]>
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> | |
49 | ||
50 | #include <linux/sched/signal.h> | |
51 | #include <linux/fs.h> | |
52 | #include <linux/file.h> | |
53 | #include <linux/fdtable.h> | |
54 | #include <linux/mm.h> | |
55 | #include <linux/mman.h> | |
56 | #include <linux/mmu_context.h> | |
57 | #include <linux/percpu.h> | |
58 | #include <linux/slab.h> | |
6c271ce2 | 59 | #include <linux/kthread.h> |
2b188cc1 | 60 | #include <linux/blkdev.h> |
edafccee | 61 | #include <linux/bvec.h> |
2b188cc1 JA |
62 | #include <linux/net.h> |
63 | #include <net/sock.h> | |
64 | #include <net/af_unix.h> | |
6b06314c | 65 | #include <net/scm.h> |
2b188cc1 JA |
66 | #include <linux/anon_inodes.h> |
67 | #include <linux/sched/mm.h> | |
68 | #include <linux/uaccess.h> | |
69 | #include <linux/nospec.h> | |
edafccee JA |
70 | #include <linux/sizes.h> |
71 | #include <linux/hugetlb.h> | |
aa4c3967 | 72 | #include <linux/highmem.h> |
15b71abe JA |
73 | #include <linux/namei.h> |
74 | #include <linux/fsnotify.h> | |
2b188cc1 | 75 | |
c826bd7a DD |
76 | #define CREATE_TRACE_POINTS |
77 | #include <trace/events/io_uring.h> | |
78 | ||
2b188cc1 JA |
79 | #include <uapi/linux/io_uring.h> |
80 | ||
81 | #include "internal.h" | |
561fb04a | 82 | #include "io-wq.h" |
2b188cc1 | 83 | |
5277deaa | 84 | #define IORING_MAX_ENTRIES 32768 |
33a107f0 | 85 | #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES) |
65e19f54 JA |
86 | |
87 | /* | |
88 | * Shift of 9 is 512 entries, or exactly one page on 64-bit archs | |
89 | */ | |
90 | #define IORING_FILE_TABLE_SHIFT 9 | |
91 | #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT) | |
92 | #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1) | |
93 | #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE) | |
2b188cc1 JA |
94 | |
95 | struct io_uring { | |
96 | u32 head ____cacheline_aligned_in_smp; | |
97 | u32 tail ____cacheline_aligned_in_smp; | |
98 | }; | |
99 | ||
1e84b97b | 100 | /* |
75b28aff HV |
101 | * This data is shared with the application through the mmap at offsets |
102 | * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING. | |
1e84b97b SB |
103 | * |
104 | * The offsets to the member fields are published through struct | |
105 | * io_sqring_offsets when calling io_uring_setup. | |
106 | */ | |
75b28aff | 107 | struct io_rings { |
1e84b97b SB |
108 | /* |
109 | * Head and tail offsets into the ring; the offsets need to be | |
110 | * masked to get valid indices. | |
111 | * | |
75b28aff HV |
112 | * The kernel controls head of the sq ring and the tail of the cq ring, |
113 | * and the application controls tail of the sq ring and the head of the | |
114 | * cq ring. | |
1e84b97b | 115 | */ |
75b28aff | 116 | struct io_uring sq, cq; |
1e84b97b | 117 | /* |
75b28aff | 118 | * Bitmasks to apply to head and tail offsets (constant, equals |
1e84b97b SB |
119 | * ring_entries - 1) |
120 | */ | |
75b28aff HV |
121 | u32 sq_ring_mask, cq_ring_mask; |
122 | /* Ring sizes (constant, power of 2) */ | |
123 | u32 sq_ring_entries, cq_ring_entries; | |
1e84b97b SB |
124 | /* |
125 | * Number of invalid entries dropped by the kernel due to | |
126 | * invalid index stored in array | |
127 | * | |
128 | * Written by the kernel, shouldn't be modified by the | |
129 | * application (i.e. get number of "new events" by comparing to | |
130 | * cached value). | |
131 | * | |
132 | * After a new SQ head value was read by the application this | |
133 | * counter includes all submissions that were dropped reaching | |
134 | * the new SQ head (and possibly more). | |
135 | */ | |
75b28aff | 136 | u32 sq_dropped; |
1e84b97b SB |
137 | /* |
138 | * Runtime flags | |
139 | * | |
140 | * Written by the kernel, shouldn't be modified by the | |
141 | * application. | |
142 | * | |
143 | * The application needs a full memory barrier before checking | |
144 | * for IORING_SQ_NEED_WAKEUP after updating the sq tail. | |
145 | */ | |
75b28aff | 146 | u32 sq_flags; |
1e84b97b SB |
147 | /* |
148 | * Number of completion events lost because the queue was full; | |
149 | * this should be avoided by the application by making sure | |
0b4295b5 | 150 | * there are not more requests pending than there is space in |
1e84b97b SB |
151 | * the completion queue. |
152 | * | |
153 | * Written by the kernel, shouldn't be modified by the | |
154 | * application (i.e. get number of "new events" by comparing to | |
155 | * cached value). | |
156 | * | |
157 | * As completion events come in out of order this counter is not | |
158 | * ordered with any other data. | |
159 | */ | |
75b28aff | 160 | u32 cq_overflow; |
1e84b97b SB |
161 | /* |
162 | * Ring buffer of completion events. | |
163 | * | |
164 | * The kernel writes completion events fresh every time they are | |
165 | * produced, so the application is allowed to modify pending | |
166 | * entries. | |
167 | */ | |
75b28aff | 168 | struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp; |
2b188cc1 JA |
169 | }; |
170 | ||
edafccee JA |
171 | struct io_mapped_ubuf { |
172 | u64 ubuf; | |
173 | size_t len; | |
174 | struct bio_vec *bvec; | |
175 | unsigned int nr_bvecs; | |
176 | }; | |
177 | ||
65e19f54 JA |
178 | struct fixed_file_table { |
179 | struct file **files; | |
31b51510 JA |
180 | }; |
181 | ||
05f3fb3c JA |
182 | enum { |
183 | FFD_F_ATOMIC, | |
184 | }; | |
185 | ||
186 | struct fixed_file_data { | |
187 | struct fixed_file_table *table; | |
188 | struct io_ring_ctx *ctx; | |
189 | ||
190 | struct percpu_ref refs; | |
191 | struct llist_head put_llist; | |
192 | unsigned long state; | |
193 | struct work_struct ref_work; | |
194 | struct completion done; | |
195 | }; | |
196 | ||
2b188cc1 JA |
197 | struct io_ring_ctx { |
198 | struct { | |
199 | struct percpu_ref refs; | |
200 | } ____cacheline_aligned_in_smp; | |
201 | ||
202 | struct { | |
203 | unsigned int flags; | |
204 | bool compat; | |
205 | bool account_mem; | |
1d7bb1d5 | 206 | bool cq_overflow_flushed; |
1b4a51b6 | 207 | bool drain_next; |
2b188cc1 | 208 | |
75b28aff HV |
209 | /* |
210 | * Ring buffer of indices into array of io_uring_sqe, which is | |
211 | * mmapped by the application using the IORING_OFF_SQES offset. | |
212 | * | |
213 | * This indirection could e.g. be used to assign fixed | |
214 | * io_uring_sqe entries to operations and only submit them to | |
215 | * the queue when needed. | |
216 | * | |
217 | * The kernel modifies neither the indices array nor the entries | |
218 | * array. | |
219 | */ | |
220 | u32 *sq_array; | |
2b188cc1 JA |
221 | unsigned cached_sq_head; |
222 | unsigned sq_entries; | |
223 | unsigned sq_mask; | |
6c271ce2 | 224 | unsigned sq_thread_idle; |
498ccd9e | 225 | unsigned cached_sq_dropped; |
206aefde | 226 | atomic_t cached_cq_overflow; |
ad3eb2c8 | 227 | unsigned long sq_check_overflow; |
de0617e4 JA |
228 | |
229 | struct list_head defer_list; | |
5262f567 | 230 | struct list_head timeout_list; |
1d7bb1d5 | 231 | struct list_head cq_overflow_list; |
fcb323cc JA |
232 | |
233 | wait_queue_head_t inflight_wait; | |
ad3eb2c8 | 234 | struct io_uring_sqe *sq_sqes; |
2b188cc1 JA |
235 | } ____cacheline_aligned_in_smp; |
236 | ||
206aefde JA |
237 | struct io_rings *rings; |
238 | ||
2b188cc1 | 239 | /* IO offload */ |
561fb04a | 240 | struct io_wq *io_wq; |
6c271ce2 | 241 | struct task_struct *sqo_thread; /* if using sq thread polling */ |
2b188cc1 | 242 | struct mm_struct *sqo_mm; |
6c271ce2 | 243 | wait_queue_head_t sqo_wait; |
75b28aff | 244 | |
6b06314c JA |
245 | /* |
246 | * If used, fixed file set. Writers must ensure that ->refs is dead, | |
247 | * readers must ensure that ->refs is alive as long as the file* is | |
248 | * used. Only updated through io_uring_register(2). | |
249 | */ | |
05f3fb3c | 250 | struct fixed_file_data *file_data; |
6b06314c JA |
251 | unsigned nr_user_files; |
252 | ||
edafccee JA |
253 | /* if used, fixed mapped user buffers */ |
254 | unsigned nr_user_bufs; | |
255 | struct io_mapped_ubuf *user_bufs; | |
256 | ||
2b188cc1 JA |
257 | struct user_struct *user; |
258 | ||
0b8c0ec7 | 259 | const struct cred *creds; |
181e448d | 260 | |
206aefde JA |
261 | /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */ |
262 | struct completion *completions; | |
263 | ||
0ddf92e8 JA |
264 | /* if all else fails... */ |
265 | struct io_kiocb *fallback_req; | |
266 | ||
206aefde JA |
267 | #if defined(CONFIG_UNIX) |
268 | struct socket *ring_sock; | |
269 | #endif | |
270 | ||
271 | struct { | |
272 | unsigned cached_cq_tail; | |
273 | unsigned cq_entries; | |
274 | unsigned cq_mask; | |
275 | atomic_t cq_timeouts; | |
ad3eb2c8 | 276 | unsigned long cq_check_overflow; |
206aefde JA |
277 | struct wait_queue_head cq_wait; |
278 | struct fasync_struct *cq_fasync; | |
279 | struct eventfd_ctx *cq_ev_fd; | |
280 | } ____cacheline_aligned_in_smp; | |
2b188cc1 JA |
281 | |
282 | struct { | |
283 | struct mutex uring_lock; | |
284 | wait_queue_head_t wait; | |
285 | } ____cacheline_aligned_in_smp; | |
286 | ||
287 | struct { | |
288 | spinlock_t completion_lock; | |
def596e9 JA |
289 | bool poll_multi_file; |
290 | /* | |
291 | * ->poll_list is protected by the ctx->uring_lock for | |
292 | * io_uring instances that don't use IORING_SETUP_SQPOLL. | |
293 | * For SQPOLL, only the single threaded io_sq_thread() will | |
294 | * manipulate the list, hence no extra locking is needed there. | |
295 | */ | |
296 | struct list_head poll_list; | |
78076bb6 JA |
297 | struct hlist_head *cancel_hash; |
298 | unsigned cancel_hash_bits; | |
31b51510 | 299 | |
fcb323cc JA |
300 | spinlock_t inflight_lock; |
301 | struct list_head inflight_list; | |
2b188cc1 | 302 | } ____cacheline_aligned_in_smp; |
2b188cc1 JA |
303 | }; |
304 | ||
09bb8394 JA |
305 | /* |
306 | * First field must be the file pointer in all the | |
307 | * iocb unions! See also 'struct kiocb' in <linux/fs.h> | |
308 | */ | |
221c5eb2 JA |
309 | struct io_poll_iocb { |
310 | struct file *file; | |
0969e783 JA |
311 | union { |
312 | struct wait_queue_head *head; | |
313 | u64 addr; | |
314 | }; | |
221c5eb2 | 315 | __poll_t events; |
8c838788 | 316 | bool done; |
221c5eb2 | 317 | bool canceled; |
392edb45 | 318 | struct wait_queue_entry wait; |
221c5eb2 JA |
319 | }; |
320 | ||
b5dba59e JA |
321 | struct io_close { |
322 | struct file *file; | |
323 | struct file *put_file; | |
324 | int fd; | |
325 | }; | |
326 | ||
ad8a48ac JA |
327 | struct io_timeout_data { |
328 | struct io_kiocb *req; | |
329 | struct hrtimer timer; | |
330 | struct timespec64 ts; | |
331 | enum hrtimer_mode mode; | |
cc42e0ac | 332 | u32 seq_offset; |
ad8a48ac JA |
333 | }; |
334 | ||
8ed8d3c3 JA |
335 | struct io_accept { |
336 | struct file *file; | |
337 | struct sockaddr __user *addr; | |
338 | int __user *addr_len; | |
339 | int flags; | |
340 | }; | |
341 | ||
342 | struct io_sync { | |
343 | struct file *file; | |
344 | loff_t len; | |
345 | loff_t off; | |
346 | int flags; | |
d63d1b5e | 347 | int mode; |
8ed8d3c3 JA |
348 | }; |
349 | ||
fbf23849 JA |
350 | struct io_cancel { |
351 | struct file *file; | |
352 | u64 addr; | |
353 | }; | |
354 | ||
b29472ee JA |
355 | struct io_timeout { |
356 | struct file *file; | |
357 | u64 addr; | |
358 | int flags; | |
26a61679 | 359 | unsigned count; |
b29472ee JA |
360 | }; |
361 | ||
9adbd45d JA |
362 | struct io_rw { |
363 | /* NOTE: kiocb has the file as the first member, so don't do it here */ | |
364 | struct kiocb kiocb; | |
365 | u64 addr; | |
366 | u64 len; | |
367 | }; | |
368 | ||
3fbb51c1 JA |
369 | struct io_connect { |
370 | struct file *file; | |
371 | struct sockaddr __user *addr; | |
372 | int addr_len; | |
373 | }; | |
374 | ||
e47293fd JA |
375 | struct io_sr_msg { |
376 | struct file *file; | |
377 | struct user_msghdr __user *msg; | |
378 | int msg_flags; | |
379 | }; | |
380 | ||
15b71abe JA |
381 | struct io_open { |
382 | struct file *file; | |
383 | int dfd; | |
eddc7ef5 JA |
384 | union { |
385 | umode_t mode; | |
386 | unsigned mask; | |
387 | }; | |
15b71abe JA |
388 | const char __user *fname; |
389 | struct filename *filename; | |
eddc7ef5 | 390 | struct statx __user *buffer; |
15b71abe JA |
391 | int flags; |
392 | }; | |
393 | ||
05f3fb3c JA |
394 | struct io_files_update { |
395 | struct file *file; | |
396 | u64 arg; | |
397 | u32 nr_args; | |
398 | u32 offset; | |
399 | }; | |
400 | ||
f499a021 JA |
401 | struct io_async_connect { |
402 | struct sockaddr_storage address; | |
403 | }; | |
404 | ||
03b1230c JA |
405 | struct io_async_msghdr { |
406 | struct iovec fast_iov[UIO_FASTIOV]; | |
407 | struct iovec *iov; | |
408 | struct sockaddr __user *uaddr; | |
409 | struct msghdr msg; | |
410 | }; | |
411 | ||
f67676d1 JA |
412 | struct io_async_rw { |
413 | struct iovec fast_iov[UIO_FASTIOV]; | |
414 | struct iovec *iov; | |
415 | ssize_t nr_segs; | |
416 | ssize_t size; | |
417 | }; | |
418 | ||
15b71abe JA |
419 | struct io_async_open { |
420 | struct filename *filename; | |
421 | }; | |
422 | ||
1a6b74fc | 423 | struct io_async_ctx { |
f67676d1 JA |
424 | union { |
425 | struct io_async_rw rw; | |
03b1230c | 426 | struct io_async_msghdr msg; |
f499a021 | 427 | struct io_async_connect connect; |
2d28390a | 428 | struct io_timeout_data timeout; |
15b71abe | 429 | struct io_async_open open; |
f67676d1 | 430 | }; |
1a6b74fc JA |
431 | }; |
432 | ||
09bb8394 JA |
433 | /* |
434 | * NOTE! Each of the iocb union members has the file pointer | |
435 | * as the first entry in their struct definition. So you can | |
436 | * access the file pointer through any of the sub-structs, | |
437 | * or directly as just 'ki_filp' in this struct. | |
438 | */ | |
2b188cc1 | 439 | struct io_kiocb { |
221c5eb2 | 440 | union { |
09bb8394 | 441 | struct file *file; |
9adbd45d | 442 | struct io_rw rw; |
221c5eb2 | 443 | struct io_poll_iocb poll; |
8ed8d3c3 JA |
444 | struct io_accept accept; |
445 | struct io_sync sync; | |
fbf23849 | 446 | struct io_cancel cancel; |
b29472ee | 447 | struct io_timeout timeout; |
3fbb51c1 | 448 | struct io_connect connect; |
e47293fd | 449 | struct io_sr_msg sr_msg; |
15b71abe | 450 | struct io_open open; |
b5dba59e | 451 | struct io_close close; |
05f3fb3c | 452 | struct io_files_update files_update; |
221c5eb2 | 453 | }; |
2b188cc1 | 454 | |
1a6b74fc | 455 | struct io_async_ctx *io; |
cf6fd4bd PB |
456 | struct file *ring_file; |
457 | int ring_fd; | |
458 | bool has_user; | |
459 | bool in_async; | |
460 | bool needs_fixed_file; | |
d625c6ee | 461 | u8 opcode; |
2b188cc1 JA |
462 | |
463 | struct io_ring_ctx *ctx; | |
eac406c6 JA |
464 | union { |
465 | struct list_head list; | |
78076bb6 | 466 | struct hlist_node hash_node; |
eac406c6 | 467 | }; |
9e645e11 | 468 | struct list_head link_list; |
2b188cc1 | 469 | unsigned int flags; |
c16361c1 | 470 | refcount_t refs; |
8449eeda | 471 | #define REQ_F_NOWAIT 1 /* must not punt to workers */ |
def596e9 | 472 | #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */ |
6b06314c | 473 | #define REQ_F_FIXED_FILE 4 /* ctx owns file */ |
4d7dd462 | 474 | #define REQ_F_LINK_NEXT 8 /* already grabbed next link */ |
e2033e33 SB |
475 | #define REQ_F_IO_DRAIN 16 /* drain existing IO first */ |
476 | #define REQ_F_IO_DRAINED 32 /* drain done */ | |
9e645e11 | 477 | #define REQ_F_LINK 64 /* linked sqes */ |
2665abfd | 478 | #define REQ_F_LINK_TIMEOUT 128 /* has linked timeout */ |
f7b76ac9 | 479 | #define REQ_F_FAIL_LINK 256 /* fail rest of links */ |
1b4a51b6 | 480 | #define REQ_F_DRAIN_LINK 512 /* link should be fully drained */ |
5262f567 | 481 | #define REQ_F_TIMEOUT 1024 /* timeout request */ |
491381ce JA |
482 | #define REQ_F_ISREG 2048 /* regular file */ |
483 | #define REQ_F_MUST_PUNT 4096 /* must be punted even for NONBLOCK */ | |
93bd25bb | 484 | #define REQ_F_TIMEOUT_NOSEQ 8192 /* no timeout sequence */ |
fb4b3d3f LT |
485 | #define REQ_F_INFLIGHT 16384 /* on inflight list */ |
486 | #define REQ_F_COMP_LOCKED 32768 /* completion under lock */ | |
4e88d6e7 | 487 | #define REQ_F_HARDLINK 65536 /* doesn't sever on completion < 0 */ |
ce35a47a | 488 | #define REQ_F_FORCE_ASYNC 131072 /* IOSQE_ASYNC */ |
2b188cc1 | 489 | u64 user_data; |
9e645e11 | 490 | u32 result; |
de0617e4 | 491 | u32 sequence; |
2b188cc1 | 492 | |
fcb323cc JA |
493 | struct list_head inflight_entry; |
494 | ||
561fb04a | 495 | struct io_wq_work work; |
2b188cc1 JA |
496 | }; |
497 | ||
498 | #define IO_PLUG_THRESHOLD 2 | |
def596e9 | 499 | #define IO_IOPOLL_BATCH 8 |
2b188cc1 | 500 | |
9a56a232 JA |
501 | struct io_submit_state { |
502 | struct blk_plug plug; | |
503 | ||
2579f913 JA |
504 | /* |
505 | * io_kiocb alloc cache | |
506 | */ | |
507 | void *reqs[IO_IOPOLL_BATCH]; | |
