1423 | } |
4d7dd462 | 1424 | } |
9e645e11 | 1425 | |
c69f8dbe JL |
1426 | static void io_free_req(struct io_kiocb *req) |
1427 | { | |
944e58bf PB |
1428 | struct io_kiocb *nxt = NULL; |
1429 | ||
1430 | io_req_find_next(req, &nxt); | |
70cf9f32 | 1431 | __io_free_req(req); |
944e58bf PB |
1432 | |
1433 | if (nxt) | |
1434 | io_queue_async_work(nxt); | |
c69f8dbe JL |
1435 | } |
1436 | ||
ba816ad6 JA |
1437 | /* |
1438 | * Drop reference to request, return next in chain (if there is one) if this | |
1439 | * was the last reference to this request. | |
1440 | */ | |
f9bd67f6 | 1441 | __attribute__((nonnull)) |
ec9c02ad | 1442 | static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr) |
e65ef56d | 1443 | { |
f9bd67f6 | 1444 | io_req_find_next(req, nxtptr); |
4d7dd462 | 1445 | |
e65ef56d | 1446 | if (refcount_dec_and_test(&req->refs)) |
4d7dd462 | 1447 | __io_free_req(req); |
2b188cc1 JA |
1448 | } |
1449 | ||
e65ef56d JA |
1450 | static void io_put_req(struct io_kiocb *req) |
1451 | { | |
1452 | if (refcount_dec_and_test(&req->refs)) | |
1453 | io_free_req(req); | |
2b188cc1 JA |
1454 | } |
1455 | ||
978db57e JA |
1456 | /* |
1457 | * Must only be used if we don't need to care about links, usually from | |
1458 | * within the completion handling itself. | |
1459 | */ | |
1460 | static void __io_double_put_req(struct io_kiocb *req) | |
78e19bbe JA |
1461 | { |
1462 | /* drop both submit and complete references */ | |
1463 | if (refcount_sub_and_test(2, &req->refs)) | |
1464 | __io_free_req(req); | |
1465 | } | |
1466 | ||
978db57e JA |
1467 | static void io_double_put_req(struct io_kiocb *req) |
1468 | { | |
1469 | /* drop both submit and complete references */ | |
1470 | if (refcount_sub_and_test(2, &req->refs)) | |
1471 | io_free_req(req); | |
1472 | } | |
1473 | ||
1d7bb1d5 | 1474 | static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush) |
a3a0e43f | 1475 | { |
84f97dc2 JA |
1476 | struct io_rings *rings = ctx->rings; |
1477 | ||
ad3eb2c8 JA |
1478 | if (test_bit(0, &ctx->cq_check_overflow)) { |
1479 | /* | |
1480 | * noflush == true is from the waitqueue handler, just ensure | |
1481 | * we wake up the task, and the next invocation will flush the | |
1482 | * entries. We cannot safely to it from here. | |
1483 | */ | |
1484 | if (noflush && !list_empty(&ctx->cq_overflow_list)) | |
1485 | return -1U; | |
1d7bb1d5 | 1486 | |
ad3eb2c8 JA |
1487 | io_cqring_overflow_flush(ctx, false); |
1488 | } | |
1d7bb1d5 | 1489 | |
a3a0e43f JA |
1490 | /* See comment at the top of this file */ |
1491 | smp_rmb(); | |
ad3eb2c8 | 1492 | return ctx->cached_cq_tail - READ_ONCE(rings->cq.head); |
a3a0e43f JA |
1493 | } |
1494 | ||
fb5ccc98 PB |
1495 | static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) |
1496 | { | |
1497 | struct io_rings *rings = ctx->rings; | |
1498 | ||
1499 | /* make sure SQ entry isn't read before tail */ | |
1500 | return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; | |
1501 | } | |
1502 | ||
8237e045 | 1503 | static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req) |
e94f141b | 1504 | { |
c6ca97b3 JA |
1505 | if ((req->flags & REQ_F_LINK) || io_is_fallback_req(req)) |
1506 | return false; | |
e94f141b | 1507 | |
c6ca97b3 JA |
1508 | if (!(req->flags & REQ_F_FIXED_FILE) || req->io) |
1509 | rb->need_iter++; | |
1510 | ||
1511 | rb->reqs[rb->to_free++] = req; | |
1512 | if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs))) | |
1513 | io_free_req_many(req->ctx, rb); | |
1514 | return true; | |
e94f141b JA |
1515 | } |
1516 | ||
def596e9 JA |
1517 | /* |
1518 | * Find and free completed poll iocbs | |
1519 | */ | |
1520 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1521 | struct list_head *done) | |
1522 | { | |
8237e045 | 1523 | struct req_batch rb; |
def596e9 | 1524 | struct io_kiocb *req; |
def596e9 | 1525 | |
c6ca97b3 | 1526 | rb.to_free = rb.need_iter = 0; |
def596e9 JA |
1527 | while (!list_empty(done)) { |
1528 | req = list_first_entry(done, struct io_kiocb, list); | |
1529 | list_del(&req->list); | |
1530 | ||
78e19bbe | 1531 | io_cqring_fill_event(req, req->result); |
def596e9 JA |
1532 | (*nr_events)++; |
1533 | ||
8237e045 JA |
1534 | if (refcount_dec_and_test(&req->refs) && |
1535 | !io_req_multi_free(&rb, req)) | |
1536 | io_free_req(req); | |
def596e9 | 1537 | } |
def596e9 | 1538 | |
09bb8394 | 1539 | io_commit_cqring(ctx); |
8237e045 | 1540 | io_free_req_many(ctx, &rb); |
def596e9 JA |
1541 | } |
1542 | ||
1543 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1544 | long min) | |
1545 | { | |
1546 | struct io_kiocb *req, *tmp; | |
1547 | LIST_HEAD(done); | |
1548 | bool spin; | |
1549 | int ret; | |
1550 | ||
1551 | /* | |
1552 | * Only spin for completions if we don't have multiple devices hanging | |
1553 | * off our complete list, and we're under the requested amount. | |
1554 | */ | |
1555 | spin = !ctx->poll_multi_file && *nr_events < min; | |
1556 | ||
1557 | ret = 0; | |
1558 | list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) { | |
9adbd45d | 1559 | struct kiocb *kiocb = &req->rw.kiocb; |
def596e9 JA |
1560 | |
1561 | /* | |
1562 | * Move completed entries to our local list. If we find a | |
1563 | * request that requires polling, break out and complete | |
1564 | * the done list first, if we have entries there. | |
1565 | */ | |
1566 | if (req->flags & REQ_F_IOPOLL_COMPLETED) { | |
1567 | list_move_tail(&req->list, &done); | |
1568 | continue; | |
1569 | } | |
1570 | if (!list_empty(&done)) | |
1571 | break; | |
1572 | ||
1573 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
1574 | if (ret < 0) | |
1575 | break; | |
1576 | ||
1577 | if (ret && spin) | |
1578 | spin = false; | |
1579 | ret = 0; | |
1580 | } | |
1581 | ||
1582 | if (!list_empty(&done)) | |
1583 | io_iopoll_complete(ctx, nr_events, &done); | |
1584 | ||
1585 | return ret; | |
1586 | } | |
1587 | ||
1588 | /* | |
d195a66e | 1589 | * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a |
def596e9 JA |
1590 | * non-spinning poll check - we'll still enter the driver poll loop, but only |
1591 | * as a non-spinning completion check. | |
1592 | */ | |
1593 | static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1594 | long min) | |
1595 | { | |
08f5439f | 1596 | while (!list_empty(&ctx->poll_list) && !need_resched()) { |
def596e9 JA |
1597 | int ret; |
1598 | ||
1599 | ret = io_do_iopoll(ctx, nr_events, min); | |
1600 | if (ret < 0) | |
1601 | return ret; | |
1602 | if (!min || *nr_events >= min) | |
1603 | return 0; | |
1604 | } | |
1605 | ||
1606 | return 1; | |
1607 | } | |
1608 | ||
1609 | /* | |
1610 | * We can't just wait for polled events to come to us, we have to actively | |
1611 | * find and complete them. | |
1612 | */ | |
1613 | static void io_iopoll_reap_events(struct io_ring_ctx *ctx) | |
1614 | { | |
1615 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
1616 | return; | |
1617 | ||
1618 | mutex_lock(&ctx->uring_lock); | |
1619 | while (!list_empty(&ctx->poll_list)) { | |
1620 | unsigned int nr_events = 0; | |
1621 | ||
1622 | io_iopoll_getevents(ctx, &nr_events, 1); | |
08f5439f JA |
1623 | |
1624 | /* | |
1625 | * Ensure we allow local-to-the-cpu processing to take place, | |
1626 | * in this case we need to ensure that we reap all events. | |
1627 | */ | |
1628 | cond_resched(); | |
def596e9 JA |
1629 | } |
1630 | mutex_unlock(&ctx->uring_lock); | |
1631 | } | |
1632 | ||
2b2ed975 JA |
1633 | static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, |
1634 | long min) | |
def596e9 | 1635 | { |
2b2ed975 | 1636 | int iters = 0, ret = 0; |
500f9fba | 1637 | |
def596e9 JA |
1638 | do { |
1639 | int tmin = 0; | |
1640 | ||
a3a0e43f JA |
1641 | /* |
1642 | * Don't enter poll loop if we already have events pending. | |
1643 | * If we do, we can potentially be spinning for commands that | |
1644 | * already triggered a CQE (eg in error). | |
1645 | */ | |
1d7bb1d5 | 1646 | if (io_cqring_events(ctx, false)) |
a3a0e43f JA |
1647 | break; |
1648 | ||
500f9fba JA |
1649 | /* |
1650 | * If a submit got punted to a workqueue, we can have the | |
1651 | * application entering polling for a command before it gets | |
1652 | * issued. That app will hold the uring_lock for the duration | |
1653 | * of the poll right here, so we need to take a breather every | |
1654 | * now and then to ensure that the issue has a chance to add | |
1655 | * the poll to the issued list. Otherwise we can spin here | |
1656 | * forever, while the workqueue is stuck trying to acquire the | |
1657 | * very same mutex. | |
1658 | */ | |
1659 | if (!(++iters & 7)) { | |
1660 | mutex_unlock(&ctx->uring_lock); | |
1661 | mutex_lock(&ctx->uring_lock); | |
1662 | } | |
1663 | ||
def596e9 JA |
1664 | if (*nr_events < min) |
1665 | tmin = min - *nr_events; | |
1666 | ||
1667 | ret = io_iopoll_getevents(ctx, nr_events, tmin); | |
1668 | if (ret <= 0) | |
1669 | break; | |
1670 | ret = 0; | |
1671 | } while (min && !*nr_events && !need_resched()); | |
1672 | ||
2b2ed975 JA |
1673 | return ret; |
1674 | } | |
1675 | ||
1676 | static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, | |
1677 | long min) | |
1678 | { | |
1679 | int ret; | |
1680 | ||
1681 | /* | |
1682 | * We disallow the app entering submit/complete with polling, but we | |
1683 | * still need to lock the ring to prevent racing with polled issue | |
1684 | * that got punted to a workqueue. | |
1685 | */ | |
1686 | mutex_lock(&ctx->uring_lock); | |
1687 | ret = __io_iopoll_check(ctx, nr_events, min); | |
500f9fba | 1688 | mutex_unlock(&ctx->uring_lock); |
def596e9 JA |
1689 | return ret; |
1690 | } | |
1691 | ||
491381ce | 1692 | static void kiocb_end_write(struct io_kiocb *req) |
2b188cc1 | 1693 | { |
491381ce JA |
1694 | /* |
1695 | * Tell lockdep we inherited freeze protection from submission | |
1696 | * thread. | |
1697 | */ | |
1698 | if (req->flags & REQ_F_ISREG) { | |
1699 | struct inode *inode = file_inode(req->file); | |
2b188cc1 | 1700 | |
491381ce | 1701 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); |
2b188cc1 | 1702 | } |
491381ce | 1703 | file_end_write(req->file); |
2b188cc1 JA |
1704 | } |
1705 | ||
4e88d6e7 JA |
1706 | static inline void req_set_fail_links(struct io_kiocb *req) |
1707 | { | |
1708 | if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK) | |
1709 | req->flags |= REQ_F_FAIL_LINK; | |
1710 | } | |
1711 | ||
ba816ad6 | 1712 | static void io_complete_rw_common(struct kiocb *kiocb, long res) |
2b188cc1 | 1713 | { |
9adbd45d | 1714 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
2b188cc1 | 1715 | |
491381ce JA |
1716 | if (kiocb->ki_flags & IOCB_WRITE) |
1717 | kiocb_end_write(req); | |
2b188cc1 | 1718 | |
4e88d6e7 JA |
1719 | if (res != req->result) |
1720 | req_set_fail_links(req); | |
78e19bbe | 1721 | io_cqring_add_event(req, res); |
ba816ad6 JA |
1722 | } |
1723 | ||
1724 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
1725 | { | |
9adbd45d | 1726 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
ba816ad6 JA |
1727 | |
1728 | io_complete_rw_common(kiocb, res); | |
e65ef56d | 1729 | io_put_req(req); |
2b188cc1 JA |
1730 | } |
1731 | ||
ba816ad6 JA |
1732 | static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res) |
1733 | { | |
9adbd45d | 1734 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
ec9c02ad | 1735 | struct io_kiocb *nxt = NULL; |
ba816ad6 JA |
1736 | |
1737 | io_complete_rw_common(kiocb, res); | |
ec9c02ad JL |
1738 | io_put_req_find_next(req, &nxt); |
1739 | ||
1740 | return nxt; | |
2b188cc1 JA |
1741 | } |
1742 | ||
def596e9 JA |
1743 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
1744 | { | |
9adbd45d | 1745 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
def596e9 | 1746 | |
491381ce JA |
1747 | if (kiocb->ki_flags & IOCB_WRITE) |
1748 | kiocb_end_write(req); | |
def596e9 | 1749 | |
4e88d6e7 JA |
1750 | if (res != req->result) |
1751 | req_set_fail_links(req); | |
9e645e11 | 1752 | req->result = res; |
def596e9 JA |
1753 | if (res != -EAGAIN) |
1754 | req->flags |= REQ_F_IOPOLL_COMPLETED; | |
1755 | } | |
1756 | ||
1757 | /* | |
1758 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1759 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
1760 | * find it from a io_iopoll_getevents() thread before the issuer is done | |
1761 | * accessing the kiocb cookie. | |
1762 | */ | |
1763 | static void io_iopoll_req_issued(struct io_kiocb *req) | |
1764 | { | |
1765 | struct io_ring_ctx *ctx = req->ctx; | |
1766 | ||
1767 | /* | |
1768 | * Track whether we have multiple files in our lists. This will impact | |
1769 | * how we do polling eventually, not spinning if we're on potentially | |
1770 | * different devices. | |
1771 | */ | |
1772 | if (list_empty(&ctx->poll_list)) { | |
1773 | ctx->poll_multi_file = false; | |
1774 | } else if (!ctx->poll_multi_file) { | |
1775 | struct io_kiocb *list_req; | |
1776 | ||
1777 | list_req = list_first_entry(&ctx->poll_list, struct io_kiocb, | |
1778 | list); | |
9adbd45d | 1779 | if (list_req->file != req->file) |
def596e9 JA |
1780 | ctx->poll_multi_file = true; |
1781 | } | |
1782 | ||
1783 | /* | |
1784 | * For fast devices, IO may have already completed. If it has, add | |
1785 | * it to the front so we find it first. | |
1786 | */ | |
1787 | if (req->flags & REQ_F_IOPOLL_COMPLETED) | |
1788 | list_add(&req->list, &ctx->poll_list); | |
1789 | else | |
1790 | list_add_tail(&req->list, &ctx->poll_list); | |
1791 | } | |
1792 | ||
3d6770fb | 1793 | static void io_file_put(struct io_submit_state *state) |
9a56a232 | 1794 | { |
3d6770fb | 1795 | if (state->file) { |
9a56a232 JA |
1796 | int diff = state->has_refs - state->used_refs; |
1797 | ||
1798 | if (diff) | |
1799 | fput_many(state->file, diff); | |
1800 | state->file = NULL; | |
1801 | } | |
1802 | } | |
1803 | ||
1804 | /* | |
1805 | * Get as many references to a file as we have IOs left in this submission, | |
1806 | * assuming most submissions are for one file, or at least that each file | |
1807 | * has more than one submission. | |
1808 | */ | |
1809 | static struct file *io_file_get(struct io_submit_state *state, int fd) | |
1810 | { | |
1811 | if (!state) | |
1812 | return fget(fd); | |
1813 | ||
1814 | if (state->file) { | |
1815 | if (state->fd == fd) { | |
1816 | state->used_refs++; | |
1817 | state->ios_left--; | |
1818 | return state->file; | |
1819 | } | |
3d6770fb | 1820 | io_file_put(state); |
9a56a232 JA |
1821 | } |
1822 | state->file = fget_many(fd, state->ios_left); | |
1823 | if (!state->file) | |
1824 | return NULL; | |
1825 | ||
1826 | state->fd = fd; | |
1827 | state->has_refs = state->ios_left; | |
1828 | state->used_refs = 1; | |
1829 | state->ios_left--; | |
1830 | return state->file; | |
1831 | } | |
1832 | ||
2b188cc1 JA |
1833 | /* |
1834 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1835 | * any file. For now, just ensure that anything potentially problematic is done | |
1836 | * inline. | |
1837 | */ | |
1838 | static bool io_file_supports_async(struct file *file) | |
1839 | { | |
1840 | umode_t mode = file_inode(file)->i_mode; | |
1841 | ||
10d59345 | 1842 | if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode)) |
2b188cc1 JA |
1843 | return true; |
1844 | if (S_ISREG(mode) && file->f_op != &io_uring_fops) | |
1845 | return true; | |
1846 | ||
1847 | return false; | |
1848 | } | |
1849 | ||
3529d8c2 JA |
1850 | static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
1851 | bool force_nonblock) | |
2b188cc1 | 1852 | { |
def596e9 | 1853 | struct io_ring_ctx *ctx = req->ctx; |
9adbd45d | 1854 | struct kiocb *kiocb = &req->rw.kiocb; |
09bb8394 JA |
1855 | unsigned ioprio; |
1856 | int ret; | |
2b188cc1 | 1857 | |
09bb8394 JA |
1858 | if (!req->file) |
1859 | return -EBADF; | |
2b188cc1 | 1860 | |
491381ce JA |
1861 | if (S_ISREG(file_inode(req->file)->i_mode)) |
1862 | req->flags |= REQ_F_ISREG; | |
1863 | ||
2b188cc1 | 1864 | kiocb->ki_pos = READ_ONCE(sqe->off); |
ba04291e JA |
1865 | if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) { |
1866 | req->flags |= REQ_F_CUR_POS; | |
1867 | kiocb->ki_pos = req->file->f_pos; | |
1868 | } | |
2b188cc1 JA |
1869 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); |
1870 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); | |
1871 | ||
1872 | ioprio = READ_ONCE(sqe->ioprio); | |
1873 | if (ioprio) { | |
1874 | ret = ioprio_check_cap(ioprio); | |
1875 | if (ret) | |
09bb8394 | 1876 | return ret; |
2b188cc1 JA |
1877 | |
1878 | kiocb->ki_ioprio = ioprio; | |
1879 | } else | |
1880 | kiocb->ki_ioprio = get_current_ioprio(); | |
1881 | ||
1882 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
1883 | if (unlikely(ret)) | |
09bb8394 | 1884 | return ret; |
8449eeda SB |
1885 | |
1886 | /* don't allow async punt if RWF_NOWAIT was requested */ | |
491381ce JA |
1887 | if ((kiocb->ki_flags & IOCB_NOWAIT) || |
1888 | (req->file->f_flags & O_NONBLOCK)) | |
8449eeda SB |
1889 | req->flags |= REQ_F_NOWAIT; |
1890 | ||
1891 | if (force_nonblock) | |
2b188cc1 | 1892 | kiocb->ki_flags |= IOCB_NOWAIT; |
8449eeda | 1893 | |
def596e9 | 1894 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
1895 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
1896 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 1897 | return -EOPNOTSUPP; |
2b188cc1 | 1898 | |
def596e9 JA |
1899 | kiocb->ki_flags |= IOCB_HIPRI; |
1900 | kiocb->ki_complete = io_complete_rw_iopoll; | |
6873e0bd | 1901 | req->result = 0; |
def596e9 | 1902 | } else { |
09bb8394 JA |
1903 | if (kiocb->ki_flags & IOCB_HIPRI) |
1904 | return -EINVAL; | |
def596e9 JA |
1905 | kiocb->ki_complete = io_complete_rw; |
1906 | } | |
9adbd45d | 1907 | |
3529d8c2 JA |
1908 | req->rw.addr = READ_ONCE(sqe->addr); |
1909 | req->rw.len = READ_ONCE(sqe->len); | |
9adbd45d JA |
1910 | /* we own ->private, reuse it for the buffer index */ |
1911 | req->rw.kiocb.private = (void *) (unsigned long) | |
3529d8c2 | 1912 | READ_ONCE(sqe->buf_index); |
2b188cc1 | 1913 | return 0; |
2b188cc1 JA |
1914 | } |
1915 | ||
1916 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
1917 | { | |
1918 | switch (ret) { | |
1919 | case -EIOCBQUEUED: | |
1920 | break; | |
1921 | case -ERESTARTSYS: | |
1922 | case -ERESTARTNOINTR: | |
1923 | case -ERESTARTNOHAND: | |
1924 | case -ERESTART_RESTARTBLOCK: | |
1925 | /* | |
1926 | * We can't just restart the syscall, since previously | |
1927 | * submitted sqes may already be in progress. Just fail this | |
1928 | * IO with EINTR. | |
1929 | */ | |
1930 | ret = -EINTR; | |
1931 | /* fall through */ | |
1932 | default: | |
1933 | kiocb->ki_complete(kiocb, ret, 0); | |
1934 | } | |
1935 | } | |
1936 | ||
ba816ad6 JA |
1937 | static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt, |
1938 | bool in_async) | |
1939 | { | |
ba04291e JA |
1940 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
1941 | ||
1942 | if (req->flags & REQ_F_CUR_POS) | |
1943 | req->file->f_pos = kiocb->ki_pos; | |
f9bd67f6 | 1944 | if (in_async && ret >= 0 && kiocb->ki_complete == io_complete_rw) |
ba816ad6 JA |
1945 | *nxt = __io_complete_rw(kiocb, ret); |
1946 | else | |
1947 | io_rw_done(kiocb, ret); | |
1948 | } | |
1949 | ||
9adbd45d | 1950 | static ssize_t io_import_fixed(struct io_kiocb *req, int rw, |
7d009165 | 1951 | struct iov_iter *iter) |
edafccee | 1952 | { |
9adbd45d JA |
1953 | struct io_ring_ctx *ctx = req->ctx; |
1954 | size_t len = req->rw.len; | |
edafccee JA |
1955 | struct io_mapped_ubuf *imu; |
1956 | unsigned index, buf_index; | |
1957 | size_t offset; | |
1958 | u64 buf_addr; | |
1959 | ||
1960 | /* attempt to use fixed buffers without having provided iovecs */ | |
1961 | if (unlikely(!ctx->user_bufs)) | |
1962 | return -EFAULT; | |
1963 | ||
9adbd45d | 1964 | buf_index = (unsigned long) req->rw.kiocb.private; |
edafccee JA |
1965 | if (unlikely(buf_index >= ctx->nr_user_bufs)) |
1966 | return -EFAULT; | |
1967 | ||
1968 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); | |
1969 | imu = &ctx->user_bufs[index]; | |
9adbd45d | 1970 | buf_addr = req->rw.addr; |
edafccee JA |
1971 | |
1972 | /* overflow */ | |
1973 | if (buf_addr + len < buf_addr) | |
1974 | return -EFAULT; | |
1975 | /* not inside the mapped region */ | |
1976 | if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len) | |
1977 | return -EFAULT; | |
1978 | ||
1979 | /* | |
1980 | * May not be a start of buffer, set size appropriately | |
1981 | * and advance us to the beginning. | |
1982 | */ | |
1983 | offset = buf_addr - imu->ubuf; | |
1984 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
bd11b3a3 JA |
1985 | |
1986 | if (offset) { | |
1987 | /* | |
1988 | * Don't use iov_iter_advance() here, as it's really slow for | |
1989 | * using the latter parts of a big fixed buffer - it iterates | |
1990 | * over each segment manually. We can cheat a bit here, because | |
1991 | * we know that: | |
1992 | * | |
1993 | * 1) it's a BVEC iter, we set it up | |
1994 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
1995 | * first and last bvec | |
1996 | * | |
1997 | * So just find our index, and adjust the iterator afterwards. | |
1998 | * If the offset is within the first bvec (or the whole first | |
1999 | * bvec, just use iov_iter_advance(). This makes it easier | |
2000 | * since we can just skip the first segment, which may not | |
2001 | * be PAGE_SIZE aligned. | |
2002 | */ | |
2003 | const struct bio_vec *bvec = imu->bvec; | |
2004 | ||
2005 | if (offset <= bvec->bv_len) { | |
2006 | iov_iter_advance(iter, offset); | |
2007 | } else { | |
2008 | unsigned long seg_skip; | |
2009 | ||
2010 | /* skip first vec */ | |
2011 | offset -= bvec->bv_len; | |
2012 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
2013 | ||
2014 | iter->bvec = bvec + seg_skip; | |
2015 | iter->nr_segs -= seg_skip; | |
99c79f66 | 2016 | iter->count -= bvec->bv_len + offset; |
bd11b3a3 | 2017 | iter->iov_offset = offset & ~PAGE_MASK; |
bd11b3a3 JA |
2018 | } |
2019 | } | |
2020 | ||
5e559561 | 2021 | return len; |
edafccee JA |
2022 | } |
2023 | ||
cf6fd4bd PB |
2024 | static ssize_t io_import_iovec(int rw, struct io_kiocb *req, |
2025 | struct iovec **iovec, struct iov_iter *iter) | |
2b188cc1 | 2026 | { |
9adbd45d JA |
2027 | void __user *buf = u64_to_user_ptr(req->rw.addr); |
2028 | size_t sqe_len = req->rw.len; | |
edafccee JA |
2029 | u8 opcode; |
2030 | ||
d625c6ee | 2031 | opcode = req->opcode; |
7d009165 | 2032 | if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) { |
edafccee | 2033 | *iovec = NULL; |
9adbd45d | 2034 | return io_import_fixed(req, rw, iter); |
edafccee | 2035 | } |
2b188cc1 | 2036 | |
9adbd45d JA |
2037 | /* buffer index only valid with fixed read/write */ |
2038 | if (req->rw.kiocb.private) | |
2039 | return -EINVAL; | |
2040 | ||
3a6820f2 JA |
2041 | if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) { |
2042 | ssize_t ret; | |
2043 | ret = import_single_range(rw, buf, sqe_len, *iovec, iter); | |
2044 | *iovec = NULL; | |
2045 | return ret; | |
2046 | } | |
2047 | ||
f67676d1 JA |
2048 | if (req->io) { |
2049 | struct io_async_rw *iorw = &req->io->rw; | |
2050 | ||
2051 | *iovec = iorw->iov; | |
2052 | iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size); | |
2053 | if (iorw->iov == iorw->fast_iov) | |
2054 | *iovec = NULL; | |
2055 | return iorw->size; | |
2056 | } | |
2057 | ||
cf6fd4bd | 2058 | if (!req->has_user) |
2b188cc1 JA |
2059 | return -EFAULT; |
2060 | ||
2061 | #ifdef CONFIG_COMPAT | |
cf6fd4bd | 2062 | if (req->ctx->compat) |
2b188cc1 JA |
2063 | return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV, |
2064 | iovec, iter); | |
2065 | #endif | |
2066 | ||
2067 | return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter); | |
2068 | } | |
2069 | ||
31b51510 | 2070 | /* |
32960613 JA |
2071 | * For files that don't have ->read_iter() and ->write_iter(), handle them |
2072 | * by looping over ->read() or ->write() manually. | |
31b51510 | 2073 | */ |
32960613 JA |
2074 | static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb, |
2075 | struct iov_iter *iter) | |
2076 | { | |
2077 | ssize_t ret = 0; | |
2078 | ||
2079 | /* | |
2080 | * Don't support polled IO through this interface, and we can't | |
2081 | * support non-blocking either. For the latter, this just causes | |
2082 | * the kiocb to be handled from an async context. | |
2083 | */ | |
2084 | if (kiocb->ki_flags & IOCB_HIPRI) | |
2085 | return -EOPNOTSUPP; | |
2086 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
2087 | return -EAGAIN; | |
2088 | ||
2089 | while (iov_iter_count(iter)) { | |
311ae9e1 | 2090 | struct iovec iovec; |
32960613 JA |
2091 | ssize_t nr; |
2092 | ||
311ae9e1 PB |
2093 | if (!iov_iter_is_bvec(iter)) { |
2094 | iovec = iov_iter_iovec(iter); | |
2095 | } else { | |
2096 | /* fixed buffers import bvec */ | |
2097 | iovec.iov_base = kmap(iter->bvec->bv_page) | |
2098 | + iter->iov_offset; | |