508 | unsigned int free_reqs; | |
509 | unsigned int cur_req; | |
510 | ||
9a56a232 JA |
511 | /* |
512 | * File reference cache | |
513 | */ | |
514 | struct file *file; | |
515 | unsigned int fd; | |
516 | unsigned int has_refs; | |
517 | unsigned int used_refs; | |
518 | unsigned int ios_left; | |
519 | }; | |
520 | ||
d3656344 JA |
521 | struct io_op_def { |
522 | /* needs req->io allocated for deferral/async */ | |
523 | unsigned async_ctx : 1; | |
524 | /* needs current->mm setup, does mm access */ | |
525 | unsigned needs_mm : 1; | |
526 | /* needs req->file assigned */ | |
527 | unsigned needs_file : 1; | |
528 | /* needs req->file assigned IFF fd is >= 0 */ | |
529 | unsigned fd_non_neg : 1; | |
530 | /* hash wq insertion if file is a regular file */ | |
531 | unsigned hash_reg_file : 1; | |
532 | /* unbound wq insertion if file is a non-regular file */ | |
533 | unsigned unbound_nonreg_file : 1; | |
534 | }; | |
535 | ||
536 | static const struct io_op_def io_op_defs[] = { | |
537 | { | |
538 | /* IORING_OP_NOP */ | |
539 | }, | |
540 | { | |
541 | /* IORING_OP_READV */ | |
542 | .async_ctx = 1, | |
543 | .needs_mm = 1, | |
544 | .needs_file = 1, | |
545 | .unbound_nonreg_file = 1, | |
546 | }, | |
547 | { | |
548 | /* IORING_OP_WRITEV */ | |
549 | .async_ctx = 1, | |
550 | .needs_mm = 1, | |
551 | .needs_file = 1, | |
552 | .hash_reg_file = 1, | |
553 | .unbound_nonreg_file = 1, | |
554 | }, | |
555 | { | |
556 | /* IORING_OP_FSYNC */ | |
557 | .needs_file = 1, | |
558 | }, | |
559 | { | |
560 | /* IORING_OP_READ_FIXED */ | |
561 | .needs_file = 1, | |
562 | .unbound_nonreg_file = 1, | |
563 | }, | |
564 | { | |
565 | /* IORING_OP_WRITE_FIXED */ | |
566 | .needs_file = 1, | |
567 | .hash_reg_file = 1, | |
568 | .unbound_nonreg_file = 1, | |
569 | }, | |
570 | { | |
571 | /* IORING_OP_POLL_ADD */ | |
572 | .needs_file = 1, | |
573 | .unbound_nonreg_file = 1, | |
574 | }, | |
575 | { | |
576 | /* IORING_OP_POLL_REMOVE */ | |
577 | }, | |
578 | { | |
579 | /* IORING_OP_SYNC_FILE_RANGE */ | |
580 | .needs_file = 1, | |
581 | }, | |
582 | { | |
583 | /* IORING_OP_SENDMSG */ | |
584 | .async_ctx = 1, | |
585 | .needs_mm = 1, | |
586 | .needs_file = 1, | |
587 | .unbound_nonreg_file = 1, | |
588 | }, | |
589 | { | |
590 | /* IORING_OP_RECVMSG */ | |
591 | .async_ctx = 1, | |
592 | .needs_mm = 1, | |
593 | .needs_file = 1, | |
594 | .unbound_nonreg_file = 1, | |
595 | }, | |
596 | { | |
597 | /* IORING_OP_TIMEOUT */ | |
598 | .async_ctx = 1, | |
599 | .needs_mm = 1, | |
600 | }, | |
601 | { | |
602 | /* IORING_OP_TIMEOUT_REMOVE */ | |
603 | }, | |
604 | { | |
605 | /* IORING_OP_ACCEPT */ | |
606 | .needs_mm = 1, | |
607 | .needs_file = 1, | |
608 | .unbound_nonreg_file = 1, | |
609 | }, | |
610 | { | |
611 | /* IORING_OP_ASYNC_CANCEL */ | |
612 | }, | |
613 | { | |
614 | /* IORING_OP_LINK_TIMEOUT */ | |
615 | .async_ctx = 1, | |
616 | .needs_mm = 1, | |
617 | }, | |
618 | { | |
619 | /* IORING_OP_CONNECT */ | |
620 | .async_ctx = 1, | |
621 | .needs_mm = 1, | |
622 | .needs_file = 1, | |
623 | .unbound_nonreg_file = 1, | |
624 | }, | |
625 | { | |
626 | /* IORING_OP_FALLOCATE */ | |
627 | .needs_file = 1, | |
628 | }, | |
629 | { | |
630 | /* IORING_OP_OPENAT */ | |
631 | .needs_file = 1, | |
632 | .fd_non_neg = 1, | |
633 | }, | |
634 | { | |
635 | /* IORING_OP_CLOSE */ | |
636 | .needs_file = 1, | |
637 | }, | |
638 | { | |
639 | /* IORING_OP_FILES_UPDATE */ | |
640 | .needs_mm = 1, | |
641 | }, | |
642 | { | |
643 | /* IORING_OP_STATX */ | |
644 | .needs_mm = 1, | |
645 | .needs_file = 1, | |
646 | .fd_non_neg = 1, | |
647 | }, | |
648 | }; | |
649 | ||
561fb04a | 650 | static void io_wq_submit_work(struct io_wq_work **workptr); |
78e19bbe | 651 | static void io_cqring_fill_event(struct io_kiocb *req, long res); |
ec9c02ad | 652 | static void io_put_req(struct io_kiocb *req); |
978db57e | 653 | static void __io_double_put_req(struct io_kiocb *req); |
94ae5e77 JA |
654 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req); |
655 | static void io_queue_linked_timeout(struct io_kiocb *req); | |
05f3fb3c JA |
656 | static int __io_sqe_files_update(struct io_ring_ctx *ctx, |
657 | struct io_uring_files_update *ip, | |
658 | unsigned nr_args); | |
de0617e4 | 659 | |
2b188cc1 JA |
660 | static struct kmem_cache *req_cachep; |
661 | ||
662 | static const struct file_operations io_uring_fops; | |
663 | ||
664 | struct sock *io_uring_get_socket(struct file *file) | |
665 | { | |
666 | #if defined(CONFIG_UNIX) | |
667 | if (file->f_op == &io_uring_fops) { | |
668 | struct io_ring_ctx *ctx = file->private_data; | |
669 | ||
670 | return ctx->ring_sock->sk; | |
671 | } | |
672 | #endif | |
673 | return NULL; | |
674 | } | |
675 | EXPORT_SYMBOL(io_uring_get_socket); | |
676 | ||
677 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) | |
678 | { | |
679 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
680 | ||
206aefde | 681 | complete(&ctx->completions[0]); |
2b188cc1 JA |
682 | } |
683 | ||
684 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) | |
685 | { | |
686 | struct io_ring_ctx *ctx; | |
78076bb6 | 687 | int hash_bits; |
2b188cc1 JA |
688 | |
689 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
690 | if (!ctx) | |
691 | return NULL; | |
692 | ||
0ddf92e8 JA |
693 | ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL); |
694 | if (!ctx->fallback_req) | |
695 | goto err; | |
696 | ||
206aefde JA |
697 | ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL); |
698 | if (!ctx->completions) | |
699 | goto err; | |
700 | ||
78076bb6 JA |
701 | /* |
702 | * Use 5 bits less than the max cq entries, that should give us around | |
703 | * 32 entries per hash list if totally full and uniformly spread. | |
704 | */ | |
705 | hash_bits = ilog2(p->cq_entries); | |
706 | hash_bits -= 5; | |
707 | if (hash_bits <= 0) | |
708 | hash_bits = 1; | |
709 | ctx->cancel_hash_bits = hash_bits; | |
710 | ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head), | |
711 | GFP_KERNEL); | |
712 | if (!ctx->cancel_hash) | |
713 | goto err; | |
714 | __hash_init(ctx->cancel_hash, 1U << hash_bits); | |
715 | ||
21482896 | 716 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
206aefde JA |
717 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) |
718 | goto err; | |
2b188cc1 JA |
719 | |
720 | ctx->flags = p->flags; | |
721 | init_waitqueue_head(&ctx->cq_wait); | |
1d7bb1d5 | 722 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
206aefde JA |
723 | init_completion(&ctx->completions[0]); |
724 | init_completion(&ctx->completions[1]); | |
2b188cc1 JA |
725 | mutex_init(&ctx->uring_lock); |
726 | init_waitqueue_head(&ctx->wait); | |
727 | spin_lock_init(&ctx->completion_lock); | |
def596e9 | 728 | INIT_LIST_HEAD(&ctx->poll_list); |
de0617e4 | 729 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 730 | INIT_LIST_HEAD(&ctx->timeout_list); |
fcb323cc JA |
731 | init_waitqueue_head(&ctx->inflight_wait); |
732 | spin_lock_init(&ctx->inflight_lock); | |
733 | INIT_LIST_HEAD(&ctx->inflight_list); | |
2b188cc1 | 734 | return ctx; |
206aefde | 735 | err: |
0ddf92e8 JA |
736 | if (ctx->fallback_req) |
737 | kmem_cache_free(req_cachep, ctx->fallback_req); | |
206aefde | 738 | kfree(ctx->completions); |
78076bb6 | 739 | kfree(ctx->cancel_hash); |
206aefde JA |
740 | kfree(ctx); |
741 | return NULL; | |
2b188cc1 JA |
742 | } |
743 | ||
9d858b21 | 744 | static inline bool __req_need_defer(struct io_kiocb *req) |
7adf4eaf | 745 | { |
a197f664 JL |
746 | struct io_ring_ctx *ctx = req->ctx; |
747 | ||
498ccd9e JA |
748 | return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped |
749 | + atomic_read(&ctx->cached_cq_overflow); | |
7adf4eaf JA |
750 | } |
751 | ||
9d858b21 | 752 | static inline bool req_need_defer(struct io_kiocb *req) |
de0617e4 | 753 | { |
9d858b21 BL |
754 | if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) == REQ_F_IO_DRAIN) |
755 | return __req_need_defer(req); | |
de0617e4 | 756 | |
9d858b21 | 757 | return false; |
de0617e4 JA |
758 | } |
759 | ||
7adf4eaf | 760 | static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx) |
de0617e4 JA |
761 | { |
762 | struct io_kiocb *req; | |
763 | ||
7adf4eaf | 764 | req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list); |
9d858b21 | 765 | if (req && !req_need_defer(req)) { |
de0617e4 JA |
766 | list_del_init(&req->list); |
767 | return req; | |
768 | } | |
769 | ||
770 | return NULL; | |
771 | } | |
772 | ||
5262f567 JA |
773 | static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx) |
774 | { | |
7adf4eaf JA |
775 | struct io_kiocb *req; |
776 | ||
777 | req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list); | |
93bd25bb JA |
778 | if (req) { |
779 | if (req->flags & REQ_F_TIMEOUT_NOSEQ) | |
780 | return NULL; | |
fb4b3d3f | 781 | if (!__req_need_defer(req)) { |
93bd25bb JA |
782 | list_del_init(&req->list); |
783 | return req; | |
784 | } | |
7adf4eaf JA |
785 | } |
786 | ||
787 | return NULL; | |
5262f567 JA |
788 | } |
789 | ||
de0617e4 | 790 | static void __io_commit_cqring(struct io_ring_ctx *ctx) |
2b188cc1 | 791 | { |
75b28aff | 792 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 793 | |
75b28aff | 794 | if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) { |
2b188cc1 | 795 | /* order cqe stores with ring update */ |
75b28aff | 796 | smp_store_release(&rings->cq.tail, ctx->cached_cq_tail); |
2b188cc1 | 797 | |
2b188cc1 JA |
798 | if (wq_has_sleeper(&ctx->cq_wait)) { |
799 | wake_up_interruptible(&ctx->cq_wait); | |
800 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
801 | } | |
802 | } | |
803 | } | |
804 | ||
94ae5e77 JA |
805 | static inline bool io_prep_async_work(struct io_kiocb *req, |
806 | struct io_kiocb **link) | |
18d9be1a | 807 | { |
d3656344 | 808 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
561fb04a | 809 | bool do_hashed = false; |
54a91f3b | 810 | |
d3656344 JA |
811 | if (req->flags & REQ_F_ISREG) { |
812 | if (def->hash_reg_file) | |
3529d8c2 | 813 | do_hashed = true; |
d3656344 JA |
814 | } else { |
815 | if (def->unbound_nonreg_file) | |
3529d8c2 | 816 | req->work.flags |= IO_WQ_WORK_UNBOUND; |
54a91f3b | 817 | } |
d3656344 | 818 | if (def->needs_mm) |
3529d8c2 | 819 | req->work.flags |= IO_WQ_WORK_NEEDS_USER; |
54a91f3b | 820 | |
94ae5e77 | 821 | *link = io_prep_linked_timeout(req); |
561fb04a JA |
822 | return do_hashed; |
823 | } | |
824 | ||
a197f664 | 825 | static inline void io_queue_async_work(struct io_kiocb *req) |
561fb04a | 826 | { |
a197f664 | 827 | struct io_ring_ctx *ctx = req->ctx; |
94ae5e77 JA |
828 | struct io_kiocb *link; |
829 | bool do_hashed; | |
830 | ||
831 | do_hashed = io_prep_async_work(req, &link); | |
561fb04a JA |
832 | |
833 | trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work, | |
834 | req->flags); | |
835 | if (!do_hashed) { | |
836 | io_wq_enqueue(ctx->io_wq, &req->work); | |
837 | } else { | |
838 | io_wq_enqueue_hashed(ctx->io_wq, &req->work, | |
839 | file_inode(req->file)); | |
840 | } | |
94ae5e77 JA |
841 | |
842 | if (link) | |
843 | io_queue_linked_timeout(link); | |
18d9be1a JA |
844 | } |
845 | ||
5262f567 JA |
846 | static void io_kill_timeout(struct io_kiocb *req) |
847 | { | |
848 | int ret; | |
849 | ||
2d28390a | 850 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); |
5262f567 JA |
851 | if (ret != -1) { |
852 | atomic_inc(&req->ctx->cq_timeouts); | |
842f9612 | 853 | list_del_init(&req->list); |
78e19bbe | 854 | io_cqring_fill_event(req, 0); |
ec9c02ad | 855 | io_put_req(req); |
5262f567 JA |
856 | } |
857 | } | |
858 | ||
859 | static void io_kill_timeouts(struct io_ring_ctx *ctx) | |
860 | { | |
861 | struct io_kiocb *req, *tmp; | |
862 | ||
863 | spin_lock_irq(&ctx->completion_lock); | |
864 | list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list) | |
865 | io_kill_timeout(req); | |
866 | spin_unlock_irq(&ctx->completion_lock); | |
867 | } | |
868 | ||
de0617e4 JA |
869 | static void io_commit_cqring(struct io_ring_ctx *ctx) |
870 | { | |
871 | struct io_kiocb *req; | |
872 | ||
5262f567 JA |
873 | while ((req = io_get_timeout_req(ctx)) != NULL) |
874 | io_kill_timeout(req); | |
875 | ||
de0617e4 JA |
876 | __io_commit_cqring(ctx); |
877 | ||
878 | while ((req = io_get_deferred_req(ctx)) != NULL) { | |
879 | req->flags |= REQ_F_IO_DRAINED; | |
a197f664 | 880 | io_queue_async_work(req); |
de0617e4 JA |
881 | } |
882 | } | |
883 | ||
2b188cc1 JA |
884 | static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx) |
885 | { | |
75b28aff | 886 | struct io_rings *rings = ctx->rings; |
2b188cc1 JA |
887 | unsigned tail; |
888 | ||
889 | tail = ctx->cached_cq_tail; | |
115e12e5 SB |
890 | /* |
891 | * writes to the cq entry need to come after reading head; the | |
892 | * control dependency is enough as we're using WRITE_ONCE to | |
893 | * fill the cq entry | |
894 | */ | |
75b28aff | 895 | if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries) |
2b188cc1 JA |
896 | return NULL; |
897 | ||
898 | ctx->cached_cq_tail++; | |
75b28aff | 899 | return &rings->cqes[tail & ctx->cq_mask]; |
2b188cc1 JA |
900 | } |
901 | ||
1d7bb1d5 JA |
902 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
903 | { | |
904 | if (waitqueue_active(&ctx->wait)) | |
905 | wake_up(&ctx->wait); | |
906 | if (waitqueue_active(&ctx->sqo_wait)) | |
907 | wake_up(&ctx->sqo_wait); | |
908 | if (ctx->cq_ev_fd) | |
909 | eventfd_signal(ctx->cq_ev_fd, 1); | |
910 | } | |
911 | ||
c4a2ed72 JA |
912 | /* Returns true if there are no backlogged entries after the flush */ |
913 | static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force) | |
1d7bb1d5 JA |
914 | { |
915 | struct io_rings *rings = ctx->rings; | |
916 | struct io_uring_cqe *cqe; | |
917 | struct io_kiocb *req; | |
918 | unsigned long flags; | |
919 | LIST_HEAD(list); | |
920 | ||
921 | if (!force) { | |
922 | if (list_empty_careful(&ctx->cq_overflow_list)) | |
c4a2ed72 | 923 | return true; |
1d7bb1d5 JA |
924 | if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) == |
925 | rings->cq_ring_entries)) | |
c4a2ed72 | 926 | return false; |
1d7bb1d5 JA |
927 | } |
928 | ||
929 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
930 | ||
931 | /* if force is set, the ring is going away. always drop after that */ | |
932 | if (force) | |
933 | ctx->cq_overflow_flushed = true; | |
934 | ||
c4a2ed72 | 935 | cqe = NULL; |
1d7bb1d5 JA |
936 | while (!list_empty(&ctx->cq_overflow_list)) { |
937 | cqe = io_get_cqring(ctx); | |
938 | if (!cqe && !force) | |
939 | break; | |
940 | ||
941 | req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb, | |
942 | list); | |
943 | list_move(&req->list, &list); | |
944 | if (cqe) { | |
945 | WRITE_ONCE(cqe->user_data, req->user_data); | |
946 | WRITE_ONCE(cqe->res, req->result); | |
947 | WRITE_ONCE(cqe->flags, 0); | |
948 | } else { | |
949 | WRITE_ONCE(ctx->rings->cq_overflow, | |
950 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
951 | } | |
952 | } | |
953 | ||
954 | io_commit_cqring(ctx); | |
ad3eb2c8 JA |
955 | if (cqe) { |
956 | clear_bit(0, &ctx->sq_check_overflow); | |
957 | clear_bit(0, &ctx->cq_check_overflow); | |
958 | } | |
1d7bb1d5 JA |
959 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
960 | io_cqring_ev_posted(ctx); | |
961 | ||
962 | while (!list_empty(&list)) { | |
963 | req = list_first_entry(&list, struct io_kiocb, list); | |
964 | list_del(&req->list); | |
ec9c02ad | 965 | io_put_req(req); |
1d7bb1d5 | 966 | } |
c4a2ed72 JA |
967 | |
968 | return cqe != NULL; | |
1d7bb1d5 JA |
969 | } |
970 | ||
78e19bbe | 971 | static void io_cqring_fill_event(struct io_kiocb *req, long res) |
2b188cc1 | 972 | { |
78e19bbe | 973 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
974 | struct io_uring_cqe *cqe; |
975 | ||
78e19bbe | 976 | trace_io_uring_complete(ctx, req->user_data, res); |
51c3ff62 | 977 | |
2b188cc1 JA |
978 | /* |
979 | * If we can't get a cq entry, userspace overflowed the | |
980 | * submission (by quite a lot). Increment the overflow count in | |
981 | * the ring. | |
982 | */ | |
983 | cqe = io_get_cqring(ctx); | |
1d7bb1d5 | 984 | if (likely(cqe)) { |
78e19bbe | 985 | WRITE_ONCE(cqe->user_data, req->user_data); |
2b188cc1 | 986 | WRITE_ONCE(cqe->res, res); |
c71ffb67 | 987 | WRITE_ONCE(cqe->flags, 0); |
1d7bb1d5 | 988 | } else if (ctx->cq_overflow_flushed) { |
498ccd9e JA |
989 | WRITE_ONCE(ctx->rings->cq_overflow, |
990 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
1d7bb1d5 | 991 | } else { |
ad3eb2c8 JA |
992 | if (list_empty(&ctx->cq_overflow_list)) { |
993 | set_bit(0, &ctx->sq_check_overflow); | |
994 | set_bit(0, &ctx->cq_check_overflow); | |
995 | } | |
1d7bb1d5 JA |
996 | refcount_inc(&req->refs); |
997 | req->result = res; | |
998 | list_add_tail(&req->list, &ctx->cq_overflow_list); | |
2b188cc1 JA |
999 | } |
1000 | } | |
1001 | ||
78e19bbe | 1002 | static void io_cqring_add_event(struct io_kiocb *req, long res) |
2b188cc1 | 1003 | { |
78e19bbe | 1004 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
1005 | unsigned long flags; |
1006 | ||
1007 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
78e19bbe | 1008 | io_cqring_fill_event(req, res); |
2b188cc1 JA |
1009 | io_commit_cqring(ctx); |
1010 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1011 | ||
8c838788 | 1012 | io_cqring_ev_posted(ctx); |
2b188cc1 JA |
1013 | } |
1014 | ||
0ddf92e8 JA |
1015 | static inline bool io_is_fallback_req(struct io_kiocb *req) |
1016 | { | |
1017 | return req == (struct io_kiocb *) | |
1018 | ((unsigned long) req->ctx->fallback_req & ~1UL); | |
1019 | } | |
1020 | ||
1021 | static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx) | |
1022 | { | |
1023 | struct io_kiocb *req; | |
1024 | ||
1025 | req = ctx->fallback_req; | |
1026 | if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req)) | |
1027 | return req; | |
1028 | ||
1029 | return NULL; | |
1030 | } | |
1031 | ||
2579f913 JA |
1032 | static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx, |
1033 | struct io_submit_state *state) | |
2b188cc1 | 1034 | { |
fd6fab2c | 1035 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
2b188cc1 JA |
1036 | struct io_kiocb *req; |
1037 | ||
1038 | if (!percpu_ref_tryget(&ctx->refs)) | |
1039 | return NULL; | |
1040 | ||
2579f913 | 1041 | if (!state) { |
fd6fab2c | 1042 | req = kmem_cache_alloc(req_cachep, gfp); |
2579f913 | 1043 | if (unlikely(!req)) |
0ddf92e8 | 1044 | goto fallback; |
2579f913 JA |
1045 | } else if (!state->free_reqs) { |
1046 | size_t sz; | |
1047 | int ret; | |
1048 | ||
1049 | sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs)); | |
fd6fab2c JA |
1050 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs); |
1051 | ||
1052 | /* | |
1053 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
1054 | * retry single alloc to be on the safe side. | |
1055 | */ | |
1056 | if (unlikely(ret <= 0)) { | |
1057 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
1058 | if (!state->reqs[0]) | |
0ddf92e8 | 1059 | goto fallback; |
fd6fab2c JA |
1060 | ret = 1; |
1061 | } | |
2579f913 JA |
1062 | state->free_reqs = ret - 1; |
1063 | state->cur_req = 1; | |
1064 | req = state->reqs[0]; | |
1065 | } else { | |
1066 | req = state->reqs[state->cur_req]; | |
1067 | state->free_reqs--; | |
1068 | state->cur_req++; | |
2b188cc1 JA |
1069 | } |
1070 | ||
0ddf92e8 | 1071 | got_it: |
1a6b74fc | 1072 | req->io = NULL; |
cf6fd4bd | 1073 | req->ring_file = NULL; |
60c112b0 | 1074 | req->file = NULL; |
2579f913 JA |
1075 | req->ctx = ctx; |
1076 | req->flags = 0; | |
e65ef56d JA |
1077 | /* one is dropped after submission, the other at completion */ |
1078 | refcount_set(&req->refs, 2); | |
9e645e11 | 1079 | req->result = 0; |
561fb04a | 1080 | INIT_IO_WORK(&req->work, io_wq_submit_work); |
2579f913 | 1081 | return req; |
0ddf92e8 JA |
1082 | fallback: |
1083 | req = io_get_fallback_req(ctx); | |
1084 | if (req) | |
1085 | goto got_it; | |
6805b32e | 1086 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
1087 | return NULL; |
1088 | } | |
1089 | ||
def596e9 JA |
1090 | static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr) |
1091 | { | |
1092 | if (*nr) { | |
1093 | kmem_cache_free_bulk(req_cachep, *nr, reqs); | |
6805b32e | 1094 | percpu_ref_put_many(&ctx->refs, *nr); |
05f3fb3c | 1095 | percpu_ref_put_many(&ctx->file_data->refs, *nr); |
def596e9 JA |
1096 | *nr = 0; |
1097 | } | |
1098 | } | |
1099 | ||
9e645e11 | 1100 | static void __io_free_req(struct io_kiocb *req) |
2b188cc1 | 1101 | { |
fcb323cc JA |
1102 | struct io_ring_ctx *ctx = req->ctx; |
1103 | ||
1a6b74fc JA |
1104 | if (req->io) |
1105 | kfree(req->io); | |
05f3fb3c JA |
1106 | if (req->file) { |
1107 | if (req->flags & REQ_F_FIXED_FILE) | |
1108 | percpu_ref_put(&ctx->file_data->refs); | |
1109 | else | |
1110 | fput(req->file); | |
1111 | } | |
fcb323cc JA |
1112 | if (req->flags & REQ_F_INFLIGHT) { |
1113 | unsigned long flags; | |
1114 | ||
1115 | spin_lock_irqsave(&ctx->inflight_lock, flags); | |
1116 | list_del(&req->inflight_entry); | |
1117 | if (waitqueue_active(&ctx->inflight_wait)) | |
1118 | wake_up(&ctx->inflight_wait); | |
1119 | spin_unlock_irqrestore(&ctx->inflight_lock, flags); | |
1120 | } | |
1121 | percpu_ref_put(&ctx->refs); | |
0ddf92e8 JA |
1122 | if (likely(!io_is_fallback_req(req))) |
1123 | kmem_cache_free(req_cachep, req); | |
1124 | else | |
1125 | clear_bit_unlock(0, (unsigned long *) ctx->fallback_req); | |
e65ef56d JA |
1126 | } |
1127 | ||
a197f664 | 1128 | static bool io_link_cancel_timeout(struct io_kiocb *req) |
2665abfd | 1129 | { |
a197f664 | 1130 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd JA |
1131 | int ret; |
1132 | ||
2d28390a | 1133 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); |
2665abfd | 1134 | if (ret != -1) { |
78e19bbe | 1135 | io_cqring_fill_event(req, -ECANCELED); |
2665abfd JA |
1136 | io_commit_cqring(ctx); |
1137 | req->flags &= ~REQ_F_LINK; | |
ec9c02ad | 1138 | io_put_req(req); |
2665abfd JA |
1139 | return true; |
1140 | } | |
1141 | ||
1142 | return false; | |
e65ef56d JA |
1143 | } |
1144 | ||
ba816ad6 | 1145 | static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr) |
9e645e11 | 1146 | { |
2665abfd | 1147 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd | 1148 | bool wake_ev = false; |
9e645e11 | 1149 | |
4d7dd462 JA |
1150 | /* Already got next link */ |
1151 | if (req->flags & REQ_F_LINK_NEXT) | |
1152 | return; | |
1153 | ||
9e645e11 JA |
1154 | /* |
1155 | * The list should never be empty when we are called here. But could | |
1156 | * potentially happen if the chain is messed up, check to be on the | |
1157 | * safe side. | |
1158 | */ | |
4493233e PB |
1159 | while (!list_empty(&req->link_list)) { |
1160 | struct io_kiocb *nxt = list_first_entry(&req->link_list, | |
1161 | struct io_kiocb, link_list); | |
94ae5e77 | 1162 | |
4493233e PB |
1163 | if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) && |
1164 | (nxt->flags & REQ_F_TIMEOUT))) { | |
1165 | list_del_init(&nxt->link_list); | |
94ae5e77 | 1166 | wake_ev |= io_link_cancel_timeout(nxt); |
94ae5e77 JA |
1167 | req->flags &= ~REQ_F_LINK_TIMEOUT; |
1168 | continue; | |
1169 | } | |
9e645e11 | 1170 | |
4493233e PB |
1171 | list_del_init(&req->link_list); |
1172 | if (!list_empty(&nxt->link_list)) | |
1173 | nxt->flags |= REQ_F_LINK; | |
b18fdf71 | 1174 | *nxtptr = nxt; |
94ae5e77 | 1175 | break; |
9e645e11 | 1176 | } |
2665abfd | 1177 | |
4d7dd462 | 1178 | req->flags |= REQ_F_LINK_NEXT; |
2665abfd JA |
1179 | if (wake_ev) |
1180 | io_cqring_ev_posted(ctx); | |
9e645e11 JA |
1181 | } |
1182 | ||
1183 | /* | |
1184 | * Called if REQ_F_LINK is set, and we fail the head request | |
1185 | */ | |
1186 | static void io_fail_links(struct io_kiocb *req) | |
1187 | { | |
2665abfd | 1188 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd JA |
1189 | unsigned long flags; |
1190 | ||
1191 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
9e645e11 JA |
1192 | |
1193 | while (!list_empty(&req->link_list)) { | |
4493233e PB |
1194 | struct io_kiocb *link = list_first_entry(&req->link_list, |
1195 | struct io_kiocb, link_list); | |
9e645e11 | 1196 | |
4493233e | 1197 | list_del_init(&link->link_list); |
c826bd7a | 1198 | trace_io_uring_fail_link(req, link); |
2665abfd JA |
1199 | |
1200 | if ((req->flags & REQ_F_LINK_TIMEOUT) && | |
d625c6ee | 1201 | link->opcode == IORING_OP_LINK_TIMEOUT) { |
a197f664 | 1202 | io_link_cancel_timeout(link); |
2665abfd | 1203 | } else { |
78e19bbe | 1204 | io_cqring_fill_event(link, -ECANCELED); |
978db57e | 1205 | __io_double_put_req(link); |
2665abfd | 1206 | } |
5d960724 | 1207 | req->flags &= ~REQ_F_LINK_TIMEOUT; |
9e645e11 | 1208 | } |
2665abfd JA |
1209 | |
1210 | io_commit_cqring(ctx); | |
1211 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1212 | io_cqring_ev_posted(ctx); | |
9e645e11 JA |
1213 | } |
1214 | ||
4d7dd462 | 1215 | static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt) |
9e645e11 | 1216 | { |
4d7dd462 | 1217 | if (likely(!(req->flags & REQ_F_LINK))) |
2665abfd | 1218 | return; |
2665abfd | 1219 | |
9e645e11 JA |
1220 | /* |
1221 | * If LINK is set, we have dependent requests in this chain. If we | |
1222 | * didn't fail this request, queue the first one up, moving any other | |
1223 | * dependencies to the next request. In case of failure, fail the rest | |
1224 | * of the chain. | |
1225 | */ | |
2665abfd JA |
1226 | if (req->flags & REQ_F_FAIL_LINK) { |
1227 | io_fail_links(req); | |
7c9e7f0f JA |
1228 | } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) == |
1229 | REQ_F_LINK_TIMEOUT) { | |
2665abfd JA |
1230 | struct io_ring_ctx *ctx = req->ctx; |
1231 | unsigned long flags; | |
1232 | ||
1233 | /* | |
1234 | * If this is a timeout link, we could be racing with the | |
1235 | * timeout timer. Grab the completion lock for this case to | |
7c9e7f0f | 1236 | * protect against that. |
2665abfd JA |
1237 | */ |
1238 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
1239 | io_req_link_next(req, nxt); | |
1240 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1241 | } else { | |
1242 | io_req_link_next(req, nxt); | |
9e645e11 | 1243 | } |
4d7dd462 | 1244 | } |
9e645e11 | 1245 | |
c69f8dbe JL |
1246 | static void io_free_req(struct io_kiocb *req) |
1247 | { | |
944e58bf PB |
1248 | struct io_kiocb *nxt = NULL; |
1249 | ||
1250 | io_req_find_next(req, &nxt); | |
70cf9f32 | 1251 | __io_free_req(req); |
944e58bf PB |
1252 | |
1253 | if (nxt) | |
1254 | io_queue_async_work(nxt); | |
c69f8dbe JL |
1255 | } |
1256 | ||
ba816ad6 JA |
1257 | /* |
1258 | * Drop reference to request, return next in chain (if there is one) if this | |
1259 | * was the last reference to this request. | |
1260 | */ | |
f9bd67f6 | 1261 | __attribute__((nonnull)) |
ec9c02ad | 1262 | static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr) |
e65ef56d | 1263 | { |
f9bd67f6 | 1264 | io_req_find_next(req, nxtptr); |
4d7dd462 | 1265 | |
e65ef56d | 1266 | if (refcount_dec_and_test(&req->refs)) |
4d7dd462 | 1267 | __io_free_req(req); |
2b188cc1 JA |
1268 | } |
1269 | ||
e65ef56d JA |
1270 | static void io_put_req(struct io_kiocb *req) |
1271 | { | |
1272 | if (refcount_dec_and_test(&req->refs)) | |
1273 | io_free_req(req); | |
2b188cc1 JA |
1274 | } |
1275 | ||
978db57e JA |
1276 | /* |
1277 | * Must only be used if we don't need to care about links, usually from | |
1278 | * within the completion handling itself. | |
1279 | */ | |
1280 | static void __io_double_put_req(struct io_kiocb *req) | |
78e19bbe JA |
1281 | { |
1282 | /* drop both submit and complete references */ | |
1283 | if (refcount_sub_and_test(2, &req->refs)) | |
1284 | __io_free_req(req); | |
1285 | } | |
1286 | ||
978db57e JA |
1287 | static void io_double_put_req(struct io_kiocb *req) |
1288 | { | |
1289 | /* drop both submit and complete references */ | |
1290 | if (refcount_sub_and_test(2, &req->refs)) | |
1291 | io_free_req(req); | |
1292 | } | |
1293 | ||
1d7bb1d5 | 1294 | static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush) |
a3a0e43f | 1295 | { |
84f97dc2 JA |
1296 | struct io_rings *rings = ctx->rings; |
1297 | ||
ad3eb2c8 JA |
1298 | if (test_bit(0, &ctx->cq_check_overflow)) { |
1299 | /* | |
1300 | * noflush == true is from the waitqueue handler, just ensure | |
1301 | * we wake up the task, and the next invocation will flush the | |
1302 | * entries. We cannot safely to it from here. | |
1303 | */ | |
1304 | if (noflush && !list_empty(&ctx->cq_overflow_list)) | |
1305 | return -1U; | |
1d7bb1d5 | 1306 | |
ad3eb2c8 JA |
1307 | io_cqring_overflow_flush(ctx, false); |
1308 | } | |
1d7bb1d5 | 1309 | |
a3a0e43f JA |
1310 | /* See comment at the top of this file */ |
1311 | smp_rmb(); | |
ad3eb2c8 | 1312 | return ctx->cached_cq_tail - READ_ONCE(rings->cq.head); |
a3a0e43f JA |
1313 | } |
1314 | ||
fb5ccc98 PB |
1315 | static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) |
1316 | { | |
1317 | struct io_rings *rings = ctx->rings; | |
1318 | ||
1319 | /* make sure SQ entry isn't read before tail */ | |
1320 | return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; | |
1321 | } | |
1322 | ||
def596e9 JA |
1323 | /* |
1324 | * Find and free completed poll iocbs | |
1325 | */ | |
1326 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1327 | struct list_head *done) | |
1328 | { | |
1329 | void *reqs[IO_IOPOLL_BATCH]; | |
1330 | struct io_kiocb *req; | |
09bb8394 | 1331 | int to_free; |
def596e9 | 1332 | |
09bb8394 | 1333 | to_free = 0; |
def596e9 JA |
1334 | while (!list_empty(done)) { |
1335 | req = list_first_entry(done, struct io_kiocb, list); | |
1336 | list_del(&req->list); | |
1337 | ||
78e19bbe | 1338 | io_cqring_fill_event(req, req->result); |
def596e9 JA |
1339 | (*nr_events)++; |
1340 | ||
09bb8394 JA |
1341 | if (refcount_dec_and_test(&req->refs)) { |
1342 | /* If we're not using fixed files, we have to pair the | |
1343 | * completion part with the file put. Use regular | |
1344 | * completions for those, only batch free for fixed | |
9e645e11 | 1345 | * file and non-linked commands. |
09bb8394 | 1346 | */ |
1a6b74fc JA |
1347 | if (((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) == |
1348 | REQ_F_FIXED_FILE) && !io_is_fallback_req(req) && | |
1349 | !req->io) { | |
09bb8394 JA |
1350 | reqs[to_free++] = req; |
1351 | if (to_free == ARRAY_SIZE(reqs)) | |
1352 | io_free_req_many(ctx, reqs, &to_free); | |
6b06314c | 1353 | } else { |
09bb8394 | 1354 | io_free_req(req); |
6b06314c | 1355 | } |
9a56a232 | 1356 | } |
def596e9 | 1357 | } |
def596e9 | 1358 | |
09bb8394 | 1359 | io_commit_cqring(ctx); |
def596e9 JA |
1360 | io_free_req_many(ctx, reqs, &to_free); |
1361 | } | |
1362 | ||
1363 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1364 | long min) | |
1365 | { | |
1366 | struct io_kiocb *req, *tmp; | |
1367 | LIST_HEAD(done); | |
1368 | bool spin; | |
1369 | int ret; | |
1370 | ||
1371 | /* | |
1372 | * Only spin for completions if we don't have multiple devices hanging | |
1373 | * off our complete list, and we're under the requested amount. | |
1374 | */ | |
1375 | spin = !ctx->poll_multi_file && *nr_events < min; | |
1376 | ||
1377 | ret = 0; | |
1378 | list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) { | |
9adbd45d | 1379 | struct kiocb *kiocb = &req->rw.kiocb; |
def596e9 JA |
1380 | |
1381 | /* | |
1382 | * Move completed entries to our local list. If we find a | |
1383 | * request that requires polling, break out and complete | |
1384 | * the done list first, if we have entries there. | |
1385 | */ | |
1386 | if (req->flags & REQ_F_IOPOLL_COMPLETED) { | |
1387 | list_move_tail(&req->list, &done); | |
1388 | continue; | |
1389 | } | |
1390 | if (!list_empty(&done)) | |
1391 | break; | |
1392 | ||
1393 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
1394 | if (ret < 0) | |
1395 | break; | |
1396 | ||
1397 | if (ret && spin) | |
1398 | spin = false; | |
1399 | ret = 0; | |
1400 | } | |
1401 | ||
1402 | if (!list_empty(&done)) | |
1403 | io_iopoll_complete(ctx, nr_events, &done); | |
1404 | ||
1405 | return ret; | |
1406 | } | |
1407 | ||
1408 | /* | |
d195a66e | 1409 | * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a |
def596e9 JA |
1410 | * non-spinning poll check - we'll still enter the driver poll loop, but only |
1411 | * as a non-spinning completion check. | |
1412 | */ | |
1413 | static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1414 | long min) | |
1415 | { | |
08f5439f | 1416 | while (!list_empty(&ctx->poll_list) && !need_resched()) { |
def596e9 JA |
1417 | int ret; |
1418 | ||
1419 | ret = io_do_iopoll(ctx, nr_events, min); | |
1420 | if (ret < 0) | |
1421 | return ret; | |
1422 | if (!min || *nr_events >= min) | |
1423 | return 0; | |
1424 | } | |
1425 | ||
1426 | return 1; | |
1427 | } | |
1428 | ||
1429 | /* | |
1430 | * We can't just wait for polled events to come to us, we have to actively | |
1431 | * find and complete them. | |
1432 | */ | |
1433 | static void io_iopoll_reap_events(struct io_ring_ctx *ctx) | |
1434 | { | |
1435 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
1436 | return; | |
1437 | ||
1438 | mutex_lock(&ctx->uring_lock); | |
1439 | while (!list_empty(&ctx->poll_list)) { | |
1440 | unsigned int nr_events = 0; | |
1441 | ||
1442 | io_iopoll_getevents(ctx, &nr_events, 1); | |
08f5439f JA |
1443 | |
1444 | /* | |