2099 | iovec.iov_len = min(iter->count, | |
2100 | iter->bvec->bv_len - iter->iov_offset); | |
2101 | } | |
2102 | ||
32960613 JA |
2103 | if (rw == READ) { |
2104 | nr = file->f_op->read(file, iovec.iov_base, | |
2105 | iovec.iov_len, &kiocb->ki_pos); | |
2106 | } else { | |
2107 | nr = file->f_op->write(file, iovec.iov_base, | |
2108 | iovec.iov_len, &kiocb->ki_pos); | |
2109 | } | |
2110 | ||
311ae9e1 PB |
2111 | if (iov_iter_is_bvec(iter)) |
2112 | kunmap(iter->bvec->bv_page); | |
2113 | ||
32960613 JA |
2114 | if (nr < 0) { |
2115 | if (!ret) | |
2116 | ret = nr; | |
2117 | break; | |
2118 | } | |
2119 | ret += nr; | |
2120 | if (nr != iovec.iov_len) | |
2121 | break; | |
2122 | iov_iter_advance(iter, nr); | |
2123 | } | |
2124 | ||
2125 | return ret; | |
2126 | } | |
2127 | ||
b7bb4f7d | 2128 | static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size, |
f67676d1 JA |
2129 | struct iovec *iovec, struct iovec *fast_iov, |
2130 | struct iov_iter *iter) | |
2131 | { | |
2132 | req->io->rw.nr_segs = iter->nr_segs; | |
2133 | req->io->rw.size = io_size; | |
2134 | req->io->rw.iov = iovec; | |
2135 | if (!req->io->rw.iov) { | |
2136 | req->io->rw.iov = req->io->rw.fast_iov; | |
2137 | memcpy(req->io->rw.iov, fast_iov, | |
2138 | sizeof(struct iovec) * iter->nr_segs); | |
2139 | } | |
2140 | } | |
2141 | ||
b7bb4f7d | 2142 | static int io_alloc_async_ctx(struct io_kiocb *req) |
f67676d1 | 2143 | { |
d3656344 JA |
2144 | if (!io_op_defs[req->opcode].async_ctx) |
2145 | return 0; | |
f67676d1 | 2146 | req->io = kmalloc(sizeof(*req->io), GFP_KERNEL); |
06b76d44 | 2147 | return req->io == NULL; |
b7bb4f7d JA |
2148 | } |
2149 | ||
2150 | static void io_rw_async(struct io_wq_work **workptr) | |
2151 | { | |
2152 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
2153 | struct iovec *iov = NULL; | |
2154 | ||
2155 | if (req->io->rw.iov != req->io->rw.fast_iov) | |
2156 | iov = req->io->rw.iov; | |
2157 | io_wq_submit_work(workptr); | |
2158 | kfree(iov); | |
2159 | } | |
2160 | ||
2161 | static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size, | |
2162 | struct iovec *iovec, struct iovec *fast_iov, | |
2163 | struct iov_iter *iter) | |
2164 | { | |
980ad263 | 2165 | if (!io_op_defs[req->opcode].async_ctx) |
74566df3 | 2166 | return 0; |
b7bb4f7d JA |
2167 | if (!req->io && io_alloc_async_ctx(req)) |
2168 | return -ENOMEM; | |
2169 | ||
2170 | io_req_map_rw(req, io_size, iovec, fast_iov, iter); | |
2171 | req->work.func = io_rw_async; | |
2172 | return 0; | |
f67676d1 JA |
2173 | } |
2174 | ||
3529d8c2 JA |
2175 | static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2176 | bool force_nonblock) | |
f67676d1 | 2177 | { |
3529d8c2 JA |
2178 | struct io_async_ctx *io; |
2179 | struct iov_iter iter; | |
f67676d1 JA |
2180 | ssize_t ret; |
2181 | ||
3529d8c2 JA |
2182 | ret = io_prep_rw(req, sqe, force_nonblock); |
2183 | if (ret) | |
2184 | return ret; | |
f67676d1 | 2185 | |
3529d8c2 JA |
2186 | if (unlikely(!(req->file->f_mode & FMODE_READ))) |
2187 | return -EBADF; | |
f67676d1 | 2188 | |
3529d8c2 JA |
2189 | if (!req->io) |
2190 | return 0; | |
2191 | ||
2192 | io = req->io; | |
2193 | io->rw.iov = io->rw.fast_iov; | |
2194 | req->io = NULL; | |
2195 | ret = io_import_iovec(READ, req, &io->rw.iov, &iter); | |
2196 | req->io = io; | |
2197 | if (ret < 0) | |
2198 | return ret; | |
2199 | ||
2200 | io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter); | |
2201 | return 0; | |
f67676d1 JA |
2202 | } |
2203 | ||
267bc904 | 2204 | static int io_read(struct io_kiocb *req, struct io_kiocb **nxt, |
8358e3a8 | 2205 | bool force_nonblock) |
2b188cc1 JA |
2206 | { |
2207 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
9adbd45d | 2208 | struct kiocb *kiocb = &req->rw.kiocb; |
2b188cc1 | 2209 | struct iov_iter iter; |
31b51510 | 2210 | size_t iov_count; |
f67676d1 | 2211 | ssize_t io_size, ret; |
2b188cc1 | 2212 | |
3529d8c2 | 2213 | ret = io_import_iovec(READ, req, &iovec, &iter); |
06b76d44 JA |
2214 | if (ret < 0) |
2215 | return ret; | |
2b188cc1 | 2216 | |
fd6c2e4c JA |
2217 | /* Ensure we clear previously set non-block flag */ |
2218 | if (!force_nonblock) | |
9adbd45d | 2219 | req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT; |
fd6c2e4c | 2220 | |
797f3f53 | 2221 | req->result = 0; |
f67676d1 | 2222 | io_size = ret; |
9e645e11 | 2223 | if (req->flags & REQ_F_LINK) |
f67676d1 JA |
2224 | req->result = io_size; |
2225 | ||
2226 | /* | |
2227 | * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so | |
2228 | * we know to async punt it even if it was opened O_NONBLOCK | |
2229 | */ | |
9adbd45d | 2230 | if (force_nonblock && !io_file_supports_async(req->file)) { |
f67676d1 JA |
2231 | req->flags |= REQ_F_MUST_PUNT; |
2232 | goto copy_iov; | |
2233 | } | |
9e645e11 | 2234 | |
31b51510 | 2235 | iov_count = iov_iter_count(&iter); |
9adbd45d | 2236 | ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count); |
2b188cc1 JA |
2237 | if (!ret) { |
2238 | ssize_t ret2; | |
2239 | ||
9adbd45d JA |
2240 | if (req->file->f_op->read_iter) |
2241 | ret2 = call_read_iter(req->file, kiocb, &iter); | |
32960613 | 2242 | else |
9adbd45d | 2243 | ret2 = loop_rw_iter(READ, req->file, kiocb, &iter); |
32960613 | 2244 | |
9d93a3f5 | 2245 | /* Catch -EAGAIN return for forced non-blocking submission */ |
f67676d1 | 2246 | if (!force_nonblock || ret2 != -EAGAIN) { |
cf6fd4bd | 2247 | kiocb_done(kiocb, ret2, nxt, req->in_async); |
f67676d1 JA |
2248 | } else { |
2249 | copy_iov: | |
b7bb4f7d | 2250 | ret = io_setup_async_rw(req, io_size, iovec, |
f67676d1 JA |
2251 | inline_vecs, &iter); |
2252 | if (ret) | |
2253 | goto out_free; | |
2254 | return -EAGAIN; | |
2255 | } | |
2b188cc1 | 2256 | } |
f67676d1 | 2257 | out_free: |
b7bb4f7d JA |
2258 | if (!io_wq_current_is_worker()) |
2259 | kfree(iovec); | |
2b188cc1 JA |
2260 | return ret; |
2261 | } | |
2262 | ||
3529d8c2 JA |
2263 | static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2264 | bool force_nonblock) | |
f67676d1 | 2265 | { |
3529d8c2 JA |
2266 | struct io_async_ctx *io; |
2267 | struct iov_iter iter; | |
f67676d1 JA |
2268 | ssize_t ret; |
2269 | ||
3529d8c2 JA |
2270 | ret = io_prep_rw(req, sqe, force_nonblock); |
2271 | if (ret) | |
2272 | return ret; | |
f67676d1 | 2273 | |
3529d8c2 JA |
2274 | if (unlikely(!(req->file->f_mode & FMODE_WRITE))) |
2275 | return -EBADF; | |
f67676d1 | 2276 | |
3529d8c2 JA |
2277 | if (!req->io) |
2278 | return 0; | |
2279 | ||
2280 | io = req->io; | |
2281 | io->rw.iov = io->rw.fast_iov; | |
2282 | req->io = NULL; | |
2283 | ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter); | |
2284 | req->io = io; | |
2285 | if (ret < 0) | |
2286 | return ret; | |
2287 | ||
2288 | io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter); | |
2289 | return 0; | |
f67676d1 JA |
2290 | } |
2291 | ||
267bc904 | 2292 | static int io_write(struct io_kiocb *req, struct io_kiocb **nxt, |
8358e3a8 | 2293 | bool force_nonblock) |
2b188cc1 JA |
2294 | { |
2295 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
9adbd45d | 2296 | struct kiocb *kiocb = &req->rw.kiocb; |
2b188cc1 | 2297 | struct iov_iter iter; |
31b51510 | 2298 | size_t iov_count; |
f67676d1 | 2299 | ssize_t ret, io_size; |
2b188cc1 | 2300 | |
3529d8c2 | 2301 | ret = io_import_iovec(WRITE, req, &iovec, &iter); |
06b76d44 JA |
2302 | if (ret < 0) |
2303 | return ret; | |
2b188cc1 | 2304 | |
fd6c2e4c JA |
2305 | /* Ensure we clear previously set non-block flag */ |
2306 | if (!force_nonblock) | |
9adbd45d | 2307 | req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT; |
fd6c2e4c | 2308 | |
797f3f53 | 2309 | req->result = 0; |
f67676d1 | 2310 | io_size = ret; |
9e645e11 | 2311 | if (req->flags & REQ_F_LINK) |
f67676d1 | 2312 | req->result = io_size; |
9e645e11 | 2313 | |
f67676d1 JA |
2314 | /* |
2315 | * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so | |
2316 | * we know to async punt it even if it was opened O_NONBLOCK | |
2317 | */ | |
2318 | if (force_nonblock && !io_file_supports_async(req->file)) { | |
2319 | req->flags |= REQ_F_MUST_PUNT; | |
2320 | goto copy_iov; | |
2321 | } | |
31b51510 | 2322 | |
10d59345 JA |
2323 | /* file path doesn't support NOWAIT for non-direct_IO */ |
2324 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) && | |
2325 | (req->flags & REQ_F_ISREG)) | |
f67676d1 | 2326 | goto copy_iov; |
31b51510 | 2327 | |
f67676d1 | 2328 | iov_count = iov_iter_count(&iter); |
9adbd45d | 2329 | ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count); |
2b188cc1 | 2330 | if (!ret) { |
9bf7933f RP |
2331 | ssize_t ret2; |
2332 | ||
2b188cc1 JA |
2333 | /* |
2334 | * Open-code file_start_write here to grab freeze protection, | |
2335 | * which will be released by another thread in | |
2336 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
2337 | * released so that it doesn't complain about the held lock when | |
2338 | * we return to userspace. | |
2339 | */ | |
491381ce | 2340 | if (req->flags & REQ_F_ISREG) { |
9adbd45d | 2341 | __sb_start_write(file_inode(req->file)->i_sb, |
2b188cc1 | 2342 | SB_FREEZE_WRITE, true); |
9adbd45d | 2343 | __sb_writers_release(file_inode(req->file)->i_sb, |
2b188cc1 JA |
2344 | SB_FREEZE_WRITE); |
2345 | } | |
2346 | kiocb->ki_flags |= IOCB_WRITE; | |
9bf7933f | 2347 | |
9adbd45d JA |
2348 | if (req->file->f_op->write_iter) |
2349 | ret2 = call_write_iter(req->file, kiocb, &iter); | |
32960613 | 2350 | else |
9adbd45d | 2351 | ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter); |
f67676d1 | 2352 | if (!force_nonblock || ret2 != -EAGAIN) { |
cf6fd4bd | 2353 | kiocb_done(kiocb, ret2, nxt, req->in_async); |
f67676d1 JA |
2354 | } else { |
2355 | copy_iov: | |
b7bb4f7d | 2356 | ret = io_setup_async_rw(req, io_size, iovec, |
f67676d1 JA |
2357 | inline_vecs, &iter); |
2358 | if (ret) | |
2359 | goto out_free; | |
2360 | return -EAGAIN; | |
2361 | } | |
2b188cc1 | 2362 | } |
31b51510 | 2363 | out_free: |
b7bb4f7d JA |
2364 | if (!io_wq_current_is_worker()) |
2365 | kfree(iovec); | |
2b188cc1 JA |
2366 | return ret; |
2367 | } | |
2368 | ||
2369 | /* | |
2370 | * IORING_OP_NOP just posts a completion event, nothing else. | |
2371 | */ | |
78e19bbe | 2372 | static int io_nop(struct io_kiocb *req) |
2b188cc1 JA |
2373 | { |
2374 | struct io_ring_ctx *ctx = req->ctx; | |
2b188cc1 | 2375 | |
def596e9 JA |
2376 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
2377 | return -EINVAL; | |
2378 | ||
78e19bbe | 2379 | io_cqring_add_event(req, 0); |
e65ef56d | 2380 | io_put_req(req); |
2b188cc1 JA |
2381 | return 0; |
2382 | } | |
2383 | ||
3529d8c2 | 2384 | static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
c992fe29 | 2385 | { |
6b06314c | 2386 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 2387 | |
09bb8394 JA |
2388 | if (!req->file) |
2389 | return -EBADF; | |
c992fe29 | 2390 | |
6b06314c | 2391 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 2392 | return -EINVAL; |
edafccee | 2393 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
2394 | return -EINVAL; |
2395 | ||
8ed8d3c3 JA |
2396 | req->sync.flags = READ_ONCE(sqe->fsync_flags); |
2397 | if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC)) | |
2398 | return -EINVAL; | |
2399 | ||
2400 | req->sync.off = READ_ONCE(sqe->off); | |
2401 | req->sync.len = READ_ONCE(sqe->len); | |
c992fe29 CH |
2402 | return 0; |
2403 | } | |
2404 | ||
8ed8d3c3 JA |
2405 | static bool io_req_cancelled(struct io_kiocb *req) |
2406 | { | |
2407 | if (req->work.flags & IO_WQ_WORK_CANCEL) { | |
2408 | req_set_fail_links(req); | |
2409 | io_cqring_add_event(req, -ECANCELED); | |
2410 | io_put_req(req); | |
2411 | return true; | |
2412 | } | |
2413 | ||
2414 | return false; | |
2415 | } | |
2416 | ||
78912934 JA |
2417 | static void io_link_work_cb(struct io_wq_work **workptr) |
2418 | { | |
2419 | struct io_wq_work *work = *workptr; | |
2420 | struct io_kiocb *link = work->data; | |
2421 | ||
2422 | io_queue_linked_timeout(link); | |
2423 | work->func = io_wq_submit_work; | |
2424 | } | |
2425 | ||
2426 | static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt) | |
2427 | { | |
2428 | struct io_kiocb *link; | |
2429 | ||
2430 | io_prep_async_work(nxt, &link); | |
2431 | *workptr = &nxt->work; | |
2432 | if (link) { | |
2433 | nxt->work.flags |= IO_WQ_WORK_CB; | |
2434 | nxt->work.func = io_link_work_cb; | |
2435 | nxt->work.data = link; | |
2436 | } | |
2437 | } | |
2438 | ||
8ed8d3c3 JA |
2439 | static void io_fsync_finish(struct io_wq_work **workptr) |
2440 | { | |
2441 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
2442 | loff_t end = req->sync.off + req->sync.len; | |
2443 | struct io_kiocb *nxt = NULL; | |
2444 | int ret; | |
2445 | ||
2446 | if (io_req_cancelled(req)) | |
2447 | return; | |
2448 | ||
9adbd45d | 2449 | ret = vfs_fsync_range(req->file, req->sync.off, |
8ed8d3c3 JA |
2450 | end > 0 ? end : LLONG_MAX, |
2451 | req->sync.flags & IORING_FSYNC_DATASYNC); | |
2452 | if (ret < 0) | |
2453 | req_set_fail_links(req); | |
2454 | io_cqring_add_event(req, ret); | |
2455 | io_put_req_find_next(req, &nxt); | |
2456 | if (nxt) | |
78912934 | 2457 | io_wq_assign_next(workptr, nxt); |
8ed8d3c3 JA |
2458 | } |
2459 | ||
fc4df999 JA |
2460 | static int io_fsync(struct io_kiocb *req, struct io_kiocb **nxt, |
2461 | bool force_nonblock) | |
c992fe29 | 2462 | { |
8ed8d3c3 | 2463 | struct io_wq_work *work, *old_work; |
c992fe29 CH |
2464 | |
2465 | /* fsync always requires a blocking context */ | |
8ed8d3c3 JA |
2466 | if (force_nonblock) { |
2467 | io_put_req(req); | |
2468 | req->work.func = io_fsync_finish; | |
c992fe29 | 2469 | return -EAGAIN; |
8ed8d3c3 | 2470 | } |
c992fe29 | 2471 | |
8ed8d3c3 JA |
2472 | work = old_work = &req->work; |
2473 | io_fsync_finish(&work); | |
2474 | if (work && work != old_work) | |
2475 | *nxt = container_of(work, struct io_kiocb, work); | |
c992fe29 CH |
2476 | return 0; |
2477 | } | |
2478 | ||
d63d1b5e | 2479 | static void io_fallocate_finish(struct io_wq_work **workptr) |
8ed8d3c3 JA |
2480 | { |
2481 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
2482 | struct io_kiocb *nxt = NULL; | |
2483 | int ret; | |
2484 | ||
d63d1b5e JA |
2485 | ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off, |
2486 | req->sync.len); | |
8ed8d3c3 JA |
2487 | if (ret < 0) |
2488 | req_set_fail_links(req); | |
2489 | io_cqring_add_event(req, ret); | |
2490 | io_put_req_find_next(req, &nxt); | |
2491 | if (nxt) | |
78912934 | 2492 | io_wq_assign_next(workptr, nxt); |
5d17b4a4 JA |
2493 | } |
2494 | ||
d63d1b5e JA |
2495 | static int io_fallocate_prep(struct io_kiocb *req, |
2496 | const struct io_uring_sqe *sqe) | |
2497 | { | |
2498 | if (sqe->ioprio || sqe->buf_index || sqe->rw_flags) | |
2499 | return -EINVAL; | |
2500 | ||
2501 | req->sync.off = READ_ONCE(sqe->off); | |
2502 | req->sync.len = READ_ONCE(sqe->addr); | |
2503 | req->sync.mode = READ_ONCE(sqe->len); | |
2504 | return 0; | |
2505 | } | |
2506 | ||
2507 | static int io_fallocate(struct io_kiocb *req, struct io_kiocb **nxt, | |
2508 | bool force_nonblock) | |
5d17b4a4 | 2509 | { |
8ed8d3c3 | 2510 | struct io_wq_work *work, *old_work; |
5d17b4a4 | 2511 | |
d63d1b5e | 2512 | /* fallocate always requiring blocking context */ |
8ed8d3c3 JA |
2513 | if (force_nonblock) { |
2514 | io_put_req(req); | |
d63d1b5e | 2515 | req->work.func = io_fallocate_finish; |
5d17b4a4 | 2516 | return -EAGAIN; |
8ed8d3c3 | 2517 | } |
5d17b4a4 | 2518 | |
8ed8d3c3 | 2519 | work = old_work = &req->work; |
d63d1b5e | 2520 | io_fallocate_finish(&work); |
8ed8d3c3 JA |
2521 | if (work && work != old_work) |
2522 | *nxt = container_of(work, struct io_kiocb, work); | |
d63d1b5e | 2523 | |
5d17b4a4 JA |
2524 | return 0; |
2525 | } | |
2526 | ||
15b71abe | 2527 | static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
b7bb4f7d | 2528 | { |
f8748881 | 2529 | const char __user *fname; |
15b71abe | 2530 | int ret; |
b7bb4f7d | 2531 | |
15b71abe JA |
2532 | if (sqe->ioprio || sqe->buf_index) |
2533 | return -EINVAL; | |
03b1230c | 2534 | |
15b71abe | 2535 | req->open.dfd = READ_ONCE(sqe->fd); |
c12cedf2 | 2536 | req->open.how.mode = READ_ONCE(sqe->len); |
f8748881 | 2537 | fname = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
c12cedf2 | 2538 | req->open.how.flags = READ_ONCE(sqe->open_flags); |
3529d8c2 | 2539 | |
f8748881 | 2540 | req->open.filename = getname(fname); |
15b71abe JA |
2541 | if (IS_ERR(req->open.filename)) { |
2542 | ret = PTR_ERR(req->open.filename); | |
2543 | req->open.filename = NULL; | |
2544 | return ret; | |
2545 | } | |
3529d8c2 | 2546 | |
15b71abe | 2547 | return 0; |
03b1230c JA |
2548 | } |
2549 | ||
cebdb986 | 2550 | static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
aa1fa28f | 2551 | { |
cebdb986 JA |
2552 | struct open_how __user *how; |
2553 | const char __user *fname; | |
2554 | size_t len; | |
0fa03c62 JA |
2555 | int ret; |
2556 | ||
cebdb986 | 2557 | if (sqe->ioprio || sqe->buf_index) |
0fa03c62 JA |
2558 | return -EINVAL; |
2559 | ||
cebdb986 JA |
2560 | req->open.dfd = READ_ONCE(sqe->fd); |
2561 | fname = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
2562 | how = u64_to_user_ptr(READ_ONCE(sqe->addr2)); | |
2563 | len = READ_ONCE(sqe->len); | |
0fa03c62 | 2564 | |
cebdb986 JA |
2565 | if (len < OPEN_HOW_SIZE_VER0) |
2566 | return -EINVAL; | |
3529d8c2 | 2567 | |
cebdb986 JA |
2568 | ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how, |
2569 | len); | |
2570 | if (ret) | |
2571 | return ret; | |
3529d8c2 | 2572 | |
cebdb986 JA |
2573 | if (!(req->open.how.flags & O_PATH) && force_o_largefile()) |
2574 | req->open.how.flags |= O_LARGEFILE; | |
0fa03c62 | 2575 | |
cebdb986 JA |
2576 | req->open.filename = getname(fname); |
2577 | if (IS_ERR(req->open.filename)) { | |
2578 | ret = PTR_ERR(req->open.filename); | |
2579 | req->open.filename = NULL; | |
2580 | return ret; | |
2581 | } | |
2582 | ||
2583 | return 0; | |
2584 | } | |
2585 | ||
2586 | static int io_openat2(struct io_kiocb *req, struct io_kiocb **nxt, | |
2587 | bool force_nonblock) | |
15b71abe JA |
2588 | { |
2589 | struct open_flags op; | |
15b71abe JA |
2590 | struct file *file; |
2591 | int ret; | |
2592 | ||
f86cd20c | 2593 | if (force_nonblock) |
15b71abe | 2594 | return -EAGAIN; |
15b71abe | 2595 | |
cebdb986 | 2596 | ret = build_open_flags(&req->open.how, &op); |
15b71abe JA |
2597 | if (ret) |
2598 | goto err; | |
2599 | ||
cebdb986 | 2600 | ret = get_unused_fd_flags(req->open.how.flags); |
15b71abe JA |
2601 | if (ret < 0) |
2602 | goto err; | |
2603 | ||
2604 | file = do_filp_open(req->open.dfd, req->open.filename, &op); | |
2605 | if (IS_ERR(file)) { | |
2606 | put_unused_fd(ret); | |
2607 | ret = PTR_ERR(file); | |
2608 | } else { | |
2609 | fsnotify_open(file); | |
2610 | fd_install(ret, file); | |
2611 | } | |
2612 | err: | |
2613 | putname(req->open.filename); | |
2614 | if (ret < 0) | |
2615 | req_set_fail_links(req); | |
2616 | io_cqring_add_event(req, ret); | |
2617 | io_put_req_find_next(req, nxt); | |
2618 | return 0; | |
2619 | } | |
2620 | ||
cebdb986 JA |
2621 | static int io_openat(struct io_kiocb *req, struct io_kiocb **nxt, |
2622 | bool force_nonblock) | |
2623 | { | |
2624 | req->open.how = build_open_how(req->open.how.flags, req->open.how.mode); | |
2625 | return io_openat2(req, nxt, force_nonblock); | |
2626 | } | |
2627 | ||
3e4827b0 JA |
2628 | static int io_epoll_ctl_prep(struct io_kiocb *req, |
2629 | const struct io_uring_sqe *sqe) | |
2630 | { | |
2631 | #if defined(CONFIG_EPOLL) | |
2632 | if (sqe->ioprio || sqe->buf_index) | |
2633 | return -EINVAL; | |
2634 | ||
2635 | req->epoll.epfd = READ_ONCE(sqe->fd); | |
2636 | req->epoll.op = READ_ONCE(sqe->len); | |
2637 | req->epoll.fd = READ_ONCE(sqe->off); | |
2638 | ||
2639 | if (ep_op_has_event(req->epoll.op)) { | |
2640 | struct epoll_event __user *ev; | |
2641 | ||
2642 | ev = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
2643 | if (copy_from_user(&req->epoll.event, ev, sizeof(*ev))) | |
2644 | return -EFAULT; | |
2645 | } | |
2646 | ||
2647 | return 0; | |
2648 | #else | |
2649 | return -EOPNOTSUPP; | |
2650 | #endif | |
2651 | } | |
2652 | ||
2653 | static int io_epoll_ctl(struct io_kiocb *req, struct io_kiocb **nxt, | |
2654 | bool force_nonblock) | |
2655 | { | |
2656 | #if defined(CONFIG_EPOLL) | |
2657 | struct io_epoll *ie = &req->epoll; | |
2658 | int ret; | |
2659 | ||
2660 | ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock); | |
2661 | if (force_nonblock && ret == -EAGAIN) | |
2662 | return -EAGAIN; | |
2663 | ||
2664 | if (ret < 0) | |
2665 | req_set_fail_links(req); | |
2666 | io_cqring_add_event(req, ret); | |
2667 | io_put_req_find_next(req, nxt); | |
2668 | return 0; | |
2669 | #else | |
2670 | return -EOPNOTSUPP; | |
2671 | #endif | |
2672 | } | |
2673 | ||
c1ca757b JA |
2674 | static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2675 | { | |
2676 | #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU) | |
2677 | if (sqe->ioprio || sqe->buf_index || sqe->off) | |
2678 | return -EINVAL; | |
2679 | ||
2680 | req->madvise.addr = READ_ONCE(sqe->addr); | |
2681 | req->madvise.len = READ_ONCE(sqe->len); | |
2682 | req->madvise.advice = READ_ONCE(sqe->fadvise_advice); | |
2683 | return 0; | |
2684 | #else | |
2685 | return -EOPNOTSUPP; | |
2686 | #endif | |
2687 | } | |
2688 | ||
2689 | static int io_madvise(struct io_kiocb *req, struct io_kiocb **nxt, | |
2690 | bool force_nonblock) | |
2691 | { | |
2692 | #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU) | |
2693 | struct io_madvise *ma = &req->madvise; | |
2694 | int ret; | |
2695 | ||
2696 | if (force_nonblock) | |
2697 | return -EAGAIN; | |
2698 | ||
2699 | ret = do_madvise(ma->addr, ma->len, ma->advice); | |
2700 | if (ret < 0) | |
2701 | req_set_fail_links(req); | |
2702 | io_cqring_add_event(req, ret); | |
2703 | io_put_req_find_next(req, nxt); | |
2704 | return 0; | |
2705 | #else | |
2706 | return -EOPNOTSUPP; | |
2707 | #endif | |
2708 | } | |
2709 | ||
4840e418 JA |
2710 | static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2711 | { | |
2712 | if (sqe->ioprio || sqe->buf_index || sqe->addr) | |
2713 | return -EINVAL; | |
2714 | ||
2715 | req->fadvise.offset = READ_ONCE(sqe->off); | |
2716 | req->fadvise.len = READ_ONCE(sqe->len); | |
2717 | req->fadvise.advice = READ_ONCE(sqe->fadvise_advice); | |
2718 | return 0; | |
2719 | } | |
2720 | ||
2721 | static int io_fadvise(struct io_kiocb *req, struct io_kiocb **nxt, | |
2722 | bool force_nonblock) | |
2723 | { | |
2724 | struct io_fadvise *fa = &req->fadvise; | |
2725 | int ret; | |
2726 | ||
2727 | /* DONTNEED may block, others _should_ not */ | |
2728 | if (fa->advice == POSIX_FADV_DONTNEED && force_nonblock) | |
2729 | return -EAGAIN; | |
2730 | ||
2731 | ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice); | |
2732 | if (ret < 0) | |
2733 | req_set_fail_links(req); | |
2734 | io_cqring_add_event(req, ret); | |
2735 | io_put_req_find_next(req, nxt); | |
2736 | return 0; | |
2737 | } | |
2738 | ||
eddc7ef5 JA |
2739 | static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2740 | { | |
f8748881 | 2741 | const char __user *fname; |
eddc7ef5 JA |
2742 | unsigned lookup_flags; |
2743 | int ret; | |
2744 | ||
2745 | if (sqe->ioprio || sqe->buf_index) | |
2746 | return -EINVAL; | |
2747 | ||
2748 | req->open.dfd = READ_ONCE(sqe->fd); | |
2749 | req->open.mask = READ_ONCE(sqe->len); | |
f8748881 | 2750 | fname = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
eddc7ef5 | 2751 | req->open.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2)); |
c12cedf2 | 2752 | req->open.how.flags = READ_ONCE(sqe->statx_flags); |
eddc7ef5 | 2753 | |
c12cedf2 | 2754 | if (vfs_stat_set_lookup_flags(&lookup_flags, req->open.how.flags)) |
eddc7ef5 JA |
2755 | return -EINVAL; |
2756 | ||
f8748881 | 2757 | req->open.filename = getname_flags(fname, lookup_flags, NULL); |
eddc7ef5 JA |
2758 | if (IS_ERR(req->open.filename)) { |
2759 | ret = PTR_ERR(req->open.filename); | |
2760 | req->open.filename = NULL; | |
2761 | return ret; | |
2762 | } | |
2763 | ||
2764 | return 0; | |
2765 | } | |
2766 | ||
2767 | static int io_statx(struct io_kiocb *req, struct io_kiocb **nxt, | |
2768 | bool force_nonblock) | |
2769 | { | |
2770 | struct io_open *ctx = &req->open; | |
2771 | unsigned lookup_flags; | |
2772 | struct path path; | |
2773 | struct kstat stat; | |
2774 | int ret; | |
2775 | ||
2776 | if (force_nonblock) | |
2777 | return -EAGAIN; | |
2778 | ||
c12cedf2 | 2779 | if (vfs_stat_set_lookup_flags(&lookup_flags, ctx->how.flags)) |
eddc7ef5 JA |
2780 | return -EINVAL; |
2781 | ||
2782 | retry: | |
2783 | /* filename_lookup() drops it, keep a reference */ | |
2784 | ctx->filename->refcnt++; | |
2785 | ||
2786 | ret = filename_lookup(ctx->dfd, ctx->filename, lookup_flags, &path, | |
2787 | NULL); | |
2788 | if (ret) | |