1445 | * Ensure we allow local-to-the-cpu processing to take place, | |
1446 | * in this case we need to ensure that we reap all events. | |
1447 | */ | |
1448 | cond_resched(); | |
def596e9 JA |
1449 | } |
1450 | mutex_unlock(&ctx->uring_lock); | |
1451 | } | |
1452 | ||
2b2ed975 JA |
1453 | static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, |
1454 | long min) | |
def596e9 | 1455 | { |
2b2ed975 | 1456 | int iters = 0, ret = 0; |
500f9fba | 1457 | |
def596e9 JA |
1458 | do { |
1459 | int tmin = 0; | |
1460 | ||
a3a0e43f JA |
1461 | /* |
1462 | * Don't enter poll loop if we already have events pending. | |
1463 | * If we do, we can potentially be spinning for commands that | |
1464 | * already triggered a CQE (eg in error). | |
1465 | */ | |
1d7bb1d5 | 1466 | if (io_cqring_events(ctx, false)) |
a3a0e43f JA |
1467 | break; |
1468 | ||
500f9fba JA |
1469 | /* |
1470 | * If a submit got punted to a workqueue, we can have the | |
1471 | * application entering polling for a command before it gets | |
1472 | * issued. That app will hold the uring_lock for the duration | |
1473 | * of the poll right here, so we need to take a breather every | |
1474 | * now and then to ensure that the issue has a chance to add | |
1475 | * the poll to the issued list. Otherwise we can spin here | |
1476 | * forever, while the workqueue is stuck trying to acquire the | |
1477 | * very same mutex. | |
1478 | */ | |
1479 | if (!(++iters & 7)) { | |
1480 | mutex_unlock(&ctx->uring_lock); | |
1481 | mutex_lock(&ctx->uring_lock); | |
1482 | } | |
1483 | ||
def596e9 JA |
1484 | if (*nr_events < min) |
1485 | tmin = min - *nr_events; | |
1486 | ||
1487 | ret = io_iopoll_getevents(ctx, nr_events, tmin); | |
1488 | if (ret <= 0) | |
1489 | break; | |
1490 | ret = 0; | |
1491 | } while (min && !*nr_events && !need_resched()); | |
1492 | ||
2b2ed975 JA |
1493 | return ret; |
1494 | } | |
1495 | ||
1496 | static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, | |
1497 | long min) | |
1498 | { | |
1499 | int ret; | |
1500 | ||
1501 | /* | |
1502 | * We disallow the app entering submit/complete with polling, but we | |
1503 | * still need to lock the ring to prevent racing with polled issue | |
1504 | * that got punted to a workqueue. | |
1505 | */ | |
1506 | mutex_lock(&ctx->uring_lock); | |
1507 | ret = __io_iopoll_check(ctx, nr_events, min); | |
500f9fba | 1508 | mutex_unlock(&ctx->uring_lock); |
def596e9 JA |
1509 | return ret; |
1510 | } | |
1511 | ||
491381ce | 1512 | static void kiocb_end_write(struct io_kiocb *req) |
2b188cc1 | 1513 | { |
491381ce JA |
1514 | /* |
1515 | * Tell lockdep we inherited freeze protection from submission | |
1516 | * thread. | |
1517 | */ | |
1518 | if (req->flags & REQ_F_ISREG) { | |
1519 | struct inode *inode = file_inode(req->file); | |
2b188cc1 | 1520 | |
491381ce | 1521 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); |
2b188cc1 | 1522 | } |
491381ce | 1523 | file_end_write(req->file); |
2b188cc1 JA |
1524 | } |
1525 | ||
4e88d6e7 JA |
1526 | static inline void req_set_fail_links(struct io_kiocb *req) |
1527 | { | |
1528 | if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK) | |
1529 | req->flags |= REQ_F_FAIL_LINK; | |
1530 | } | |
1531 | ||
ba816ad6 | 1532 | static void io_complete_rw_common(struct kiocb *kiocb, long res) |
2b188cc1 | 1533 | { |
9adbd45d | 1534 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
2b188cc1 | 1535 | |
491381ce JA |
1536 | if (kiocb->ki_flags & IOCB_WRITE) |
1537 | kiocb_end_write(req); | |
2b188cc1 | 1538 | |
4e88d6e7 JA |
1539 | if (res != req->result) |
1540 | req_set_fail_links(req); | |
78e19bbe | 1541 | io_cqring_add_event(req, res); |
ba816ad6 JA |
1542 | } |
1543 | ||
1544 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
1545 | { | |
9adbd45d | 1546 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
ba816ad6 JA |
1547 | |
1548 | io_complete_rw_common(kiocb, res); | |
e65ef56d | 1549 | io_put_req(req); |
2b188cc1 JA |
1550 | } |
1551 | ||
ba816ad6 JA |
1552 | static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res) |
1553 | { | |
9adbd45d | 1554 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
ec9c02ad | 1555 | struct io_kiocb *nxt = NULL; |
ba816ad6 JA |
1556 | |
1557 | io_complete_rw_common(kiocb, res); | |
ec9c02ad JL |
1558 | io_put_req_find_next(req, &nxt); |
1559 | ||
1560 | return nxt; | |
2b188cc1 JA |
1561 | } |
1562 | ||
def596e9 JA |
1563 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
1564 | { | |
9adbd45d | 1565 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
def596e9 | 1566 | |
491381ce JA |
1567 | if (kiocb->ki_flags & IOCB_WRITE) |
1568 | kiocb_end_write(req); | |
def596e9 | 1569 | |
4e88d6e7 JA |
1570 | if (res != req->result) |
1571 | req_set_fail_links(req); | |
9e645e11 | 1572 | req->result = res; |
def596e9 JA |
1573 | if (res != -EAGAIN) |
1574 | req->flags |= REQ_F_IOPOLL_COMPLETED; | |
1575 | } | |
1576 | ||
1577 | /* | |
1578 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1579 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
1580 | * find it from a io_iopoll_getevents() thread before the issuer is done | |
1581 | * accessing the kiocb cookie. | |
1582 | */ | |
1583 | static void io_iopoll_req_issued(struct io_kiocb *req) | |
1584 | { | |
1585 | struct io_ring_ctx *ctx = req->ctx; | |
1586 | ||
1587 | /* | |
1588 | * Track whether we have multiple files in our lists. This will impact | |
1589 | * how we do polling eventually, not spinning if we're on potentially | |
1590 | * different devices. | |
1591 | */ | |
1592 | if (list_empty(&ctx->poll_list)) { | |
1593 | ctx->poll_multi_file = false; | |
1594 | } else if (!ctx->poll_multi_file) { | |
1595 | struct io_kiocb *list_req; | |
1596 | ||
1597 | list_req = list_first_entry(&ctx->poll_list, struct io_kiocb, | |
1598 | list); | |
9adbd45d | 1599 | if (list_req->file != req->file) |
def596e9 JA |
1600 | ctx->poll_multi_file = true; |
1601 | } | |
1602 | ||
1603 | /* | |
1604 | * For fast devices, IO may have already completed. If it has, add | |
1605 | * it to the front so we find it first. | |
1606 | */ | |
1607 | if (req->flags & REQ_F_IOPOLL_COMPLETED) | |
1608 | list_add(&req->list, &ctx->poll_list); | |
1609 | else | |
1610 | list_add_tail(&req->list, &ctx->poll_list); | |
1611 | } | |
1612 | ||
3d6770fb | 1613 | static void io_file_put(struct io_submit_state *state) |
9a56a232 | 1614 | { |
3d6770fb | 1615 | if (state->file) { |
9a56a232 JA |
1616 | int diff = state->has_refs - state->used_refs; |
1617 | ||
1618 | if (diff) | |
1619 | fput_many(state->file, diff); | |
1620 | state->file = NULL; | |
1621 | } | |
1622 | } | |
1623 | ||
1624 | /* | |
1625 | * Get as many references to a file as we have IOs left in this submission, | |
1626 | * assuming most submissions are for one file, or at least that each file | |
1627 | * has more than one submission. | |
1628 | */ | |
1629 | static struct file *io_file_get(struct io_submit_state *state, int fd) | |
1630 | { | |
1631 | if (!state) | |
1632 | return fget(fd); | |
1633 | ||
1634 | if (state->file) { | |
1635 | if (state->fd == fd) { | |
1636 | state->used_refs++; | |
1637 | state->ios_left--; | |
1638 | return state->file; | |
1639 | } | |
3d6770fb | 1640 | io_file_put(state); |
9a56a232 JA |
1641 | } |
1642 | state->file = fget_many(fd, state->ios_left); | |
1643 | if (!state->file) | |
1644 | return NULL; | |
1645 | ||
1646 | state->fd = fd; | |
1647 | state->has_refs = state->ios_left; | |
1648 | state->used_refs = 1; | |
1649 | state->ios_left--; | |
1650 | return state->file; | |
1651 | } | |
1652 | ||
2b188cc1 JA |
1653 | /* |
1654 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1655 | * any file. For now, just ensure that anything potentially problematic is done | |
1656 | * inline. | |
1657 | */ | |
1658 | static bool io_file_supports_async(struct file *file) | |
1659 | { | |
1660 | umode_t mode = file_inode(file)->i_mode; | |
1661 | ||
10d59345 | 1662 | if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode)) |
2b188cc1 JA |
1663 | return true; |
1664 | if (S_ISREG(mode) && file->f_op != &io_uring_fops) | |
1665 | return true; | |
1666 | ||
1667 | return false; | |
1668 | } | |
1669 | ||
3529d8c2 JA |
1670 | static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
1671 | bool force_nonblock) | |
2b188cc1 | 1672 | { |
def596e9 | 1673 | struct io_ring_ctx *ctx = req->ctx; |
9adbd45d | 1674 | struct kiocb *kiocb = &req->rw.kiocb; |
09bb8394 JA |
1675 | unsigned ioprio; |
1676 | int ret; | |
2b188cc1 | 1677 | |
09bb8394 JA |
1678 | if (!req->file) |
1679 | return -EBADF; | |
2b188cc1 | 1680 | |
491381ce JA |
1681 | if (S_ISREG(file_inode(req->file)->i_mode)) |
1682 | req->flags |= REQ_F_ISREG; | |
1683 | ||
2b188cc1 JA |
1684 | kiocb->ki_pos = READ_ONCE(sqe->off); |
1685 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); | |
1686 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); | |
1687 | ||
1688 | ioprio = READ_ONCE(sqe->ioprio); | |
1689 | if (ioprio) { | |
1690 | ret = ioprio_check_cap(ioprio); | |
1691 | if (ret) | |
09bb8394 | 1692 | return ret; |
2b188cc1 JA |
1693 | |
1694 | kiocb->ki_ioprio = ioprio; | |
1695 | } else | |
1696 | kiocb->ki_ioprio = get_current_ioprio(); | |
1697 | ||
1698 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
1699 | if (unlikely(ret)) | |
09bb8394 | 1700 | return ret; |
8449eeda SB |
1701 | |
1702 | /* don't allow async punt if RWF_NOWAIT was requested */ | |
491381ce JA |
1703 | if ((kiocb->ki_flags & IOCB_NOWAIT) || |
1704 | (req->file->f_flags & O_NONBLOCK)) | |
8449eeda SB |
1705 | req->flags |= REQ_F_NOWAIT; |
1706 | ||
1707 | if (force_nonblock) | |
2b188cc1 | 1708 | kiocb->ki_flags |= IOCB_NOWAIT; |
8449eeda | 1709 | |
def596e9 | 1710 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
1711 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
1712 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 1713 | return -EOPNOTSUPP; |
2b188cc1 | 1714 | |
def596e9 JA |
1715 | kiocb->ki_flags |= IOCB_HIPRI; |
1716 | kiocb->ki_complete = io_complete_rw_iopoll; | |
6873e0bd | 1717 | req->result = 0; |
def596e9 | 1718 | } else { |
09bb8394 JA |
1719 | if (kiocb->ki_flags & IOCB_HIPRI) |
1720 | return -EINVAL; | |
def596e9 JA |
1721 | kiocb->ki_complete = io_complete_rw; |
1722 | } | |
9adbd45d | 1723 | |
3529d8c2 JA |
1724 | req->rw.addr = READ_ONCE(sqe->addr); |
1725 | req->rw.len = READ_ONCE(sqe->len); | |
9adbd45d JA |
1726 | /* we own ->private, reuse it for the buffer index */ |
1727 | req->rw.kiocb.private = (void *) (unsigned long) | |
3529d8c2 | 1728 | READ_ONCE(sqe->buf_index); |
2b188cc1 | 1729 | return 0; |
2b188cc1 JA |
1730 | } |
1731 | ||
1732 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
1733 | { | |
1734 | switch (ret) { | |
1735 | case -EIOCBQUEUED: | |
1736 | break; | |
1737 | case -ERESTARTSYS: | |
1738 | case -ERESTARTNOINTR: | |
1739 | case -ERESTARTNOHAND: | |
1740 | case -ERESTART_RESTARTBLOCK: | |
1741 | /* | |
1742 | * We can't just restart the syscall, since previously | |
1743 | * submitted sqes may already be in progress. Just fail this | |
1744 | * IO with EINTR. | |
1745 | */ | |
1746 | ret = -EINTR; | |
1747 | /* fall through */ | |
1748 | default: | |
1749 | kiocb->ki_complete(kiocb, ret, 0); | |
1750 | } | |
1751 | } | |
1752 | ||
ba816ad6 JA |
1753 | static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt, |
1754 | bool in_async) | |
1755 | { | |
f9bd67f6 | 1756 | if (in_async && ret >= 0 && kiocb->ki_complete == io_complete_rw) |
ba816ad6 JA |
1757 | *nxt = __io_complete_rw(kiocb, ret); |
1758 | else | |
1759 | io_rw_done(kiocb, ret); | |
1760 | } | |
1761 | ||
9adbd45d | 1762 | static ssize_t io_import_fixed(struct io_kiocb *req, int rw, |
7d009165 | 1763 | struct iov_iter *iter) |
edafccee | 1764 | { |
9adbd45d JA |
1765 | struct io_ring_ctx *ctx = req->ctx; |
1766 | size_t len = req->rw.len; | |
edafccee JA |
1767 | struct io_mapped_ubuf *imu; |
1768 | unsigned index, buf_index; | |
1769 | size_t offset; | |
1770 | u64 buf_addr; | |
1771 | ||
1772 | /* attempt to use fixed buffers without having provided iovecs */ | |
1773 | if (unlikely(!ctx->user_bufs)) | |
1774 | return -EFAULT; | |
1775 | ||
9adbd45d | 1776 | buf_index = (unsigned long) req->rw.kiocb.private; |
edafccee JA |
1777 | if (unlikely(buf_index >= ctx->nr_user_bufs)) |
1778 | return -EFAULT; | |
1779 | ||
1780 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); | |
1781 | imu = &ctx->user_bufs[index]; | |
9adbd45d | 1782 | buf_addr = req->rw.addr; |
edafccee JA |
1783 | |
1784 | /* overflow */ | |
1785 | if (buf_addr + len < buf_addr) | |
1786 | return -EFAULT; | |
1787 | /* not inside the mapped region */ | |
1788 | if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len) | |
1789 | return -EFAULT; | |
1790 | ||
1791 | /* | |
1792 | * May not be a start of buffer, set size appropriately | |
1793 | * and advance us to the beginning. | |
1794 | */ | |
1795 | offset = buf_addr - imu->ubuf; | |
1796 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
bd11b3a3 JA |
1797 | |
1798 | if (offset) { | |
1799 | /* | |
1800 | * Don't use iov_iter_advance() here, as it's really slow for | |
1801 | * using the latter parts of a big fixed buffer - it iterates | |
1802 | * over each segment manually. We can cheat a bit here, because | |
1803 | * we know that: | |
1804 | * | |
1805 | * 1) it's a BVEC iter, we set it up | |
1806 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
1807 | * first and last bvec | |
1808 | * | |
1809 | * So just find our index, and adjust the iterator afterwards. | |
1810 | * If the offset is within the first bvec (or the whole first | |
1811 | * bvec, just use iov_iter_advance(). This makes it easier | |
1812 | * since we can just skip the first segment, which may not | |
1813 | * be PAGE_SIZE aligned. | |
1814 | */ | |
1815 | const struct bio_vec *bvec = imu->bvec; | |
1816 | ||
1817 | if (offset <= bvec->bv_len) { | |
1818 | iov_iter_advance(iter, offset); | |
1819 | } else { | |
1820 | unsigned long seg_skip; | |
1821 | ||
1822 | /* skip first vec */ | |
1823 | offset -= bvec->bv_len; | |
1824 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
1825 | ||
1826 | iter->bvec = bvec + seg_skip; | |
1827 | iter->nr_segs -= seg_skip; | |
99c79f66 | 1828 | iter->count -= bvec->bv_len + offset; |
bd11b3a3 | 1829 | iter->iov_offset = offset & ~PAGE_MASK; |
bd11b3a3 JA |
1830 | } |
1831 | } | |
1832 | ||
5e559561 | 1833 | return len; |
edafccee JA |
1834 | } |
1835 | ||
cf6fd4bd PB |
1836 | static ssize_t io_import_iovec(int rw, struct io_kiocb *req, |
1837 | struct iovec **iovec, struct iov_iter *iter) | |
2b188cc1 | 1838 | { |
9adbd45d JA |
1839 | void __user *buf = u64_to_user_ptr(req->rw.addr); |
1840 | size_t sqe_len = req->rw.len; | |
edafccee JA |
1841 | u8 opcode; |
1842 | ||
d625c6ee | 1843 | opcode = req->opcode; |
7d009165 | 1844 | if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) { |
edafccee | 1845 | *iovec = NULL; |
9adbd45d | 1846 | return io_import_fixed(req, rw, iter); |
edafccee | 1847 | } |
2b188cc1 | 1848 | |
9adbd45d JA |
1849 | /* buffer index only valid with fixed read/write */ |
1850 | if (req->rw.kiocb.private) | |
1851 | return -EINVAL; | |
1852 | ||
f67676d1 JA |
1853 | if (req->io) { |
1854 | struct io_async_rw *iorw = &req->io->rw; | |
1855 | ||
1856 | *iovec = iorw->iov; | |
1857 | iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size); | |
1858 | if (iorw->iov == iorw->fast_iov) | |
1859 | *iovec = NULL; | |
1860 | return iorw->size; | |
1861 | } | |
1862 | ||
cf6fd4bd | 1863 | if (!req->has_user) |
2b188cc1 JA |
1864 | return -EFAULT; |
1865 | ||
1866 | #ifdef CONFIG_COMPAT | |
cf6fd4bd | 1867 | if (req->ctx->compat) |
2b188cc1 JA |
1868 | return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV, |
1869 | iovec, iter); | |
1870 | #endif | |
1871 | ||
1872 | return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter); | |
1873 | } | |
1874 | ||
31b51510 | 1875 | /* |
32960613 JA |
1876 | * For files that don't have ->read_iter() and ->write_iter(), handle them |
1877 | * by looping over ->read() or ->write() manually. | |
31b51510 | 1878 | */ |
32960613 JA |
1879 | static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb, |
1880 | struct iov_iter *iter) | |
1881 | { | |
1882 | ssize_t ret = 0; | |
1883 | ||
1884 | /* | |
1885 | * Don't support polled IO through this interface, and we can't | |
1886 | * support non-blocking either. For the latter, this just causes | |
1887 | * the kiocb to be handled from an async context. | |
1888 | */ | |
1889 | if (kiocb->ki_flags & IOCB_HIPRI) | |
1890 | return -EOPNOTSUPP; | |