2789 | goto err; | |
2790 | ||
c12cedf2 | 2791 | ret = vfs_getattr(&path, &stat, ctx->mask, ctx->how.flags); |
eddc7ef5 JA |
2792 | path_put(&path); |
2793 | if (retry_estale(ret, lookup_flags)) { | |
2794 | lookup_flags |= LOOKUP_REVAL; | |
2795 | goto retry; | |
2796 | } | |
2797 | if (!ret) | |
2798 | ret = cp_statx(&stat, ctx->buffer); | |
2799 | err: | |
2800 | putname(ctx->filename); | |
2801 | if (ret < 0) | |
2802 | req_set_fail_links(req); | |
2803 | io_cqring_add_event(req, ret); | |
2804 | io_put_req_find_next(req, nxt); | |
2805 | return 0; | |
2806 | } | |
2807 | ||
b5dba59e JA |
2808 | static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2809 | { | |
2810 | /* | |
2811 | * If we queue this for async, it must not be cancellable. That would | |
2812 | * leave the 'file' in an undeterminate state. | |
2813 | */ | |
2814 | req->work.flags |= IO_WQ_WORK_NO_CANCEL; | |
2815 | ||
2816 | if (sqe->ioprio || sqe->off || sqe->addr || sqe->len || | |
2817 | sqe->rw_flags || sqe->buf_index) | |
2818 | return -EINVAL; | |
2819 | if (sqe->flags & IOSQE_FIXED_FILE) | |
2820 | return -EINVAL; | |
2821 | ||
2822 | req->close.fd = READ_ONCE(sqe->fd); | |
2823 | if (req->file->f_op == &io_uring_fops || | |
b14cca0c | 2824 | req->close.fd == req->ctx->ring_fd) |
b5dba59e JA |
2825 | return -EBADF; |
2826 | ||
2827 | return 0; | |
2828 | } | |
2829 | ||
2830 | static void io_close_finish(struct io_wq_work **workptr) | |
2831 | { | |
2832 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
2833 | struct io_kiocb *nxt = NULL; | |
2834 | ||
2835 | /* Invoked with files, we need to do the close */ | |
2836 | if (req->work.files) { | |
2837 | int ret; | |
2838 | ||
2839 | ret = filp_close(req->close.put_file, req->work.files); | |
2840 | if (ret < 0) { | |
2841 | req_set_fail_links(req); | |
2842 | } | |
2843 | io_cqring_add_event(req, ret); | |
2844 | } | |
2845 | ||
2846 | fput(req->close.put_file); | |
2847 | ||
2848 | /* we bypassed the re-issue, drop the submission reference */ | |
2849 | io_put_req(req); | |
2850 | io_put_req_find_next(req, &nxt); | |
2851 | if (nxt) | |
2852 | io_wq_assign_next(workptr, nxt); | |
2853 | } | |
2854 | ||
2855 | static int io_close(struct io_kiocb *req, struct io_kiocb **nxt, | |
2856 | bool force_nonblock) | |
2857 | { | |
2858 | int ret; | |
2859 | ||
2860 | req->close.put_file = NULL; | |
2861 | ret = __close_fd_get_file(req->close.fd, &req->close.put_file); | |
2862 | if (ret < 0) | |
2863 | return ret; | |
2864 | ||
2865 | /* if the file has a flush method, be safe and punt to async */ | |
f86cd20c | 2866 | if (req->close.put_file->f_op->flush && !io_wq_current_is_worker()) |
b5dba59e | 2867 | goto eagain; |
b5dba59e JA |
2868 | |
2869 | /* | |
2870 | * No ->flush(), safely close from here and just punt the | |
2871 | * fput() to async context. | |
2872 | */ | |
2873 | ret = filp_close(req->close.put_file, current->files); | |
2874 | ||
2875 | if (ret < 0) | |
2876 | req_set_fail_links(req); | |
2877 | io_cqring_add_event(req, ret); | |
2878 | ||
2879 | if (io_wq_current_is_worker()) { | |
2880 | struct io_wq_work *old_work, *work; | |
2881 | ||
2882 | old_work = work = &req->work; | |
2883 | io_close_finish(&work); | |
2884 | if (work && work != old_work) | |
2885 | *nxt = container_of(work, struct io_kiocb, work); | |
2886 | return 0; | |
2887 | } | |
2888 | ||
2889 | eagain: | |
2890 | req->work.func = io_close_finish; | |
2891 | return -EAGAIN; | |
2892 | } | |
2893 | ||
3529d8c2 | 2894 | static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
5d17b4a4 JA |
2895 | { |
2896 | struct io_ring_ctx *ctx = req->ctx; | |
5d17b4a4 JA |
2897 | |
2898 | if (!req->file) | |
2899 | return -EBADF; | |
5d17b4a4 JA |
2900 | |
2901 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
2902 | return -EINVAL; | |
2903 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) | |
2904 | return -EINVAL; | |
2905 | ||
8ed8d3c3 JA |
2906 | req->sync.off = READ_ONCE(sqe->off); |
2907 | req->sync.len = READ_ONCE(sqe->len); | |
2908 | req->sync.flags = READ_ONCE(sqe->sync_range_flags); | |
8ed8d3c3 JA |
2909 | return 0; |
2910 | } | |
2911 | ||
2912 | static void io_sync_file_range_finish(struct io_wq_work **workptr) | |
2913 | { | |
2914 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
2915 | struct io_kiocb *nxt = NULL; | |
2916 | int ret; | |
2917 | ||
2918 | if (io_req_cancelled(req)) | |
2919 | return; | |
2920 | ||
9adbd45d | 2921 | ret = sync_file_range(req->file, req->sync.off, req->sync.len, |
8ed8d3c3 JA |
2922 | req->sync.flags); |
2923 | if (ret < 0) | |
2924 | req_set_fail_links(req); | |
2925 | io_cqring_add_event(req, ret); | |
2926 | io_put_req_find_next(req, &nxt); | |
2927 | if (nxt) | |
78912934 | 2928 | io_wq_assign_next(workptr, nxt); |
5d17b4a4 JA |
2929 | } |
2930 | ||
fc4df999 | 2931 | static int io_sync_file_range(struct io_kiocb *req, struct io_kiocb **nxt, |
5d17b4a4 JA |
2932 | bool force_nonblock) |
2933 | { | |
8ed8d3c3 | 2934 | struct io_wq_work *work, *old_work; |
5d17b4a4 JA |
2935 | |
2936 | /* sync_file_range always requires a blocking context */ | |
8ed8d3c3 JA |
2937 | if (force_nonblock) { |
2938 | io_put_req(req); | |
2939 | req->work.func = io_sync_file_range_finish; | |
5d17b4a4 | 2940 | return -EAGAIN; |
8ed8d3c3 | 2941 | } |
5d17b4a4 | 2942 | |
8ed8d3c3 JA |
2943 | work = old_work = &req->work; |
2944 | io_sync_file_range_finish(&work); | |
2945 | if (work && work != old_work) | |
2946 | *nxt = container_of(work, struct io_kiocb, work); | |
5d17b4a4 JA |
2947 | return 0; |
2948 | } | |
2949 | ||
b7bb4f7d JA |
2950 | #if defined(CONFIG_NET) |
2951 | static void io_sendrecv_async(struct io_wq_work **workptr) | |
2952 | { | |
2953 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
2954 | struct iovec *iov = NULL; | |
2955 | ||
2956 | if (req->io->rw.iov != req->io->rw.fast_iov) | |
2957 | iov = req->io->msg.iov; | |
2958 | io_wq_submit_work(workptr); | |
2959 | kfree(iov); | |
2960 | } | |
2961 | #endif | |
2962 | ||
3529d8c2 | 2963 | static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
03b1230c | 2964 | { |
0fa03c62 | 2965 | #if defined(CONFIG_NET) |
e47293fd | 2966 | struct io_sr_msg *sr = &req->sr_msg; |
3529d8c2 | 2967 | struct io_async_ctx *io = req->io; |
03b1230c | 2968 | |
e47293fd JA |
2969 | sr->msg_flags = READ_ONCE(sqe->msg_flags); |
2970 | sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
fddaface | 2971 | sr->len = READ_ONCE(sqe->len); |
3529d8c2 | 2972 | |
fddaface | 2973 | if (!io || req->opcode == IORING_OP_SEND) |
3529d8c2 JA |
2974 | return 0; |
2975 | ||
d9688565 | 2976 | io->msg.iov = io->msg.fast_iov; |
3529d8c2 | 2977 | return sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags, |
e47293fd | 2978 | &io->msg.iov); |
03b1230c | 2979 | #else |
e47293fd | 2980 | return -EOPNOTSUPP; |
03b1230c JA |
2981 | #endif |
2982 | } | |
2983 | ||
fc4df999 JA |
2984 | static int io_sendmsg(struct io_kiocb *req, struct io_kiocb **nxt, |
2985 | bool force_nonblock) | |
aa1fa28f | 2986 | { |
03b1230c | 2987 | #if defined(CONFIG_NET) |
0b416c3e | 2988 | struct io_async_msghdr *kmsg = NULL; |
0fa03c62 JA |
2989 | struct socket *sock; |
2990 | int ret; | |
2991 | ||
2992 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
2993 | return -EINVAL; | |
2994 | ||
2995 | sock = sock_from_file(req->file, &ret); | |
2996 | if (sock) { | |
b7bb4f7d | 2997 | struct io_async_ctx io; |
03b1230c | 2998 | struct sockaddr_storage addr; |
0fa03c62 JA |
2999 | unsigned flags; |
3000 | ||
03b1230c | 3001 | if (req->io) { |
0b416c3e JA |
3002 | kmsg = &req->io->msg; |
3003 | kmsg->msg.msg_name = &addr; | |
3004 | /* if iov is set, it's allocated already */ | |
3005 | if (!kmsg->iov) | |
3006 | kmsg->iov = kmsg->fast_iov; | |
3007 | kmsg->msg.msg_iter.iov = kmsg->iov; | |
03b1230c | 3008 | } else { |
3529d8c2 JA |
3009 | struct io_sr_msg *sr = &req->sr_msg; |
3010 | ||
0b416c3e JA |
3011 | kmsg = &io.msg; |
3012 | kmsg->msg.msg_name = &addr; | |
3529d8c2 JA |
3013 | |
3014 | io.msg.iov = io.msg.fast_iov; | |
3015 | ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg, | |
3016 | sr->msg_flags, &io.msg.iov); | |
03b1230c | 3017 | if (ret) |
3529d8c2 | 3018 | return ret; |
03b1230c | 3019 | } |
0fa03c62 | 3020 | |
e47293fd JA |
3021 | flags = req->sr_msg.msg_flags; |
3022 | if (flags & MSG_DONTWAIT) | |
3023 | req->flags |= REQ_F_NOWAIT; | |
3024 | else if (force_nonblock) | |
3025 | flags |= MSG_DONTWAIT; | |
3026 | ||
0b416c3e | 3027 | ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags); |
03b1230c | 3028 | if (force_nonblock && ret == -EAGAIN) { |
b7bb4f7d JA |
3029 | if (req->io) |
3030 | return -EAGAIN; | |
3031 | if (io_alloc_async_ctx(req)) | |
3032 | return -ENOMEM; | |
3033 | memcpy(&req->io->msg, &io.msg, sizeof(io.msg)); | |
3034 | req->work.func = io_sendrecv_async; | |
0b416c3e | 3035 | return -EAGAIN; |
03b1230c | 3036 | } |
441cdbd5 JA |
3037 | if (ret == -ERESTARTSYS) |
3038 | ret = -EINTR; | |
0fa03c62 JA |
3039 | } |
3040 | ||
b7bb4f7d | 3041 | if (!io_wq_current_is_worker() && kmsg && kmsg->iov != kmsg->fast_iov) |
0b416c3e | 3042 | kfree(kmsg->iov); |
78e19bbe | 3043 | io_cqring_add_event(req, ret); |
4e88d6e7 JA |
3044 | if (ret < 0) |
3045 | req_set_fail_links(req); | |
ec9c02ad | 3046 | io_put_req_find_next(req, nxt); |
5d17b4a4 | 3047 | return 0; |
03b1230c JA |
3048 | #else |
3049 | return -EOPNOTSUPP; | |
aa1fa28f | 3050 | #endif |
03b1230c | 3051 | } |
aa1fa28f | 3052 | |
fddaface JA |
3053 | static int io_send(struct io_kiocb *req, struct io_kiocb **nxt, |
3054 | bool force_nonblock) | |
3055 | { | |
3056 | #if defined(CONFIG_NET) | |
3057 | struct socket *sock; | |
3058 | int ret; | |
3059 | ||
3060 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
3061 | return -EINVAL; | |
3062 | ||
3063 | sock = sock_from_file(req->file, &ret); | |
3064 | if (sock) { | |
3065 | struct io_sr_msg *sr = &req->sr_msg; | |
3066 | struct msghdr msg; | |
3067 | struct iovec iov; | |
3068 | unsigned flags; | |
3069 | ||
3070 | ret = import_single_range(WRITE, sr->buf, sr->len, &iov, | |
3071 | &msg.msg_iter); | |
3072 | if (ret) | |
3073 | return ret; | |
3074 | ||
3075 | msg.msg_name = NULL; | |
3076 | msg.msg_control = NULL; | |
3077 | msg.msg_controllen = 0; | |
3078 | msg.msg_namelen = 0; | |
3079 | ||
3080 | flags = req->sr_msg.msg_flags; | |
3081 | if (flags & MSG_DONTWAIT) | |
3082 | req->flags |= REQ_F_NOWAIT; | |
3083 | else if (force_nonblock) | |
3084 | flags |= MSG_DONTWAIT; | |
3085 | ||
3086 | ret = __sys_sendmsg_sock(sock, &msg, flags); | |
3087 | if (force_nonblock && ret == -EAGAIN) | |
3088 | return -EAGAIN; | |
3089 | if (ret == -ERESTARTSYS) | |
3090 | ret = -EINTR; | |
3091 | } | |
3092 | ||
3093 | io_cqring_add_event(req, ret); | |
3094 | if (ret < 0) | |
3095 | req_set_fail_links(req); | |
3096 | io_put_req_find_next(req, nxt); | |
3097 | return 0; | |
3098 | #else | |
3099 | return -EOPNOTSUPP; | |
3100 | #endif | |
3101 | } | |
3102 | ||
3529d8c2 JA |
3103 | static int io_recvmsg_prep(struct io_kiocb *req, |
3104 | const struct io_uring_sqe *sqe) | |
aa1fa28f JA |
3105 | { |
3106 | #if defined(CONFIG_NET) | |
e47293fd | 3107 | struct io_sr_msg *sr = &req->sr_msg; |
3529d8c2 JA |
3108 | struct io_async_ctx *io = req->io; |
3109 | ||
3110 | sr->msg_flags = READ_ONCE(sqe->msg_flags); | |
3111 | sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
06b76d44 | 3112 | |
fddaface | 3113 | if (!io || req->opcode == IORING_OP_RECV) |
06b76d44 | 3114 | return 0; |
03b1230c | 3115 | |
d9688565 | 3116 | io->msg.iov = io->msg.fast_iov; |
3529d8c2 | 3117 | return recvmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags, |
e47293fd | 3118 | &io->msg.uaddr, &io->msg.iov); |
aa1fa28f | 3119 | #else |
e47293fd | 3120 | return -EOPNOTSUPP; |
aa1fa28f JA |
3121 | #endif |
3122 | } | |
3123 | ||
fc4df999 JA |
3124 | static int io_recvmsg(struct io_kiocb *req, struct io_kiocb **nxt, |
3125 | bool force_nonblock) | |
aa1fa28f JA |
3126 | { |
3127 | #if defined(CONFIG_NET) | |
0b416c3e | 3128 | struct io_async_msghdr *kmsg = NULL; |
03b1230c JA |
3129 | struct socket *sock; |
3130 | int ret; | |
3131 | ||
3132 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
3133 | return -EINVAL; | |
3134 | ||
3135 | sock = sock_from_file(req->file, &ret); | |
3136 | if (sock) { | |
b7bb4f7d | 3137 | struct io_async_ctx io; |
03b1230c | 3138 | struct sockaddr_storage addr; |
03b1230c JA |
3139 | unsigned flags; |
3140 | ||
03b1230c | 3141 | if (req->io) { |
0b416c3e JA |
3142 | kmsg = &req->io->msg; |
3143 | kmsg->msg.msg_name = &addr; | |
3144 | /* if iov is set, it's allocated already */ | |
3145 | if (!kmsg->iov) | |
3146 | kmsg->iov = kmsg->fast_iov; | |
3147 | kmsg->msg.msg_iter.iov = kmsg->iov; | |
03b1230c | 3148 | } else { |
3529d8c2 JA |
3149 | struct io_sr_msg *sr = &req->sr_msg; |
3150 | ||
0b416c3e JA |
3151 | kmsg = &io.msg; |
3152 | kmsg->msg.msg_name = &addr; | |
3529d8c2 JA |
3153 | |
3154 | io.msg.iov = io.msg.fast_iov; | |
3155 | ret = recvmsg_copy_msghdr(&io.msg.msg, sr->msg, | |
3156 | sr->msg_flags, &io.msg.uaddr, | |
3157 | &io.msg.iov); | |
03b1230c | 3158 | if (ret) |
3529d8c2 | 3159 | return ret; |
03b1230c JA |
3160 | } |
3161 | ||
e47293fd JA |
3162 | flags = req->sr_msg.msg_flags; |
3163 | if (flags & MSG_DONTWAIT) | |
3164 | req->flags |= REQ_F_NOWAIT; | |
3165 | else if (force_nonblock) | |
3166 | flags |= MSG_DONTWAIT; | |
3167 | ||
3168 | ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg, | |
3169 | kmsg->uaddr, flags); | |
03b1230c | 3170 | if (force_nonblock && ret == -EAGAIN) { |
b7bb4f7d JA |
3171 | if (req->io) |
3172 | return -EAGAIN; | |
3173 | if (io_alloc_async_ctx(req)) | |
3174 | return -ENOMEM; | |
3175 | memcpy(&req->io->msg, &io.msg, sizeof(io.msg)); | |
3176 | req->work.func = io_sendrecv_async; | |
0b416c3e | 3177 | return -EAGAIN; |
03b1230c JA |
3178 | } |
3179 | if (ret == -ERESTARTSYS) | |
3180 | ret = -EINTR; | |
3181 | } | |
3182 | ||
b7bb4f7d | 3183 | if (!io_wq_current_is_worker() && kmsg && kmsg->iov != kmsg->fast_iov) |
0b416c3e | 3184 | kfree(kmsg->iov); |
03b1230c | 3185 | io_cqring_add_event(req, ret); |
4e88d6e7 JA |
3186 | if (ret < 0) |
3187 | req_set_fail_links(req); | |
03b1230c JA |
3188 | io_put_req_find_next(req, nxt); |
3189 | return 0; | |
0fa03c62 JA |
3190 | #else |
3191 | return -EOPNOTSUPP; | |
3192 | #endif | |
3193 | } | |
5d17b4a4 | 3194 | |
fddaface JA |
3195 | static int io_recv(struct io_kiocb *req, struct io_kiocb **nxt, |
3196 | bool force_nonblock) | |
3197 | { | |
3198 | #if defined(CONFIG_NET) | |
3199 | struct socket *sock; | |
3200 | int ret; | |
3201 | ||
3202 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
3203 | return -EINVAL; | |
3204 | ||
3205 | sock = sock_from_file(req->file, &ret); | |
3206 | if (sock) { | |
3207 | struct io_sr_msg *sr = &req->sr_msg; | |
3208 | struct msghdr msg; | |
3209 | struct iovec iov; | |
3210 | unsigned flags; | |
3211 | ||
3212 | ret = import_single_range(READ, sr->buf, sr->len, &iov, | |
3213 | &msg.msg_iter); | |
3214 | if (ret) | |
3215 | return ret; | |
3216 | ||
3217 | msg.msg_name = NULL; | |
3218 | msg.msg_control = NULL; | |
3219 | msg.msg_controllen = 0; | |
3220 | msg.msg_namelen = 0; | |
3221 | msg.msg_iocb = NULL; | |
3222 | msg.msg_flags = 0; | |
3223 | ||
3224 | flags = req->sr_msg.msg_flags; | |
3225 | if (flags & MSG_DONTWAIT) | |
3226 | req->flags |= REQ_F_NOWAIT; | |
3227 | else if (force_nonblock) | |
3228 | flags |= MSG_DONTWAIT; | |
3229 | ||
3230 | ret = __sys_recvmsg_sock(sock, &msg, NULL, NULL, flags); | |
3231 | if (force_nonblock && ret == -EAGAIN) | |
3232 | return -EAGAIN; | |
3233 | if (ret == -ERESTARTSYS) | |
3234 | ret = -EINTR; | |
3235 | } | |
3236 | ||
3237 | io_cqring_add_event(req, ret); | |
3238 | if (ret < 0) | |
3239 | req_set_fail_links(req); | |
3240 | io_put_req_find_next(req, nxt); | |
3241 | return 0; | |
3242 | #else | |
3243 | return -EOPNOTSUPP; | |
3244 | #endif | |
3245 | } | |
3246 | ||
3247 | ||
3529d8c2 | 3248 | static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
17f2fe35 JA |
3249 | { |
3250 | #if defined(CONFIG_NET) | |
8ed8d3c3 JA |
3251 | struct io_accept *accept = &req->accept; |
3252 | ||
17f2fe35 JA |
3253 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) |
3254 | return -EINVAL; | |
8042d6ce | 3255 | if (sqe->ioprio || sqe->len || sqe->buf_index) |
17f2fe35 JA |
3256 | return -EINVAL; |
3257 | ||
d55e5f5b JA |
3258 | accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
3259 | accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2)); | |
8ed8d3c3 | 3260 | accept->flags = READ_ONCE(sqe->accept_flags); |
8ed8d3c3 JA |
3261 | return 0; |
3262 | #else | |
3263 | return -EOPNOTSUPP; | |
3264 | #endif | |
3265 | } | |
17f2fe35 | 3266 | |
8ed8d3c3 JA |
3267 | #if defined(CONFIG_NET) |
3268 | static int __io_accept(struct io_kiocb *req, struct io_kiocb **nxt, | |
3269 | bool force_nonblock) | |
3270 | { | |
3271 | struct io_accept *accept = &req->accept; | |
3272 | unsigned file_flags; | |
3273 | int ret; | |
3274 | ||
3275 | file_flags = force_nonblock ? O_NONBLOCK : 0; | |
3276 | ret = __sys_accept4_file(req->file, file_flags, accept->addr, | |
3277 | accept->addr_len, accept->flags); | |
3278 | if (ret == -EAGAIN && force_nonblock) | |
17f2fe35 | 3279 | return -EAGAIN; |
8e3cca12 JA |
3280 | if (ret == -ERESTARTSYS) |
3281 | ret = -EINTR; | |
4e88d6e7 JA |
3282 | if (ret < 0) |
3283 | req_set_fail_links(req); | |
78e19bbe | 3284 | io_cqring_add_event(req, ret); |
ec9c02ad | 3285 | io_put_req_find_next(req, nxt); |
17f2fe35 | 3286 | return 0; |
8ed8d3c3 JA |
3287 | } |
3288 | ||
3289 | static void io_accept_finish(struct io_wq_work **workptr) | |
3290 | { | |
3291 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
3292 | struct io_kiocb *nxt = NULL; | |
3293 | ||
3294 | if (io_req_cancelled(req)) | |
3295 | return; | |
3296 | __io_accept(req, &nxt, false); | |
3297 | if (nxt) | |
78912934 | 3298 | io_wq_assign_next(workptr, nxt); |
8ed8d3c3 JA |
3299 | } |
3300 | #endif | |
3301 | ||
3302 | static int io_accept(struct io_kiocb *req, struct io_kiocb **nxt, | |
3303 | bool force_nonblock) | |
3304 | { | |
3305 | #if defined(CONFIG_NET) | |
3306 | int ret; | |
3307 | ||
8ed8d3c3 JA |
3308 | ret = __io_accept(req, nxt, force_nonblock); |
3309 | if (ret == -EAGAIN && force_nonblock) { | |
3310 | req->work.func = io_accept_finish; | |
8ed8d3c3 JA |
3311 | io_put_req(req); |
3312 | return -EAGAIN; | |
3313 | } | |
3314 | return 0; | |
0fa03c62 JA |
3315 | #else |
3316 | return -EOPNOTSUPP; | |
3317 | #endif | |
3318 | } | |
5d17b4a4 | 3319 | |
3529d8c2 | 3320 | static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f499a021 JA |
3321 | { |
3322 | #if defined(CONFIG_NET) | |
3529d8c2 JA |
3323 | struct io_connect *conn = &req->connect; |
3324 | struct io_async_ctx *io = req->io; | |
f499a021 | 3325 | |
3fbb51c1 JA |
3326 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) |
3327 | return -EINVAL; | |
3328 | if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags) | |
3329 | return -EINVAL; | |
3330 | ||
3529d8c2 JA |
3331 | conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
3332 | conn->addr_len = READ_ONCE(sqe->addr2); | |
3333 | ||
3334 | if (!io) | |
3335 | return 0; | |
3336 | ||
3337 | return move_addr_to_kernel(conn->addr, conn->addr_len, | |
3fbb51c1 | 3338 | &io->connect.address); |
f499a021 | 3339 | #else |
3fbb51c1 | 3340 | return -EOPNOTSUPP; |
f499a021 JA |
3341 | #endif |
3342 | } | |
3343 | ||
fc4df999 JA |
3344 | static int io_connect(struct io_kiocb *req, struct io_kiocb **nxt, |
3345 | bool force_nonblock) | |
f8e85cf2 JA |
3346 | { |
3347 | #if defined(CONFIG_NET) | |
f499a021 | 3348 | struct io_async_ctx __io, *io; |
f8e85cf2 | 3349 | unsigned file_flags; |
3fbb51c1 | 3350 | int ret; |
f8e85cf2 | 3351 | |
f499a021 JA |
3352 | if (req->io) { |
3353 | io = req->io; | |
3354 | } else { | |
3529d8c2 JA |
3355 | ret = move_addr_to_kernel(req->connect.addr, |
3356 | req->connect.addr_len, | |
3357 | &__io.connect.address); | |
f499a021 JA |
3358 | if (ret) |
3359 | goto out; | |
3360 | io = &__io; | |
3361 | } | |
3362 | ||
3fbb51c1 JA |
3363 | file_flags = force_nonblock ? O_NONBLOCK : 0; |
3364 | ||
3365 | ret = __sys_connect_file(req->file, &io->connect.address, | |
3366 | req->connect.addr_len, file_flags); | |
87f80d62 | 3367 | if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) { |
b7bb4f7d JA |
3368 | if (req->io) |
3369 | return -EAGAIN; | |
3370 | if (io_alloc_async_ctx(req)) { | |
f499a021 JA |
3371 | ret = -ENOMEM; |
3372 | goto out; | |
3373 | } | |
b7bb4f7d | 3374 | memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect)); |
f8e85cf2 | 3375 | return -EAGAIN; |
f499a021 | 3376 | } |
f8e85cf2 JA |
3377 | if (ret == -ERESTARTSYS) |
3378 | ret = -EINTR; | |
f499a021 | 3379 | out: |
4e88d6e7 JA |
3380 | if (ret < 0) |
3381 | req_set_fail_links(req); | |
f8e85cf2 JA |
3382 | io_cqring_add_event(req, ret); |
3383 | io_put_req_find_next(req, nxt); | |
3384 | return 0; | |
3385 | #else | |
3386 | return -EOPNOTSUPP; | |
3387 | #endif | |
3388 | } | |
3389 | ||
221c5eb2 JA |
3390 | static void io_poll_remove_one(struct io_kiocb *req) |
3391 | { | |
3392 | struct io_poll_iocb *poll = &req->poll; | |
3393 | ||
3394 | spin_lock(&poll->head->lock); | |
3395 | WRITE_ONCE(poll->canceled, true); | |
392edb45 JA |
3396 | if (!list_empty(&poll->wait.entry)) { |
3397 | list_del_init(&poll->wait.entry); | |
a197f664 | 3398 | io_queue_async_work(req); |
221c5eb2 JA |
3399 | } |
3400 | spin_unlock(&poll->head->lock); | |
78076bb6 | 3401 | hash_del(&req->hash_node); |
221c5eb2 JA |
3402 | } |
3403 | ||
3404 | static void io_poll_remove_all(struct io_ring_ctx *ctx) | |
3405 | { | |
78076bb6 | 3406 | struct hlist_node *tmp; |
221c5eb2 | 3407 | struct io_kiocb *req; |
78076bb6 | 3408 | int i; |
221c5eb2 JA |
3409 | |
3410 | spin_lock_irq(&ctx->completion_lock); | |
78076bb6 JA |
3411 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { |
3412 | struct hlist_head *list; | |
3413 | ||
3414 | list = &ctx->cancel_hash[i]; | |
3415 | hlist_for_each_entry_safe(req, tmp, list, hash_node) | |
3416 | io_poll_remove_one(req); | |
221c5eb2 JA |
3417 | } |
3418 | spin_unlock_irq(&ctx->completion_lock); | |
3419 | } | |
3420 | ||
47f46768 JA |
3421 | static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr) |
3422 | { | |
78076bb6 | 3423 | struct hlist_head *list; |
47f46768 JA |
3424 | struct io_kiocb *req; |
3425 | ||
78076bb6 JA |
3426 | list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)]; |
3427 | hlist_for_each_entry(req, list, hash_node) { | |
3428 | if (sqe_addr == req->user_data) { | |
eac406c6 JA |
3429 | io_poll_remove_one(req); |
3430 | return 0; | |
3431 | } | |
47f46768 JA |
3432 | } |
3433 | ||
3434 | return -ENOENT; | |
3435 | } | |
3436 | ||
3529d8c2 JA |
3437 | static int io_poll_remove_prep(struct io_kiocb *req, |
3438 | const struct io_uring_sqe *sqe) | |
0969e783 | 3439 | { |
0969e783 JA |
3440 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3441 | return -EINVAL; | |
3442 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index || | |
3443 | sqe->poll_events) | |
3444 | return -EINVAL; | |
3445 | ||
3446 | req->poll.addr = READ_ONCE(sqe->addr); | |
0969e783 JA |
3447 | return 0; |
3448 | } | |
3449 | ||
221c5eb2 JA |
3450 | /* |
3451 | * Find a running poll command that matches one specified in sqe->addr, | |
3452 | * and remove it if found. | |
3453 | */ | |
fc4df999 | 3454 | static int io_poll_remove(struct io_kiocb *req) |
221c5eb2 JA |
3455 | { |
3456 | struct io_ring_ctx *ctx = req->ctx; | |
0969e783 | 3457 | u64 addr; |
47f46768 | 3458 | int ret; |
221c5eb2 | 3459 | |
0969e783 | 3460 | addr = req->poll.addr; |
221c5eb2 | 3461 | spin_lock_irq(&ctx->completion_lock); |
0969e783 | 3462 | ret = io_poll_cancel(ctx, addr); |
221c5eb2 JA |
3463 | spin_unlock_irq(&ctx->completion_lock); |
3464 | ||
78e19bbe | 3465 | io_cqring_add_event(req, ret); |
4e88d6e7 JA |
3466 | if (ret < 0) |
3467 | req_set_fail_links(req); | |
e65ef56d | 3468 | io_put_req(req); |
221c5eb2 JA |
3469 | return 0; |
3470 | } | |
3471 | ||
b0dd8a41 | 3472 | static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error) |
221c5eb2 | 3473 | { |
a197f664 JL |
3474 | struct io_ring_ctx *ctx = req->ctx; |
3475 | ||
8c838788 | 3476 | req->poll.done = true; |
b0dd8a41 JA |
3477 | if (error) |
3478 | io_cqring_fill_event(req, error); | |
3479 | else | |
3480 | io_cqring_fill_event(req, mangle_poll(mask)); | |
8c838788 | 3481 | io_commit_cqring(ctx); |
221c5eb2 JA |
3482 | } |
3483 | ||
561fb04a | 3484 | static void io_poll_complete_work(struct io_wq_work **workptr) |
221c5eb2 | 3485 | { |
561fb04a | 3486 | struct io_wq_work *work = *workptr; |
221c5eb2 JA |
3487 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
3488 | struct io_poll_iocb *poll = &req->poll; | |
3489 | struct poll_table_struct pt = { ._key = poll->events }; | |
3490 | struct io_ring_ctx *ctx = req->ctx; | |
89723d0b | 3491 | struct io_kiocb *nxt = NULL; |
221c5eb2 | 3492 | __poll_t mask = 0; |
b0dd8a41 | 3493 | int ret = 0; |
221c5eb2 | 3494 | |
b0dd8a41 | 3495 | if (work->flags & IO_WQ_WORK_CANCEL) { |
561fb04a | 3496 | WRITE_ONCE(poll->canceled, true); |
b0dd8a41 JA |
3497 | ret = -ECANCELED; |
3498 | } else if (READ_ONCE(poll->canceled)) { | |
3499 | ret = -ECANCELED; | |
3500 | } | |
561fb04a | 3501 | |
b0dd8a41 | 3502 | if (ret != -ECANCELED) |
221c5eb2 JA |
3503 | mask = vfs_poll(poll->file, &pt) & poll->events; |
3504 | ||
3505 | /* | |
3506 | * Note that ->ki_cancel callers also delete iocb from active_reqs after | |
3507 | * calling ->ki_cancel. We need the ctx_lock roundtrip here to | |
3508 | * synchronize with them. In the cancellation case the list_del_init | |
3509 | * itself is not actually needed, but harmless so we keep it in to | |
3510 | * avoid further branches in the fast path. | |
3511 | */ | |
3512 | spin_lock_irq(&ctx->completion_lock); | |
b0dd8a41 | 3513 | if (!mask && ret != -ECANCELED) { |
392edb45 | 3514 | add_wait_queue(poll->head, &poll->wait); |
221c5eb2 JA |
3515 | spin_unlock_irq(&ctx->completion_lock); |
3516 | return; | |
3517 | } | |
78076bb6 | 3518 | hash_del(&req->hash_node); |
b0dd8a41 | 3519 | io_poll_complete(req, mask, ret); |
221c5eb2 JA |
3520 | spin_unlock_irq(&ctx->completion_lock); |
3521 | ||
8c838788 | 3522 | io_cqring_ev_posted(ctx); |
89723d0b | 3523 | |
4e88d6e7 JA |
3524 | if (ret < 0) |
3525 | req_set_fail_links(req); | |
ec9c02ad | 3526 | io_put_req_find_next(req, &nxt); |
89723d0b | 3527 | if (nxt) |
78912934 | 3528 | io_wq_assign_next(workptr, nxt); |
221c5eb2 JA |
3529 | } |
3530 | ||
e94f141b JA |
3531 | static void __io_poll_flush(struct io_ring_ctx *ctx, struct llist_node *nodes) |
3532 | { | |
e94f141b | 3533 | struct io_kiocb *req, *tmp; |
8237e045 | 3534 | struct req_batch rb; |
e94f141b | 3535 | |
c6ca97b3 | 3536 | rb.to_free = rb.need_iter = 0; |
e94f141b JA |
3537 | spin_lock_irq(&ctx->completion_lock); |
3538 | llist_for_each_entry_safe(req, tmp, nodes, llist_node) { | |
3539 | hash_del(&req->hash_node); | |
3540 | io_poll_complete(req, req->result, 0); | |
3541 | ||
8237e045 JA |
3542 | if (refcount_dec_and_test(&req->refs) && |
3543 | !io_req_multi_free(&rb, req)) { | |
3544 | req->flags |= REQ_F_COMP_LOCKED; | |
3545 | io_free_req(req); | |
e94f141b JA |
3546 | } |
3547 | } | |
3548 | spin_unlock_irq(&ctx->completion_lock); | |
3549 | ||
3550 | io_cqring_ev_posted(ctx); | |
8237e045 | 3551 | io_free_req_many(ctx, &rb); |
e94f141b JA |
3552 | } |
3553 | ||
3554 | static void io_poll_flush(struct io_wq_work **workptr) | |
3555 | { | |
3556 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
3557 | struct llist_node *nodes; | |
3558 | ||
3559 | nodes = llist_del_all(&req->ctx->poll_llist); | |
3560 | if (nodes) | |
3561 | __io_poll_flush(req->ctx, nodes); | |
3562 | } | |
3563 | ||
221c5eb2 JA |
3564 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, |
3565 | void *key) | |
3566 | { | |
e944475e | 3567 | struct io_poll_iocb *poll = wait->private; |
221c5eb2 JA |
3568 | struct io_kiocb *req = container_of(poll, struct io_kiocb, poll); |
3569 | struct io_ring_ctx *ctx = req->ctx; | |
3570 | __poll_t mask = key_to_poll(key); | |
221c5eb2 JA |
3571 | |
3572 | /* for instances that support it check for an event match first: */ | |
8c838788 JA |
3573 | if (mask && !(mask & poll->events)) |
3574 | return 0; | |
221c5eb2 | 3575 | |
392edb45 | 3576 | list_del_init(&poll->wait.entry); |
221c5eb2 | 3577 | |
7c9e7f0f JA |
3578 | /* |
3579 | * Run completion inline if we can. We're using trylock here because | |
3580 | * we are violating the completion_lock -> poll wq lock ordering. | |
3581 | * If we have a link timeout we're going to need the completion_lock | |
3582 | * for finalizing the request, mark us as having grabbed that already. | |
3583 | */ | |
e94f141b JA |
3584 | if (mask) { |
3585 | unsigned long flags; | |
221c5eb2 | 3586 | |
e94f141b JA |
3587 | if (llist_empty(&ctx->poll_llist) && |
3588 | spin_trylock_irqsave(&ctx->completion_lock, flags)) { | |
3589 | hash_del(&req->hash_node); | |
3590 | io_poll_complete(req, mask, 0); | |
3591 | req->flags |= REQ_F_COMP_LOCKED; | |
3592 | io_put_req(req); | |
3593 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
3594 | ||
3595 | io_cqring_ev_posted(ctx); | |
3596 | req = NULL; | |
3597 | } else { | |
3598 | req->result = mask; | |
3599 | req->llist_node.next = NULL; | |
3600 | /* if the list wasn't empty, we're done */ | |
3601 | if (!llist_add(&req->llist_node, &ctx->poll_llist)) | |
3602 | req = NULL; | |
3603 | else | |
3604 | req->work.func = io_poll_flush; | |
3605 | } | |
221c5eb2 | 3606 | } |
e94f141b JA |
3607 | if (req) |
3608 | io_queue_async_work(req); | |
221c5eb2 | 3609 | |
221c5eb2 JA |
3610 | return 1; |
3611 | } | |
3612 | ||
3613 | struct io_poll_table { | |
3614 | struct poll_table_struct pt; | |
3615 | struct io_kiocb *req; | |
3616 | int error; | |
3617 | }; | |
3618 | ||
3619 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
3620 | struct poll_table_struct *p) | |
3621 | { | |
3622 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
3623 | ||
3624 | if (unlikely(pt->req->poll.head)) { | |
3625 | pt->error = -EINVAL; | |
3626 | return; | |
3627 | } | |
3628 | ||
3629 | pt->error = 0; | |
3630 | pt->req->poll.head = head; | |
392edb45 | 3631 | add_wait_queue(head, &pt->req->poll.wait); |
221c5eb2 JA |
3632 | } |
3633 | ||
eac406c6 JA |
3634 | static void io_poll_req_insert(struct io_kiocb *req) |
3635 | { | |
3636 | struct io_ring_ctx *ctx = req->ctx; | |
78076bb6 JA |
3637 | struct hlist_head *list; |
3638 | ||
3639 | list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)]; | |
3640 | hlist_add_head(&req->hash_node, list); | |
eac406c6 JA |
3641 | } |
3642 | ||
3529d8c2 | 3643 | static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
221c5eb2 JA |
3644 | { |
3645 | struct io_poll_iocb *poll = &req->poll; | |
221c5eb2 | 3646 | u16 events; |
221c5eb2 JA |
3647 | |
3648 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
3649 | return -EINVAL; | |
3650 | if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
3651 | return -EINVAL; | |
09bb8394 JA |
3652 | if (!poll->file) |
3653 | return -EBADF; | |
221c5eb2 | 3654 | |
221c5eb2 JA |
3655 | events = READ_ONCE(sqe->poll_events); |
3656 | poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP; | |
0969e783 JA |
3657 | return 0; |
3658 | } | |
3659 | ||
3660 | static int io_poll_add(struct io_kiocb *req, struct io_kiocb **nxt) | |
3661 | { | |
3662 | struct io_poll_iocb *poll = &req->poll; | |
3663 | struct io_ring_ctx *ctx = req->ctx; | |
3664 | struct io_poll_table ipt; | |
3665 | bool cancel = false; | |
3666 | __poll_t mask; | |
0969e783 JA |
3667 | |
3668 | INIT_IO_WORK(&req->work, io_poll_complete_work); | |
78076bb6 | 3669 | INIT_HLIST_NODE(&req->hash_node); |
221c5eb2 | 3670 | |
221c5eb2 | 3671 | poll->head = NULL; |
8c838788 | 3672 | poll->done = false; |
221c5eb2 JA |
3673 | poll->canceled = false; |
3674 | ||
3675 | ipt.pt._qproc = io_poll_queue_proc; | |
3676 | ipt.pt._key = poll->events; | |
3677 | ipt.req = req; | |
3678 | ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */ | |
3679 | ||
3680 | /* initialized the list so that we can do list_empty checks */ | |
392edb45 JA |
3681 | INIT_LIST_HEAD(&poll->wait.entry); |
3682 | init_waitqueue_func_entry(&poll->wait, io_poll_wake); | |
3683 | poll->wait.private = poll; | |
221c5eb2 | 3684 | |
36703247 JA |
3685 | INIT_LIST_HEAD(&req->list); |
3686 | ||
221c5eb2 | 3687 | mask = vfs_poll(poll->file, &ipt.pt) & poll->events; |
221c5eb2 JA |
3688 | |
3689 | spin_lock_irq(&ctx->completion_lock); | |
8c838788 JA |
3690 | if (likely(poll->head)) { |
3691 | spin_lock(&poll->head->lock); | |
392edb45 | 3692 | if (unlikely(list_empty(&poll->wait.entry))) { |
8c838788 JA |
3693 | if (ipt.error) |
3694 | cancel = true; | |
3695 | ipt.error = 0; | |
3696 | mask = 0; | |
3697 | } | |
3698 | if (mask || ipt.error) | |
392edb45 | 3699 | list_del_init(&poll->wait.entry); |
8c838788 JA |
3700 | else if (cancel) |
3701 | WRITE_ONCE(poll->canceled, true); | |
3702 | else if (!poll->done) /* actually waiting for an event */ | |
eac406c6 | 3703 | io_poll_req_insert(req); |
8c838788 JA |
3704 | spin_unlock(&poll->head->lock); |
3705 | } | |
3706 | if (mask) { /* no async, we'd stolen it */ | |
221c5eb2 | 3707 | ipt.error = 0; |
b0dd8a41 | 3708 | io_poll_complete(req, mask, 0); |
221c5eb2 | 3709 | } |
221c5eb2 JA |
3710 | spin_unlock_irq(&ctx->completion_lock); |
3711 | ||
8c838788 JA |
3712 | if (mask) { |
3713 | io_cqring_ev_posted(ctx); | |
ec9c02ad | 3714 | io_put_req_find_next(req, nxt); |
221c5eb2 | 3715 | } |
8c838788 | 3716 | return ipt.error; |
221c5eb2 JA |
3717 | } |
3718 | ||
5262f567 JA |
3719 | static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer) |
3720 | { | |
ad8a48ac JA |
3721 | struct io_timeout_data *data = container_of(timer, |
3722 | struct io_timeout_data, timer); | |
3723 | struct io_kiocb *req = data->req; | |
3724 | struct io_ring_ctx *ctx = req->ctx; | |
5262f567 JA |
3725 | unsigned long flags; |
3726 | ||
5262f567 JA |
3727 | atomic_inc(&ctx->cq_timeouts); |
3728 | ||
3729 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
ef03681a | 3730 | /* |
11365043 JA |
3731 | * We could be racing with timeout deletion. If the list is empty, |
3732 | * then timeout lookup already found it and will be handling it. | |
ef03681a | 3733 | */ |
842f9612 | 3734 | if (!list_empty(&req->list)) { |
11365043 | 3735 | struct io_kiocb *prev; |
5262f567 | 3736 | |
11365043 JA |
3737 | /* |
3738 | * Adjust the reqs sequence before the current one because it | |
d195a66e | 3739 | * will consume a slot in the cq_ring and the cq_tail |
11365043 JA |
3740 | * pointer will be increased, otherwise other timeout reqs may |
3741 | * return in advance without waiting for enough wait_nr. | |
3742 | */ | |
3743 | prev = req; | |
3744 | list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list) | |
3745 | prev->sequence++; | |
11365043 | 3746 | list_del_init(&req->list); |
11365043 | 3747 | } |
5262f567 | 3748 | |
78e19bbe | 3749 | io_cqring_fill_event(req, -ETIME); |
5262f567 JA |
3750 | io_commit_cqring(ctx); |
3751 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
3752 | ||
3753 | io_cqring_ev_posted(ctx); | |
4e88d6e7 | 3754 | req_set_fail_links(req); |
5262f567 JA |
3755 | io_put_req(req); |
3756 | return HRTIMER_NORESTART; | |
3757 | } | |
3758 | ||
47f46768 JA |
3759 | static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data) |
3760 | { | |
3761 | struct io_kiocb *req; | |
3762 | int ret = -ENOENT; | |
3763 | ||
3764 | list_for_each_entry(req, &ctx->timeout_list, list) { | |
3765 | if (user_data == req->user_data) { | |
3766 | list_del_init(&req->list); | |
3767 | ret = 0; | |
3768 | break; | |
3769 | } | |
3770 | } | |
3771 | ||
3772 | if (ret == -ENOENT) | |
3773 | return ret; | |
3774 | ||
2d28390a | 3775 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); |
47f46768 JA |
3776 | if (ret == -1) |
3777 | return -EALREADY; | |
3778 | ||
4e88d6e7 | 3779 | req_set_fail_links(req); |
47f46768 JA |
3780 | io_cqring_fill_event(req, -ECANCELED); |
3781 | io_put_req(req); | |
3782 | return 0; | |
3783 | } | |
3784 | ||
3529d8c2 JA |
3785 | static int io_timeout_remove_prep(struct io_kiocb *req, |
3786 | const struct io_uring_sqe *sqe) | |
b29472ee | 3787 | { |
b29472ee JA |
3788 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3789 | return -EINVAL; | |
3790 | if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len) | |
3791 | return -EINVAL; | |
3792 | ||
3793 | req->timeout.addr = READ_ONCE(sqe->addr); | |
3794 | req->timeout.flags = READ_ONCE(sqe->timeout_flags); | |
3795 | if (req->timeout.flags) | |
3796 | return -EINVAL; | |
3797 | ||
b29472ee JA |
3798 | return 0; |
3799 | } | |
3800 | ||
11365043 JA |
3801 | /* |
3802 | * Remove or update an existing timeout command | |
3803 | */ | |
fc4df999 | 3804 | static int io_timeout_remove(struct io_kiocb *req) |
11365043 JA |
3805 | { |
3806 | struct io_ring_ctx *ctx = req->ctx; | |
47f46768 | 3807 | int ret; |
11365043 | 3808 | |
11365043 | 3809 | spin_lock_irq(&ctx->completion_lock); |
b29472ee | 3810 | ret = io_timeout_cancel(ctx, req->timeout.addr); |
11365043 | 3811 | |
47f46768 | 3812 | io_cqring_fill_event(req, ret); |
11365043 JA |
3813 | io_commit_cqring(ctx); |
3814 | spin_unlock_irq(&ctx->completion_lock); | |
5262f567 | 3815 | io_cqring_ev_posted(ctx); |
4e88d6e7 JA |
3816 | if (ret < 0) |
3817 | req_set_fail_links(req); | |
ec9c02ad | 3818 | io_put_req(req); |
11365043 | 3819 | return 0; |
5262f567 JA |
3820 | } |
3821 | ||
3529d8c2 | 3822 | static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2d28390a | 3823 | bool is_timeout_link) |
5262f567 | 3824 | { |
ad8a48ac | 3825 | struct io_timeout_data *data; |
a41525ab | 3826 | unsigned flags; |
5262f567 | 3827 | |
ad8a48ac | 3828 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5262f567 | 3829 | return -EINVAL; |
ad8a48ac | 3830 | if (sqe->ioprio || sqe->buf_index || sqe->len != 1) |
a41525ab | 3831 | return -EINVAL; |
2d28390a JA |
3832 | if (sqe->off && is_timeout_link) |
3833 | return -EINVAL; | |
a41525ab JA |
3834 | flags = READ_ONCE(sqe->timeout_flags); |
3835 | if (flags & ~IORING_TIMEOUT_ABS) | |
5262f567 | 3836 | return -EINVAL; |
bdf20073 | 3837 | |
26a61679 JA |
3838 | req->timeout.count = READ_ONCE(sqe->off); |
3839 | ||
3529d8c2 | 3840 | if (!req->io && io_alloc_async_ctx(req)) |
26a61679 JA |
3841 | return -ENOMEM; |
3842 | ||
3843 | data = &req->io->timeout; | |
ad8a48ac | 3844 | data->req = req; |
ad8a48ac JA |
3845 | req->flags |= REQ_F_TIMEOUT; |
3846 | ||
3847 | if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr))) | |
5262f567 JA |
3848 | return -EFAULT; |
3849 | ||
11365043 | 3850 | if (flags & IORING_TIMEOUT_ABS) |
ad8a48ac | 3851 | data->mode = HRTIMER_MODE_ABS; |
11365043 | 3852 | else |
ad8a48ac | 3853 | data->mode = HRTIMER_MODE_REL; |
11365043 | 3854 | |
ad8a48ac JA |
3855 | hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode); |
3856 | return 0; | |
3857 | } | |
3858 | ||
fc4df999 | 3859 | static int io_timeout(struct io_kiocb *req) |
ad8a48ac JA |
3860 | { |
3861 | unsigned count; | |
3862 | struct io_ring_ctx *ctx = req->ctx; | |
3863 | struct io_timeout_data *data; | |
3864 | struct list_head *entry; | |
3865 | unsigned span = 0; | |
ad8a48ac | 3866 | |
2d28390a | 3867 | data = &req->io->timeout; |
93bd25bb | 3868 | |
5262f567 JA |
3869 | /* |
3870 | * sqe->off holds how many events that need to occur for this | |
93bd25bb JA |
3871 | * timeout event to be satisfied. If it isn't set, then this is |
3872 | * a pure timeout request, sequence isn't used. | |
5262f567 | 3873 | */ |
26a61679 | 3874 | count = req->timeout.count; |
93bd25bb JA |
3875 | if (!count) { |
3876 | req->flags |= REQ_F_TIMEOUT_NOSEQ; | |
3877 | spin_lock_irq(&ctx->completion_lock); | |
3878 | entry = ctx->timeout_list.prev; | |
3879 | goto add; | |
3880 | } | |
5262f567 JA |
3881 | |
3882 | req->sequence = ctx->cached_sq_head + count - 1; | |
2d28390a | 3883 | data->seq_offset = count; |
5262f567 JA |
3884 | |
3885 | /* | |
3886 | * Insertion sort, ensuring the first entry in the list is always | |
3887 | * the one we need first. | |
3888 | */ | |
5262f567 JA |
3889 | spin_lock_irq(&ctx->completion_lock); |
3890 | list_for_each_prev(entry, &ctx->timeout_list) { | |
3891 | struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list); | |
5da0fb1a | 3892 | unsigned nxt_sq_head; |
3893 | long long tmp, tmp_nxt; | |
2d28390a | 3894 | u32 nxt_offset = nxt->io->timeout.seq_offset; |
5262f567 | 3895 | |
93bd25bb JA |
3896 | if (nxt->flags & REQ_F_TIMEOUT_NOSEQ) |
3897 | continue; | |
3898 | ||
5da0fb1a | 3899 | /* |
3900 | * Since cached_sq_head + count - 1 can overflow, use type long | |
3901 | * long to store it. | |
3902 | */ | |
3903 | tmp = (long long)ctx->cached_sq_head + count - 1; | |
cc42e0ac PB |
3904 | nxt_sq_head = nxt->sequence - nxt_offset + 1; |
3905 | tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1; | |
5da0fb1a | 3906 | |
3907 | /* | |
3908 | * cached_sq_head may overflow, and it will never overflow twice | |
3909 | * once there is some timeout req still be valid. | |
3910 | */ | |
3911 | if (ctx->cached_sq_head < nxt_sq_head) | |
8b07a65a | 3912 | tmp += UINT_MAX; |
5da0fb1a | 3913 | |
a1f58ba4 | 3914 | if (tmp > tmp_nxt) |
5262f567 | 3915 | break; |
a1f58ba4 | 3916 | |
3917 | /* | |
3918 | * Sequence of reqs after the insert one and itself should | |
3919 | * be adjusted because each timeout req consumes a slot. | |
3920 | */ | |
3921 | span++; | |
3922 | nxt->sequence++; | |
5262f567 | 3923 | } |
a1f58ba4 | 3924 | req->sequence -= span; |
93bd25bb | 3925 | add: |
5262f567 | 3926 | list_add(&req->list, entry); |
ad8a48ac JA |
3927 | data->timer.function = io_timeout_fn; |
3928 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode); | |
5262f567 | 3929 | spin_unlock_irq(&ctx->completion_lock); |
5262f567 JA |
3930 | return 0; |
3931 | } | |
5262f567 | 3932 | |
62755e35 JA |
3933 | static bool io_cancel_cb(struct io_wq_work *work, void *data) |
3934 | { | |
3935 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
3936 | ||
3937 | return req->user_data == (unsigned long) data; | |
3938 | } | |
3939 | ||
e977d6d3 | 3940 | static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr) |
62755e35 | 3941 | { |
62755e35 | 3942 | enum io_wq_cancel cancel_ret; |
62755e35 JA |
3943 | int ret = 0; |
3944 | ||
62755e35 JA |
3945 | cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr); |
3946 | switch (cancel_ret) { | |
3947 | case IO_WQ_CANCEL_OK: | |
3948 | ret = 0; | |
3949 | break; | |
3950 | case IO_WQ_CANCEL_RUNNING: | |
3951 | ret = -EALREADY; | |
3952 | break; | |
3953 | case IO_WQ_CANCEL_NOTFOUND: | |
3954 | ret = -ENOENT; | |
3955 | break; | |
3956 | } | |
3957 | ||
e977d6d3 JA |
3958 | return ret; |
3959 | } | |
3960 | ||
47f46768 JA |
3961 | static void io_async_find_and_cancel(struct io_ring_ctx *ctx, |
3962 | struct io_kiocb *req, __u64 sqe_addr, | |
b0dd8a41 | 3963 | struct io_kiocb **nxt, int success_ret) |
47f46768 JA |
3964 | { |
3965 | unsigned long flags; | |
3966 | int ret; | |
3967 | ||
3968 | ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr); | |
3969 | if (ret != -ENOENT) { | |
3970 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
3971 | goto done; | |
3972 | } | |
3973 | ||
3974 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
3975 | ret = io_timeout_cancel(ctx, sqe_addr); | |
3976 | if (ret != -ENOENT) | |
3977 | goto done; | |
3978 | ret = io_poll_cancel(ctx, sqe_addr); | |
3979 | done: | |
b0dd8a41 JA |
3980 | if (!ret) |
3981 | ret = success_ret; | |
47f46768 JA |
3982 | io_cqring_fill_event(req, ret); |
3983 | io_commit_cqring(ctx); | |
3984 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
3985 | io_cqring_ev_posted(ctx); | |
3986 | ||
4e88d6e7 JA |
3987 | if (ret < 0) |
3988 | req_set_fail_links(req); | |
47f46768 JA |
3989 | io_put_req_find_next(req, nxt); |
3990 | } | |
3991 | ||
3529d8c2 JA |
3992 | static int io_async_cancel_prep(struct io_kiocb *req, |
3993 | const struct io_uring_sqe *sqe) | |
e977d6d3 | 3994 | { |
fbf23849 | 3995 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
e977d6d3 JA |
3996 | return -EINVAL; |
3997 | if (sqe->flags || sqe->ioprio || sqe->off || sqe->len || | |
3998 | sqe->cancel_flags) | |
3999 | return -EINVAL; | |
4000 | ||
fbf23849 JA |
4001 | req->cancel.addr = READ_ONCE(sqe->addr); |
4002 | return 0; | |
4003 | } | |
4004 | ||
4005 | static int io_async_cancel(struct io_kiocb *req, struct io_kiocb **nxt) | |
4006 | { | |
4007 | struct io_ring_ctx *ctx = req->ctx; | |
fbf23849 JA |
4008 | |
4009 | io_async_find_and_cancel(ctx, req, req->cancel.addr, nxt, 0); | |
5262f567 JA |
4010 | return 0; |
4011 | } | |
4012 | ||
05f3fb3c JA |
4013 | static int io_files_update_prep(struct io_kiocb *req, |
4014 | const struct io_uring_sqe *sqe) | |
4015 | { | |
4016 | if (sqe->flags || sqe->ioprio || sqe->rw_flags) | |
4017 | return -EINVAL; | |
4018 | ||
4019 | req->files_update.offset = READ_ONCE(sqe->off); | |
4020 | req->files_update.nr_args = READ_ONCE(sqe->len); | |
4021 | if (!req->files_update.nr_args) | |
4022 | return -EINVAL; | |
4023 | req->files_update.arg = READ_ONCE(sqe->addr); | |
4024 | return 0; | |
4025 | } | |
4026 | ||
4027 | static int io_files_update(struct io_kiocb *req, bool force_nonblock) | |
fbf23849 JA |
4028 | { |
4029 | struct io_ring_ctx *ctx = req->ctx; | |
05f3fb3c JA |
4030 | struct io_uring_files_update up; |
4031 | int ret; | |
fbf23849 | 4032 | |
f86cd20c | 4033 | if (force_nonblock) |
05f3fb3c | 4034 | return -EAGAIN; |
05f3fb3c JA |
4035 | |
4036 | up.offset = req->files_update.offset; | |
4037 | up.fds = req->files_update.arg; | |
4038 | ||
4039 | mutex_lock(&ctx->uring_lock); | |
4040 | ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args); | |
4041 | mutex_unlock(&ctx->uring_lock); | |
4042 | ||
4043 | if (ret < 0) | |
4044 | req_set_fail_links(req); | |
4045 | io_cqring_add_event(req, ret); | |
4046 | io_put_req(req); | |
5262f567 JA |
4047 | return 0; |
4048 | } | |
4049 | ||
3529d8c2 JA |
4050 | static int io_req_defer_prep(struct io_kiocb *req, |
4051 | const struct io_uring_sqe *sqe) | |
f67676d1 | 4052 | { |
e781573e | 4053 | ssize_t ret = 0; |
f67676d1 | 4054 | |
f86cd20c JA |
4055 | if (io_op_defs[req->opcode].file_table) { |
4056 | ret = io_grab_files(req); | |
4057 | if (unlikely(ret)) | |
4058 | return ret; | |
4059 | } | |
4060 | ||
cccf0ee8 JA |
4061 | io_req_work_grab_env(req, &io_op_defs[req->opcode]); |
4062 | ||
d625c6ee | 4063 | switch (req->opcode) { |
e781573e JA |
4064 | case IORING_OP_NOP: |
4065 | break; | |
f67676d1 JA |
4066 | case IORING_OP_READV: |
4067 | case IORING_OP_READ_FIXED: | |
3a6820f2 | 4068 | case IORING_OP_READ: |
3529d8c2 | 4069 | ret = io_read_prep(req, sqe, true); |
f67676d1 JA |
4070 | break; |
4071 | case IORING_OP_WRITEV: | |
4072 | case IORING_OP_WRITE_FIXED: | |
3a6820f2 | 4073 | case IORING_OP_WRITE: |
3529d8c2 | 4074 | ret = io_write_prep(req, sqe, true); |
f67676d1 | 4075 | break; |
0969e783 | 4076 | case IORING_OP_POLL_ADD: |
3529d8c2 | 4077 | ret = io_poll_add_prep(req, sqe); |
0969e783 JA |
4078 | break; |
4079 | case IORING_OP_POLL_REMOVE: | |
3529d8c2 | 4080 | ret = io_poll_remove_prep(req, sqe); |
0969e783 | 4081 | break; |
8ed8d3c3 | 4082 | case IORING_OP_FSYNC: |
3529d8c2 | 4083 | ret = io_prep_fsync(req, sqe); |
8ed8d3c3 JA |
4084 | break; |
4085 | case IORING_OP_SYNC_FILE_RANGE: | |
3529d8c2 | 4086 | ret = io_prep_sfr(req, sqe); |
8ed8d3c3 | 4087 | break; |
03b1230c | 4088 | case IORING_OP_SENDMSG: |
fddaface | 4089 | case IORING_OP_SEND: |
3529d8c2 | 4090 | ret = io_sendmsg_prep(req, sqe); |
03b1230c JA |
4091 | break; |
4092 | case IORING_OP_RECVMSG: | |
fddaface | 4093 | case IORING_OP_RECV: |
3529d8c2 | 4094 | ret = io_recvmsg_prep(req, sqe); |
03b1230c | 4095 | break; |
f499a021 | 4096 | case IORING_OP_CONNECT: |
3529d8c2 | 4097 | ret = io_connect_prep(req, sqe); |
f499a021 | 4098 | break; |
2d28390a | 4099 | case IORING_OP_TIMEOUT: |
3529d8c2 | 4100 | ret = io_timeout_prep(req, sqe, false); |
b7bb4f7d | 4101 | break; |
b29472ee | 4102 | case IORING_OP_TIMEOUT_REMOVE: |
3529d8c2 | 4103 | ret = io_timeout_remove_prep(req, sqe); |
b29472ee | 4104 | break; |
fbf23849 | 4105 | case IORING_OP_ASYNC_CANCEL: |
3529d8c2 | 4106 | ret = io_async_cancel_prep(req, sqe); |
fbf23849 | 4107 | break; |
2d28390a | 4108 | case IORING_OP_LINK_TIMEOUT: |
3529d8c2 | 4109 | ret = io_timeout_prep(req, sqe, true); |
b7bb4f7d | 4110 | break; |
8ed8d3c3 | 4111 | case IORING_OP_ACCEPT: |
3529d8c2 | 4112 | ret = io_accept_prep(req, sqe); |
8ed8d3c3 | 4113 | break; |
d63d1b5e JA |
4114 | case IORING_OP_FALLOCATE: |
4115 | ret = io_fallocate_prep(req, sqe); | |
4116 | break; | |
15b71abe JA |
4117 | case IORING_OP_OPENAT: |
4118 | ret = io_openat_prep(req, sqe); | |
4119 | break; | |
b5dba59e JA |
4120 | case IORING_OP_CLOSE: |
4121 | ret = io_close_prep(req, sqe); | |
4122 | break; | |
05f3fb3c JA |
4123 | case IORING_OP_FILES_UPDATE: |
4124 | ret = io_files_update_prep(req, sqe); | |
4125 | break; | |
eddc7ef5 JA |
4126 | case IORING_OP_STATX: |
4127 | ret = io_statx_prep(req, sqe); | |
4128 | break; | |
4840e418 JA |
4129 | case IORING_OP_FADVISE: |
4130 | ret = io_fadvise_prep(req, sqe); | |
4131 | break; | |
c1ca757b JA |
4132 | case IORING_OP_MADVISE: |
4133 | ret = io_madvise_prep(req, sqe); | |
4134 | break; | |
cebdb986 JA |
4135 | case IORING_OP_OPENAT2: |
4136 | ret = io_openat2_prep(req, sqe); | |