1891 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
1892 | return -EAGAIN; | |
1893 | ||
1894 | while (iov_iter_count(iter)) { | |
311ae9e1 | 1895 | struct iovec iovec; |
32960613 JA |
1896 | ssize_t nr; |
1897 | ||
311ae9e1 PB |
1898 | if (!iov_iter_is_bvec(iter)) { |
1899 | iovec = iov_iter_iovec(iter); | |
1900 | } else { | |
1901 | /* fixed buffers import bvec */ | |
1902 | iovec.iov_base = kmap(iter->bvec->bv_page) | |
1903 | + iter->iov_offset; | |
1904 | iovec.iov_len = min(iter->count, | |
1905 | iter->bvec->bv_len - iter->iov_offset); | |
1906 | } | |
1907 | ||
32960613 JA |
1908 | if (rw == READ) { |
1909 | nr = file->f_op->read(file, iovec.iov_base, | |
1910 | iovec.iov_len, &kiocb->ki_pos); | |
1911 | } else { | |
1912 | nr = file->f_op->write(file, iovec.iov_base, | |
1913 | iovec.iov_len, &kiocb->ki_pos); | |
1914 | } | |
1915 | ||
311ae9e1 PB |
1916 | if (iov_iter_is_bvec(iter)) |
1917 | kunmap(iter->bvec->bv_page); | |
1918 | ||
32960613 JA |
1919 | if (nr < 0) { |
1920 | if (!ret) | |
1921 | ret = nr; | |
1922 | break; | |
1923 | } | |
1924 | ret += nr; | |
1925 | if (nr != iovec.iov_len) | |
1926 | break; | |
1927 | iov_iter_advance(iter, nr); | |
1928 | } | |
1929 | ||
1930 | return ret; | |
1931 | } | |
1932 | ||
b7bb4f7d | 1933 | static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size, |
f67676d1 JA |
1934 | struct iovec *iovec, struct iovec *fast_iov, |
1935 | struct iov_iter *iter) | |
1936 | { | |
1937 | req->io->rw.nr_segs = iter->nr_segs; | |
1938 | req->io->rw.size = io_size; | |
1939 | req->io->rw.iov = iovec; | |
1940 | if (!req->io->rw.iov) { | |
1941 | req->io->rw.iov = req->io->rw.fast_iov; | |
1942 | memcpy(req->io->rw.iov, fast_iov, | |
1943 | sizeof(struct iovec) * iter->nr_segs); | |
1944 | } | |
1945 | } | |
1946 | ||
b7bb4f7d | 1947 | static int io_alloc_async_ctx(struct io_kiocb *req) |
f67676d1 | 1948 | { |
d3656344 JA |
1949 | if (!io_op_defs[req->opcode].async_ctx) |
1950 | return 0; | |
f67676d1 | 1951 | req->io = kmalloc(sizeof(*req->io), GFP_KERNEL); |
06b76d44 | 1952 | return req->io == NULL; |
b7bb4f7d JA |
1953 | } |
1954 | ||
1955 | static void io_rw_async(struct io_wq_work **workptr) | |
1956 | { | |
1957 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
1958 | struct iovec *iov = NULL; | |
1959 | ||
1960 | if (req->io->rw.iov != req->io->rw.fast_iov) | |
1961 | iov = req->io->rw.iov; | |
1962 | io_wq_submit_work(workptr); | |
1963 | kfree(iov); | |
1964 | } | |
1965 | ||
1966 | static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size, | |
1967 | struct iovec *iovec, struct iovec *fast_iov, | |
1968 | struct iov_iter *iter) | |
1969 | { | |
74566df3 JA |
1970 | if (req->opcode == IORING_OP_READ_FIXED || |
1971 | req->opcode == IORING_OP_WRITE_FIXED) | |
1972 | return 0; | |
b7bb4f7d JA |
1973 | if (!req->io && io_alloc_async_ctx(req)) |
1974 | return -ENOMEM; | |
1975 | ||
1976 | io_req_map_rw(req, io_size, iovec, fast_iov, iter); | |
1977 | req->work.func = io_rw_async; | |
1978 | return 0; | |
f67676d1 JA |
1979 | } |
1980 | ||
3529d8c2 JA |
1981 | static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
1982 | bool force_nonblock) | |
f67676d1 | 1983 | { |
3529d8c2 JA |
1984 | struct io_async_ctx *io; |
1985 | struct iov_iter iter; | |
f67676d1 JA |
1986 | ssize_t ret; |
1987 | ||
3529d8c2 JA |
1988 | ret = io_prep_rw(req, sqe, force_nonblock); |
1989 | if (ret) | |
1990 | return ret; | |
f67676d1 | 1991 | |
3529d8c2 JA |
1992 | if (unlikely(!(req->file->f_mode & FMODE_READ))) |
1993 | return -EBADF; | |
f67676d1 | 1994 | |
3529d8c2 JA |
1995 | if (!req->io) |
1996 | return 0; | |
1997 | ||
1998 | io = req->io; | |
1999 | io->rw.iov = io->rw.fast_iov; | |
2000 | req->io = NULL; | |
2001 | ret = io_import_iovec(READ, req, &io->rw.iov, &iter); | |
2002 | req->io = io; | |
2003 | if (ret < 0) | |
2004 | return ret; | |
2005 | ||
2006 | io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter); | |
2007 | return 0; | |
f67676d1 JA |
2008 | } |
2009 | ||
267bc904 | 2010 | static int io_read(struct io_kiocb *req, struct io_kiocb **nxt, |
8358e3a8 | 2011 | bool force_nonblock) |
2b188cc1 JA |
2012 | { |
2013 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
9adbd45d | 2014 | struct kiocb *kiocb = &req->rw.kiocb; |
2b188cc1 | 2015 | struct iov_iter iter; |
31b51510 | 2016 | size_t iov_count; |
f67676d1 | 2017 | ssize_t io_size, ret; |
2b188cc1 | 2018 | |
3529d8c2 | 2019 | ret = io_import_iovec(READ, req, &iovec, &iter); |
06b76d44 JA |
2020 | if (ret < 0) |
2021 | return ret; | |
2b188cc1 | 2022 | |
fd6c2e4c JA |
2023 | /* Ensure we clear previously set non-block flag */ |
2024 | if (!force_nonblock) | |
9adbd45d | 2025 | req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT; |
fd6c2e4c | 2026 | |
797f3f53 | 2027 | req->result = 0; |
f67676d1 | 2028 | io_size = ret; |
9e645e11 | 2029 | if (req->flags & REQ_F_LINK) |
f67676d1 JA |
2030 | req->result = io_size; |
2031 | ||
2032 | /* | |
2033 | * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so | |
2034 | * we know to async punt it even if it was opened O_NONBLOCK | |
2035 | */ | |
9adbd45d | 2036 | if (force_nonblock && !io_file_supports_async(req->file)) { |
f67676d1 JA |
2037 | req->flags |= REQ_F_MUST_PUNT; |
2038 | goto copy_iov; | |
2039 | } | |
9e645e11 | 2040 | |
31b51510 | 2041 | iov_count = iov_iter_count(&iter); |
9adbd45d | 2042 | ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count); |
2b188cc1 JA |
2043 | if (!ret) { |
2044 | ssize_t ret2; | |
2045 | ||
9adbd45d JA |
2046 | if (req->file->f_op->read_iter) |
2047 | ret2 = call_read_iter(req->file, kiocb, &iter); | |
32960613 | 2048 | else |
9adbd45d | 2049 | ret2 = loop_rw_iter(READ, req->file, kiocb, &iter); |
32960613 | 2050 | |
9d93a3f5 | 2051 | /* Catch -EAGAIN return for forced non-blocking submission */ |
f67676d1 | 2052 | if (!force_nonblock || ret2 != -EAGAIN) { |
cf6fd4bd | 2053 | kiocb_done(kiocb, ret2, nxt, req->in_async); |
f67676d1 JA |
2054 | } else { |
2055 | copy_iov: | |
b7bb4f7d | 2056 | ret = io_setup_async_rw(req, io_size, iovec, |
f67676d1 JA |
2057 | inline_vecs, &iter); |
2058 | if (ret) | |
2059 | goto out_free; | |
2060 | return -EAGAIN; | |
2061 | } | |
2b188cc1 | 2062 | } |
f67676d1 | 2063 | out_free: |
b7bb4f7d JA |
2064 | if (!io_wq_current_is_worker()) |
2065 | kfree(iovec); | |
2b188cc1 JA |
2066 | return ret; |
2067 | } | |
2068 | ||
3529d8c2 JA |
2069 | static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2070 | bool force_nonblock) | |
f67676d1 | 2071 | { |
3529d8c2 JA |
2072 | struct io_async_ctx *io; |
2073 | struct iov_iter iter; | |
f67676d1 JA |
2074 | ssize_t ret; |
2075 | ||
3529d8c2 JA |
2076 | ret = io_prep_rw(req, sqe, force_nonblock); |
2077 | if (ret) | |
2078 | return ret; | |
f67676d1 | 2079 | |
3529d8c2 JA |
2080 | if (unlikely(!(req->file->f_mode & FMODE_WRITE))) |
2081 | return -EBADF; | |
f67676d1 | 2082 | |
3529d8c2 JA |
2083 | if (!req->io) |
2084 | return 0; | |
2085 | ||
2086 | io = req->io; | |
2087 | io->rw.iov = io->rw.fast_iov; | |
2088 | req->io = NULL; | |
2089 | ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter); | |
2090 | req->io = io; | |
2091 | if (ret < 0) | |
2092 | return ret; | |
2093 | ||
2094 | io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter); | |
2095 | return 0; | |
f67676d1 JA |
2096 | } |
2097 | ||
267bc904 | 2098 | static int io_write(struct io_kiocb *req, struct io_kiocb **nxt, |
8358e3a8 | 2099 | bool force_nonblock) |
2b188cc1 JA |
2100 | { |
2101 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
9adbd45d | 2102 | struct kiocb *kiocb = &req->rw.kiocb; |
2b188cc1 | 2103 | struct iov_iter iter; |
31b51510 | 2104 | size_t iov_count; |
f67676d1 | 2105 | ssize_t ret, io_size; |
2b188cc1 | 2106 | |
3529d8c2 | 2107 | ret = io_import_iovec(WRITE, req, &iovec, &iter); |
06b76d44 JA |
2108 | if (ret < 0) |
2109 | return ret; | |
2b188cc1 | 2110 | |
fd6c2e4c JA |
2111 | /* Ensure we clear previously set non-block flag */ |
2112 | if (!force_nonblock) | |
9adbd45d | 2113 | req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT; |
fd6c2e4c | 2114 | |
797f3f53 | 2115 | req->result = 0; |
f67676d1 | 2116 | io_size = ret; |
9e645e11 | 2117 | if (req->flags & REQ_F_LINK) |
f67676d1 | 2118 | req->result = io_size; |
9e645e11 | 2119 | |
f67676d1 JA |
2120 | /* |
2121 | * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so | |
2122 | * we know to async punt it even if it was opened O_NONBLOCK | |
2123 | */ | |
2124 | if (force_nonblock && !io_file_supports_async(req->file)) { | |
2125 | req->flags |= REQ_F_MUST_PUNT; | |
2126 | goto copy_iov; | |
2127 | } | |
31b51510 | 2128 | |
10d59345 JA |
2129 | /* file path doesn't support NOWAIT for non-direct_IO */ |
2130 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) && | |
2131 | (req->flags & REQ_F_ISREG)) | |
f67676d1 | 2132 | goto copy_iov; |
31b51510 | 2133 | |
f67676d1 | 2134 | iov_count = iov_iter_count(&iter); |
9adbd45d | 2135 | ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count); |
2b188cc1 | 2136 | if (!ret) { |
9bf7933f RP |
2137 | ssize_t ret2; |
2138 | ||
2b188cc1 JA |
2139 | /* |
2140 | * Open-code file_start_write here to grab freeze protection, | |
2141 | * which will be released by another thread in | |
2142 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
2143 | * released so that it doesn't complain about the held lock when | |
2144 | * we return to userspace. | |
2145 | */ | |
491381ce | 2146 | if (req->flags & REQ_F_ISREG) { |
9adbd45d | 2147 | __sb_start_write(file_inode(req->file)->i_sb, |
2b188cc1 | 2148 | SB_FREEZE_WRITE, true); |
9adbd45d | 2149 | __sb_writers_release(file_inode(req->file)->i_sb, |
2b188cc1 JA |
2150 | SB_FREEZE_WRITE); |
2151 | } | |
2152 | kiocb->ki_flags |= IOCB_WRITE; | |
9bf7933f | 2153 | |
9adbd45d JA |
2154 | if (req->file->f_op->write_iter) |
2155 | ret2 = call_write_iter(req->file, kiocb, &iter); | |
32960613 | 2156 | else |
9adbd45d | 2157 | ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter); |
f67676d1 | 2158 | if (!force_nonblock || ret2 != -EAGAIN) { |
cf6fd4bd | 2159 | kiocb_done(kiocb, ret2, nxt, req->in_async); |
f67676d1 JA |
2160 | } else { |
2161 | copy_iov: | |
b7bb4f7d | 2162 | ret = io_setup_async_rw(req, io_size, iovec, |
f67676d1 JA |
2163 | inline_vecs, &iter); |
2164 | if (ret) | |
2165 | goto out_free; | |
2166 | return -EAGAIN; | |
2167 | } | |
2b188cc1 | 2168 | } |
31b51510 | 2169 | out_free: |
b7bb4f7d JA |
2170 | if (!io_wq_current_is_worker()) |
2171 | kfree(iovec); | |
2b188cc1 JA |
2172 | return ret; |
2173 | } | |
2174 | ||
2175 | /* | |
2176 | * IORING_OP_NOP just posts a completion event, nothing else. | |
2177 | */ | |
78e19bbe | 2178 | static int io_nop(struct io_kiocb *req) |
2b188cc1 JA |
2179 | { |
2180 | struct io_ring_ctx *ctx = req->ctx; | |
2b188cc1 | 2181 | |
def596e9 JA |
2182 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
2183 | return -EINVAL; | |
2184 | ||
78e19bbe | 2185 | io_cqring_add_event(req, 0); |
e65ef56d | 2186 | io_put_req(req); |
2b188cc1 JA |
2187 | return 0; |
2188 | } | |
2189 | ||
3529d8c2 | 2190 | static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
c992fe29 | 2191 | { |
6b06314c | 2192 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 2193 | |
09bb8394 JA |
2194 | if (!req->file) |
2195 | return -EBADF; | |
c992fe29 | 2196 | |
6b06314c | 2197 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 2198 | return -EINVAL; |
edafccee | 2199 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
2200 | return -EINVAL; |
2201 | ||
8ed8d3c3 JA |
2202 | req->sync.flags = READ_ONCE(sqe->fsync_flags); |
2203 | if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC)) | |
2204 | return -EINVAL; | |
2205 | ||
2206 | req->sync.off = READ_ONCE(sqe->off); | |
2207 | req->sync.len = READ_ONCE(sqe->len); | |
c992fe29 CH |
2208 | return 0; |
2209 | } | |
2210 | ||
8ed8d3c3 JA |
2211 | static bool io_req_cancelled(struct io_kiocb *req) |
2212 | { | |
2213 | if (req->work.flags & IO_WQ_WORK_CANCEL) { | |
2214 | req_set_fail_links(req); | |
2215 | io_cqring_add_event(req, -ECANCELED); | |
2216 | io_put_req(req); | |
2217 | return true; | |
2218 | } | |
2219 | ||
2220 | return false; | |
2221 | } | |
2222 | ||
78912934 JA |
2223 | static void io_link_work_cb(struct io_wq_work **workptr) |
2224 | { | |
2225 | struct io_wq_work *work = *workptr; | |
2226 | struct io_kiocb *link = work->data; | |
2227 | ||
2228 | io_queue_linked_timeout(link); | |
2229 | work->func = io_wq_submit_work; | |
2230 | } | |
2231 | ||
2232 | static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt) | |
2233 | { | |
2234 | struct io_kiocb *link; | |
2235 | ||
2236 | io_prep_async_work(nxt, &link); | |
2237 | *workptr = &nxt->work; | |
2238 | if (link) { | |
2239 | nxt->work.flags |= IO_WQ_WORK_CB; | |
2240 | nxt->work.func = io_link_work_cb; | |
2241 | nxt->work.data = link; | |
2242 | } | |
2243 | } | |
2244 | ||
8ed8d3c3 JA |
2245 | static void io_fsync_finish(struct io_wq_work **workptr) |
2246 | { | |
2247 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
2248 | loff_t end = req->sync.off + req->sync.len; | |
2249 | struct io_kiocb *nxt = NULL; | |
2250 | int ret; | |
2251 | ||
2252 | if (io_req_cancelled(req)) | |
2253 | return; | |
2254 | ||
9adbd45d | 2255 | ret = vfs_fsync_range(req->file, req->sync.off, |
8ed8d3c3 JA |
2256 | end > 0 ? end : LLONG_MAX, |
2257 | req->sync.flags & IORING_FSYNC_DATASYNC); | |
2258 | if (ret < 0) | |
2259 | req_set_fail_links(req); | |
2260 | io_cqring_add_event(req, ret); | |
2261 | io_put_req_find_next(req, &nxt); | |
2262 | if (nxt) | |
78912934 | 2263 | io_wq_assign_next(workptr, nxt); |
8ed8d3c3 JA |
2264 | } |
2265 | ||
fc4df999 JA |
2266 | static int io_fsync(struct io_kiocb *req, struct io_kiocb **nxt, |
2267 | bool force_nonblock) | |
c992fe29 | 2268 | { |
8ed8d3c3 | 2269 | struct io_wq_work *work, *old_work; |
c992fe29 CH |
2270 | |
2271 | /* fsync always requires a blocking context */ | |
8ed8d3c3 JA |
2272 | if (force_nonblock) { |
2273 | io_put_req(req); | |
2274 | req->work.func = io_fsync_finish; | |
c992fe29 | 2275 | return -EAGAIN; |
8ed8d3c3 | 2276 | } |
c992fe29 | 2277 | |
8ed8d3c3 JA |
2278 | work = old_work = &req->work; |
2279 | io_fsync_finish(&work); | |
2280 | if (work && work != old_work) | |
2281 | *nxt = container_of(work, struct io_kiocb, work); | |
c992fe29 CH |
2282 | return 0; |
2283 | } | |
2284 | ||
d63d1b5e JA |
2285 | static void io_fallocate_finish(struct io_wq_work **workptr) |
2286 | { | |
2287 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
2288 | struct io_kiocb *nxt = NULL; | |
2289 | int ret; | |
2290 | ||
2291 | ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off, | |
2292 | req->sync.len); | |
2293 | if (ret < 0) | |
2294 | req_set_fail_links(req); | |
2295 | io_cqring_add_event(req, ret); | |
2296 | io_put_req_find_next(req, &nxt); | |
2297 | if (nxt) | |
2298 | io_wq_assign_next(workptr, nxt); | |
2299 | } | |
2300 | ||
2301 | static int io_fallocate_prep(struct io_kiocb *req, | |
2302 | const struct io_uring_sqe *sqe) | |
2303 | { | |
2304 | if (sqe->ioprio || sqe->buf_index || sqe->rw_flags) | |
2305 | return -EINVAL; | |
2306 | ||
2307 | req->sync.off = READ_ONCE(sqe->off); | |
2308 | req->sync.len = READ_ONCE(sqe->addr); | |
2309 | req->sync.mode = READ_ONCE(sqe->len); | |
2310 | return 0; | |
2311 | } | |
2312 | ||
2313 | static int io_fallocate(struct io_kiocb *req, struct io_kiocb **nxt, | |
2314 | bool force_nonblock) | |
2315 | { | |
2316 | struct io_wq_work *work, *old_work; | |
2317 | ||
2318 | /* fallocate always requiring blocking context */ | |
2319 | if (force_nonblock) { | |
2320 | io_put_req(req); | |
2321 | req->work.func = io_fallocate_finish; | |
2322 | return -EAGAIN; | |
2323 | } | |
2324 | ||
2325 | work = old_work = &req->work; | |
2326 | io_fallocate_finish(&work); | |
2327 | if (work && work != old_work) | |
2328 | *nxt = container_of(work, struct io_kiocb, work); | |
2329 | ||
2330 | return 0; | |
2331 | } | |
2332 | ||
15b71abe JA |
2333 | static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2334 | { | |
2335 | int ret; | |
2336 | ||
2337 | if (sqe->ioprio || sqe->buf_index) | |
2338 | return -EINVAL; | |
2339 | ||
2340 | req->open.dfd = READ_ONCE(sqe->fd); | |
2341 | req->open.mode = READ_ONCE(sqe->len); | |
2342 | req->open.fname = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
2343 | req->open.flags = READ_ONCE(sqe->open_flags); | |
2344 | ||
2345 | req->open.filename = getname(req->open.fname); | |
2346 | if (IS_ERR(req->open.filename)) { | |
2347 | ret = PTR_ERR(req->open.filename); | |
2348 | req->open.filename = NULL; | |