4137 | break; | |
3e4827b0 JA |
4138 | case IORING_OP_EPOLL_CTL: |
4139 | ret = io_epoll_ctl_prep(req, sqe); | |
4140 | break; | |
f67676d1 | 4141 | default: |
e781573e JA |
4142 | printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n", |
4143 | req->opcode); | |
4144 | ret = -EINVAL; | |
b7bb4f7d | 4145 | break; |
f67676d1 JA |
4146 | } |
4147 | ||
b7bb4f7d | 4148 | return ret; |
f67676d1 JA |
4149 | } |
4150 | ||
3529d8c2 | 4151 | static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
de0617e4 | 4152 | { |
a197f664 | 4153 | struct io_ring_ctx *ctx = req->ctx; |
f67676d1 | 4154 | int ret; |
de0617e4 | 4155 | |
9d858b21 BL |
4156 | /* Still need defer if there is pending req in defer list. */ |
4157 | if (!req_need_defer(req) && list_empty(&ctx->defer_list)) | |
de0617e4 JA |
4158 | return 0; |
4159 | ||
3529d8c2 | 4160 | if (!req->io && io_alloc_async_ctx(req)) |
de0617e4 JA |
4161 | return -EAGAIN; |
4162 | ||
3529d8c2 | 4163 | ret = io_req_defer_prep(req, sqe); |
b7bb4f7d | 4164 | if (ret < 0) |
2d28390a | 4165 | return ret; |
2d28390a | 4166 | |
de0617e4 | 4167 | spin_lock_irq(&ctx->completion_lock); |
9d858b21 | 4168 | if (!req_need_defer(req) && list_empty(&ctx->defer_list)) { |
de0617e4 | 4169 | spin_unlock_irq(&ctx->completion_lock); |
de0617e4 JA |
4170 | return 0; |
4171 | } | |
4172 | ||
915967f6 | 4173 | trace_io_uring_defer(ctx, req, req->user_data); |
de0617e4 JA |
4174 | list_add_tail(&req->list, &ctx->defer_list); |
4175 | spin_unlock_irq(&ctx->completion_lock); | |
4176 | return -EIOCBQUEUED; | |
4177 | } | |
4178 | ||
3529d8c2 JA |
4179 | static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
4180 | struct io_kiocb **nxt, bool force_nonblock) | |
2b188cc1 | 4181 | { |
a197f664 | 4182 | struct io_ring_ctx *ctx = req->ctx; |
d625c6ee | 4183 | int ret; |
2b188cc1 | 4184 | |
d625c6ee | 4185 | switch (req->opcode) { |
2b188cc1 | 4186 | case IORING_OP_NOP: |
78e19bbe | 4187 | ret = io_nop(req); |
2b188cc1 JA |
4188 | break; |
4189 | case IORING_OP_READV: | |
edafccee | 4190 | case IORING_OP_READ_FIXED: |
3a6820f2 | 4191 | case IORING_OP_READ: |
3529d8c2 JA |
4192 | if (sqe) { |
4193 | ret = io_read_prep(req, sqe, force_nonblock); | |
4194 | if (ret < 0) | |
4195 | break; | |
4196 | } | |
267bc904 | 4197 | ret = io_read(req, nxt, force_nonblock); |
edafccee | 4198 | break; |
3529d8c2 | 4199 | case IORING_OP_WRITEV: |
edafccee | 4200 | case IORING_OP_WRITE_FIXED: |
3a6820f2 | 4201 | case IORING_OP_WRITE: |
3529d8c2 JA |
4202 | if (sqe) { |
4203 | ret = io_write_prep(req, sqe, force_nonblock); | |
4204 | if (ret < 0) | |
4205 | break; | |
4206 | } | |
267bc904 | 4207 | ret = io_write(req, nxt, force_nonblock); |
2b188cc1 | 4208 | break; |
c992fe29 | 4209 | case IORING_OP_FSYNC: |
3529d8c2 JA |
4210 | if (sqe) { |
4211 | ret = io_prep_fsync(req, sqe); | |
4212 | if (ret < 0) | |
4213 | break; | |
4214 | } | |
fc4df999 | 4215 | ret = io_fsync(req, nxt, force_nonblock); |
c992fe29 | 4216 | break; |
221c5eb2 | 4217 | case IORING_OP_POLL_ADD: |
3529d8c2 JA |
4218 | if (sqe) { |
4219 | ret = io_poll_add_prep(req, sqe); | |
4220 | if (ret) | |
4221 | break; | |
4222 | } | |
fc4df999 | 4223 | ret = io_poll_add(req, nxt); |
221c5eb2 JA |
4224 | break; |
4225 | case IORING_OP_POLL_REMOVE: | |
3529d8c2 JA |
4226 | if (sqe) { |
4227 | ret = io_poll_remove_prep(req, sqe); | |
4228 | if (ret < 0) | |
4229 | break; | |
4230 | } | |
fc4df999 | 4231 | ret = io_poll_remove(req); |
221c5eb2 | 4232 | break; |
5d17b4a4 | 4233 | case IORING_OP_SYNC_FILE_RANGE: |
3529d8c2 JA |
4234 | if (sqe) { |
4235 | ret = io_prep_sfr(req, sqe); | |
4236 | if (ret < 0) | |
4237 | break; | |
4238 | } | |
fc4df999 | 4239 | ret = io_sync_file_range(req, nxt, force_nonblock); |
5d17b4a4 | 4240 | break; |
0fa03c62 | 4241 | case IORING_OP_SENDMSG: |
fddaface | 4242 | case IORING_OP_SEND: |
3529d8c2 JA |
4243 | if (sqe) { |
4244 | ret = io_sendmsg_prep(req, sqe); | |
4245 | if (ret < 0) | |
4246 | break; | |
4247 | } | |
fddaface JA |
4248 | if (req->opcode == IORING_OP_SENDMSG) |
4249 | ret = io_sendmsg(req, nxt, force_nonblock); | |
4250 | else | |
4251 | ret = io_send(req, nxt, force_nonblock); | |
0fa03c62 | 4252 | break; |
aa1fa28f | 4253 | case IORING_OP_RECVMSG: |
fddaface | 4254 | case IORING_OP_RECV: |
3529d8c2 JA |
4255 | if (sqe) { |
4256 | ret = io_recvmsg_prep(req, sqe); | |
4257 | if (ret) | |
4258 | break; | |
4259 | } | |
fddaface JA |
4260 | if (req->opcode == IORING_OP_RECVMSG) |
4261 | ret = io_recvmsg(req, nxt, force_nonblock); | |
4262 | else | |
4263 | ret = io_recv(req, nxt, force_nonblock); | |
aa1fa28f | 4264 | break; |
5262f567 | 4265 | case IORING_OP_TIMEOUT: |
3529d8c2 JA |
4266 | if (sqe) { |
4267 | ret = io_timeout_prep(req, sqe, false); | |
4268 | if (ret) | |
4269 | break; | |
4270 | } | |
fc4df999 | 4271 | ret = io_timeout(req); |
5262f567 | 4272 | break; |
11365043 | 4273 | case IORING_OP_TIMEOUT_REMOVE: |
3529d8c2 JA |
4274 | if (sqe) { |
4275 | ret = io_timeout_remove_prep(req, sqe); | |
4276 | if (ret) | |
4277 | break; | |
4278 | } | |
fc4df999 | 4279 | ret = io_timeout_remove(req); |
11365043 | 4280 | break; |
17f2fe35 | 4281 | case IORING_OP_ACCEPT: |
3529d8c2 JA |
4282 | if (sqe) { |
4283 | ret = io_accept_prep(req, sqe); | |
4284 | if (ret) | |
4285 | break; | |
4286 | } | |
fc4df999 | 4287 | ret = io_accept(req, nxt, force_nonblock); |
17f2fe35 | 4288 | break; |
f8e85cf2 | 4289 | case IORING_OP_CONNECT: |
3529d8c2 JA |
4290 | if (sqe) { |
4291 | ret = io_connect_prep(req, sqe); | |
4292 | if (ret) | |
4293 | break; | |
4294 | } | |
fc4df999 | 4295 | ret = io_connect(req, nxt, force_nonblock); |
f8e85cf2 | 4296 | break; |
62755e35 | 4297 | case IORING_OP_ASYNC_CANCEL: |
3529d8c2 JA |
4298 | if (sqe) { |
4299 | ret = io_async_cancel_prep(req, sqe); | |
4300 | if (ret) | |
4301 | break; | |
4302 | } | |
fc4df999 | 4303 | ret = io_async_cancel(req, nxt); |
62755e35 | 4304 | break; |
d63d1b5e JA |
4305 | case IORING_OP_FALLOCATE: |
4306 | if (sqe) { | |
4307 | ret = io_fallocate_prep(req, sqe); | |
4308 | if (ret) | |
4309 | break; | |
4310 | } | |
4311 | ret = io_fallocate(req, nxt, force_nonblock); | |
4312 | break; | |
15b71abe JA |
4313 | case IORING_OP_OPENAT: |
4314 | if (sqe) { | |
4315 | ret = io_openat_prep(req, sqe); | |
4316 | if (ret) | |
4317 | break; | |
4318 | } | |
4319 | ret = io_openat(req, nxt, force_nonblock); | |
4320 | break; | |
b5dba59e JA |
4321 | case IORING_OP_CLOSE: |
4322 | if (sqe) { | |
4323 | ret = io_close_prep(req, sqe); | |
4324 | if (ret) | |
4325 | break; | |
4326 | } | |
4327 | ret = io_close(req, nxt, force_nonblock); | |
4328 | break; | |
05f3fb3c JA |
4329 | case IORING_OP_FILES_UPDATE: |
4330 | if (sqe) { | |
4331 | ret = io_files_update_prep(req, sqe); | |
4332 | if (ret) | |
4333 | break; | |
4334 | } | |
4335 | ret = io_files_update(req, force_nonblock); | |
4336 | break; | |
eddc7ef5 JA |
4337 | case IORING_OP_STATX: |
4338 | if (sqe) { | |
4339 | ret = io_statx_prep(req, sqe); | |
4340 | if (ret) | |
4341 | break; | |
4342 | } | |
4343 | ret = io_statx(req, nxt, force_nonblock); | |
4344 | break; | |
4840e418 JA |
4345 | case IORING_OP_FADVISE: |
4346 | if (sqe) { | |
4347 | ret = io_fadvise_prep(req, sqe); | |
4348 | if (ret) | |
4349 | break; | |
4350 | } | |
4351 | ret = io_fadvise(req, nxt, force_nonblock); | |
4352 | break; | |
c1ca757b JA |
4353 | case IORING_OP_MADVISE: |
4354 | if (sqe) { | |
4355 | ret = io_madvise_prep(req, sqe); | |
4356 | if (ret) | |
4357 | break; | |
4358 | } | |
4359 | ret = io_madvise(req, nxt, force_nonblock); | |
4360 | break; | |
cebdb986 JA |
4361 | case IORING_OP_OPENAT2: |
4362 | if (sqe) { | |
4363 | ret = io_openat2_prep(req, sqe); | |
4364 | if (ret) | |
4365 | break; | |
4366 | } | |
4367 | ret = io_openat2(req, nxt, force_nonblock); | |
4368 | break; | |
3e4827b0 JA |
4369 | case IORING_OP_EPOLL_CTL: |
4370 | if (sqe) { | |
4371 | ret = io_epoll_ctl_prep(req, sqe); | |
4372 | if (ret) | |
4373 | break; | |
4374 | } | |
4375 | ret = io_epoll_ctl(req, nxt, force_nonblock); | |
4376 | break; | |
2b188cc1 JA |
4377 | default: |
4378 | ret = -EINVAL; | |
4379 | break; | |
4380 | } | |
4381 | ||
def596e9 JA |
4382 | if (ret) |
4383 | return ret; | |
4384 | ||
4385 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
11ba820b JA |
4386 | const bool in_async = io_wq_current_is_worker(); |
4387 | ||
9e645e11 | 4388 | if (req->result == -EAGAIN) |
def596e9 JA |
4389 | return -EAGAIN; |
4390 | ||
11ba820b JA |
4391 | /* workqueue context doesn't hold uring_lock, grab it now */ |
4392 | if (in_async) | |
4393 | mutex_lock(&ctx->uring_lock); | |
4394 | ||
def596e9 | 4395 | io_iopoll_req_issued(req); |
11ba820b JA |
4396 | |
4397 | if (in_async) | |
4398 | mutex_unlock(&ctx->uring_lock); | |
def596e9 JA |
4399 | } |
4400 | ||
4401 | return 0; | |
2b188cc1 JA |
4402 | } |
4403 | ||
561fb04a | 4404 | static void io_wq_submit_work(struct io_wq_work **workptr) |
2b188cc1 | 4405 | { |
561fb04a | 4406 | struct io_wq_work *work = *workptr; |
2b188cc1 | 4407 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
561fb04a JA |
4408 | struct io_kiocb *nxt = NULL; |
4409 | int ret = 0; | |
2b188cc1 | 4410 | |
0c9d5ccd JA |
4411 | /* if NO_CANCEL is set, we must still run the work */ |
4412 | if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) == | |
4413 | IO_WQ_WORK_CANCEL) { | |
561fb04a | 4414 | ret = -ECANCELED; |
0c9d5ccd | 4415 | } |
31b51510 | 4416 | |
561fb04a | 4417 | if (!ret) { |
cf6fd4bd PB |
4418 | req->has_user = (work->flags & IO_WQ_WORK_HAS_MM) != 0; |
4419 | req->in_async = true; | |
561fb04a | 4420 | do { |
3529d8c2 | 4421 | ret = io_issue_sqe(req, NULL, &nxt, false); |
561fb04a JA |
4422 | /* |
4423 | * We can get EAGAIN for polled IO even though we're | |
4424 | * forcing a sync submission from here, since we can't | |
4425 | * wait for request slots on the block side. | |
4426 | */ | |
4427 | if (ret != -EAGAIN) | |
4428 | break; | |
4429 | cond_resched(); | |
4430 | } while (1); | |
4431 | } | |
31b51510 | 4432 | |
561fb04a | 4433 | /* drop submission reference */ |
ec9c02ad | 4434 | io_put_req(req); |
817869d2 | 4435 | |
561fb04a | 4436 | if (ret) { |
4e88d6e7 | 4437 | req_set_fail_links(req); |
78e19bbe | 4438 | io_cqring_add_event(req, ret); |
817869d2 | 4439 | io_put_req(req); |
edafccee | 4440 | } |
2b188cc1 | 4441 | |
561fb04a | 4442 | /* if a dependent link is ready, pass it back */ |
78912934 JA |
4443 | if (!ret && nxt) |
4444 | io_wq_assign_next(workptr, nxt); | |
2b188cc1 JA |
4445 | } |
4446 | ||
15b71abe | 4447 | static int io_req_needs_file(struct io_kiocb *req, int fd) |
9e3aa61a | 4448 | { |
d3656344 | 4449 | if (!io_op_defs[req->opcode].needs_file) |
9e3aa61a | 4450 | return 0; |
d3656344 JA |
4451 | if (fd == -1 && io_op_defs[req->opcode].fd_non_neg) |
4452 | return 0; | |
4453 | return 1; | |
09bb8394 JA |
4454 | } |
4455 | ||
65e19f54 JA |
4456 | static inline struct file *io_file_from_index(struct io_ring_ctx *ctx, |
4457 | int index) | |
4458 | { | |
4459 | struct fixed_file_table *table; | |
4460 | ||
05f3fb3c JA |
4461 | table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT]; |
4462 | return table->files[index & IORING_FILE_TABLE_MASK];; | |
65e19f54 JA |
4463 | } |
4464 | ||
3529d8c2 JA |
4465 | static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req, |
4466 | const struct io_uring_sqe *sqe) | |
09bb8394 | 4467 | { |
a197f664 | 4468 | struct io_ring_ctx *ctx = req->ctx; |
09bb8394 | 4469 | unsigned flags; |
d3656344 | 4470 | int fd; |
09bb8394 | 4471 | |
3529d8c2 JA |
4472 | flags = READ_ONCE(sqe->flags); |
4473 | fd = READ_ONCE(sqe->fd); | |
09bb8394 | 4474 | |
d3656344 JA |
4475 | if (!io_req_needs_file(req, fd)) |
4476 | return 0; | |
09bb8394 JA |
4477 | |
4478 | if (flags & IOSQE_FIXED_FILE) { | |
05f3fb3c | 4479 | if (unlikely(!ctx->file_data || |
09bb8394 JA |
4480 | (unsigned) fd >= ctx->nr_user_files)) |
4481 | return -EBADF; | |
b7620121 | 4482 | fd = array_index_nospec(fd, ctx->nr_user_files); |
65e19f54 JA |
4483 | req->file = io_file_from_index(ctx, fd); |
4484 | if (!req->file) | |
08a45173 | 4485 | return -EBADF; |
09bb8394 | 4486 | req->flags |= REQ_F_FIXED_FILE; |
05f3fb3c | 4487 | percpu_ref_get(&ctx->file_data->refs); |
09bb8394 | 4488 | } else { |
cf6fd4bd | 4489 | if (req->needs_fixed_file) |
09bb8394 | 4490 | return -EBADF; |
c826bd7a | 4491 | trace_io_uring_file_get(ctx, fd); |
09bb8394 JA |
4492 | req->file = io_file_get(state, fd); |
4493 | if (unlikely(!req->file)) | |
4494 | return -EBADF; | |
4495 | } | |
4496 | ||
4497 | return 0; | |
4498 | } | |
4499 | ||
a197f664 | 4500 | static int io_grab_files(struct io_kiocb *req) |
fcb323cc JA |
4501 | { |
4502 | int ret = -EBADF; | |
a197f664 | 4503 | struct io_ring_ctx *ctx = req->ctx; |
fcb323cc | 4504 | |
f86cd20c JA |
4505 | if (req->work.files) |
4506 | return 0; | |
b14cca0c | 4507 | if (!ctx->ring_file) |
b5dba59e JA |
4508 | return -EBADF; |
4509 | ||
fcb323cc JA |
4510 | rcu_read_lock(); |
4511 | spin_lock_irq(&ctx->inflight_lock); | |
4512 | /* | |
4513 | * We use the f_ops->flush() handler to ensure that we can flush | |
4514 | * out work accessing these files if the fd is closed. Check if | |
4515 | * the fd has changed since we started down this path, and disallow | |
4516 | * this operation if it has. | |
4517 | */ | |
b14cca0c | 4518 | if (fcheck(ctx->ring_fd) == ctx->ring_file) { |
fcb323cc JA |
4519 | list_add(&req->inflight_entry, &ctx->inflight_list); |
4520 | req->flags |= REQ_F_INFLIGHT; | |
4521 | req->work.files = current->files; | |
4522 | ret = 0; | |
4523 | } | |
4524 | spin_unlock_irq(&ctx->inflight_lock); | |
4525 | rcu_read_unlock(); | |
4526 | ||
4527 | return ret; | |
4528 | } | |
4529 | ||
2665abfd | 4530 | static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer) |
2b188cc1 | 4531 | { |
ad8a48ac JA |
4532 | struct io_timeout_data *data = container_of(timer, |
4533 | struct io_timeout_data, timer); | |
4534 | struct io_kiocb *req = data->req; | |
2665abfd JA |
4535 | struct io_ring_ctx *ctx = req->ctx; |
4536 | struct io_kiocb *prev = NULL; | |
4537 | unsigned long flags; | |
2665abfd JA |
4538 | |
4539 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
4540 | ||
4541 | /* | |
4542 | * We don't expect the list to be empty, that will only happen if we | |
4543 | * race with the completion of the linked work. | |
4544 | */ | |
4493233e PB |
4545 | if (!list_empty(&req->link_list)) { |
4546 | prev = list_entry(req->link_list.prev, struct io_kiocb, | |
4547 | link_list); | |
5d960724 | 4548 | if (refcount_inc_not_zero(&prev->refs)) { |
4493233e | 4549 | list_del_init(&req->link_list); |
5d960724 JA |
4550 | prev->flags &= ~REQ_F_LINK_TIMEOUT; |
4551 | } else | |
76a46e06 | 4552 | prev = NULL; |
2665abfd JA |
4553 | } |
4554 | ||
4555 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
4556 | ||
4557 | if (prev) { | |
4e88d6e7 | 4558 | req_set_fail_links(prev); |
b0dd8a41 JA |
4559 | io_async_find_and_cancel(ctx, req, prev->user_data, NULL, |
4560 | -ETIME); | |
76a46e06 | 4561 | io_put_req(prev); |
47f46768 JA |
4562 | } else { |
4563 | io_cqring_add_event(req, -ETIME); | |
4564 | io_put_req(req); | |
2665abfd | 4565 | } |
2665abfd JA |
4566 | return HRTIMER_NORESTART; |
4567 | } | |
4568 | ||
ad8a48ac | 4569 | static void io_queue_linked_timeout(struct io_kiocb *req) |
2665abfd | 4570 | { |
76a46e06 | 4571 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd | 4572 | |
76a46e06 JA |
4573 | /* |
4574 | * If the list is now empty, then our linked request finished before | |
4575 | * we got a chance to setup the timer | |
4576 | */ | |
4577 | spin_lock_irq(&ctx->completion_lock); | |
4493233e | 4578 | if (!list_empty(&req->link_list)) { |
2d28390a | 4579 | struct io_timeout_data *data = &req->io->timeout; |
94ae5e77 | 4580 | |
ad8a48ac JA |
4581 | data->timer.function = io_link_timeout_fn; |
4582 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), | |
4583 | data->mode); | |
2665abfd | 4584 | } |
76a46e06 | 4585 | spin_unlock_irq(&ctx->completion_lock); |
2665abfd | 4586 | |
2665abfd | 4587 | /* drop submission reference */ |
76a46e06 JA |
4588 | io_put_req(req); |
4589 | } | |
2665abfd | 4590 | |
ad8a48ac | 4591 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) |
2665abfd JA |
4592 | { |
4593 | struct io_kiocb *nxt; | |
4594 | ||
4595 | if (!(req->flags & REQ_F_LINK)) | |
4596 | return NULL; | |
4597 | ||
4493233e PB |
4598 | nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, |
4599 | link_list); | |
d625c6ee | 4600 | if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT) |
76a46e06 | 4601 | return NULL; |
2665abfd | 4602 | |
76a46e06 | 4603 | req->flags |= REQ_F_LINK_TIMEOUT; |
76a46e06 | 4604 | return nxt; |
2665abfd JA |
4605 | } |
4606 | ||
3529d8c2 | 4607 | static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2b188cc1 | 4608 | { |
4a0a7a18 | 4609 | struct io_kiocb *linked_timeout; |
f9bd67f6 | 4610 | struct io_kiocb *nxt = NULL; |
e0c5c576 | 4611 | int ret; |
2b188cc1 | 4612 | |
4a0a7a18 JA |
4613 | again: |
4614 | linked_timeout = io_prep_linked_timeout(req); | |
4615 | ||
3529d8c2 | 4616 | ret = io_issue_sqe(req, sqe, &nxt, true); |
491381ce JA |
4617 | |
4618 | /* | |
4619 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
4620 | * doesn't support non-blocking read/write attempts | |
4621 | */ | |
4622 | if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) || | |
4623 | (req->flags & REQ_F_MUST_PUNT))) { | |
86a761f8 | 4624 | punt: |
f86cd20c | 4625 | if (io_op_defs[req->opcode].file_table) { |
bbad27b2 PB |
4626 | ret = io_grab_files(req); |
4627 | if (ret) | |
4628 | goto err; | |
2b188cc1 | 4629 | } |
bbad27b2 PB |
4630 | |
4631 | /* | |
4632 | * Queued up for async execution, worker will release | |
4633 | * submit reference when the iocb is actually submitted. | |
4634 | */ | |
4635 | io_queue_async_work(req); | |
4a0a7a18 | 4636 | goto done_req; |
2b188cc1 | 4637 | } |
e65ef56d | 4638 | |
fcb323cc | 4639 | err: |
76a46e06 | 4640 | /* drop submission reference */ |
ec9c02ad | 4641 | io_put_req(req); |
e65ef56d | 4642 | |
f9bd67f6 | 4643 | if (linked_timeout) { |
76a46e06 | 4644 | if (!ret) |
f9bd67f6 | 4645 | io_queue_linked_timeout(linked_timeout); |
76a46e06 | 4646 | else |
f9bd67f6 | 4647 | io_put_req(linked_timeout); |
76a46e06 JA |
4648 | } |
4649 | ||
e65ef56d | 4650 | /* and drop final reference, if we failed */ |
9e645e11 | 4651 | if (ret) { |
78e19bbe | 4652 | io_cqring_add_event(req, ret); |
4e88d6e7 | 4653 | req_set_fail_links(req); |
e65ef56d | 4654 | io_put_req(req); |
9e645e11 | 4655 | } |
4a0a7a18 JA |
4656 | done_req: |
4657 | if (nxt) { | |
4658 | req = nxt; | |
4659 | nxt = NULL; | |
86a761f8 PB |
4660 | |
4661 | if (req->flags & REQ_F_FORCE_ASYNC) | |
4662 | goto punt; | |
4a0a7a18 JA |
4663 | goto again; |
4664 | } | |
2b188cc1 JA |
4665 | } |
4666 | ||
3529d8c2 | 4667 | static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4fe2c963 JL |
4668 | { |
4669 | int ret; | |
4670 | ||
3529d8c2 | 4671 | ret = io_req_defer(req, sqe); |
4fe2c963 JL |
4672 | if (ret) { |
4673 | if (ret != -EIOCBQUEUED) { | |
1118591a | 4674 | fail_req: |
78e19bbe | 4675 | io_cqring_add_event(req, ret); |
4e88d6e7 | 4676 | req_set_fail_links(req); |
78e19bbe | 4677 | io_double_put_req(req); |
4fe2c963 | 4678 | } |
2550878f | 4679 | } else if (req->flags & REQ_F_FORCE_ASYNC) { |
1118591a PB |
4680 | ret = io_req_defer_prep(req, sqe); |
4681 | if (unlikely(ret < 0)) | |
4682 | goto fail_req; | |
ce35a47a JA |
4683 | /* |
4684 | * Never try inline submit of IOSQE_ASYNC is set, go straight | |
4685 | * to async execution. | |
4686 | */ | |
4687 | req->work.flags |= IO_WQ_WORK_CONCURRENT; | |
4688 | io_queue_async_work(req); | |
4689 | } else { | |
3529d8c2 | 4690 | __io_queue_sqe(req, sqe); |
ce35a47a | 4691 | } |
4fe2c963 JL |
4692 | } |
4693 | ||
1b4a51b6 | 4694 | static inline void io_queue_link_head(struct io_kiocb *req) |
4fe2c963 | 4695 | { |
94ae5e77 | 4696 | if (unlikely(req->flags & REQ_F_FAIL_LINK)) { |
1b4a51b6 PB |
4697 | io_cqring_add_event(req, -ECANCELED); |
4698 | io_double_put_req(req); | |
4699 | } else | |
3529d8c2 | 4700 | io_queue_sqe(req, NULL); |
4fe2c963 JL |
4701 | } |
4702 | ||
4e88d6e7 | 4703 | #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \ |
ce35a47a | 4704 | IOSQE_IO_HARDLINK | IOSQE_ASYNC) |
9e645e11 | 4705 | |
3529d8c2 JA |
4706 | static bool io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
4707 | struct io_submit_state *state, struct io_kiocb **link) | |
9e645e11 | 4708 | { |
75c6a039 | 4709 | const struct cred *old_creds = NULL; |
a197f664 | 4710 | struct io_ring_ctx *ctx = req->ctx; |
32fe525b | 4711 | unsigned int sqe_flags; |
75c6a039 | 4712 | int ret, id; |
9e645e11 | 4713 | |
32fe525b | 4714 | sqe_flags = READ_ONCE(sqe->flags); |
9e645e11 JA |
4715 | |
4716 | /* enforce forwards compatibility on users */ | |
32fe525b | 4717 | if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) { |
9e645e11 | 4718 | ret = -EINVAL; |
196be95c | 4719 | goto err_req; |
9e645e11 JA |
4720 | } |
4721 | ||
75c6a039 JA |
4722 | id = READ_ONCE(sqe->personality); |
4723 | if (id) { | |
4724 | const struct cred *personality_creds; | |
4725 | ||
4726 | personality_creds = idr_find(&ctx->personality_idr, id); | |
4727 | if (unlikely(!personality_creds)) { | |
4728 | ret = -EINVAL; | |
4729 | goto err_req; | |
4730 | } | |
4731 | old_creds = override_creds(personality_creds); | |
4732 | } | |
4733 | ||
6b47ee6e PB |
4734 | /* same numerical values with corresponding REQ_F_*, safe to copy */ |
4735 | req->flags |= sqe_flags & (IOSQE_IO_DRAIN|IOSQE_IO_HARDLINK| | |
4736 | IOSQE_ASYNC); | |
9e645e11 | 4737 | |
3529d8c2 | 4738 | ret = io_req_set_file(state, req, sqe); |
9e645e11 JA |
4739 | if (unlikely(ret)) { |
4740 | err_req: | |
78e19bbe JA |
4741 | io_cqring_add_event(req, ret); |
4742 | io_double_put_req(req); | |
75c6a039 JA |
4743 | if (old_creds) |
4744 | revert_creds(old_creds); | |
2e6e1fde | 4745 | return false; |
9e645e11 JA |
4746 | } |
4747 | ||
9e645e11 JA |
4748 | /* |
4749 | * If we already have a head request, queue this one for async | |
4750 | * submittal once the head completes. If we don't have a head but | |
4751 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
4752 | * submitted sync once the chain is complete. If none of those | |
4753 | * conditions are true (normal request), then just queue it. | |
4754 | */ | |
4755 | if (*link) { | |
9d76377f | 4756 | struct io_kiocb *head = *link; |
4e88d6e7 | 4757 | |
8cdf2193 PB |
4758 | /* |
4759 | * Taking sequential execution of a link, draining both sides | |
4760 | * of the link also fullfils IOSQE_IO_DRAIN semantics for all | |
4761 | * requests in the link. So, it drains the head and the | |
4762 | * next after the link request. The last one is done via | |
4763 | * drain_next flag to persist the effect across calls. | |
4764 | */ | |
711be031 PB |
4765 | if (sqe_flags & IOSQE_IO_DRAIN) { |
4766 | head->flags |= REQ_F_IO_DRAIN; | |
4767 | ctx->drain_next = 1; | |
4768 | } | |
b7bb4f7d | 4769 | if (io_alloc_async_ctx(req)) { |
9e645e11 JA |
4770 | ret = -EAGAIN; |
4771 | goto err_req; | |
4772 | } | |
4773 | ||
3529d8c2 | 4774 | ret = io_req_defer_prep(req, sqe); |
2d28390a | 4775 | if (ret) { |
4e88d6e7 | 4776 | /* fail even hard links since we don't submit */ |
9d76377f | 4777 | head->flags |= REQ_F_FAIL_LINK; |
f67676d1 | 4778 | goto err_req; |
2d28390a | 4779 | } |
9d76377f PB |
4780 | trace_io_uring_link(ctx, req, head); |
4781 | list_add_tail(&req->link_list, &head->link_list); | |
32fe525b PB |
4782 | |
4783 | /* last request of a link, enqueue the link */ | |
4784 | if (!(sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK))) { | |
4785 | io_queue_link_head(head); | |
4786 | *link = NULL; | |
4787 | } | |
9e645e11 | 4788 | } else { |
711be031 PB |
4789 | if (unlikely(ctx->drain_next)) { |
4790 | req->flags |= REQ_F_IO_DRAIN; | |
4791 | req->ctx->drain_next = 0; | |
4792 | } | |
4793 | if (sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) { | |
4794 | req->flags |= REQ_F_LINK; | |
711be031 PB |
4795 | INIT_LIST_HEAD(&req->link_list); |
4796 | ret = io_req_defer_prep(req, sqe); | |
4797 | if (ret) | |
4798 | req->flags |= REQ_F_FAIL_LINK; | |
4799 | *link = req; | |
4800 | } else { | |
4801 | io_queue_sqe(req, sqe); | |
4802 | } | |
9e645e11 | 4803 | } |
2e6e1fde | 4804 | |
75c6a039 JA |
4805 | if (old_creds) |
4806 | revert_creds(old_creds); | |
2e6e1fde | 4807 | return true; |
9e645e11 JA |
4808 | } |
4809 | ||
9a56a232 JA |
4810 | /* |
4811 | * Batched submission is done, ensure local IO is flushed out. | |
4812 | */ | |
4813 | static void io_submit_state_end(struct io_submit_state *state) | |
4814 | { | |
4815 | blk_finish_plug(&state->plug); | |
3d6770fb | 4816 | io_file_put(state); |
2579f913 JA |