2349 | return ret; | |
2350 | } | |
2351 | ||
2352 | return 0; | |
2353 | } | |
2354 | ||
2355 | static int io_openat(struct io_kiocb *req, struct io_kiocb **nxt, | |
2356 | bool force_nonblock) | |
2357 | { | |
2358 | struct open_flags op; | |
2359 | struct open_how how; | |
2360 | struct file *file; | |
2361 | int ret; | |
2362 | ||
2363 | if (force_nonblock) { | |
2364 | req->work.flags |= IO_WQ_WORK_NEEDS_FILES; | |
2365 | return -EAGAIN; | |
2366 | } | |
2367 | ||
2368 | how = build_open_how(req->open.flags, req->open.mode); | |
2369 | ret = build_open_flags(&how, &op); | |
2370 | if (ret) | |
2371 | goto err; | |
2372 | ||
2373 | ret = get_unused_fd_flags(how.flags); | |
2374 | if (ret < 0) | |
2375 | goto err; | |
2376 | ||
2377 | file = do_filp_open(req->open.dfd, req->open.filename, &op); | |
2378 | if (IS_ERR(file)) { | |
2379 | put_unused_fd(ret); | |
2380 | ret = PTR_ERR(file); | |
2381 | } else { | |
2382 | fsnotify_open(file); | |
2383 | fd_install(ret, file); | |
2384 | } | |
2385 | err: | |
2386 | putname(req->open.filename); | |
2387 | if (ret < 0) | |
2388 | req_set_fail_links(req); | |
2389 | io_cqring_add_event(req, ret); | |
2390 | io_put_req_find_next(req, nxt); | |
2391 | return 0; | |
2392 | } | |
2393 | ||
eddc7ef5 JA |
2394 | static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2395 | { | |
2396 | unsigned lookup_flags; | |
2397 | int ret; | |
2398 | ||
2399 | if (sqe->ioprio || sqe->buf_index) | |
2400 | return -EINVAL; | |
2401 | ||
2402 | req->open.dfd = READ_ONCE(sqe->fd); | |
2403 | req->open.mask = READ_ONCE(sqe->len); | |
2404 | req->open.fname = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
2405 | req->open.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2)); | |
2406 | req->open.flags = READ_ONCE(sqe->statx_flags); | |
2407 | ||
2408 | if (vfs_stat_set_lookup_flags(&lookup_flags, req->open.flags)) | |
2409 | return -EINVAL; | |
2410 | ||
2411 | req->open.filename = getname_flags(req->open.fname, lookup_flags, NULL); | |
2412 | if (IS_ERR(req->open.filename)) { | |
2413 | ret = PTR_ERR(req->open.filename); | |
2414 | req->open.filename = NULL; | |
2415 | return ret; | |
2416 | } | |
2417 | ||
2418 | return 0; | |
2419 | } | |
2420 | ||
2421 | static int io_statx(struct io_kiocb *req, struct io_kiocb **nxt, | |
2422 | bool force_nonblock) | |
2423 | { | |
2424 | struct io_open *ctx = &req->open; | |
2425 | unsigned lookup_flags; | |
2426 | struct path path; | |
2427 | struct kstat stat; | |
2428 | int ret; | |
2429 | ||
2430 | if (force_nonblock) | |
2431 | return -EAGAIN; | |
2432 | ||
2433 | if (vfs_stat_set_lookup_flags(&lookup_flags, ctx->flags)) | |
2434 | return -EINVAL; | |
2435 | ||
2436 | retry: | |
2437 | /* filename_lookup() drops it, keep a reference */ | |
2438 | ctx->filename->refcnt++; | |
2439 | ||
2440 | ret = filename_lookup(ctx->dfd, ctx->filename, lookup_flags, &path, | |
2441 | NULL); | |
2442 | if (ret) | |
2443 | goto err; | |
2444 | ||
2445 | ret = vfs_getattr(&path, &stat, ctx->mask, ctx->flags); | |
2446 | path_put(&path); | |
2447 | if (retry_estale(ret, lookup_flags)) { | |
2448 | lookup_flags |= LOOKUP_REVAL; | |
2449 | goto retry; | |
2450 | } | |
2451 | if (!ret) | |
2452 | ret = cp_statx(&stat, ctx->buffer); | |
2453 | err: | |
2454 | putname(ctx->filename); | |
2455 | if (ret < 0) | |
2456 | req_set_fail_links(req); | |
2457 | io_cqring_add_event(req, ret); | |
2458 | io_put_req_find_next(req, nxt); | |
2459 | return 0; | |
2460 | } | |
2461 | ||
b5dba59e JA |
2462 | static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2463 | { | |
2464 | /* | |
2465 | * If we queue this for async, it must not be cancellable. That would | |
2466 | * leave the 'file' in an undeterminate state. | |
2467 | */ | |
2468 | req->work.flags |= IO_WQ_WORK_NO_CANCEL; | |
2469 | ||
2470 | if (sqe->ioprio || sqe->off || sqe->addr || sqe->len || | |
2471 | sqe->rw_flags || sqe->buf_index) | |
2472 | return -EINVAL; | |
2473 | if (sqe->flags & IOSQE_FIXED_FILE) | |
2474 | return -EINVAL; | |
2475 | ||
2476 | req->close.fd = READ_ONCE(sqe->fd); | |
2477 | if (req->file->f_op == &io_uring_fops || | |
2478 | req->close.fd == req->ring_fd) | |
2479 | return -EBADF; | |
2480 | ||
2481 | return 0; | |
2482 | } | |
2483 | ||
2484 | static void io_close_finish(struct io_wq_work **workptr) | |
2485 | { | |
2486 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
2487 | struct io_kiocb *nxt = NULL; | |
2488 | ||
2489 | /* Invoked with files, we need to do the close */ | |
2490 | if (req->work.files) { | |
2491 | int ret; | |
2492 | ||
2493 | ret = filp_close(req->close.put_file, req->work.files); | |
2494 | if (ret < 0) { | |
2495 | req_set_fail_links(req); | |
2496 | } | |
2497 | io_cqring_add_event(req, ret); | |
2498 | } | |
2499 | ||
2500 | fput(req->close.put_file); | |
2501 | ||
2502 | /* we bypassed the re-issue, drop the submission reference */ | |
2503 | io_put_req(req); | |
2504 | io_put_req_find_next(req, &nxt); | |
2505 | if (nxt) | |
2506 | io_wq_assign_next(workptr, nxt); | |
2507 | } | |
2508 | ||
2509 | static int io_close(struct io_kiocb *req, struct io_kiocb **nxt, | |
2510 | bool force_nonblock) | |
2511 | { | |
2512 | int ret; | |
2513 | ||
2514 | req->close.put_file = NULL; | |
2515 | ret = __close_fd_get_file(req->close.fd, &req->close.put_file); | |
2516 | if (ret < 0) | |
2517 | return ret; | |
2518 | ||
2519 | /* if the file has a flush method, be safe and punt to async */ | |
2520 | if (req->close.put_file->f_op->flush && !io_wq_current_is_worker()) { | |
2521 | req->work.flags |= IO_WQ_WORK_NEEDS_FILES; | |
2522 | goto eagain; | |
2523 | } | |
2524 | ||
2525 | /* | |
2526 | * No ->flush(), safely close from here and just punt the | |
2527 | * fput() to async context. | |
2528 | */ | |
2529 | ret = filp_close(req->close.put_file, current->files); | |
2530 | ||
2531 | if (ret < 0) | |
2532 | req_set_fail_links(req); | |
2533 | io_cqring_add_event(req, ret); | |
2534 | ||
2535 | if (io_wq_current_is_worker()) { | |
2536 | struct io_wq_work *old_work, *work; | |
2537 | ||
2538 | old_work = work = &req->work; | |
2539 | io_close_finish(&work); | |
2540 | if (work && work != old_work) | |
2541 | *nxt = container_of(work, struct io_kiocb, work); | |
2542 | return 0; | |
2543 | } | |
2544 | ||
2545 | eagain: | |
2546 | req->work.func = io_close_finish; | |
2547 | return -EAGAIN; | |
2548 | } | |
2549 | ||
3529d8c2 | 2550 | static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
5d17b4a4 JA |
2551 | { |
2552 | struct io_ring_ctx *ctx = req->ctx; | |
5d17b4a4 JA |
2553 | |
2554 | if (!req->file) | |
2555 | return -EBADF; | |
5d17b4a4 JA |
2556 | |
2557 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
2558 | return -EINVAL; | |
2559 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) | |
2560 | return -EINVAL; | |
2561 | ||
8ed8d3c3 JA |
2562 | req->sync.off = READ_ONCE(sqe->off); |
2563 | req->sync.len = READ_ONCE(sqe->len); | |
2564 | req->sync.flags = READ_ONCE(sqe->sync_range_flags); | |
8ed8d3c3 JA |
2565 | return 0; |
2566 | } | |
2567 | ||
2568 | static void io_sync_file_range_finish(struct io_wq_work **workptr) | |
2569 | { | |
2570 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
2571 | struct io_kiocb *nxt = NULL; | |
2572 | int ret; | |
2573 | ||
2574 | if (io_req_cancelled(req)) | |
2575 | return; | |
2576 | ||
9adbd45d | 2577 | ret = sync_file_range(req->file, req->sync.off, req->sync.len, |
8ed8d3c3 JA |
2578 | req->sync.flags); |
2579 | if (ret < 0) | |
2580 | req_set_fail_links(req); | |
2581 | io_cqring_add_event(req, ret); | |
2582 | io_put_req_find_next(req, &nxt); | |
2583 | if (nxt) | |
78912934 | 2584 | io_wq_assign_next(workptr, nxt); |
5d17b4a4 JA |
2585 | } |
2586 | ||
fc4df999 | 2587 | static int io_sync_file_range(struct io_kiocb *req, struct io_kiocb **nxt, |
5d17b4a4 JA |
2588 | bool force_nonblock) |
2589 | { | |
8ed8d3c3 | 2590 | struct io_wq_work *work, *old_work; |
5d17b4a4 JA |
2591 | |
2592 | /* sync_file_range always requires a blocking context */ | |
8ed8d3c3 JA |
2593 | if (force_nonblock) { |
2594 | io_put_req(req); | |
2595 | req->work.func = io_sync_file_range_finish; | |
5d17b4a4 | 2596 | return -EAGAIN; |
8ed8d3c3 | 2597 | } |
5d17b4a4 | 2598 | |
8ed8d3c3 JA |
2599 | work = old_work = &req->work; |
2600 | io_sync_file_range_finish(&work); | |
2601 | if (work && work != old_work) | |
2602 | *nxt = container_of(work, struct io_kiocb, work); | |
5d17b4a4 JA |
2603 | return 0; |
2604 | } | |
2605 | ||
b7bb4f7d JA |
2606 | #if defined(CONFIG_NET) |
2607 | static void io_sendrecv_async(struct io_wq_work **workptr) | |
2608 | { | |
2609 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
2610 | struct iovec *iov = NULL; | |
2611 | ||
2612 | if (req->io->rw.iov != req->io->rw.fast_iov) | |
2613 | iov = req->io->msg.iov; | |
2614 | io_wq_submit_work(workptr); | |
2615 | kfree(iov); | |
2616 | } | |
2617 | #endif | |
2618 | ||
3529d8c2 | 2619 | static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
03b1230c | 2620 | { |
0fa03c62 | 2621 | #if defined(CONFIG_NET) |
e47293fd | 2622 | struct io_sr_msg *sr = &req->sr_msg; |
3529d8c2 | 2623 | struct io_async_ctx *io = req->io; |
03b1230c | 2624 | |
e47293fd JA |
2625 | sr->msg_flags = READ_ONCE(sqe->msg_flags); |
2626 | sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
3529d8c2 JA |
2627 | |
2628 | if (!io) | |
2629 | return 0; | |
2630 | ||
d9688565 | 2631 | io->msg.iov = io->msg.fast_iov; |
3529d8c2 | 2632 | return sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags, |
e47293fd | 2633 | &io->msg.iov); |
03b1230c | 2634 | #else |
e47293fd | 2635 | return -EOPNOTSUPP; |
03b1230c JA |
2636 | #endif |
2637 | } | |
2638 | ||
fc4df999 JA |
2639 | static int io_sendmsg(struct io_kiocb *req, struct io_kiocb **nxt, |
2640 | bool force_nonblock) | |
aa1fa28f | 2641 | { |
03b1230c | 2642 | #if defined(CONFIG_NET) |
0b416c3e | 2643 | struct io_async_msghdr *kmsg = NULL; |
0fa03c62 JA |
2644 | struct socket *sock; |
2645 | int ret; | |
2646 | ||
2647 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
2648 | return -EINVAL; | |
2649 | ||
2650 | sock = sock_from_file(req->file, &ret); | |
2651 | if (sock) { | |
b7bb4f7d | 2652 | struct io_async_ctx io; |
03b1230c | 2653 | struct sockaddr_storage addr; |
0fa03c62 JA |
2654 | unsigned flags; |
2655 | ||
03b1230c | 2656 | if (req->io) { |
0b416c3e JA |
2657 | kmsg = &req->io->msg; |
2658 | kmsg->msg.msg_name = &addr; | |
2659 | /* if iov is set, it's allocated already */ | |
2660 | if (!kmsg->iov) | |
2661 | kmsg->iov = kmsg->fast_iov; | |
2662 | kmsg->msg.msg_iter.iov = kmsg->iov; | |
03b1230c | 2663 | } else { |
3529d8c2 JA |
2664 | struct io_sr_msg *sr = &req->sr_msg; |
2665 | ||
0b416c3e JA |
2666 | kmsg = &io.msg; |
2667 | kmsg->msg.msg_name = &addr; | |
3529d8c2 JA |
2668 | |
2669 | io.msg.iov = io.msg.fast_iov; | |
2670 | ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg, | |
2671 | sr->msg_flags, &io.msg.iov); | |
03b1230c | 2672 | if (ret) |
3529d8c2 | 2673 | return ret; |
03b1230c | 2674 | } |
0fa03c62 | 2675 | |
e47293fd JA |
2676 | flags = req->sr_msg.msg_flags; |
2677 | if (flags & MSG_DONTWAIT) | |
2678 | req->flags |= REQ_F_NOWAIT; | |
2679 | else if (force_nonblock) | |
2680 | flags |= MSG_DONTWAIT; | |
2681 | ||
0b416c3e | 2682 | ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags); |
03b1230c | 2683 | if (force_nonblock && ret == -EAGAIN) { |
b7bb4f7d JA |
2684 | if (req->io) |
2685 | return -EAGAIN; | |
2686 | if (io_alloc_async_ctx(req)) | |
2687 | return -ENOMEM; | |
2688 | memcpy(&req->io->msg, &io.msg, sizeof(io.msg)); | |
2689 | req->work.func = io_sendrecv_async; | |
0b416c3e | 2690 | return -EAGAIN; |
03b1230c | 2691 | } |
441cdbd5 JA |
2692 | if (ret == -ERESTARTSYS) |
2693 | ret = -EINTR; | |
0fa03c62 JA |
2694 | } |
2695 | ||
b7bb4f7d | 2696 | if (!io_wq_current_is_worker() && kmsg && kmsg->iov != kmsg->fast_iov) |
0b416c3e | 2697 | kfree(kmsg->iov); |
78e19bbe | 2698 | io_cqring_add_event(req, ret); |
4e88d6e7 JA |
2699 | if (ret < 0) |
2700 | req_set_fail_links(req); | |
ec9c02ad | 2701 | io_put_req_find_next(req, nxt); |
5d17b4a4 | 2702 | return 0; |
03b1230c JA |
2703 | #else |
2704 | return -EOPNOTSUPP; | |
aa1fa28f | 2705 | #endif |
03b1230c | 2706 | } |
aa1fa28f | 2707 | |
3529d8c2 JA |
2708 | static int io_recvmsg_prep(struct io_kiocb *req, |
2709 | const struct io_uring_sqe *sqe) | |
aa1fa28f JA |
2710 | { |
2711 | #if defined(CONFIG_NET) | |
e47293fd | 2712 | struct io_sr_msg *sr = &req->sr_msg; |
3529d8c2 JA |
2713 | struct io_async_ctx *io = req->io; |
2714 | ||
2715 | sr->msg_flags = READ_ONCE(sqe->msg_flags); | |
2716 | sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
06b76d44 | 2717 | |
3529d8c2 | 2718 | if (!io) |
06b76d44 | 2719 | return 0; |
03b1230c | 2720 | |
d9688565 | 2721 | io->msg.iov = io->msg.fast_iov; |
3529d8c2 | 2722 | return recvmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags, |
e47293fd | 2723 | &io->msg.uaddr, &io->msg.iov); |
aa1fa28f | 2724 | #else |
e47293fd | 2725 | return -EOPNOTSUPP; |
aa1fa28f JA |
2726 | #endif |
2727 | } | |
2728 | ||
fc4df999 JA |
2729 | static int io_recvmsg(struct io_kiocb *req, struct io_kiocb **nxt, |
2730 | bool force_nonblock) | |
aa1fa28f JA |
2731 | { |
2732 | #if defined(CONFIG_NET) | |
0b416c3e | 2733 | struct io_async_msghdr *kmsg = NULL; |
03b1230c JA |
2734 | struct socket *sock; |
2735 | int ret; | |
2736 | ||
2737 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
2738 | return -EINVAL; | |
2739 | ||
2740 | sock = sock_from_file(req->file, &ret); | |
2741 | if (sock) { | |
b7bb4f7d | 2742 | struct io_async_ctx io; |
03b1230c | 2743 | struct sockaddr_storage addr; |
03b1230c JA |
2744 | unsigned flags; |
2745 | ||
03b1230c | 2746 | if (req->io) { |
0b416c3e JA |
2747 | kmsg = &req->io->msg; |
2748 | kmsg->msg.msg_name = &addr; | |
2749 | /* if iov is set, it's allocated already */ | |
2750 | if (!kmsg->iov) | |
2751 | kmsg->iov = kmsg->fast_iov; | |
2752 | kmsg->msg.msg_iter.iov = kmsg->iov; | |
03b1230c | 2753 | } else { |
3529d8c2 JA |
2754 | struct io_sr_msg *sr = &req->sr_msg; |
2755 | ||
0b416c3e JA |
2756 | kmsg = &io.msg; |
2757 | kmsg->msg.msg_name = &addr; | |
3529d8c2 JA |
2758 | |
2759 | io.msg.iov = io.msg.fast_iov; | |
2760 | ret = recvmsg_copy_msghdr(&io.msg.msg, sr->msg, | |
2761 | sr->msg_flags, &io.msg.uaddr, | |
2762 | &io.msg.iov); | |
03b1230c | 2763 | if (ret) |
3529d8c2 | 2764 | return ret; |
03b1230c JA |
2765 | } |
2766 | ||
e47293fd JA |
2767 | flags = req->sr_msg.msg_flags; |
2768 | if (flags & MSG_DONTWAIT) | |
2769 | req->flags |= REQ_F_NOWAIT; | |
2770 | else if (force_nonblock) | |
2771 | flags |= MSG_DONTWAIT; | |
2772 | ||
2773 | ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg, | |
2774 | kmsg->uaddr, flags); | |
03b1230c | 2775 | if (force_nonblock && ret == -EAGAIN) { |
b7bb4f7d JA |
2776 | if (req->io) |
2777 | return -EAGAIN; | |
2778 | if (io_alloc_async_ctx(req)) | |
2779 | return -ENOMEM; | |
2780 | memcpy(&req->io->msg, &io.msg, sizeof(io.msg)); | |
2781 | req->work.func = io_sendrecv_async; | |
0b416c3e | 2782 | return -EAGAIN; |
03b1230c JA |
2783 | } |
2784 | if (ret == -ERESTARTSYS) | |
2785 | ret = -EINTR; | |
2786 | } | |
2787 | ||
b7bb4f7d | 2788 | if (!io_wq_current_is_worker() && kmsg && kmsg->iov != kmsg->fast_iov) |
0b416c3e | 2789 | kfree(kmsg->iov); |
03b1230c | 2790 | io_cqring_add_event(req, ret); |
4e88d6e7 JA |
2791 | if (ret < 0) |
2792 | req_set_fail_links(req); | |
03b1230c JA |
2793 | io_put_req_find_next(req, nxt); |
2794 | return 0; | |
0fa03c62 JA |
2795 | #else |
2796 | return -EOPNOTSUPP; | |
2797 | #endif | |
2798 | } | |
5d17b4a4 | 2799 | |
3529d8c2 | 2800 | static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
17f2fe35 JA |
2801 | { |
2802 | #if defined(CONFIG_NET) | |
8ed8d3c3 JA |
2803 | struct io_accept *accept = &req->accept; |
2804 | ||
17f2fe35 JA |
2805 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) |
2806 | return -EINVAL; | |
8042d6ce | 2807 | if (sqe->ioprio || sqe->len || sqe->buf_index) |
17f2fe35 JA |
2808 | return -EINVAL; |
2809 | ||
d55e5f5b JA |
2810 | accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
2811 | accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2)); | |
8ed8d3c3 | 2812 | accept->flags = READ_ONCE(sqe->accept_flags); |
8ed8d3c3 JA |
2813 | return 0; |
2814 | #else | |
2815 | return -EOPNOTSUPP; | |
2816 | #endif | |
2817 | } | |
17f2fe35 | 2818 | |
8ed8d3c3 JA |
2819 | #if defined(CONFIG_NET) |
2820 | static int __io_accept(struct io_kiocb *req, struct io_kiocb **nxt, | |
2821 | bool force_nonblock) | |
2822 | { | |
2823 | struct io_accept *accept = &req->accept; | |
2824 | unsigned file_flags; | |
2825 | int ret; | |
2826 | ||
2827 | file_flags = force_nonblock ? O_NONBLOCK : 0; | |
2828 | ret = __sys_accept4_file(req->file, file_flags, accept->addr, | |
2829 | accept->addr_len, accept->flags); | |
2830 | if (ret == -EAGAIN && force_nonblock) | |
17f2fe35 | 2831 | return -EAGAIN; |
8e3cca12 JA |
2832 | if (ret == -ERESTARTSYS) |
2833 | ret = -EINTR; | |
4e88d6e7 JA |
2834 | if (ret < 0) |
2835 | req_set_fail_links(req); | |
78e19bbe | 2836 | io_cqring_add_event(req, ret); |
ec9c02ad | 2837 | io_put_req_find_next(req, nxt); |
17f2fe35 | 2838 | return 0; |
8ed8d3c3 JA |
2839 | } |
2840 | ||
2841 | static void io_accept_finish(struct io_wq_work **workptr) | |
2842 | { | |
2843 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
2844 | struct io_kiocb *nxt = NULL; | |
2845 | ||
2846 | if (io_req_cancelled(req)) | |
2847 | return; | |
2848 | __io_accept(req, &nxt, false); | |
2849 | if (nxt) | |
78912934 | 2850 | io_wq_assign_next(workptr, nxt); |
8ed8d3c3 JA |
2851 | } |
2852 | #endif | |
2853 | ||
2854 | static int io_accept(struct io_kiocb *req, struct io_kiocb **nxt, | |
2855 | bool force_nonblock) | |
2856 | { | |
2857 | #if defined(CONFIG_NET) | |
2858 | int ret; | |
2859 | ||
8ed8d3c3 JA |
2860 | ret = __io_accept(req, nxt, force_nonblock); |
2861 | if (ret == -EAGAIN && force_nonblock) { | |
2862 | req->work.func = io_accept_finish; | |
2863 | req->work.flags |= IO_WQ_WORK_NEEDS_FILES; | |
2864 | io_put_req(req); | |
2865 | return -EAGAIN; | |
2866 | } | |
2867 | return 0; | |
0fa03c62 JA |
2868 | #else |
2869 | return -EOPNOTSUPP; | |
2870 | #endif | |
2871 | } | |
5d17b4a4 | 2872 | |
3529d8c2 | 2873 | static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f499a021 JA |
2874 | { |
2875 | #if defined(CONFIG_NET) | |
3529d8c2 JA |
2876 | struct io_connect *conn = &req->connect; |
2877 | struct io_async_ctx *io = req->io; | |
f499a021 | 2878 | |
3fbb51c1 JA |
2879 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) |
2880 | return -EINVAL; | |
2881 | if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags) | |
2882 | return -EINVAL; | |
2883 | ||
3529d8c2 JA |
2884 | conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
2885 | conn->addr_len = READ_ONCE(sqe->addr2); | |
2886 | ||
2887 | if (!io) | |
2888 | return 0; | |
2889 | ||
2890 | return move_addr_to_kernel(conn->addr, conn->addr_len, | |
3fbb51c1 | 2891 | &io->connect.address); |
f499a021 | 2892 | #else |
3fbb51c1 | 2893 | return -EOPNOTSUPP; |
f499a021 JA |
2894 | #endif |
2895 | } | |
2896 | ||
fc4df999 JA |
2897 | static int io_connect(struct io_kiocb *req, struct io_kiocb **nxt, |
2898 | bool force_nonblock) | |
f8e85cf2 JA |
2899 | { |
2900 | #if defined(CONFIG_NET) | |
f499a021 | 2901 | struct io_async_ctx __io, *io; |
f8e85cf2 | 2902 | unsigned file_flags; |
3fbb51c1 | 2903 | int ret; |
f8e85cf2 | 2904 | |
f499a021 JA |
2905 | if (req->io) { |
2906 | io = req->io; | |
2907 | } else { | |
3529d8c2 JA |
2908 | ret = move_addr_to_kernel(req->connect.addr, |
2909 | req->connect.addr_len, | |
2910 | &__io.connect.address); | |
f499a021 JA |
2911 | if (ret) |
2912 | goto out; | |
2913 | io = &__io; | |
2914 | } | |
2915 | ||
3fbb51c1 JA |
2916 | file_flags = force_nonblock ? O_NONBLOCK : 0; |
2917 | ||
2918 | ret = __sys_connect_file(req->file, &io->connect.address, | |
2919 | req->connect.addr_len, file_flags); | |
87f80d62 | 2920 | if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) { |
b7bb4f7d JA |
2921 | if (req->io) |
2922 | return -EAGAIN; | |
2923 | if (io_alloc_async_ctx(req)) { | |
f499a021 JA |
2924 | ret = -ENOMEM; |
2925 | goto out; | |
2926 | } | |
b7bb4f7d | 2927 | memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect)); |
f8e85cf2 | 2928 | return -EAGAIN; |
f499a021 | 2929 | } |
f8e85cf2 JA |
2930 | if (ret == -ERESTARTSYS) |
2931 | ret = -EINTR; | |
f499a021 | 2932 | out: |
4e88d6e7 JA |
2933 | if (ret < 0) |
2934 | req_set_fail_links(req); | |
f8e85cf2 JA |
2935 | io_cqring_add_event(req, ret); |
2936 | io_put_req_find_next(req, nxt); | |
2937 | return 0; | |
2938 | #else | |
2939 | return -EOPNOTSUPP; | |
2940 | #endif | |
2941 | } | |
2942 | ||
221c5eb2 JA |
2943 | static void io_poll_remove_one(struct io_kiocb *req) |
2944 | { | |
2945 | struct io_poll_iocb *poll = &req->poll; | |
2946 | ||
2947 | spin_lock(&poll->head->lock); | |
2948 | WRITE_ONCE(poll->canceled, true); | |
392edb45 JA |
2949 | if (!list_empty(&poll->wait.entry)) { |
2950 | list_del_init(&poll->wait.entry); | |
a197f664 | 2951 | io_queue_async_work(req); |
221c5eb2 JA |
2952 | } |
2953 | spin_unlock(&poll->head->lock); | |
78076bb6 | 2954 | hash_del(&req->hash_node); |
221c5eb2 JA |
2955 | } |
2956 | ||
2957 | static void io_poll_remove_all(struct io_ring_ctx *ctx) | |
2958 | { | |
78076bb6 | 2959 | struct hlist_node *tmp; |
221c5eb2 | 2960 | struct io_kiocb *req; |
78076bb6 | 2961 | int i; |
221c5eb2 JA |
2962 | |
2963 | spin_lock_irq(&ctx->completion_lock); | |
78076bb6 JA |
2964 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { |
2965 | struct hlist_head *list; | |
2966 | ||
2967 | list = &ctx->cancel_hash[i]; | |
2968 | hlist_for_each_entry_safe(req, tmp, list, hash_node) | |
2969 | io_poll_remove_one(req); | |
221c5eb2 JA |
2970 | } |
2971 | spin_unlock_irq(&ctx->completion_lock); | |
2972 | } | |
2973 | ||
47f46768 JA |
2974 | static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr) |
2975 | { | |
78076bb6 | 2976 | struct hlist_head *list; |
47f46768 JA |
2977 | struct io_kiocb *req; |
2978 | ||
78076bb6 JA |
2979 | list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)]; |
2980 | hlist_for_each_entry(req, list, hash_node) { | |
2981 | if (sqe_addr == req->user_data) { | |
eac406c6 JA |
2982 | io_poll_remove_one(req); |
2983 | return 0; | |
2984 | } | |
47f46768 JA |
2985 | } |
2986 | ||
2987 | return -ENOENT; | |
2988 | } | |
2989 | ||
3529d8c2 JA |
2990 | static int io_poll_remove_prep(struct io_kiocb *req, |
2991 | const struct io_uring_sqe *sqe) | |
0969e783 | 2992 | { |
0969e783 JA |
2993 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
2994 | return -EINVAL; | |
2995 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index || | |
2996 | sqe->poll_events) | |
2997 | return -EINVAL; | |
2998 | ||
2999 | req->poll.addr = READ_ONCE(sqe->addr); | |
0969e783 JA |
3000 | return 0; |
3001 | } | |
3002 | ||
221c5eb2 JA |
3003 | /* |
3004 | * Find a running poll command that matches one specified in sqe->addr, | |
3005 | * and remove it if found. | |
3006 | */ | |
fc4df999 | 3007 | static int io_poll_remove(struct io_kiocb *req) |
221c5eb2 JA |
3008 | { |
3009 | struct io_ring_ctx *ctx = req->ctx; | |
0969e783 | 3010 | u64 addr; |
47f46768 | 3011 | int ret; |
221c5eb2 | 3012 | |
0969e783 | 3013 | addr = req->poll.addr; |
221c5eb2 | 3014 | spin_lock_irq(&ctx->completion_lock); |
0969e783 | 3015 | ret = io_poll_cancel(ctx, addr); |
221c5eb2 JA |
3016 | spin_unlock_irq(&ctx->completion_lock); |
3017 | ||
78e19bbe | 3018 | io_cqring_add_event(req, ret); |
4e88d6e7 JA |
3019 | if (ret < 0) |
3020 | req_set_fail_links(req); | |
e65ef56d | 3021 | io_put_req(req); |
221c5eb2 JA |
3022 | return 0; |
3023 | } | |
3024 | ||
b0dd8a41 | 3025 | static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error) |
221c5eb2 | 3026 | { |
a197f664 JL |
3027 | struct io_ring_ctx *ctx = req->ctx; |
3028 | ||
8c838788 | 3029 | req->poll.done = true; |
b0dd8a41 JA |
3030 | if (error) |
3031 | io_cqring_fill_event(req, error); | |
3032 | else | |
3033 | io_cqring_fill_event(req, mangle_poll(mask)); | |
8c838788 | 3034 | io_commit_cqring(ctx); |
221c5eb2 JA |
3035 | } |
3036 | ||
561fb04a | 3037 | static void io_poll_complete_work(struct io_wq_work **workptr) |
221c5eb2 | 3038 | { |
561fb04a | 3039 | struct io_wq_work *work = *workptr; |
221c5eb2 JA |
3040 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
3041 | struct io_poll_iocb *poll = &req->poll; | |
3042 | struct poll_table_struct pt = { ._key = poll->events }; | |
3043 | struct io_ring_ctx *ctx = req->ctx; | |
89723d0b | 3044 | struct io_kiocb *nxt = NULL; |
221c5eb2 | 3045 | __poll_t mask = 0; |
b0dd8a41 | 3046 | int ret = 0; |
221c5eb2 | 3047 | |
b0dd8a41 | 3048 | if (work->flags & IO_WQ_WORK_CANCEL) { |
561fb04a | 3049 | WRITE_ONCE(poll->canceled, true); |
b0dd8a41 JA |
3050 | ret = -ECANCELED; |
3051 | } else if (READ_ONCE(poll->canceled)) { | |
3052 | ret = -ECANCELED; | |
3053 | } | |
561fb04a | 3054 | |
b0dd8a41 | 3055 | if (ret != -ECANCELED) |
221c5eb2 JA |
3056 | mask = vfs_poll(poll->file, &pt) & poll->events; |
3057 | ||
3058 | /* | |
3059 | * Note that ->ki_cancel callers also delete iocb from active_reqs after | |
3060 | * calling ->ki_cancel. We need the ctx_lock roundtrip here to | |
3061 | * synchronize with them. In the cancellation case the list_del_init | |
3062 | * itself is not actually needed, but harmless so we keep it in to | |
3063 | * avoid further branches in the fast path. | |
3064 | */ | |
3065 | spin_lock_irq(&ctx->completion_lock); | |
b0dd8a41 | 3066 | if (!mask && ret != -ECANCELED) { |
392edb45 | 3067 | add_wait_queue(poll->head, &poll->wait); |
221c5eb2 JA |
3068 | spin_unlock_irq(&ctx->completion_lock); |
3069 | return; | |
3070 | } | |
78076bb6 | 3071 | hash_del(&req->hash_node); |
b0dd8a41 | 3072 | io_poll_complete(req, mask, ret); |
221c5eb2 JA |
3073 | spin_unlock_irq(&ctx->completion_lock); |
3074 | ||
8c838788 | 3075 | io_cqring_ev_posted(ctx); |
89723d0b | 3076 | |
4e88d6e7 JA |
3077 | if (ret < 0) |
3078 | req_set_fail_links(req); | |
ec9c02ad | 3079 | io_put_req_find_next(req, &nxt); |
89723d0b | 3080 | if (nxt) |
78912934 | 3081 | io_wq_assign_next(workptr, nxt); |
221c5eb2 JA |
3082 | } |
3083 | ||
3084 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
3085 | void *key) | |
3086 | { | |
e944475e | 3087 | struct io_poll_iocb *poll = wait->private; |
221c5eb2 JA |
3088 | struct io_kiocb *req = container_of(poll, struct io_kiocb, poll); |
3089 | struct io_ring_ctx *ctx = req->ctx; | |
3090 | __poll_t mask = key_to_poll(key); | |
8c838788 | 3091 | unsigned long flags; |
221c5eb2 JA |
3092 | |
3093 | /* for instances that support it check for an event match first: */ | |
8c838788 JA |
3094 | if (mask && !(mask & poll->events)) |
3095 | return 0; | |
221c5eb2 | 3096 | |
392edb45 | 3097 | list_del_init(&poll->wait.entry); |
221c5eb2 | 3098 | |
7c9e7f0f JA |
3099 | /* |
3100 | * Run completion inline if we can. We're using trylock here because | |
3101 | * we are violating the completion_lock -> poll wq lock ordering. | |
3102 | * If we have a link timeout we're going to need the completion_lock | |
3103 | * for finalizing the request, mark us as having grabbed that already. | |
3104 | */ | |
8c838788 | 3105 | if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) { |
78076bb6 | 3106 | hash_del(&req->hash_node); |
b0dd8a41 | 3107 | io_poll_complete(req, mask, 0); |
7c9e7f0f JA |
3108 | req->flags |= REQ_F_COMP_LOCKED; |
3109 | io_put_req(req); | |
8c838788 | 3110 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
221c5eb2 | 3111 | |
8c838788 | 3112 | io_cqring_ev_posted(ctx); |
8c838788 | 3113 | } else { |
a197f664 | 3114 | io_queue_async_work(req); |
221c5eb2 JA |
3115 | } |
3116 | ||
221c5eb2 JA |
3117 | return 1; |
3118 | } | |
3119 | ||
3120 | struct io_poll_table { | |
3121 | struct poll_table_struct pt; | |
3122 | struct io_kiocb *req; | |
3123 | int error; | |
3124 | }; | |
3125 | ||
3126 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
3127 | struct poll_table_struct *p) | |
3128 | { | |
3129 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
3130 | ||
3131 | if (unlikely(pt->req->poll.head)) { | |
3132 | pt->error = -EINVAL; | |
3133 | return; | |
3134 | } | |
3135 | ||
3136 | pt->error = 0; | |
3137 | pt->req->poll.head = head; | |
392edb45 | 3138 | add_wait_queue(head, &pt->req->poll.wait); |
221c5eb2 JA |
3139 | } |
3140 | ||
eac406c6 JA |
3141 | static void io_poll_req_insert(struct io_kiocb *req) |
3142 | { | |
3143 | struct io_ring_ctx *ctx = req->ctx; | |
78076bb6 JA |
3144 | struct hlist_head *list; |
3145 | ||
3146 | list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)]; | |
3147 | hlist_add_head(&req->hash_node, list); | |
eac406c6 JA |
3148 | } |
3149 | ||
3529d8c2 | 3150 | static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
221c5eb2 JA |
3151 | { |
3152 | struct io_poll_iocb *poll = &req->poll; | |
221c5eb2 | 3153 | u16 events; |
221c5eb2 JA |
3154 | |
3155 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
3156 | return -EINVAL; | |
3157 | if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
3158 | return -EINVAL; | |
09bb8394 JA |
3159 | if (!poll->file) |
3160 | return -EBADF; | |
221c5eb2 | 3161 | |
221c5eb2 JA |
3162 | events = READ_ONCE(sqe->poll_events); |
3163 | poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP; | |
0969e783 JA |
3164 | return 0; |
3165 | } | |
3166 | ||
3167 | static int io_poll_add(struct io_kiocb *req, struct io_kiocb **nxt) | |
3168 | { | |
3169 | struct io_poll_iocb *poll = &req->poll; | |
3170 | struct io_ring_ctx *ctx = req->ctx; | |
3171 | struct io_poll_table ipt; | |
3172 | bool cancel = false; | |
3173 | __poll_t mask; | |
0969e783 JA |
3174 | |
3175 | INIT_IO_WORK(&req->work, io_poll_complete_work); | |
78076bb6 | 3176 | INIT_HLIST_NODE(&req->hash_node); |
221c5eb2 | 3177 | |
221c5eb2 | 3178 | poll->head = NULL; |
8c838788 | 3179 | poll->done = false; |
221c5eb2 JA |
3180 | poll->canceled = false; |
3181 | ||
3182 | ipt.pt._qproc = io_poll_queue_proc; | |
3183 | ipt.pt._key = poll->events; | |
3184 | ipt.req = req; | |
3185 | ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */ | |
3186 | ||
3187 | /* initialized the list so that we can do list_empty checks */ | |
392edb45 JA |
3188 | INIT_LIST_HEAD(&poll->wait.entry); |
3189 | init_waitqueue_func_entry(&poll->wait, io_poll_wake); | |
3190 | poll->wait.private = poll; | |
221c5eb2 | 3191 | |
36703247 JA |
3192 | INIT_LIST_HEAD(&req->list); |
3193 | ||
221c5eb2 | 3194 | mask = vfs_poll(poll->file, &ipt.pt) & poll->events; |
221c5eb2 JA |
3195 | |
3196 | spin_lock_irq(&ctx->completion_lock); | |
8c838788 JA |
3197 | if (likely(poll->head)) { |
3198 | spin_lock(&poll->head->lock); | |
392edb45 | 3199 | if (unlikely(list_empty(&poll->wait.entry))) { |
8c838788 JA |
3200 | if (ipt.error) |
3201 | cancel = true; | |
3202 | ipt.error = 0; | |
3203 | mask = 0; | |
3204 | } | |
3205 | if (mask || ipt.error) | |
392edb45 | 3206 | list_del_init(&poll->wait.entry); |
8c838788 JA |
3207 | else if (cancel) |
3208 | WRITE_ONCE(poll->canceled, true); | |
3209 | else if (!poll->done) /* actually waiting for an event */ | |
eac406c6 | 3210 | io_poll_req_insert(req); |
8c838788 JA |
3211 | spin_unlock(&poll->head->lock); |
3212 | } | |
3213 | if (mask) { /* no async, we'd stolen it */ | |
221c5eb2 | 3214 | ipt.error = 0; |
b0dd8a41 | 3215 | io_poll_complete(req, mask, 0); |
221c5eb2 | 3216 | } |
221c5eb2 JA |
3217 | spin_unlock_irq(&ctx->completion_lock); |
3218 | ||
8c838788 JA |
3219 | if (mask) { |
3220 | io_cqring_ev_posted(ctx); | |
ec9c02ad | 3221 | io_put_req_find_next(req, nxt); |
221c5eb2 | 3222 | } |
8c838788 | 3223 | return ipt.error; |
221c5eb2 JA |
3224 | } |
3225 | ||
5262f567 JA |
3226 | static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer) |
3227 | { | |
ad8a48ac JA |
3228 | struct io_timeout_data *data = container_of(timer, |
3229 | struct io_timeout_data, timer); | |
3230 | struct io_kiocb *req = data->req; | |
3231 | struct io_ring_ctx *ctx = req->ctx; | |
5262f567 JA |
3232 | unsigned long flags; |
3233 | ||
5262f567 JA |
3234 | atomic_inc(&ctx->cq_timeouts); |
3235 | ||
3236 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
ef03681a | 3237 | /* |
11365043 JA |
3238 | * We could be racing with timeout deletion. If the list is empty, |
3239 | * then timeout lookup already found it and will be handling it. | |
ef03681a | 3240 | */ |
842f9612 | 3241 | if (!list_empty(&req->list)) { |
11365043 | 3242 | struct io_kiocb *prev; |
5262f567 | 3243 | |
11365043 JA |
3244 | /* |
3245 | * Adjust the reqs sequence before the current one because it | |
d195a66e | 3246 | * will consume a slot in the cq_ring and the cq_tail |
11365043 JA |
3247 | * pointer will be increased, otherwise other timeout reqs may |
3248 | * return in advance without waiting for enough wait_nr. | |
3249 | */ | |
3250 | prev = req; | |
3251 | list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list) | |
3252 | prev->sequence++; | |
11365043 | 3253 | list_del_init(&req->list); |
11365043 | 3254 | } |
5262f567 | 3255 | |