4817 | if (state->free_reqs) |
4818 | kmem_cache_free_bulk(req_cachep, state->free_reqs, | |
4819 | &state->reqs[state->cur_req]); | |
9a56a232 JA |
4820 | } |
4821 | ||
4822 | /* | |
4823 | * Start submission side cache. | |
4824 | */ | |
4825 | static void io_submit_state_start(struct io_submit_state *state, | |
22efde59 | 4826 | unsigned int max_ios) |
9a56a232 JA |
4827 | { |
4828 | blk_start_plug(&state->plug); | |
2579f913 | 4829 | state->free_reqs = 0; |
9a56a232 JA |
4830 | state->file = NULL; |
4831 | state->ios_left = max_ios; | |
4832 | } | |
4833 | ||
2b188cc1 JA |
4834 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
4835 | { | |
75b28aff | 4836 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 4837 | |
caf582c6 PB |
4838 | /* |
4839 | * Ensure any loads from the SQEs are done at this point, | |
4840 | * since once we write the new head, the application could | |
4841 | * write new data to them. | |
4842 | */ | |
4843 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
4844 | } |
4845 | ||
2b188cc1 | 4846 | /* |
3529d8c2 | 4847 | * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory |
2b188cc1 JA |
4848 | * that is mapped by userspace. This means that care needs to be taken to |
4849 | * ensure that reads are stable, as we cannot rely on userspace always | |
4850 | * being a good citizen. If members of the sqe are validated and then later | |
4851 | * used, it's important that those reads are done through READ_ONCE() to | |
4852 | * prevent a re-load down the line. | |
4853 | */ | |
3529d8c2 JA |
4854 | static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req, |
4855 | const struct io_uring_sqe **sqe_ptr) | |
2b188cc1 | 4856 | { |
75b28aff | 4857 | u32 *sq_array = ctx->sq_array; |
2b188cc1 JA |
4858 | unsigned head; |
4859 | ||
4860 | /* | |
4861 | * The cached sq head (or cq tail) serves two purposes: | |
4862 | * | |
4863 | * 1) allows us to batch the cost of updating the user visible | |
4864 | * head updates. | |
4865 | * 2) allows the kernel side to track the head on its own, even | |
4866 | * though the application is the one updating it. | |
4867 | */ | |
ee7d46d9 | 4868 | head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]); |
9835d6fa | 4869 | if (likely(head < ctx->sq_entries)) { |
cf6fd4bd PB |
4870 | /* |
4871 | * All io need record the previous position, if LINK vs DARIN, | |
4872 | * it can be used to mark the position of the first IO in the | |
4873 | * link list. | |
4874 | */ | |
4875 | req->sequence = ctx->cached_sq_head; | |
3529d8c2 JA |
4876 | *sqe_ptr = &ctx->sq_sqes[head]; |
4877 | req->opcode = READ_ONCE((*sqe_ptr)->opcode); | |
4878 | req->user_data = READ_ONCE((*sqe_ptr)->user_data); | |
2b188cc1 JA |
4879 | ctx->cached_sq_head++; |
4880 | return true; | |
4881 | } | |
4882 | ||
4883 | /* drop invalid entries */ | |
4884 | ctx->cached_sq_head++; | |
498ccd9e | 4885 | ctx->cached_sq_dropped++; |
ee7d46d9 | 4886 | WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped); |
2b188cc1 JA |
4887 | return false; |
4888 | } | |
4889 | ||
fb5ccc98 | 4890 | static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr, |
ae9428ca PB |
4891 | struct file *ring_file, int ring_fd, |
4892 | struct mm_struct **mm, bool async) | |
6c271ce2 JA |
4893 | { |
4894 | struct io_submit_state state, *statep = NULL; | |
9e645e11 | 4895 | struct io_kiocb *link = NULL; |
9e645e11 | 4896 | int i, submitted = 0; |
95a1b3ff | 4897 | bool mm_fault = false; |
6c271ce2 | 4898 | |
c4a2ed72 | 4899 | /* if we have a backlog and couldn't flush it all, return BUSY */ |
ad3eb2c8 JA |
4900 | if (test_bit(0, &ctx->sq_check_overflow)) { |
4901 | if (!list_empty(&ctx->cq_overflow_list) && | |
4902 | !io_cqring_overflow_flush(ctx, false)) | |
4903 | return -EBUSY; | |
4904 | } | |
6c271ce2 | 4905 | |
ee7d46d9 PB |
4906 | /* make sure SQ entry isn't read before tail */ |
4907 | nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx)); | |
9ef4f124 | 4908 | |
2b85edfc PB |
4909 | if (!percpu_ref_tryget_many(&ctx->refs, nr)) |
4910 | return -EAGAIN; | |
6c271ce2 JA |
4911 | |
4912 | if (nr > IO_PLUG_THRESHOLD) { | |
22efde59 | 4913 | io_submit_state_start(&state, nr); |
6c271ce2 JA |
4914 | statep = &state; |
4915 | } | |
4916 | ||
b14cca0c PB |
4917 | ctx->ring_fd = ring_fd; |
4918 | ctx->ring_file = ring_file; | |
4919 | ||
6c271ce2 | 4920 | for (i = 0; i < nr; i++) { |
3529d8c2 | 4921 | const struct io_uring_sqe *sqe; |
196be95c | 4922 | struct io_kiocb *req; |
fb5ccc98 | 4923 | |
196be95c PB |
4924 | req = io_get_req(ctx, statep); |
4925 | if (unlikely(!req)) { | |
4926 | if (!submitted) | |
4927 | submitted = -EAGAIN; | |
fb5ccc98 | 4928 | break; |
196be95c | 4929 | } |
3529d8c2 | 4930 | if (!io_get_sqring(ctx, req, &sqe)) { |
2b85edfc | 4931 | __io_req_do_free(req); |
196be95c PB |
4932 | break; |
4933 | } | |
fb5ccc98 | 4934 | |
d3656344 JA |
4935 | /* will complete beyond this point, count as submitted */ |
4936 | submitted++; | |
4937 | ||
4938 | if (unlikely(req->opcode >= IORING_OP_LAST)) { | |
4939 | io_cqring_add_event(req, -EINVAL); | |
4940 | io_double_put_req(req); | |
196be95c PB |
4941 | break; |
4942 | } | |
fb5ccc98 | 4943 | |
d3656344 | 4944 | if (io_op_defs[req->opcode].needs_mm && !*mm) { |
95a1b3ff PB |
4945 | mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm); |
4946 | if (!mm_fault) { | |
4947 | use_mm(ctx->sqo_mm); | |
4948 | *mm = ctx->sqo_mm; | |
4949 | } | |
9e645e11 | 4950 | } |
9e645e11 | 4951 | |
cf6fd4bd PB |
4952 | req->has_user = *mm != NULL; |
4953 | req->in_async = async; | |
4954 | req->needs_fixed_file = async; | |
354420f7 JA |
4955 | trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data, |
4956 | true, async); | |
3529d8c2 | 4957 | if (!io_submit_sqe(req, sqe, statep, &link)) |
2e6e1fde | 4958 | break; |
6c271ce2 JA |
4959 | } |
4960 | ||
9466f437 PB |
4961 | if (unlikely(submitted != nr)) { |
4962 | int ref_used = (submitted == -EAGAIN) ? 0 : submitted; | |
4963 | ||
4964 | percpu_ref_put_many(&ctx->refs, nr - ref_used); | |
4965 | } | |
9e645e11 | 4966 | if (link) |
1b4a51b6 | 4967 | io_queue_link_head(link); |
6c271ce2 JA |
4968 | if (statep) |
4969 | io_submit_state_end(&state); | |
4970 | ||
ae9428ca PB |
4971 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
4972 | io_commit_sqring(ctx); | |
4973 | ||
6c271ce2 JA |
4974 | return submitted; |
4975 | } | |
4976 | ||
4977 | static int io_sq_thread(void *data) | |
4978 | { | |
6c271ce2 JA |
4979 | struct io_ring_ctx *ctx = data; |
4980 | struct mm_struct *cur_mm = NULL; | |
181e448d | 4981 | const struct cred *old_cred; |
6c271ce2 JA |
4982 | mm_segment_t old_fs; |
4983 | DEFINE_WAIT(wait); | |
4984 | unsigned inflight; | |
4985 | unsigned long timeout; | |
c1edbf5f | 4986 | int ret; |
6c271ce2 | 4987 | |
206aefde | 4988 | complete(&ctx->completions[1]); |
a4c0b3de | 4989 | |
6c271ce2 JA |
4990 | old_fs = get_fs(); |
4991 | set_fs(USER_DS); | |
181e448d | 4992 | old_cred = override_creds(ctx->creds); |
6c271ce2 | 4993 | |
c1edbf5f | 4994 | ret = timeout = inflight = 0; |
2bbcd6d3 | 4995 | while (!kthread_should_park()) { |
fb5ccc98 | 4996 | unsigned int to_submit; |
6c271ce2 JA |
4997 | |
4998 | if (inflight) { | |
4999 | unsigned nr_events = 0; | |
5000 | ||
5001 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
2b2ed975 JA |
5002 | /* |
5003 | * inflight is the count of the maximum possible | |
5004 | * entries we submitted, but it can be smaller | |
5005 | * if we dropped some of them. If we don't have | |
5006 | * poll entries available, then we know that we | |
5007 | * have nothing left to poll for. Reset the | |
5008 | * inflight count to zero in that case. | |
5009 | */ | |
5010 | mutex_lock(&ctx->uring_lock); | |
5011 | if (!list_empty(&ctx->poll_list)) | |
5012 | __io_iopoll_check(ctx, &nr_events, 0); | |
5013 | else | |
5014 | inflight = 0; | |
5015 | mutex_unlock(&ctx->uring_lock); | |
6c271ce2 JA |
5016 | } else { |
5017 | /* | |
5018 | * Normal IO, just pretend everything completed. | |
5019 | * We don't have to poll completions for that. | |
5020 | */ | |
5021 | nr_events = inflight; | |
5022 | } | |
5023 | ||
5024 | inflight -= nr_events; | |
5025 | if (!inflight) | |
5026 | timeout = jiffies + ctx->sq_thread_idle; | |
5027 | } | |
5028 | ||
fb5ccc98 | 5029 | to_submit = io_sqring_entries(ctx); |
c1edbf5f JA |
5030 | |
5031 | /* | |
5032 | * If submit got -EBUSY, flag us as needing the application | |
5033 | * to enter the kernel to reap and flush events. | |
5034 | */ | |
5035 | if (!to_submit || ret == -EBUSY) { | |
6c271ce2 JA |
5036 | /* |
5037 | * We're polling. If we're within the defined idle | |
5038 | * period, then let us spin without work before going | |
c1edbf5f JA |
5039 | * to sleep. The exception is if we got EBUSY doing |
5040 | * more IO, we should wait for the application to | |
5041 | * reap events and wake us up. | |
6c271ce2 | 5042 | */ |
c1edbf5f JA |
5043 | if (inflight || |
5044 | (!time_after(jiffies, timeout) && ret != -EBUSY)) { | |
9831a90c | 5045 | cond_resched(); |
6c271ce2 JA |
5046 | continue; |
5047 | } | |
5048 | ||
5049 | /* | |
5050 | * Drop cur_mm before scheduling, we can't hold it for | |
5051 | * long periods (or over schedule()). Do this before | |
5052 | * adding ourselves to the waitqueue, as the unuse/drop | |
5053 | * may sleep. | |
5054 | */ | |
5055 | if (cur_mm) { | |
5056 | unuse_mm(cur_mm); | |
5057 | mmput(cur_mm); | |
5058 | cur_mm = NULL; | |
5059 | } | |
5060 | ||
5061 | prepare_to_wait(&ctx->sqo_wait, &wait, | |
5062 | TASK_INTERRUPTIBLE); | |
5063 | ||
5064 | /* Tell userspace we may need a wakeup call */ | |
75b28aff | 5065 | ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP; |
0d7bae69 SB |
5066 | /* make sure to read SQ tail after writing flags */ |
5067 | smp_mb(); | |
6c271ce2 | 5068 | |
fb5ccc98 | 5069 | to_submit = io_sqring_entries(ctx); |
c1edbf5f | 5070 | if (!to_submit || ret == -EBUSY) { |
2bbcd6d3 | 5071 | if (kthread_should_park()) { |
6c271ce2 JA |
5072 | finish_wait(&ctx->sqo_wait, &wait); |
5073 | break; | |
5074 | } | |
5075 | if (signal_pending(current)) | |
5076 | flush_signals(current); | |
5077 | schedule(); | |
5078 | finish_wait(&ctx->sqo_wait, &wait); | |
5079 | ||
75b28aff | 5080 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
5081 | continue; |
5082 | } | |
5083 | finish_wait(&ctx->sqo_wait, &wait); | |
5084 | ||
75b28aff | 5085 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
5086 | } |
5087 | ||
8a4955ff | 5088 | mutex_lock(&ctx->uring_lock); |
1d7bb1d5 | 5089 | ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true); |
8a4955ff | 5090 | mutex_unlock(&ctx->uring_lock); |
1d7bb1d5 JA |
5091 | if (ret > 0) |
5092 | inflight += ret; | |
6c271ce2 JA |
5093 | } |
5094 | ||
5095 | set_fs(old_fs); | |
5096 | if (cur_mm) { | |
5097 | unuse_mm(cur_mm); | |
5098 | mmput(cur_mm); | |
5099 | } | |
181e448d | 5100 | revert_creds(old_cred); |
06058632 | 5101 | |
2bbcd6d3 | 5102 | kthread_parkme(); |
06058632 | 5103 | |
6c271ce2 JA |
5104 | return 0; |
5105 | } | |
5106 | ||
bda52162 JA |
5107 | struct io_wait_queue { |
5108 | struct wait_queue_entry wq; | |
5109 | struct io_ring_ctx *ctx; | |
5110 | unsigned to_wait; | |
5111 | unsigned nr_timeouts; | |
5112 | }; | |
5113 | ||
1d7bb1d5 | 5114 | static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush) |
bda52162 JA |
5115 | { |
5116 | struct io_ring_ctx *ctx = iowq->ctx; | |
5117 | ||
5118 | /* | |
d195a66e | 5119 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
5120 | * started waiting. For timeouts, we always want to return to userspace, |
5121 | * regardless of event count. | |
5122 | */ | |
1d7bb1d5 | 5123 | return io_cqring_events(ctx, noflush) >= iowq->to_wait || |
bda52162 JA |
5124 | atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
5125 | } | |
5126 | ||
5127 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
5128 | int wake_flags, void *key) | |
5129 | { | |
5130 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
5131 | wq); | |
5132 | ||
1d7bb1d5 JA |
5133 | /* use noflush == true, as we can't safely rely on locking context */ |
5134 | if (!io_should_wake(iowq, true)) | |
bda52162 JA |
5135 | return -1; |
5136 | ||
5137 | return autoremove_wake_function(curr, mode, wake_flags, key); | |
5138 | } | |
5139 | ||
2b188cc1 JA |
5140 | /* |
5141 | * Wait until events become available, if we don't already have some. The | |
5142 | * application must reap them itself, as they reside on the shared cq ring. | |
5143 | */ | |
5144 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
5145 | const sigset_t __user *sig, size_t sigsz) | |
5146 | { | |
bda52162 JA |
5147 | struct io_wait_queue iowq = { |
5148 | .wq = { | |
5149 | .private = current, | |
5150 | .func = io_wake_function, | |
5151 | .entry = LIST_HEAD_INIT(iowq.wq.entry), | |
5152 | }, | |
5153 | .ctx = ctx, | |
5154 | .to_wait = min_events, | |
5155 | }; | |
75b28aff | 5156 | struct io_rings *rings = ctx->rings; |
e9ffa5c2 | 5157 | int ret = 0; |
2b188cc1 | 5158 | |
1d7bb1d5 | 5159 | if (io_cqring_events(ctx, false) >= min_events) |
2b188cc1 JA |
5160 | return 0; |
5161 | ||
5162 | if (sig) { | |
9e75ad5d AB |
5163 | #ifdef CONFIG_COMPAT |
5164 | if (in_compat_syscall()) | |
5165 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 5166 | sigsz); |
9e75ad5d AB |
5167 | else |
5168 | #endif | |
b772434b | 5169 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 5170 | |
2b188cc1 JA |
5171 | if (ret) |
5172 | return ret; | |
5173 | } | |
5174 | ||
bda52162 | 5175 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
c826bd7a | 5176 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 JA |
5177 | do { |
5178 | prepare_to_wait_exclusive(&ctx->wait, &iowq.wq, | |
5179 | TASK_INTERRUPTIBLE); | |
1d7bb1d5 | 5180 | if (io_should_wake(&iowq, false)) |
bda52162 JA |
5181 | break; |
5182 | schedule(); | |
5183 | if (signal_pending(current)) { | |
e9ffa5c2 | 5184 | ret = -EINTR; |
bda52162 JA |
5185 | break; |
5186 | } | |
5187 | } while (1); | |
5188 | finish_wait(&ctx->wait, &iowq.wq); | |
5189 | ||
e9ffa5c2 | 5190 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 5191 | |
75b28aff | 5192 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
5193 | } |
5194 | ||
6b06314c JA |
5195 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
5196 | { | |
5197 | #if defined(CONFIG_UNIX) | |
5198 | if (ctx->ring_sock) { | |
5199 | struct sock *sock = ctx->ring_sock->sk; | |
5200 | struct sk_buff *skb; | |
5201 | ||
5202 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
5203 | kfree_skb(skb); | |
5204 | } | |
5205 | #else | |
5206 | int i; | |
5207 | ||
65e19f54 JA |
5208 | for (i = 0; i < ctx->nr_user_files; i++) { |
5209 | struct file *file; | |
5210 | ||
5211 | file = io_file_from_index(ctx, i); | |
5212 | if (file) | |
5213 | fput(file); | |
5214 | } | |
6b06314c JA |
5215 | #endif |
5216 | } | |
5217 | ||
05f3fb3c JA |
5218 | static void io_file_ref_kill(struct percpu_ref *ref) |
5219 | { | |
5220 | struct fixed_file_data *data; | |
5221 | ||
5222 | data = container_of(ref, struct fixed_file_data, refs); | |
5223 | complete(&data->done); | |
5224 | } | |
5225 | ||
6b06314c JA |
5226 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) |
5227 | { | |
05f3fb3c | 5228 | struct fixed_file_data *data = ctx->file_data; |
65e19f54 JA |
5229 | unsigned nr_tables, i; |
5230 | ||
05f3fb3c | 5231 | if (!data) |
6b06314c JA |
5232 | return -ENXIO; |
5233 | ||
05f3fb3c | 5234 | /* protect against inflight atomic switch, which drops the ref */ |
05f3fb3c | 5235 | percpu_ref_get(&data->refs); |
e46a7950 JA |
5236 | /* wait for existing switches */ |
5237 | flush_work(&data->ref_work); | |
05f3fb3c JA |
5238 | percpu_ref_kill_and_confirm(&data->refs, io_file_ref_kill); |
5239 | wait_for_completion(&data->done); | |
5240 | percpu_ref_put(&data->refs); | |
e46a7950 JA |
5241 | /* flush potential new switch */ |
5242 | flush_work(&data->ref_work); | |
05f3fb3c JA |
5243 | percpu_ref_exit(&data->refs); |
5244 | ||
6b06314c | 5245 | __io_sqe_files_unregister(ctx); |
65e19f54 JA |
5246 | nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE); |
5247 | for (i = 0; i < nr_tables; i++) | |
05f3fb3c JA |
5248 | kfree(data->table[i].files); |
5249 | kfree(data->table); | |
5250 | kfree(data); | |
5251 | ctx->file_data = NULL; | |
6b06314c JA |
5252 | ctx->nr_user_files = 0; |
5253 | return 0; | |
5254 | } | |
5255 | ||
6c271ce2 JA |
5256 | static void io_sq_thread_stop(struct io_ring_ctx *ctx) |
5257 | { | |
5258 | if (ctx->sqo_thread) { | |
206aefde | 5259 | wait_for_completion(&ctx->completions[1]); |
2bbcd6d3 RP |
5260 | /* |
5261 | * The park is a bit of a work-around, without it we get | |
5262 | * warning spews on shutdown with SQPOLL set and affinity | |
5263 | * set to a single CPU. | |
5264 | */ | |
06058632 | 5265 | kthread_park(ctx->sqo_thread); |
6c271ce2 JA |
5266 | kthread_stop(ctx->sqo_thread); |
5267 | ctx->sqo_thread = NULL; | |
5268 | } | |
5269 | } | |
5270 | ||
6b06314c JA |
5271 | static void io_finish_async(struct io_ring_ctx *ctx) |
5272 | { | |
6c271ce2 JA |
5273 | io_sq_thread_stop(ctx); |
5274 | ||
561fb04a JA |
5275 | if (ctx->io_wq) { |
5276 | io_wq_destroy(ctx->io_wq); | |
5277 | ctx->io_wq = NULL; | |
6b06314c JA |
5278 | } |
5279 | } | |
5280 | ||
5281 | #if defined(CONFIG_UNIX) | |
6b06314c JA |
5282 | /* |
5283 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
5284 | * the io_uring can be safely unregistered on process exit, even if we have | |
5285 | * loops in the file referencing. | |
5286 | */ | |
5287 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
5288 | { | |
5289 | struct sock *sk = ctx->ring_sock->sk; | |
5290 | struct scm_fp_list *fpl; | |
5291 | struct sk_buff *skb; | |
08a45173 | 5292 | int i, nr_files; |
6b06314c JA |
5293 | |
5294 | if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) { | |
5295 | unsigned long inflight = ctx->user->unix_inflight + nr; | |
5296 | ||
5297 | if (inflight > task_rlimit(current, RLIMIT_NOFILE)) | |
5298 | return -EMFILE; | |
5299 | } | |
5300 | ||
5301 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); | |
5302 | if (!fpl) | |
5303 | return -ENOMEM; | |
5304 | ||
5305 | skb = alloc_skb(0, GFP_KERNEL); | |
5306 | if (!skb) { | |
5307 | kfree(fpl); | |
5308 | return -ENOMEM; | |
5309 | } | |
5310 | ||
5311 | skb->sk = sk; | |
6b06314c | 5312 | |
08a45173 | 5313 | nr_files = 0; |
6b06314c JA |
5314 | fpl->user = get_uid(ctx->user); |
5315 | for (i = 0; i < nr; i++) { | |
65e19f54 JA |
5316 | struct file *file = io_file_from_index(ctx, i + offset); |
5317 | ||
5318 | if (!file) | |
08a45173 | 5319 | continue; |
65e19f54 | 5320 | fpl->fp[nr_files] = get_file(file); |
08a45173 JA |
5321 | unix_inflight(fpl->user, fpl->fp[nr_files]); |
5322 | nr_files++; | |
6b06314c JA |
5323 | } |
5324 | ||
08a45173 JA |
5325 | if (nr_files) { |
5326 | fpl->max = SCM_MAX_FD; | |
5327 | fpl->count = nr_files; | |
5328 | UNIXCB(skb).fp = fpl; | |
05f3fb3c | 5329 | skb->destructor = unix_destruct_scm; |
08a45173 JA |
5330 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); |
5331 | skb_queue_head(&sk->sk_receive_queue, skb); | |
6b06314c | 5332 | |
08a45173 JA |
5333 | for (i = 0; i < nr_files; i++) |
5334 | fput(fpl->fp[i]); | |
5335 | } else { | |
5336 | kfree_skb(skb); | |
5337 | kfree(fpl); | |
5338 | } | |
6b06314c JA |
5339 | |
5340 | return 0; | |
5341 | } | |
5342 | ||
5343 | /* | |
5344 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
5345 | * causes regular reference counting to break down. We rely on the UNIX | |
5346 | * garbage collection to take care of this problem for us. | |
5347 | */ | |
5348 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
5349 | { | |
5350 | unsigned left, total; | |
5351 | int ret = 0; | |
5352 | ||
5353 | total = 0; | |
5354 | left = ctx->nr_user_files; | |
5355 | while (left) { | |
5356 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
5357 | |
5358 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
5359 | if (ret) | |
5360 | break; | |
5361 | left -= this_files; | |
5362 | total += this_files; | |
5363 | } | |
5364 | ||
5365 | if (!ret) | |
5366 | return 0; | |
5367 | ||
5368 | while (total < ctx->nr_user_files) { | |
65e19f54 JA |
5369 | struct file *file = io_file_from_index(ctx, total); |
5370 | ||
5371 | if (file) | |
5372 | fput(file); | |
6b06314c JA |
5373 | total++; |
5374 | } | |
5375 | ||
5376 | return ret; | |
5377 | } | |
5378 | #else | |
5379 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
5380 | { | |
5381 | return 0; | |
5382 | } | |
5383 | #endif | |
5384 | ||
65e19f54 JA |
5385 | static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables, |
5386 | unsigned nr_files) | |
5387 | { | |
5388 | int i; | |
5389 | ||
5390 | for (i = 0; i < nr_tables; i++) { | |
05f3fb3c | 5391 | struct fixed_file_table *table = &ctx->file_data->table[i]; |
65e19f54 JA |
5392 | unsigned this_files; |
5393 | ||
5394 | this_files = min(nr_files, IORING_MAX_FILES_TABLE); | |
5395 | table->files = kcalloc(this_files, sizeof(struct file *), | |
5396 | GFP_KERNEL); | |
5397 | if (!table->files) | |
5398 | break; | |
5399 | nr_files -= this_files; | |
5400 | } | |
5401 | ||
5402 | if (i == nr_tables) | |
5403 | return 0; | |
5404 | ||
5405 | for (i = 0; i < nr_tables; i++) { | |
05f3fb3c | 5406 | struct fixed_file_table *table = &ctx->file_data->table[i]; |
65e19f54 JA |
5407 | kfree(table->files); |
5408 | } | |
5409 | return 1; | |
5410 | } | |
5411 | ||
05f3fb3c JA |
5412 | static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file) |
5413 | { | |
5414 | #if defined(CONFIG_UNIX) | |
5415 | struct sock *sock = ctx->ring_sock->sk; | |
5416 | struct sk_buff_head list, *head = &sock->sk_receive_queue; | |
5417 | struct sk_buff *skb; | |
5418 | int i; | |
5419 | ||
5420 | __skb_queue_head_init(&list); | |
5421 | ||
5422 | /* | |
5423 | * Find the skb that holds this file in its SCM_RIGHTS. When found, | |
5424 | * remove this entry and rearrange the file array. | |
5425 | */ | |
5426 | skb = skb_dequeue(head); | |
5427 | while (skb) { | |
5428 | struct scm_fp_list *fp; | |
5429 | ||
5430 | fp = UNIXCB(skb).fp; | |
5431 | for (i = 0; i < fp->count; i++) { | |
5432 | int left; | |
5433 | ||
5434 | if (fp->fp[i] != file) | |
5435 | continue; | |
5436 | ||
5437 | unix_notinflight(fp->user, fp->fp[i]); | |
5438 | left = fp->count - 1 - i; | |
5439 | if (left) { | |
5440 | memmove(&fp->fp[i], &fp->fp[i + 1], | |
5441 | left * sizeof(struct file *)); | |
5442 | } | |
5443 | fp->count--; | |
5444 | if (!fp->count) { | |
5445 | kfree_skb(skb); | |
5446 | skb = NULL; | |
5447 | } else { | |
5448 | __skb_queue_tail(&list, skb); | |
5449 | } | |
5450 | fput(file); | |
5451 | file = NULL; | |
5452 | break; | |
5453 | } | |
5454 | ||
5455 | if (!file) | |
5456 | break; | |
5457 | ||
5458 | __skb_queue_tail(&list, skb); | |
5459 | ||
5460 | skb = skb_dequeue(head); | |
5461 | } | |
5462 | ||
5463 | if (skb_peek(&list)) { | |
5464 | spin_lock_irq(&head->lock); | |
5465 | while ((skb = __skb_dequeue(&list)) != NULL) | |
5466 | __skb_queue_tail(head, skb); | |
5467 | spin_unlock_irq(&head->lock); | |
5468 | } | |
5469 | #else | |
5470 | fput(file); | |
5471 | #endif | |
5472 | } | |
5473 | ||
5474 | struct io_file_put { | |
5475 | struct llist_node llist; | |
5476 | struct file *file; | |
5477 | struct completion *done; | |
5478 | }; | |
5479 | ||
5480 | static void io_ring_file_ref_switch(struct work_struct *work) | |
65e19f54 | 5481 | { |
05f3fb3c JA |
5482 | struct io_file_put *pfile, *tmp; |
5483 | struct fixed_file_data *data; | |
5484 | struct llist_node *node; | |
65e19f54 | 5485 | |
05f3fb3c | 5486 | data = container_of(work, struct fixed_file_data, ref_work); |
65e19f54 | 5487 | |
05f3fb3c JA |
5488 | while ((node = llist_del_all(&data->put_llist)) != NULL) { |
5489 | llist_for_each_entry_safe(pfile, tmp, node, llist) { | |
5490 | io_ring_file_put(data->ctx, pfile->file); | |
5491 | if (pfile->done) | |
5492 | complete(pfile->done); | |
5493 | else | |
5494 | kfree(pfile); | |
5495 | } | |
65e19f54 JA |
5496 | } |
5497 | ||
05f3fb3c JA |
5498 | percpu_ref_get(&data->refs); |
5499 | percpu_ref_switch_to_percpu(&data->refs); | |
5500 | } | |
65e19f54 | 5501 | |
05f3fb3c JA |
5502 | static void io_file_data_ref_zero(struct percpu_ref *ref) |
5503 | { | |
5504 | struct fixed_file_data *data; | |
5505 | ||
5506 | data = container_of(ref, struct fixed_file_data, refs); | |
5507 | ||
5508 | /* we can't safely switch from inside this context, punt to wq */ | |
5509 | queue_work(system_wq, &data->ref_work); | |
65e19f54 JA |
5510 | } |
5511 | ||
6b06314c JA |
5512 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, |
5513 | unsigned nr_args) | |
5514 | { | |
5515 | __s32 __user *fds = (__s32 __user *) arg; | |
65e19f54 | 5516 | unsigned nr_tables; |
05f3fb3c | 5517 | struct file *file; |
6b06314c JA |
5518 | int fd, ret = 0; |
5519 | unsigned i; | |
5520 | ||
05f3fb3c | 5521 | if (ctx->file_data) |
6b06314c JA |
5522 | return -EBUSY; |
5523 | if (!nr_args) | |
5524 | return -EINVAL; | |
5525 | if (nr_args > IORING_MAX_FIXED_FILES) | |
5526 | return -EMFILE; | |
5527 | ||
05f3fb3c JA |
5528 | ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL); |
5529 | if (!ctx->file_data) | |
5530 | return -ENOMEM; | |
5531 | ctx->file_data->ctx = ctx; | |
5532 | init_completion(&ctx->file_data->done); | |
5533 | ||
65e19f54 | 5534 | nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE); |
05f3fb3c JA |
5535 | ctx->file_data->table = kcalloc(nr_tables, |
5536 | sizeof(struct fixed_file_table), | |
65e19f54 | 5537 | GFP_KERNEL); |
05f3fb3c JA |
5538 | if (!ctx->file_data->table) { |
5539 | kfree(ctx->file_data); | |
5540 | ctx->file_data = NULL; | |
6b06314c | 5541 | return -ENOMEM; |
05f3fb3c JA |
5542 | } |
5543 | ||
5544 | if (percpu_ref_init(&ctx->file_data->refs, io_file_data_ref_zero, | |
5545 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) { | |
5546 | kfree(ctx->file_data->table); | |
5547 | kfree(ctx->file_data); | |
5548 | ctx->file_data = NULL; | |
6b06314c | 5549 | return -ENOMEM; |
05f3fb3c JA |
5550 | } |
5551 | ctx->file_data->put_llist.first = NULL; | |
5552 | INIT_WORK(&ctx->file_data->ref_work, io_ring_file_ref_switch); | |
6b06314c | 5553 | |
65e19f54 | 5554 | if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) { |
05f3fb3c JA |
5555 | percpu_ref_exit(&ctx->file_data->refs); |
5556 | kfree(ctx->file_data->table); | |
5557 | kfree(ctx->file_data); | |
5558 | ctx->file_data = NULL; | |
65e19f54 JA |
5559 | return -ENOMEM; |
5560 | } | |
5561 | ||
08a45173 | 5562 | for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { |
65e19f54 JA |
5563 | struct fixed_file_table *table; |
5564 | unsigned index; | |
5565 | ||
6b06314c JA |
5566 | ret = -EFAULT; |
5567 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) | |
5568 | break; | |
08a45173 JA |
5569 | /* allow sparse sets */ |
5570 | if (fd == -1) { | |
5571 | ret = 0; | |
5572 | continue; | |
5573 | } | |
6b06314c | 5574 | |
05f3fb3c | 5575 | table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT]; |
65e19f54 | 5576 | index = i & IORING_FILE_TABLE_MASK; |
05f3fb3c | 5577 | file = fget(fd); |
6b06314c JA |
5578 | |
5579 | ret = -EBADF; | |
05f3fb3c | 5580 | if (!file) |
6b06314c | 5581 | break; |
05f3fb3c | 5582 | |
6b06314c JA |
5583 | /* |
5584 | * Don't allow io_uring instances to be registered. If UNIX | |
5585 | * isn't enabled, then this causes a reference cycle and this | |
5586 | * instance can never get freed. If UNIX is enabled we'll | |
5587 | * handle it just fine, but there's still no point in allowing | |
5588 | * a ring fd as it doesn't support regular read/write anyway. | |
5589 | */ | |
05f3fb3c JA |
5590 | if (file->f_op == &io_uring_fops) { |
5591 | fput(file); | |
6b06314c JA |
5592 | break; |
5593 | } | |
6b06314c | 5594 | ret = 0; |
05f3fb3c | 5595 | table->files[index] = file; |
6b06314c JA |
5596 | } |
5597 | ||
5598 | if (ret) { | |
65e19f54 | 5599 | for (i = 0; i < ctx->nr_user_files; i++) { |
65e19f54 JA |
5600 | file = io_file_from_index(ctx, i); |
5601 | if (file) | |
5602 | fput(file); | |
5603 | } | |
5604 | for (i = 0; i < nr_tables; i++) | |
05f3fb3c | 5605 | kfree(ctx->file_data->table[i].files); |
6b06314c | 5606 | |
05f3fb3c JA |
5607 | kfree(ctx->file_data->table); |
5608 | kfree(ctx->file_data); | |
5609 | ctx->file_data = NULL; | |
6b06314c JA |
5610 | ctx->nr_user_files = 0; |
5611 | return ret; | |
5612 | } | |
5613 | ||
5614 | ret = io_sqe_files_scm(ctx); | |
5615 | if (ret) | |
5616 | io_sqe_files_unregister(ctx); | |
5617 | ||
5618 | return ret; | |
5619 | } | |
5620 | ||
c3a31e60 JA |
5621 | static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file, |
5622 | int index) | |
5623 | { | |
5624 | #if defined(CONFIG_UNIX) | |
5625 | struct sock *sock = ctx->ring_sock->sk; | |
5626 | struct sk_buff_head *head = &sock->sk_receive_queue; | |
5627 | struct sk_buff *skb; | |
5628 | ||
5629 | /* | |
5630 | * See if we can merge this file into an existing skb SCM_RIGHTS | |
5631 | * file set. If there's no room, fall back to allocating a new skb | |
5632 | * and filling it in. | |
5633 | */ | |
5634 | spin_lock_irq(&head->lock); | |
5635 | skb = skb_peek(head); | |
5636 | if (skb) { | |
5637 | struct scm_fp_list *fpl = UNIXCB(skb).fp; | |
5638 | ||
5639 | if (fpl->count < SCM_MAX_FD) { | |
5640 | __skb_unlink(skb, head); | |
5641 | spin_unlock_irq(&head->lock); | |
5642 | fpl->fp[fpl->count] = get_file(file); | |
5643 | unix_inflight(fpl->user, fpl->fp[fpl->count]); | |
5644 | fpl->count++; | |
5645 | spin_lock_irq(&head->lock); | |
5646 | __skb_queue_head(head, skb); | |
5647 | } else { | |
5648 | skb = NULL; | |
5649 | } | |
5650 | } | |
5651 | spin_unlock_irq(&head->lock); | |
5652 | ||
5653 | if (skb) { | |
5654 | fput(file); | |
5655 | return 0; | |
5656 | } | |
5657 | ||
5658 | return __io_sqe_files_scm(ctx, 1, index); | |
5659 | #else | |
5660 | return 0; | |
5661 | #endif | |
5662 | } | |
5663 | ||
05f3fb3c | 5664 | static void io_atomic_switch(struct percpu_ref *ref) |
c3a31e60 | 5665 | { |
05f3fb3c JA |
5666 | struct fixed_file_data *data; |
5667 | ||
5668 | data = container_of(ref, struct fixed_file_data, refs); | |
5669 | clear_bit(FFD_F_ATOMIC, &data->state); | |
5670 | } | |
5671 | ||
5672 | static bool io_queue_file_removal(struct fixed_file_data *data, | |
5673 | struct file *file) | |
5674 | { | |
5675 | struct io_file_put *pfile, pfile_stack; | |
5676 | DECLARE_COMPLETION_ONSTACK(done); | |
5677 | ||
5678 | /* | |
5679 | * If we fail allocating the struct we need for doing async reomval | |
5680 | * of this file, just punt to sync and wait for it. | |
5681 | */ | |
5682 | pfile = kzalloc(sizeof(*pfile), GFP_KERNEL); | |
5683 | if (!pfile) { | |
5684 | pfile = &pfile_stack; | |
5685 | pfile->done = &done; | |
5686 | } | |
5687 | ||
5688 | pfile->file = file; | |
5689 | llist_add(&pfile->llist, &data->put_llist); | |
5690 | ||
5691 | if (pfile == &pfile_stack) { | |
5692 | if (!test_and_set_bit(FFD_F_ATOMIC, &data->state)) { | |
5693 | percpu_ref_put(&data->refs); | |
5694 | percpu_ref_switch_to_atomic(&data->refs, | |
5695 | io_atomic_switch); | |
5696 | } | |
5697 | wait_for_completion(&done); | |
5698 | flush_work(&data->ref_work); | |
5699 | return false; | |
5700 | } | |
5701 | ||
5702 | return true; | |
5703 | } | |
5704 | ||
5705 | static int __io_sqe_files_update(struct io_ring_ctx *ctx, | |
5706 | struct io_uring_files_update *up, | |
5707 | unsigned nr_args) | |
5708 | { | |
5709 | struct fixed_file_data *data = ctx->file_data; | |
5710 | bool ref_switch = false; | |
5711 | struct file *file; | |
c3a31e60 JA |
5712 | __s32 __user *fds; |
5713 | int fd, i, err; | |
5714 | __u32 done; | |
5715 | ||
05f3fb3c | 5716 | if (check_add_overflow(up->offset, nr_args, &done)) |
c3a31e60 JA |
5717 | return -EOVERFLOW; |
5718 | if (done > ctx->nr_user_files) | |
5719 | return -EINVAL; | |
5720 | ||
5721 | done = 0; | |
05f3fb3c | 5722 | fds = u64_to_user_ptr(up->fds); |
c3a31e60 | 5723 | while (nr_args) { |
65e19f54 JA |
5724 | struct fixed_file_table *table; |
5725 | unsigned index; | |
5726 | ||
c3a31e60 JA |
5727 | err = 0; |
5728 | if (copy_from_user(&fd, &fds[done], sizeof(fd))) { | |
5729 | err = -EFAULT; | |
5730 | break; | |
5731 | } | |
05f3fb3c JA |
5732 | i = array_index_nospec(up->offset, ctx->nr_user_files); |
5733 | table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT]; | |
65e19f54 JA |
5734 | index = i & IORING_FILE_TABLE_MASK; |
5735 | if (table->files[index]) { | |
05f3fb3c | 5736 | file = io_file_from_index(ctx, index); |
65e19f54 | 5737 | table->files[index] = NULL; |
05f3fb3c JA |
5738 | if (io_queue_file_removal(data, file)) |
5739 | ref_switch = true; | |
c3a31e60 JA |
5740 | } |
5741 | if (fd != -1) { | |
c3a31e60 JA |
5742 | file = fget(fd); |
5743 | if (!file) { | |
5744 | err = -EBADF; | |
5745 | break; | |
5746 | } | |
5747 | /* | |
5748 | * Don't allow io_uring instances to be registered. If | |
5749 | * UNIX isn't enabled, then this causes a reference | |
5750 | * cycle and this instance can never get freed. If UNIX | |
5751 | * is enabled we'll handle it just fine, but there's | |
5752 | * still no point in allowing a ring fd as it doesn't | |
5753 | * support regular read/write anyway. | |
5754 | */ | |
5755 | if (file->f_op == &io_uring_fops) { | |
5756 | fput(file); | |
5757 | err = -EBADF; | |
5758 | break; | |
5759 | } | |
65e19f54 | 5760 | table->files[index] = file; |
c3a31e60 JA |
5761 | err = io_sqe_file_register(ctx, file, i); |
5762 | if (err) | |
5763 | break; | |
5764 | } | |
5765 | nr_args--; | |
5766 | done++; | |
05f3fb3c JA |
5767 | up->offset++; |
5768 | } | |
5769 | ||
5770 | if (ref_switch && !test_and_set_bit(FFD_F_ATOMIC, &data->state)) { | |
5771 | percpu_ref_put(&data->refs); | |
5772 | percpu_ref_switch_to_atomic(&data->refs, io_atomic_switch); | |
c3a31e60 JA |
5773 | } |
5774 | ||
5775 | return done ? done : err; | |
5776 | } | |
05f3fb3c JA |
5777 | static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg, |
5778 | unsigned nr_args) | |
5779 | { | |
5780 | struct io_uring_files_update up; | |
5781 | ||
5782 | if (!ctx->file_data) | |
5783 | return -ENXIO; | |
5784 | if (!nr_args) | |
5785 | return -EINVAL; | |
5786 | if (copy_from_user(&up, arg, sizeof(up))) | |
5787 | return -EFAULT; | |
5788 | if (up.resv) | |
5789 | return -EINVAL; | |
5790 | ||
5791 | return __io_sqe_files_update(ctx, &up, nr_args); | |
5792 | } | |
c3a31e60 | 5793 | |
7d723065 JA |
5794 | static void io_put_work(struct io_wq_work *work) |
5795 | { | |
5796 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
5797 | ||
5798 | io_put_req(req); | |
5799 | } | |
5800 | ||
5801 | static void io_get_work(struct io_wq_work *work) | |
5802 | { | |
5803 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
5804 | ||
5805 | refcount_inc(&req->refs); | |
5806 | } | |
5807 | ||
24369c2e PB |
5808 | static int io_init_wq_offload(struct io_ring_ctx *ctx, |
5809 | struct io_uring_params *p) | |
5810 | { | |
5811 | struct io_wq_data data; | |
5812 | struct fd f; | |
5813 | struct io_ring_ctx *ctx_attach; | |
5814 | unsigned int concurrency; | |
5815 | int ret = 0; | |
5816 | ||
5817 | data.user = ctx->user; | |
5818 | data.get_work = io_get_work; | |
5819 | data.put_work = io_put_work; | |
5820 | ||
5821 | if (!(p->flags & IORING_SETUP_ATTACH_WQ)) { | |
5822 | /* Do QD, or 4 * CPUS, whatever is smallest */ | |
5823 | concurrency = min(ctx->sq_entries, 4 * num_online_cpus()); | |
5824 | ||
5825 | ctx->io_wq = io_wq_create(concurrency, &data); | |
5826 | if (IS_ERR(ctx->io_wq)) { | |
5827 | ret = PTR_ERR(ctx->io_wq); | |
5828 | ctx->io_wq = NULL; | |
5829 | } | |
5830 | return ret; | |
5831 | } | |
5832 | ||
5833 | f = fdget(p->wq_fd); | |
5834 | if (!f.file) | |
5835 | return -EBADF; | |
5836 | ||
5837 | if (f.file->f_op != &io_uring_fops) { | |
5838 | ret = -EINVAL; | |
5839 | goto out_fput; | |
5840 | } | |
5841 | ||
5842 | ctx_attach = f.file->private_data; | |
5843 | /* @io_wq is protected by holding the fd */ | |
5844 | if (!io_wq_get(ctx_attach->io_wq, &data)) { | |
5845 | ret = -EINVAL; | |
5846 | goto out_fput; | |
5847 | } | |
5848 | ||
5849 | ctx->io_wq = ctx_attach->io_wq; | |
5850 | out_fput: | |
5851 | fdput(f); | |
5852 | return ret; | |
5853 | } | |
5854 | ||
6c271ce2 JA |
5855 | static int io_sq_offload_start(struct io_ring_ctx *ctx, |
5856 | struct io_uring_params *p) | |
2b188cc1 JA |
5857 | { |
5858 | int ret; | |
5859 | ||
6c271ce2 | 5860 | init_waitqueue_head(&ctx->sqo_wait); |
2b188cc1 JA |
5861 | mmgrab(current->mm); |
5862 | ctx->sqo_mm = current->mm; | |
5863 | ||
6c271ce2 | 5864 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3ec482d1 JA |
5865 | ret = -EPERM; |
5866 | if (!capable(CAP_SYS_ADMIN)) | |
5867 | goto err; | |
5868 | ||
917257da JA |
5869 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
5870 | if (!ctx->sq_thread_idle) | |
5871 | ctx->sq_thread_idle = HZ; | |
5872 | ||
6c271ce2 | 5873 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 5874 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 5875 | |
917257da | 5876 | ret = -EINVAL; |
44a9bd18 JA |
5877 | if (cpu >= nr_cpu_ids) |
5878 | goto err; | |
7889f44d | 5879 | if (!cpu_online(cpu)) |
917257da JA |
5880 | goto err; |
5881 | ||
6c271ce2 JA |
5882 | ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread, |
5883 | ctx, cpu, | |
5884 | "io_uring-sq"); | |
5885 | } else { | |
5886 | ctx->sqo_thread = kthread_create(io_sq_thread, ctx, | |
5887 | "io_uring-sq"); | |
5888 | } | |
5889 | if (IS_ERR(ctx->sqo_thread)) { | |
5890 | ret = PTR_ERR(ctx->sqo_thread); | |
5891 | ctx->sqo_thread = NULL; | |
5892 | goto err; | |
5893 | } | |
5894 | wake_up_process(ctx->sqo_thread); | |
5895 | } else if (p->flags & IORING_SETUP_SQ_AFF) { | |
5896 | /* Can't have SQ_AFF without SQPOLL */ | |
5897 | ret = -EINVAL; | |
5898 | goto err; | |
5899 | } | |
5900 | ||
24369c2e PB |
5901 | ret = io_init_wq_offload(ctx, p); |
5902 | if (ret) | |
2b188cc1 | 5903 | goto err; |
2b188cc1 JA |
5904 | |
5905 | return 0; | |
5906 | err: | |
54a91f3b | 5907 | io_finish_async(ctx); |
2b188cc1 JA |
5908 | mmdrop(ctx->sqo_mm); |
5909 | ctx->sqo_mm = NULL; | |
5910 | return ret; | |
5911 | } | |
5912 | ||
5913 | static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages) | |
5914 | { | |
5915 | atomic_long_sub(nr_pages, &user->locked_vm); | |
5916 | } | |
5917 | ||
5918 | static int io_account_mem(struct user_struct *user, unsigned long nr_pages) | |
5919 | { | |
5920 | unsigned long page_limit, cur_pages, new_pages; | |
5921 | ||
5922 | /* Don't allow more pages than we can safely lock */ | |
5923 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
5924 | ||
5925 | do { | |
5926 | cur_pages = atomic_long_read(&user->locked_vm); | |
5927 | new_pages = cur_pages + nr_pages; | |
5928 | if (new_pages > page_limit) | |
5929 | return -ENOMEM; | |
5930 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
5931 | new_pages) != cur_pages); | |
5932 | ||
5933 | return 0; | |
5934 | } | |
5935 | ||
5936 | static void io_mem_free(void *ptr) | |
5937 | { | |
52e04ef4 MR |
5938 | struct page *page; |
5939 | ||
5940 | if (!ptr) | |
5941 | return; | |
2b188cc1 | 5942 | |
52e04ef4 | 5943 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
5944 | if (put_page_testzero(page)) |
5945 | free_compound_page(page); | |
5946 | } | |
5947 | ||
5948 | static void *io_mem_alloc(size_t size) | |
5949 | { | |
5950 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
5951 | __GFP_NORETRY; | |
5952 | ||
5953 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
5954 | } | |
5955 | ||
75b28aff HV |
5956 | static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries, |
5957 | size_t *sq_offset) | |
5958 | { | |
5959 | struct io_rings *rings; | |
5960 | size_t off, sq_array_size; | |
5961 | ||
5962 | off = struct_size(rings, cqes, cq_entries); | |
5963 | if (off == SIZE_MAX) | |
5964 | return SIZE_MAX; | |
5965 | ||
5966 | #ifdef CONFIG_SMP | |
5967 | off = ALIGN(off, SMP_CACHE_BYTES); | |
5968 | if (off == 0) | |
5969 | return SIZE_MAX; | |
5970 | #endif | |
5971 | ||
5972 | sq_array_size = array_size(sizeof(u32), sq_entries); | |
5973 | if (sq_array_size == SIZE_MAX) | |
5974 | return SIZE_MAX; | |
5975 | ||
5976 | if (check_add_overflow(off, sq_array_size, &off)) | |
5977 | return SIZE_MAX; | |
5978 | ||
5979 | if (sq_offset) | |
5980 | *sq_offset = off; | |
5981 | ||
5982 | return off; | |
5983 | } | |
5984 | ||
2b188cc1 JA |
5985 | static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries) |
5986 | { | |
75b28aff | 5987 | size_t pages; |
2b188cc1 | 5988 | |
75b28aff HV |
5989 | pages = (size_t)1 << get_order( |
5990 | rings_size(sq_entries, cq_entries, NULL)); | |
5991 | pages += (size_t)1 << get_order( | |
5992 | array_size(sizeof(struct io_uring_sqe), sq_entries)); | |
2b188cc1 | 5993 | |
75b28aff | 5994 | return pages; |
2b188cc1 JA |
5995 | } |
5996 | ||
edafccee JA |
5997 | static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx) |
5998 | { | |
5999 | int i, j; | |
6000 | ||
6001 | if (!ctx->user_bufs) | |
6002 | return -ENXIO; | |
6003 | ||
6004 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
6005 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
6006 | ||
6007 | for (j = 0; j < imu->nr_bvecs; j++) | |
27c4d3a3 | 6008 | put_user_page(imu->bvec[j].bv_page); |
edafccee JA |
6009 | |
6010 | if (ctx->account_mem) | |
6011 | io_unaccount_mem(ctx->user, imu->nr_bvecs); | |
d4ef6475 | 6012 | kvfree(imu->bvec); |
edafccee JA |
6013 | imu->nr_bvecs = 0; |
6014 | } | |
6015 | ||
6016 | kfree(ctx->user_bufs); | |
6017 | ctx->user_bufs = NULL; | |
6018 | ctx->nr_user_bufs = 0; | |
6019 | return 0; | |
6020 | } | |
6021 | ||
6022 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
6023 | void __user *arg, unsigned index) | |
6024 | { | |
6025 | struct iovec __user *src; | |
6026 | ||
6027 | #ifdef CONFIG_COMPAT | |
6028 | if (ctx->compat) { | |
6029 | struct compat_iovec __user *ciovs; | |
6030 | struct compat_iovec ciov; | |
6031 | ||
6032 | ciovs = (struct compat_iovec __user *) arg; | |
6033 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
6034 | return -EFAULT; | |
6035 | ||
d55e5f5b | 6036 | dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base); |
edafccee JA |
6037 | dst->iov_len = ciov.iov_len; |
6038 | return 0; | |
6039 | } | |
6040 | #endif | |
6041 | src = (struct iovec __user *) arg; | |
6042 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
6043 | return -EFAULT; | |
6044 | return 0; | |
6045 | } | |
6046 | ||
6047 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, | |
6048 | unsigned nr_args) | |
6049 | { | |
6050 | struct vm_area_struct **vmas = NULL; | |
6051 | struct page **pages = NULL; | |
6052 | int i, j, got_pages = 0; | |
6053 | int ret = -EINVAL; | |
6054 | ||
6055 | if (ctx->user_bufs) | |
6056 | return -EBUSY; | |
6057 | if (!nr_args || nr_args > UIO_MAXIOV) | |
6058 | return -EINVAL; | |
6059 | ||
6060 | ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf), | |
6061 | GFP_KERNEL); | |
6062 | if (!ctx->user_bufs) | |
6063 | return -ENOMEM; | |
6064 | ||
6065 | for (i = 0; i < nr_args; i++) { | |
6066 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
6067 | unsigned long off, start, end, ubuf; | |
6068 | int pret, nr_pages; | |
6069 | struct iovec iov; | |
6070 | size_t size; | |
6071 | ||
6072 | ret = io_copy_iov(ctx, &iov, arg, i); | |
6073 | if (ret) | |
a278682d | 6074 | goto err; |
edafccee JA |
6075 | |
6076 | /* | |
6077 | * Don't impose further limits on the size and buffer | |
6078 | * constraints here, we'll -EINVAL later when IO is | |
6079 | * submitted if they are wrong. | |
6080 | */ | |
6081 | ret = -EFAULT; | |
6082 | if (!iov.iov_base || !iov.iov_len) | |
6083 | goto err; | |
6084 | ||
6085 | /* arbitrary limit, but we need something */ | |
6086 | if (iov.iov_len > SZ_1G) | |
6087 | goto err; | |
6088 | ||
6089 | ubuf = (unsigned long) iov.iov_base; | |
6090 | end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
6091 | start = ubuf >> PAGE_SHIFT; | |
6092 | nr_pages = end - start; | |
6093 | ||
6094 | if (ctx->account_mem) { | |
6095 | ret = io_account_mem(ctx->user, nr_pages); | |
6096 | if (ret) | |
6097 | goto err; | |
6098 | } | |
6099 | ||
6100 | ret = 0; | |
6101 | if (!pages || nr_pages > got_pages) { | |
6102 | kfree(vmas); | |
6103 | kfree(pages); | |
d4ef6475 | 6104 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), |
edafccee | 6105 | GFP_KERNEL); |
d4ef6475 | 6106 | vmas = kvmalloc_array(nr_pages, |
edafccee JA |
6107 | sizeof(struct vm_area_struct *), |
6108 | GFP_KERNEL); | |
6109 | if (!pages || !vmas) { | |
6110 | ret = -ENOMEM; | |
6111 | if (ctx->account_mem) | |
6112 | io_unaccount_mem(ctx->user, nr_pages); | |
6113 | goto err; | |
6114 | } | |
6115 | got_pages = nr_pages; | |
6116 | } | |
6117 | ||
d4ef6475 | 6118 | imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec), |
edafccee JA |
6119 | GFP_KERNEL); |
6120 | ret = -ENOMEM; | |
6121 | if (!imu->bvec) { | |
6122 | if (ctx->account_mem) | |
6123 | io_unaccount_mem(ctx->user, nr_pages); | |
6124 | goto err; | |
6125 | } | |
6126 | ||
6127 | ret = 0; | |
6128 | down_read(¤t->mm->mmap_sem); | |
932f4a63 IW |
6129 | pret = get_user_pages(ubuf, nr_pages, |
6130 | FOLL_WRITE | FOLL_LONGTERM, | |
6131 | pages, vmas); | |
edafccee JA |
6132 | if (pret == nr_pages) { |
6133 | /* don't support file backed memory */ | |
6134 | for (j = 0; j < nr_pages; j++) { | |
6135 | struct vm_area_struct *vma = vmas[j]; | |
6136 | ||
6137 | if (vma->vm_file && | |
6138 | !is_file_hugepages(vma->vm_file)) { | |
6139 | ret = -EOPNOTSUPP; | |
6140 | break; | |
6141 | } | |
6142 | } | |
6143 | } else { | |
6144 | ret = pret < 0 ? pret : -EFAULT; | |
6145 | } | |
6146 | up_read(¤t->mm->mmap_sem); | |
6147 | if (ret) { | |
6148 | /* | |
6149 | * if we did partial map, or found file backed vmas, | |
6150 | * release any pages we did get | |
6151 | */ | |
27c4d3a3 JH |
6152 | if (pret > 0) |
6153 | put_user_pages(pages, pret); | |
edafccee JA |
6154 | if (ctx->account_mem) |
6155 | io_unaccount_mem(ctx->user, nr_pages); | |
d4ef6475 | 6156 | kvfree(imu->bvec); |
edafccee JA |
6157 | goto err; |
6158 | } | |
6159 | ||
6160 | off = ubuf & ~PAGE_MASK; | |
6161 | size = iov.iov_len; | |
6162 | for (j = 0; j < nr_pages; j++) { | |
6163 | size_t vec_len; | |
6164 | ||
6165 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
6166 | imu->bvec[j].bv_page = pages[j]; | |
6167 | imu->bvec[j].bv_len = vec_len; | |
6168 | imu->bvec[j].bv_offset = off; | |
6169 | off = 0; | |
6170 | size -= vec_len; | |
6171 | } | |
6172 | /* store original address for later verification */ | |
6173 | imu->ubuf = ubuf; | |
6174 | imu->len = iov.iov_len; | |
6175 | imu->nr_bvecs = nr_pages; | |
6176 | ||
6177 | ctx->nr_user_bufs++; | |
6178 | } | |
d4ef6475 MR |
6179 | kvfree(pages); |
6180 | kvfree(vmas); | |
edafccee JA |
6181 | return 0; |
6182 | err: | |
d4ef6475 MR |
6183 | kvfree(pages); |
6184 | kvfree(vmas); | |
edafccee JA |
6185 | io_sqe_buffer_unregister(ctx); |
6186 | return ret; | |
6187 | } | |
6188 | ||
9b402849 JA |
6189 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
6190 | { | |
6191 | __s32 __user *fds = arg; | |
6192 | int fd; | |
6193 | ||
6194 | if (ctx->cq_ev_fd) | |
6195 | return -EBUSY; | |
6196 | ||
6197 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
6198 | return -EFAULT; | |
6199 | ||
6200 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
6201 | if (IS_ERR(ctx->cq_ev_fd)) { | |
6202 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
6203 | ctx->cq_ev_fd = NULL; | |
6204 | return ret; | |
6205 | } | |
6206 | ||
6207 | return 0; | |
6208 | } | |
6209 | ||
6210 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
6211 | { | |
6212 | if (ctx->cq_ev_fd) { | |
6213 | eventfd_ctx_put(ctx->cq_ev_fd); | |
6214 | ctx->cq_ev_fd = NULL; | |
6215 | return 0; | |
6216 | } | |
6217 | ||
6218 | return -ENXIO; | |
6219 | } | |
6220 | ||
2b188cc1 JA |
6221 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
6222 | { | |
6b06314c | 6223 | io_finish_async(ctx); |
2b188cc1 JA |
6224 | if (ctx->sqo_mm) |
6225 | mmdrop(ctx->sqo_mm); | |
def596e9 JA |
6226 | |
6227 | io_iopoll_reap_events(ctx); | |
edafccee | 6228 | io_sqe_buffer_unregister(ctx); |
6b06314c | 6229 | io_sqe_files_unregister(ctx); |
9b402849 | 6230 | io_eventfd_unregister(ctx); |
def596e9 | 6231 | |
2b188cc1 | 6232 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
6233 | if (ctx->ring_sock) { |
6234 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 6235 | sock_release(ctx->ring_sock); |
355e8d26 | 6236 | } |
2b188cc1 JA |
6237 | #endif |
6238 | ||
75b28aff | 6239 | io_mem_free(ctx->rings); |
2b188cc1 | 6240 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
6241 | |
6242 | percpu_ref_exit(&ctx->refs); | |
6243 | if (ctx->account_mem) | |
6244 | io_unaccount_mem(ctx->user, | |
6245 | ring_pages(ctx->sq_entries, ctx->cq_entries)); | |
6246 | free_uid(ctx->user); | |
181e448d | 6247 | put_cred(ctx->creds); |
206aefde | 6248 | kfree(ctx->completions); |
78076bb6 | 6249 | kfree(ctx->cancel_hash); |
0ddf92e8 | 6250 | kmem_cache_free(req_cachep, ctx->fallback_req); |
2b188cc1 JA |
6251 | kfree(ctx); |
6252 | } | |
6253 | ||
6254 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
6255 | { | |
6256 | struct io_ring_ctx *ctx = file->private_data; | |
6257 | __poll_t mask = 0; | |
6258 | ||
6259 | poll_wait(file, &ctx->cq_wait, wait); | |
4f7067c3 SB |
6260 | /* |
6261 | * synchronizes with barrier from wq_has_sleeper call in | |
6262 | * io_commit_cqring | |
6263 | */ | |
2b188cc1 | 6264 | smp_rmb(); |
75b28aff HV |
6265 | if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head != |
6266 | ctx->rings->sq_ring_entries) | |
2b188cc1 | 6267 | mask |= EPOLLOUT | EPOLLWRNORM; |
daa5de54 | 6268 | if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail) |
2b188cc1 JA |
6269 | mask |= EPOLLIN | EPOLLRDNORM; |
6270 | ||
6271 | return mask; | |
6272 | } | |
6273 | ||
6274 | static int io_uring_fasync(int fd, struct file *file, int on) | |
6275 | { | |
6276 | struct io_ring_ctx *ctx = file->private_data; | |
6277 | ||
6278 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
6279 | } | |
6280 | ||
071698e1 JA |
6281 | static int io_remove_personalities(int id, void *p, void *data) |
6282 | { | |
6283 | struct io_ring_ctx *ctx = data; | |
6284 | const struct cred *cred; | |
6285 | ||
6286 | cred = idr_remove(&ctx->personality_idr, id); | |
6287 | if (cred) | |
6288 | put_cred(cred); | |
6289 | return 0; | |
6290 | } | |
6291 | ||
2b188cc1 JA |
6292 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
6293 | { | |
6294 | mutex_lock(&ctx->uring_lock); | |
6295 | percpu_ref_kill(&ctx->refs); | |
6296 | mutex_unlock(&ctx->uring_lock); | |
6297 | ||
5262f567 | 6298 | io_kill_timeouts(ctx); |
221c5eb2 | 6299 | io_poll_remove_all(ctx); |
561fb04a JA |
6300 | |
6301 | if (ctx->io_wq) | |
6302 | io_wq_cancel_all(ctx->io_wq); | |
6303 | ||
def596e9 | 6304 | io_iopoll_reap_events(ctx); |
15dff286 JA |
6305 | /* if we failed setting up the ctx, we might not have any rings */ |
6306 | if (ctx->rings) | |
6307 | io_cqring_overflow_flush(ctx, true); | |
071698e1 | 6308 | idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx); |
206aefde | 6309 | wait_for_completion(&ctx->completions[0]); |
2b188cc1 JA |
6310 | io_ring_ctx_free(ctx); |
6311 | } | |
6312 | ||
6313 | static int io_uring_release(struct inode *inode, struct file *file) | |
6314 | { | |
6315 | struct io_ring_ctx *ctx = file->private_data; | |
6316 | ||
6317 | file->private_data = NULL; | |
6318 | io_ring_ctx_wait_and_kill(ctx); | |
6319 | return 0; | |
6320 | } | |
6321 | ||
fcb323cc JA |
6322 | static void io_uring_cancel_files(struct io_ring_ctx *ctx, |
6323 | struct files_struct *files) | |
6324 | { | |
6325 | struct io_kiocb *req; | |
6326 | DEFINE_WAIT(wait); | |
6327 | ||
6328 | while (!list_empty_careful(&ctx->inflight_list)) { | |
768134d4 | 6329 | struct io_kiocb *cancel_req = NULL; |
fcb323cc JA |
6330 | |
6331 | spin_lock_irq(&ctx->inflight_lock); | |
6332 | list_for_each_entry(req, &ctx->inflight_list, inflight_entry) { | |
768134d4 JA |
6333 | if (req->work.files != files) |
6334 | continue; | |
6335 | /* req is being completed, ignore */ | |
6336 | if (!refcount_inc_not_zero(&req->refs)) | |
6337 | continue; | |
6338 | cancel_req = req; | |
6339 | break; | |
fcb323cc | 6340 | } |
768134d4 | 6341 | if (cancel_req) |
fcb323cc | 6342 | prepare_to_wait(&ctx->inflight_wait, &wait, |
768134d4 | 6343 | TASK_UNINTERRUPTIBLE); |
fcb323cc JA |
6344 | spin_unlock_irq(&ctx->inflight_lock); |
6345 | ||
768134d4 JA |
6346 | /* We need to keep going until we don't find a matching req */ |
6347 | if (!cancel_req) | |
fcb323cc | 6348 | break; |
2f6d9b9d BL |
6349 | |
6350 | io_wq_cancel_work(ctx->io_wq, &cancel_req->work); | |
6351 | io_put_req(cancel_req); | |
fcb323cc JA |
6352 | schedule(); |
6353 | } | |
768134d4 | 6354 | finish_wait(&ctx->inflight_wait, &wait); |
fcb323cc JA |
6355 | } |
6356 | ||
6357 | static int io_uring_flush(struct file *file, void *data) | |
6358 | { | |
6359 | struct io_ring_ctx *ctx = file->private_data; | |
6360 | ||
6361 | io_uring_cancel_files(ctx, data); | |
fcb323cc JA |
6362 | return 0; |
6363 | } | |
6364 | ||
6c5c240e RP |
6365 | static void *io_uring_validate_mmap_request(struct file *file, |
6366 | loff_t pgoff, size_t sz) | |
2b188cc1 | 6367 | { |
2b188cc1 | 6368 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 6369 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
6370 | struct page *page; |
6371 | void *ptr; | |
6372 | ||
6373 | switch (offset) { | |
6374 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
6375 | case IORING_OFF_CQ_RING: |
6376 | ptr = ctx->rings; | |
2b188cc1 JA |
6377 | break; |
6378 | case IORING_OFF_SQES: | |
6379 | ptr = ctx->sq_sqes; | |
6380 | break; | |
2b188cc1 | 6381 | default: |
6c5c240e | 6382 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
6383 | } |
6384 | ||
6385 | page = virt_to_head_page(ptr); | |
a50b854e | 6386 | if (sz > page_size(page)) |
6c5c240e RP |
6387 | return ERR_PTR(-EINVAL); |
6388 | ||
6389 | return ptr; | |
6390 | } | |
6391 | ||
6392 | #ifdef CONFIG_MMU | |
6393 | ||
6394 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
6395 | { | |
6396 | size_t sz = vma->vm_end - vma->vm_start; | |
6397 | unsigned long pfn; | |
6398 | void *ptr; | |
6399 | ||
6400 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
6401 | if (IS_ERR(ptr)) | |
6402 | return PTR_ERR(ptr); | |
2b188cc1 JA |
6403 | |
6404 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
6405 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
6406 | } | |
6407 | ||
6c5c240e RP |
6408 | #else /* !CONFIG_MMU */ |
6409 | ||
6410 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
6411 | { | |
6412 | return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL; | |
6413 | } | |
6414 | ||
6415 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
6416 | { | |
6417 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
6418 | } | |
6419 | ||
6420 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
6421 | unsigned long addr, unsigned long len, | |
6422 | unsigned long pgoff, unsigned long flags) | |
6423 | { | |
6424 | void *ptr; | |
6425 | ||
6426 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
6427 | if (IS_ERR(ptr)) | |
6428 | return PTR_ERR(ptr); | |
6429 | ||
6430 | return (unsigned long) ptr; | |
6431 | } | |
6432 | ||
6433 | #endif /* !CONFIG_MMU */ | |
6434 | ||
2b188cc1 JA |
6435 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
6436 | u32, min_complete, u32, flags, const sigset_t __user *, sig, | |
6437 | size_t, sigsz) | |
6438 | { | |
6439 | struct io_ring_ctx *ctx; | |
6440 | long ret = -EBADF; | |
6441 | int submitted = 0; | |
6442 | struct fd f; | |
6443 | ||
6c271ce2 | 6444 | if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP)) |
2b188cc1 JA |
6445 | return -EINVAL; |
6446 | ||
6447 | f = fdget(fd); | |
6448 | if (!f.file) | |
6449 | return -EBADF; | |
6450 | ||
6451 | ret = -EOPNOTSUPP; | |
6452 | if (f.file->f_op != &io_uring_fops) | |
6453 | goto out_fput; | |
6454 | ||
6455 | ret = -ENXIO; | |
6456 | ctx = f.file->private_data; | |
6457 | if (!percpu_ref_tryget(&ctx->refs)) | |
6458 | goto out_fput; | |
6459 | ||
6c271ce2 JA |
6460 | /* |
6461 | * For SQ polling, the thread will do all submissions and completions. | |
6462 | * Just return the requested submit count, and wake the thread if | |
6463 | * we were asked to. | |
6464 | */ | |
b2a9eada | 6465 | ret = 0; |
6c271ce2 | 6466 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
c1edbf5f JA |
6467 | if (!list_empty_careful(&ctx->cq_overflow_list)) |
6468 | io_cqring_overflow_flush(ctx, false); | |
6c271ce2 JA |
6469 | if (flags & IORING_ENTER_SQ_WAKEUP) |
6470 | wake_up(&ctx->sqo_wait); | |
6471 | submitted = to_submit; | |
b2a9eada | 6472 | } else if (to_submit) { |
ae9428ca | 6473 | struct mm_struct *cur_mm; |
2b188cc1 JA |
6474 | |
6475 | mutex_lock(&ctx->uring_lock); | |
ae9428ca PB |
6476 | /* already have mm, so io_submit_sqes() won't try to grab it */ |
6477 | cur_mm = ctx->sqo_mm; | |
6478 | submitted = io_submit_sqes(ctx, to_submit, f.file, fd, | |
6479 | &cur_mm, false); | |
2b188cc1 | 6480 | mutex_unlock(&ctx->uring_lock); |
7c504e65 PB |
6481 | |
6482 | if (submitted != to_submit) | |
6483 | goto out; | |
2b188cc1 JA |
6484 | } |
6485 | if (flags & IORING_ENTER_GETEVENTS) { | |
def596e9 JA |
6486 | unsigned nr_events = 0; |
6487 | ||
2b188cc1 JA |
6488 | min_complete = min(min_complete, ctx->cq_entries); |
6489 | ||
def596e9 | 6490 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 | 6491 | ret = io_iopoll_check(ctx, &nr_events, min_complete); |
def596e9 JA |
6492 | } else { |
6493 | ret = io_cqring_wait(ctx, min_complete, sig, sigsz); | |
6494 | } | |
2b188cc1 JA |
6495 | } |
6496 | ||
7c504e65 | 6497 | out: |
6805b32e | 6498 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
6499 | out_fput: |
6500 | fdput(f); | |
6501 | return submitted ? submitted : ret; | |
6502 | } | |
6503 | ||
6504 | static const struct file_operations io_uring_fops = { | |
6505 | .release = io_uring_release, | |
fcb323cc | 6506 | .flush = io_uring_flush, |
2b188cc1 | 6507 | .mmap = io_uring_mmap, |
6c5c240e RP |
6508 | #ifndef CONFIG_MMU |
6509 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
6510 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
6511 | #endif | |
2b188cc1 JA |
6512 | .poll = io_uring_poll, |
6513 | .fasync = io_uring_fasync, | |
6514 | }; | |
6515 | ||
6516 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
6517 | struct io_uring_params *p) | |
6518 | { | |
75b28aff HV |
6519 | struct io_rings *rings; |
6520 | size_t size, sq_array_offset; | |
2b188cc1 | 6521 | |
75b28aff HV |
6522 | size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset); |
6523 | if (size == SIZE_MAX) | |
6524 | return -EOVERFLOW; | |
6525 | ||
6526 | rings = io_mem_alloc(size); | |
6527 | if (!rings) | |
2b188cc1 JA |
6528 | return -ENOMEM; |
6529 | ||
75b28aff HV |
6530 | ctx->rings = rings; |
6531 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
6532 | rings->sq_ring_mask = p->sq_entries - 1; | |
6533 | rings->cq_ring_mask = p->cq_entries - 1; | |
6534 | rings->sq_ring_entries = p->sq_entries; | |
6535 | rings->cq_ring_entries = p->cq_entries; | |
6536 | ctx->sq_mask = rings->sq_ring_mask; | |
6537 | ctx->cq_mask = rings->cq_ring_mask; | |
6538 | ctx->sq_entries = rings->sq_ring_entries; | |
6539 | ctx->cq_entries = rings->cq_ring_entries; | |
2b188cc1 JA |
6540 | |
6541 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 JA |
6542 | if (size == SIZE_MAX) { |
6543 | io_mem_free(ctx->rings); | |
6544 | ctx->rings = NULL; | |
2b188cc1 | 6545 | return -EOVERFLOW; |
eb065d30 | 6546 | } |
2b188cc1 JA |
6547 | |
6548 | ctx->sq_sqes = io_mem_alloc(size); | |
eb065d30 JA |
6549 | if (!ctx->sq_sqes) { |
6550 | io_mem_free(ctx->rings); | |
6551 | ctx->rings = NULL; | |
2b188cc1 | 6552 | return -ENOMEM; |
eb065d30 | 6553 | } |
2b188cc1 | 6554 | |
2b188cc1 JA |
6555 | return 0; |
6556 | } | |
6557 | ||
6558 | /* | |
6559 | * Allocate an anonymous fd, this is what constitutes the application | |
6560 | * visible backing of an io_uring instance. The application mmaps this | |
6561 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
6562 | * we have to tie this fd to a socket for file garbage collection purposes. | |
6563 | */ | |
6564 | static int io_uring_get_fd(struct io_ring_ctx *ctx) | |
6565 | { | |
6566 | struct file *file; | |
6567 | int ret; | |
6568 | ||
6569 | #if defined(CONFIG_UNIX) | |
6570 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, | |
6571 | &ctx->ring_sock); | |
6572 | if (ret) | |
6573 | return ret; | |
6574 | #endif | |
6575 | ||
6576 | ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
6577 | if (ret < 0) | |
6578 | goto err; | |
6579 | ||
6580 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, | |
6581 | O_RDWR | O_CLOEXEC); | |
6582 | if (IS_ERR(file)) { | |
6583 | put_unused_fd(ret); | |
6584 | ret = PTR_ERR(file); | |
6585 | goto err; | |
6586 | } | |
6587 | ||
6588 | #if defined(CONFIG_UNIX) | |
6589 | ctx->ring_sock->file = file; | |
6590 | #endif | |
6591 | fd_install(ret, file); | |
6592 | return ret; | |
6593 | err: | |
6594 | #if defined(CONFIG_UNIX) | |
6595 | sock_release(ctx->ring_sock); | |
6596 | ctx->ring_sock = NULL; | |
6597 | #endif | |
6598 | return ret; | |
6599 | } | |
6600 | ||
6601 | static int io_uring_create(unsigned entries, struct io_uring_params *p) | |
6602 | { | |
6603 | struct user_struct *user = NULL; | |
6604 | struct io_ring_ctx *ctx; | |
6605 | bool account_mem; | |
6606 | int ret; | |
6607 | ||
8110c1a6 | 6608 | if (!entries) |
2b188cc1 | 6609 | return -EINVAL; |
8110c1a6 JA |
6610 | if (entries > IORING_MAX_ENTRIES) { |
6611 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
6612 | return -EINVAL; | |
6613 | entries = IORING_MAX_ENTRIES; | |
6614 | } | |
2b188cc1 JA |
6615 | |
6616 | /* | |
6617 | * Use twice as many entries for the CQ ring. It's possible for the | |
6618 | * application to drive a higher depth than the size of the SQ ring, | |
6619 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
6620 | * some flexibility in overcommitting a bit. If the application has |
6621 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
6622 | * of CQ ring entries manually. | |
2b188cc1 JA |
6623 | */ |
6624 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
6625 | if (p->flags & IORING_SETUP_CQSIZE) { |
6626 | /* | |
6627 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
6628 | * to a power-of-two, if it isn't already. We do NOT impose | |
6629 | * any cq vs sq ring sizing. | |
6630 | */ | |
8110c1a6 | 6631 | if (p->cq_entries < p->sq_entries) |
33a107f0 | 6632 | return -EINVAL; |
8110c1a6 JA |
6633 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
6634 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
6635 | return -EINVAL; | |
6636 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
6637 | } | |
33a107f0 JA |
6638 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
6639 | } else { | |
6640 | p->cq_entries = 2 * p->sq_entries; | |
6641 | } | |
2b188cc1 JA |
6642 | |
6643 | user = get_uid(current_user()); | |
6644 | account_mem = !capable(CAP_IPC_LOCK); | |
6645 | ||
6646 | if (account_mem) { | |
6647 | ret = io_account_mem(user, | |
6648 | ring_pages(p->sq_entries, p->cq_entries)); | |
6649 | if (ret) { | |
6650 | free_uid(user); | |
6651 | return ret; | |
6652 | } | |
6653 | } | |
6654 | ||
6655 | ctx = io_ring_ctx_alloc(p); | |
6656 | if (!ctx) { | |
6657 | if (account_mem) | |
6658 | io_unaccount_mem(user, ring_pages(p->sq_entries, | |
6659 | p->cq_entries)); | |
6660 | free_uid(user); | |
6661 | return -ENOMEM; | |
6662 | } | |
6663 | ctx->compat = in_compat_syscall(); | |
6664 | ctx->account_mem = account_mem; | |
6665 | ctx->user = user; | |
0b8c0ec7 | 6666 | ctx->creds = get_current_cred(); |
2b188cc1 JA |
6667 | |
6668 | ret = io_allocate_scq_urings(ctx, p); | |
6669 | if (ret) | |
6670 | goto err; | |
6671 | ||
6c271ce2 | 6672 | ret = io_sq_offload_start(ctx, p); |
2b188cc1 JA |
6673 | if (ret) |
6674 | goto err; | |
6675 | ||
2b188cc1 | 6676 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
6677 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
6678 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
6679 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
6680 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
6681 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
6682 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
6683 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
6684 | |
6685 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
6686 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
6687 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
6688 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
6689 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
6690 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
6691 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
ac90f249 | 6692 | |
044c1ab3 JA |
6693 | /* |
6694 | * Install ring fd as the very last thing, so we don't risk someone | |
6695 | * having closed it before we finish setup | |
6696 | */ | |
6697 | ret = io_uring_get_fd(ctx); | |
6698 | if (ret < 0) | |
6699 | goto err; | |
6700 | ||
da8c9690 | 6701 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
cccf0ee8 JA |
6702 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | |
6703 | IORING_FEAT_CUR_PERSONALITY; | |
c826bd7a | 6704 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
6705 | return ret; |
6706 | err: | |
6707 | io_ring_ctx_wait_and_kill(ctx); | |
6708 | return ret; | |
6709 | } | |
6710 | ||
6711 | /* | |
6712 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
6713 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
6714 | * params structure passed in. | |
6715 | */ | |
6716 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
6717 | { | |
6718 | struct io_uring_params p; | |
6719 | long ret; | |
6720 | int i; | |
6721 | ||
6722 | if (copy_from_user(&p, params, sizeof(p))) | |
6723 | return -EFAULT; | |
6724 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
6725 | if (p.resv[i]) | |
6726 | return -EINVAL; | |
6727 | } | |
6728 | ||
6c271ce2 | 6729 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 6730 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
24369c2e | 6731 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ)) |
2b188cc1 JA |
6732 | return -EINVAL; |
6733 | ||
6734 | ret = io_uring_create(entries, &p); | |
6735 | if (ret < 0) | |
6736 | return ret; | |
6737 | ||
6738 | if (copy_to_user(params, &p, sizeof(p))) | |
6739 | return -EFAULT; | |
6740 | ||
6741 | return ret; | |
6742 | } | |
6743 | ||
6744 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
6745 | struct io_uring_params __user *, params) | |
6746 | { | |
6747 | return io_uring_setup(entries, params); | |
6748 | } | |
6749 | ||
66f4af93 JA |
6750 | static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args) |
6751 | { | |
6752 | struct io_uring_probe *p; | |
6753 | size_t size; | |
6754 | int i, ret; | |
6755 | ||
6756 | size = struct_size(p, ops, nr_args); | |
6757 | if (size == SIZE_MAX) | |
6758 | return -EOVERFLOW; | |
6759 | p = kzalloc(size, GFP_KERNEL); | |
6760 | if (!p) | |
6761 | return -ENOMEM; | |
6762 | ||
6763 | ret = -EFAULT; | |
6764 | if (copy_from_user(p, arg, size)) | |
6765 | goto out; | |
6766 | ret = -EINVAL; | |
6767 | if (memchr_inv(p, 0, size)) | |
6768 | goto out; | |
6769 | ||
6770 | p->last_op = IORING_OP_LAST - 1; | |
6771 | if (nr_args > IORING_OP_LAST) | |
6772 | nr_args = IORING_OP_LAST; | |
6773 | ||
6774 | for (i = 0; i < nr_args; i++) { | |
6775 | p->ops[i].op = i; | |
6776 | if (!io_op_defs[i].not_supported) | |
6777 | p->ops[i].flags = IO_URING_OP_SUPPORTED; | |
6778 | } | |
6779 | p->ops_len = i; | |
6780 | ||
6781 | ret = 0; | |
6782 | if (copy_to_user(arg, p, size)) | |
6783 | ret = -EFAULT; | |
6784 | out: | |
6785 | kfree(p); | |
6786 | return ret; | |
6787 | } | |
6788 | ||
071698e1 JA |
6789 | static int io_register_personality(struct io_ring_ctx *ctx) |
6790 | { | |
6791 | const struct cred *creds = get_current_cred(); | |
6792 | int id; | |
6793 | ||
6794 | id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1, | |
6795 | USHRT_MAX, GFP_KERNEL); | |
6796 | if (id < 0) | |
6797 | put_cred(creds); | |
6798 | return id; | |
6799 | } | |
6800 | ||
6801 | static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) | |
6802 | { | |
6803 | const struct cred *old_creds; | |
6804 | ||
6805 | old_creds = idr_remove(&ctx->personality_idr, id); | |
6806 | if (old_creds) { | |
6807 | put_cred(old_creds); | |
6808 | return 0; | |
6809 | } | |
6810 | ||
6811 | return -EINVAL; | |
6812 | } | |
6813 | ||
6814 | static bool io_register_op_must_quiesce(int op) | |
6815 | { | |
6816 | switch (op) { | |
6817 | case IORING_UNREGISTER_FILES: | |
6818 | case IORING_REGISTER_FILES_UPDATE: | |
6819 | case IORING_REGISTER_PROBE: | |
6820 | case IORING_REGISTER_PERSONALITY: | |
6821 | case IORING_UNREGISTER_PERSONALITY: | |
6822 | return false; | |
6823 | default: | |
6824 | return true; | |
6825 | } | |
6826 | } | |
6827 | ||
edafccee JA |
6828 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
6829 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
6830 | __releases(ctx->uring_lock) |
6831 | __acquires(ctx->uring_lock) | |
edafccee JA |
6832 | { |
6833 | int ret; | |
6834 | ||
35fa71a0 JA |
6835 | /* |
6836 | * We're inside the ring mutex, if the ref is already dying, then | |
6837 | * someone else killed the ctx or is already going through | |
6838 | * io_uring_register(). | |
6839 | */ | |
6840 | if (percpu_ref_is_dying(&ctx->refs)) | |
6841 | return -ENXIO; | |
6842 | ||
071698e1 | 6843 | if (io_register_op_must_quiesce(opcode)) { |
05f3fb3c | 6844 | percpu_ref_kill(&ctx->refs); |
b19062a5 | 6845 | |
05f3fb3c JA |
6846 | /* |
6847 | * Drop uring mutex before waiting for references to exit. If | |
6848 | * another thread is currently inside io_uring_enter() it might | |
6849 | * need to grab the uring_lock to make progress. If we hold it | |
6850 | * here across the drain wait, then we can deadlock. It's safe | |
6851 | * to drop the mutex here, since no new references will come in | |
6852 | * after we've killed the percpu ref. | |
6853 | */ | |
6854 | mutex_unlock(&ctx->uring_lock); | |
c150368b | 6855 | ret = wait_for_completion_interruptible(&ctx->completions[0]); |
05f3fb3c | 6856 | mutex_lock(&ctx->uring_lock); |
c150368b JA |
6857 | if (ret) { |
6858 | percpu_ref_resurrect(&ctx->refs); | |
6859 | ret = -EINTR; | |
6860 | goto out; | |
6861 | } | |
05f3fb3c | 6862 | } |
edafccee JA |
6863 | |
6864 | switch (opcode) { | |
6865 | case IORING_REGISTER_BUFFERS: | |
6866 | ret = io_sqe_buffer_register(ctx, arg, nr_args); | |
6867 | break; | |
6868 | case IORING_UNREGISTER_BUFFERS: | |
6869 | ret = -EINVAL; | |
6870 | if (arg || nr_args) | |
6871 | break; | |
6872 | ret = io_sqe_buffer_unregister(ctx); | |
6873 | break; | |
6b06314c JA |
6874 | case IORING_REGISTER_FILES: |
6875 | ret = io_sqe_files_register(ctx, arg, nr_args); | |
6876 | break; | |
6877 | case IORING_UNREGISTER_FILES: | |
6878 | ret = -EINVAL; | |
6879 | if (arg || nr_args) | |
6880 | break; | |
6881 | ret = io_sqe_files_unregister(ctx); | |
6882 | break; | |
c3a31e60 JA |
6883 | case IORING_REGISTER_FILES_UPDATE: |
6884 | ret = io_sqe_files_update(ctx, arg, nr_args); | |
6885 | break; | |
9b402849 | 6886 | case IORING_REGISTER_EVENTFD: |
f2842ab5 | 6887 | case IORING_REGISTER_EVENTFD_ASYNC: |
9b402849 JA |
6888 | ret = -EINVAL; |
6889 | if (nr_args != 1) | |
6890 | break; | |
6891 | ret = io_eventfd_register(ctx, arg); | |
f2842ab5 JA |
6892 | if (ret) |
6893 | break; | |
6894 | if (opcode == IORING_REGISTER_EVENTFD_ASYNC) | |
6895 | ctx->eventfd_async = 1; | |
6896 | else | |
6897 | ctx->eventfd_async = 0; | |
9b402849 JA |
6898 | break; |
6899 | case IORING_UNREGISTER_EVENTFD: | |
6900 | ret = -EINVAL; | |
6901 | if (arg || nr_args) | |
6902 | break; | |
6903 | ret = io_eventfd_unregister(ctx); | |
6904 | break; | |
66f4af93 JA |
6905 | case IORING_REGISTER_PROBE: |
6906 | ret = -EINVAL; | |
6907 | if (!arg || nr_args > 256) | |
6908 | break; | |
6909 | ret = io_probe(ctx, arg, nr_args); | |
6910 | break; | |
071698e1 JA |
6911 | case IORING_REGISTER_PERSONALITY: |
6912 | ret = -EINVAL; | |
6913 | if (arg || nr_args) | |
6914 | break; | |
6915 | ret = io_register_personality(ctx); | |
6916 | break; | |
6917 | case IORING_UNREGISTER_PERSONALITY: | |
6918 | ret = -EINVAL; | |
6919 | if (arg) | |
6920 | break; | |
6921 | ret = io_unregister_personality(ctx, nr_args); | |
6922 | break; | |
edafccee JA |
6923 | default: |
6924 | ret = -EINVAL; | |
6925 | break; | |
6926 | } | |
6927 | ||
071698e1 | 6928 | if (io_register_op_must_quiesce(opcode)) { |
05f3fb3c | 6929 | /* bring the ctx back to life */ |
05f3fb3c | 6930 | percpu_ref_reinit(&ctx->refs); |
c150368b JA |
6931 | out: |
6932 | reinit_completion(&ctx->completions[0]); | |
05f3fb3c | 6933 | } |
edafccee JA |
6934 | return ret; |
6935 | } | |
6936 | ||
6937 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
6938 | void __user *, arg, unsigned int, nr_args) | |
6939 | { | |
6940 | struct io_ring_ctx *ctx; | |
6941 | long ret = -EBADF; | |
6942 | struct fd f; | |
6943 | ||
6944 | f = fdget(fd); | |
6945 | if (!f.file) | |
6946 | return -EBADF; | |
6947 | ||
6948 | ret = -EOPNOTSUPP; | |
6949 | if (f.file->f_op != &io_uring_fops) | |
6950 | goto out_fput; | |
6951 | ||
6952 | ctx = f.file->private_data; | |
6953 | ||
6954 | mutex_lock(&ctx->uring_lock); | |
6955 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
6956 | mutex_unlock(&ctx->uring_lock); | |
c826bd7a DD |
6957 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, |
6958 | ctx->cq_ev_fd != NULL, ret); | |
edafccee JA |
6959 | out_fput: |
6960 | fdput(f); | |
6961 | return ret; | |
6962 | } | |
6963 | ||
2b188cc1 JA |
6964 | static int __init io_uring_init(void) |
6965 | { | |
d3656344 | 6966 | BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST); |
2b188cc1 JA |
6967 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC); |
6968 | return 0; | |
6969 | }; | |
6970 | __initcall(io_uring_init); |