78e19bbe | 3256 | io_cqring_fill_event(req, -ETIME); |
5262f567 JA |
3257 | io_commit_cqring(ctx); |
3258 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
3259 | ||
3260 | io_cqring_ev_posted(ctx); | |
4e88d6e7 | 3261 | req_set_fail_links(req); |
5262f567 JA |
3262 | io_put_req(req); |
3263 | return HRTIMER_NORESTART; | |
3264 | } | |
3265 | ||
47f46768 JA |
3266 | static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data) |
3267 | { | |
3268 | struct io_kiocb *req; | |
3269 | int ret = -ENOENT; | |
3270 | ||
3271 | list_for_each_entry(req, &ctx->timeout_list, list) { | |
3272 | if (user_data == req->user_data) { | |
3273 | list_del_init(&req->list); | |
3274 | ret = 0; | |
3275 | break; | |
3276 | } | |
3277 | } | |
3278 | ||
3279 | if (ret == -ENOENT) | |
3280 | return ret; | |
3281 | ||
2d28390a | 3282 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); |
47f46768 JA |
3283 | if (ret == -1) |
3284 | return -EALREADY; | |
3285 | ||
4e88d6e7 | 3286 | req_set_fail_links(req); |
47f46768 JA |
3287 | io_cqring_fill_event(req, -ECANCELED); |
3288 | io_put_req(req); | |
3289 | return 0; | |
3290 | } | |
3291 | ||
3529d8c2 JA |
3292 | static int io_timeout_remove_prep(struct io_kiocb *req, |
3293 | const struct io_uring_sqe *sqe) | |
b29472ee | 3294 | { |
b29472ee JA |
3295 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3296 | return -EINVAL; | |
3297 | if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len) | |
3298 | return -EINVAL; | |
3299 | ||
3300 | req->timeout.addr = READ_ONCE(sqe->addr); | |
3301 | req->timeout.flags = READ_ONCE(sqe->timeout_flags); | |
3302 | if (req->timeout.flags) | |
3303 | return -EINVAL; | |
3304 | ||
b29472ee JA |
3305 | return 0; |
3306 | } | |
3307 | ||
11365043 JA |
3308 | /* |
3309 | * Remove or update an existing timeout command | |
3310 | */ | |
fc4df999 | 3311 | static int io_timeout_remove(struct io_kiocb *req) |
11365043 JA |
3312 | { |
3313 | struct io_ring_ctx *ctx = req->ctx; | |
47f46768 | 3314 | int ret; |
11365043 | 3315 | |
11365043 | 3316 | spin_lock_irq(&ctx->completion_lock); |
b29472ee | 3317 | ret = io_timeout_cancel(ctx, req->timeout.addr); |
11365043 | 3318 | |
47f46768 | 3319 | io_cqring_fill_event(req, ret); |
11365043 JA |
3320 | io_commit_cqring(ctx); |
3321 | spin_unlock_irq(&ctx->completion_lock); | |
5262f567 | 3322 | io_cqring_ev_posted(ctx); |
4e88d6e7 JA |
3323 | if (ret < 0) |
3324 | req_set_fail_links(req); | |
ec9c02ad | 3325 | io_put_req(req); |
11365043 | 3326 | return 0; |
5262f567 JA |
3327 | } |
3328 | ||
3529d8c2 | 3329 | static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2d28390a | 3330 | bool is_timeout_link) |
5262f567 | 3331 | { |
ad8a48ac | 3332 | struct io_timeout_data *data; |
a41525ab | 3333 | unsigned flags; |
5262f567 | 3334 | |
ad8a48ac | 3335 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5262f567 | 3336 | return -EINVAL; |
ad8a48ac | 3337 | if (sqe->ioprio || sqe->buf_index || sqe->len != 1) |
a41525ab | 3338 | return -EINVAL; |
2d28390a JA |
3339 | if (sqe->off && is_timeout_link) |
3340 | return -EINVAL; | |
a41525ab JA |
3341 | flags = READ_ONCE(sqe->timeout_flags); |
3342 | if (flags & ~IORING_TIMEOUT_ABS) | |
5262f567 | 3343 | return -EINVAL; |
bdf20073 | 3344 | |
26a61679 JA |
3345 | req->timeout.count = READ_ONCE(sqe->off); |
3346 | ||
3529d8c2 | 3347 | if (!req->io && io_alloc_async_ctx(req)) |
26a61679 JA |
3348 | return -ENOMEM; |
3349 | ||
3350 | data = &req->io->timeout; | |
ad8a48ac | 3351 | data->req = req; |
ad8a48ac JA |
3352 | req->flags |= REQ_F_TIMEOUT; |
3353 | ||
3354 | if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr))) | |
5262f567 JA |
3355 | return -EFAULT; |
3356 | ||
11365043 | 3357 | if (flags & IORING_TIMEOUT_ABS) |
ad8a48ac | 3358 | data->mode = HRTIMER_MODE_ABS; |
11365043 | 3359 | else |
ad8a48ac | 3360 | data->mode = HRTIMER_MODE_REL; |
11365043 | 3361 | |
ad8a48ac JA |
3362 | hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode); |
3363 | return 0; | |
3364 | } | |
3365 | ||
fc4df999 | 3366 | static int io_timeout(struct io_kiocb *req) |
ad8a48ac JA |
3367 | { |
3368 | unsigned count; | |
3369 | struct io_ring_ctx *ctx = req->ctx; | |
3370 | struct io_timeout_data *data; | |
3371 | struct list_head *entry; | |
3372 | unsigned span = 0; | |
ad8a48ac | 3373 | |
2d28390a | 3374 | data = &req->io->timeout; |
93bd25bb | 3375 | |
5262f567 JA |
3376 | /* |
3377 | * sqe->off holds how many events that need to occur for this | |
93bd25bb JA |
3378 | * timeout event to be satisfied. If it isn't set, then this is |
3379 | * a pure timeout request, sequence isn't used. | |
5262f567 | 3380 | */ |
26a61679 | 3381 | count = req->timeout.count; |
93bd25bb JA |
3382 | if (!count) { |
3383 | req->flags |= REQ_F_TIMEOUT_NOSEQ; | |
3384 | spin_lock_irq(&ctx->completion_lock); | |
3385 | entry = ctx->timeout_list.prev; | |
3386 | goto add; | |
3387 | } | |
5262f567 JA |
3388 | |
3389 | req->sequence = ctx->cached_sq_head + count - 1; | |
2d28390a | 3390 | data->seq_offset = count; |
5262f567 JA |
3391 | |
3392 | /* | |
3393 | * Insertion sort, ensuring the first entry in the list is always | |
3394 | * the one we need first. | |
3395 | */ | |
5262f567 JA |
3396 | spin_lock_irq(&ctx->completion_lock); |
3397 | list_for_each_prev(entry, &ctx->timeout_list) { | |
3398 | struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list); | |
5da0fb1a | 3399 | unsigned nxt_sq_head; |
3400 | long long tmp, tmp_nxt; | |
2d28390a | 3401 | u32 nxt_offset = nxt->io->timeout.seq_offset; |
5262f567 | 3402 | |
93bd25bb JA |
3403 | if (nxt->flags & REQ_F_TIMEOUT_NOSEQ) |
3404 | continue; | |
3405 | ||
5da0fb1a | 3406 | /* |
3407 | * Since cached_sq_head + count - 1 can overflow, use type long | |
3408 | * long to store it. | |
3409 | */ | |
3410 | tmp = (long long)ctx->cached_sq_head + count - 1; | |
cc42e0ac PB |
3411 | nxt_sq_head = nxt->sequence - nxt_offset + 1; |
3412 | tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1; | |
5da0fb1a | 3413 | |
3414 | /* | |
3415 | * cached_sq_head may overflow, and it will never overflow twice | |
3416 | * once there is some timeout req still be valid. | |
3417 | */ | |
3418 | if (ctx->cached_sq_head < nxt_sq_head) | |
8b07a65a | 3419 | tmp += UINT_MAX; |
5da0fb1a | 3420 | |
a1f58ba4 | 3421 | if (tmp > tmp_nxt) |
5262f567 | 3422 | break; |
a1f58ba4 | 3423 | |
3424 | /* | |
3425 | * Sequence of reqs after the insert one and itself should | |
3426 | * be adjusted because each timeout req consumes a slot. | |
3427 | */ | |
3428 | span++; | |
3429 | nxt->sequence++; | |
5262f567 | 3430 | } |
a1f58ba4 | 3431 | req->sequence -= span; |
93bd25bb | 3432 | add: |
5262f567 | 3433 | list_add(&req->list, entry); |
ad8a48ac JA |
3434 | data->timer.function = io_timeout_fn; |
3435 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode); | |
5262f567 | 3436 | spin_unlock_irq(&ctx->completion_lock); |
5262f567 JA |
3437 | return 0; |
3438 | } | |
5262f567 | 3439 | |
62755e35 JA |
3440 | static bool io_cancel_cb(struct io_wq_work *work, void *data) |
3441 | { | |
3442 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
3443 | ||
3444 | return req->user_data == (unsigned long) data; | |
3445 | } | |
3446 | ||
e977d6d3 | 3447 | static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr) |
62755e35 | 3448 | { |
62755e35 | 3449 | enum io_wq_cancel cancel_ret; |
62755e35 JA |
3450 | int ret = 0; |
3451 | ||
62755e35 JA |
3452 | cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr); |
3453 | switch (cancel_ret) { | |
3454 | case IO_WQ_CANCEL_OK: | |
3455 | ret = 0; | |
3456 | break; | |
3457 | case IO_WQ_CANCEL_RUNNING: | |
3458 | ret = -EALREADY; | |
3459 | break; | |
3460 | case IO_WQ_CANCEL_NOTFOUND: | |
3461 | ret = -ENOENT; | |
3462 | break; | |
3463 | } | |
3464 | ||
e977d6d3 JA |
3465 | return ret; |
3466 | } | |
3467 | ||
47f46768 JA |
3468 | static void io_async_find_and_cancel(struct io_ring_ctx *ctx, |
3469 | struct io_kiocb *req, __u64 sqe_addr, | |
b0dd8a41 | 3470 | struct io_kiocb **nxt, int success_ret) |
47f46768 JA |
3471 | { |
3472 | unsigned long flags; | |
3473 | int ret; | |
3474 | ||
3475 | ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr); | |
3476 | if (ret != -ENOENT) { | |
3477 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
3478 | goto done; | |
3479 | } | |
3480 | ||
3481 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
3482 | ret = io_timeout_cancel(ctx, sqe_addr); | |
3483 | if (ret != -ENOENT) | |
3484 | goto done; | |
3485 | ret = io_poll_cancel(ctx, sqe_addr); | |
3486 | done: | |
b0dd8a41 JA |
3487 | if (!ret) |
3488 | ret = success_ret; | |
47f46768 JA |
3489 | io_cqring_fill_event(req, ret); |
3490 | io_commit_cqring(ctx); | |
3491 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
3492 | io_cqring_ev_posted(ctx); | |
3493 | ||
4e88d6e7 JA |
3494 | if (ret < 0) |
3495 | req_set_fail_links(req); | |
47f46768 JA |
3496 | io_put_req_find_next(req, nxt); |
3497 | } | |
3498 | ||
3529d8c2 JA |
3499 | static int io_async_cancel_prep(struct io_kiocb *req, |
3500 | const struct io_uring_sqe *sqe) | |
e977d6d3 | 3501 | { |
fbf23849 | 3502 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
e977d6d3 JA |
3503 | return -EINVAL; |
3504 | if (sqe->flags || sqe->ioprio || sqe->off || sqe->len || | |
3505 | sqe->cancel_flags) | |
3506 | return -EINVAL; | |
3507 | ||
fbf23849 JA |
3508 | req->cancel.addr = READ_ONCE(sqe->addr); |
3509 | return 0; | |
3510 | } | |
3511 | ||
3512 | static int io_async_cancel(struct io_kiocb *req, struct io_kiocb **nxt) | |
3513 | { | |
3514 | struct io_ring_ctx *ctx = req->ctx; | |
fbf23849 JA |
3515 | |
3516 | io_async_find_and_cancel(ctx, req, req->cancel.addr, nxt, 0); | |
5262f567 JA |
3517 | return 0; |
3518 | } | |
3519 | ||
05f3fb3c JA |
3520 | static int io_files_update_prep(struct io_kiocb *req, |
3521 | const struct io_uring_sqe *sqe) | |
3522 | { | |
3523 | if (sqe->flags || sqe->ioprio || sqe->rw_flags) | |
3524 | return -EINVAL; | |
3525 | ||
3526 | req->files_update.offset = READ_ONCE(sqe->off); | |
3527 | req->files_update.nr_args = READ_ONCE(sqe->len); | |
3528 | if (!req->files_update.nr_args) | |
3529 | return -EINVAL; | |
3530 | req->files_update.arg = READ_ONCE(sqe->addr); | |
3531 | return 0; | |
3532 | } | |
3533 | ||
3534 | static int io_files_update(struct io_kiocb *req, bool force_nonblock) | |
3535 | { | |
3536 | struct io_ring_ctx *ctx = req->ctx; | |
3537 | struct io_uring_files_update up; | |
3538 | int ret; | |
3539 | ||
3540 | if (force_nonblock) { | |
3541 | req->work.flags |= IO_WQ_WORK_NEEDS_FILES; | |
3542 | return -EAGAIN; | |
3543 | } | |
3544 | ||
3545 | up.offset = req->files_update.offset; | |
3546 | up.fds = req->files_update.arg; | |
3547 | ||
3548 | mutex_lock(&ctx->uring_lock); | |
3549 | ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args); | |
3550 | mutex_unlock(&ctx->uring_lock); | |
3551 | ||
3552 | if (ret < 0) | |
3553 | req_set_fail_links(req); | |
3554 | io_cqring_add_event(req, ret); | |
3555 | io_put_req(req); | |
3556 | return 0; | |
3557 | } | |
3558 | ||
3529d8c2 JA |
3559 | static int io_req_defer_prep(struct io_kiocb *req, |
3560 | const struct io_uring_sqe *sqe) | |
f67676d1 | 3561 | { |
e781573e | 3562 | ssize_t ret = 0; |
f67676d1 | 3563 | |
d625c6ee | 3564 | switch (req->opcode) { |
e781573e JA |
3565 | case IORING_OP_NOP: |
3566 | break; | |
f67676d1 JA |
3567 | case IORING_OP_READV: |
3568 | case IORING_OP_READ_FIXED: | |
3529d8c2 | 3569 | ret = io_read_prep(req, sqe, true); |
f67676d1 JA |
3570 | break; |
3571 | case IORING_OP_WRITEV: | |
3572 | case IORING_OP_WRITE_FIXED: | |
3529d8c2 | 3573 | ret = io_write_prep(req, sqe, true); |
f67676d1 | 3574 | break; |
0969e783 | 3575 | case IORING_OP_POLL_ADD: |
3529d8c2 | 3576 | ret = io_poll_add_prep(req, sqe); |
0969e783 JA |
3577 | break; |
3578 | case IORING_OP_POLL_REMOVE: | |
3529d8c2 | 3579 | ret = io_poll_remove_prep(req, sqe); |
0969e783 | 3580 | break; |
8ed8d3c3 | 3581 | case IORING_OP_FSYNC: |
3529d8c2 | 3582 | ret = io_prep_fsync(req, sqe); |
8ed8d3c3 JA |
3583 | break; |
3584 | case IORING_OP_SYNC_FILE_RANGE: | |
3529d8c2 | 3585 | ret = io_prep_sfr(req, sqe); |
8ed8d3c3 | 3586 | break; |
03b1230c | 3587 | case IORING_OP_SENDMSG: |
3529d8c2 | 3588 | ret = io_sendmsg_prep(req, sqe); |
03b1230c JA |
3589 | break; |
3590 | case IORING_OP_RECVMSG: | |
3529d8c2 | 3591 | ret = io_recvmsg_prep(req, sqe); |
03b1230c | 3592 | break; |
f499a021 | 3593 | case IORING_OP_CONNECT: |
3529d8c2 | 3594 | ret = io_connect_prep(req, sqe); |
f499a021 | 3595 | break; |
2d28390a | 3596 | case IORING_OP_TIMEOUT: |
3529d8c2 | 3597 | ret = io_timeout_prep(req, sqe, false); |
b7bb4f7d | 3598 | break; |
b29472ee | 3599 | case IORING_OP_TIMEOUT_REMOVE: |
3529d8c2 | 3600 | ret = io_timeout_remove_prep(req, sqe); |
b29472ee | 3601 | break; |
fbf23849 | 3602 | case IORING_OP_ASYNC_CANCEL: |
3529d8c2 | 3603 | ret = io_async_cancel_prep(req, sqe); |
fbf23849 | 3604 | break; |
2d28390a | 3605 | case IORING_OP_LINK_TIMEOUT: |
3529d8c2 | 3606 | ret = io_timeout_prep(req, sqe, true); |
b7bb4f7d | 3607 | break; |
8ed8d3c3 | 3608 | case IORING_OP_ACCEPT: |
3529d8c2 | 3609 | ret = io_accept_prep(req, sqe); |
8ed8d3c3 | 3610 | break; |
d63d1b5e JA |
3611 | case IORING_OP_FALLOCATE: |
3612 | ret = io_fallocate_prep(req, sqe); | |
3613 | break; | |
15b71abe JA |
3614 | case IORING_OP_OPENAT: |
3615 | ret = io_openat_prep(req, sqe); | |
3616 | break; | |
b5dba59e JA |
3617 | case IORING_OP_CLOSE: |
3618 | ret = io_close_prep(req, sqe); | |
3619 | break; | |
05f3fb3c JA |
3620 | case IORING_OP_FILES_UPDATE: |
3621 | ret = io_files_update_prep(req, sqe); | |
3622 | break; | |
eddc7ef5 JA |
3623 | case IORING_OP_STATX: |
3624 | ret = io_statx_prep(req, sqe); | |
3625 | break; | |
f67676d1 | 3626 | default: |
e781573e JA |
3627 | printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n", |
3628 | req->opcode); | |
3629 | ret = -EINVAL; | |
b7bb4f7d | 3630 | break; |
f67676d1 JA |
3631 | } |
3632 | ||
b7bb4f7d | 3633 | return ret; |
f67676d1 JA |
3634 | } |
3635 | ||
3529d8c2 | 3636 | static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
de0617e4 | 3637 | { |
a197f664 | 3638 | struct io_ring_ctx *ctx = req->ctx; |
f67676d1 | 3639 | int ret; |
de0617e4 | 3640 | |
9d858b21 BL |
3641 | /* Still need defer if there is pending req in defer list. */ |
3642 | if (!req_need_defer(req) && list_empty(&ctx->defer_list)) | |
de0617e4 JA |
3643 | return 0; |
3644 | ||
3529d8c2 | 3645 | if (!req->io && io_alloc_async_ctx(req)) |
de0617e4 JA |
3646 | return -EAGAIN; |
3647 | ||
3529d8c2 | 3648 | ret = io_req_defer_prep(req, sqe); |
b7bb4f7d | 3649 | if (ret < 0) |
2d28390a | 3650 | return ret; |
2d28390a | 3651 | |
de0617e4 | 3652 | spin_lock_irq(&ctx->completion_lock); |
9d858b21 | 3653 | if (!req_need_defer(req) && list_empty(&ctx->defer_list)) { |
de0617e4 | 3654 | spin_unlock_irq(&ctx->completion_lock); |
de0617e4 JA |
3655 | return 0; |
3656 | } | |
3657 | ||
915967f6 | 3658 | trace_io_uring_defer(ctx, req, req->user_data); |
de0617e4 JA |
3659 | list_add_tail(&req->list, &ctx->defer_list); |
3660 | spin_unlock_irq(&ctx->completion_lock); | |
3661 | return -EIOCBQUEUED; | |
3662 | } | |
3663 | ||
3529d8c2 JA |
3664 | static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
3665 | struct io_kiocb **nxt, bool force_nonblock) | |
2b188cc1 | 3666 | { |
a197f664 | 3667 | struct io_ring_ctx *ctx = req->ctx; |
d625c6ee | 3668 | int ret; |
2b188cc1 | 3669 | |
d625c6ee | 3670 | switch (req->opcode) { |
2b188cc1 | 3671 | case IORING_OP_NOP: |
78e19bbe | 3672 | ret = io_nop(req); |
2b188cc1 JA |
3673 | break; |
3674 | case IORING_OP_READV: | |
edafccee | 3675 | case IORING_OP_READ_FIXED: |
3529d8c2 JA |
3676 | if (sqe) { |
3677 | ret = io_read_prep(req, sqe, force_nonblock); | |
3678 | if (ret < 0) | |
3679 | break; | |
3680 | } | |
267bc904 | 3681 | ret = io_read(req, nxt, force_nonblock); |
edafccee | 3682 | break; |
3529d8c2 | 3683 | case IORING_OP_WRITEV: |
edafccee | 3684 | case IORING_OP_WRITE_FIXED: |
3529d8c2 JA |
3685 | if (sqe) { |
3686 | ret = io_write_prep(req, sqe, force_nonblock); | |
3687 | if (ret < 0) | |
3688 | break; | |
3689 | } | |
267bc904 | 3690 | ret = io_write(req, nxt, force_nonblock); |
2b188cc1 | 3691 | break; |
c992fe29 | 3692 | case IORING_OP_FSYNC: |
3529d8c2 JA |
3693 | if (sqe) { |
3694 | ret = io_prep_fsync(req, sqe); | |
3695 | if (ret < 0) | |
3696 | break; | |
3697 | } | |
fc4df999 | 3698 | ret = io_fsync(req, nxt, force_nonblock); |
c992fe29 | 3699 | break; |
221c5eb2 | 3700 | case IORING_OP_POLL_ADD: |
3529d8c2 JA |
3701 | if (sqe) { |
3702 | ret = io_poll_add_prep(req, sqe); | |
3703 | if (ret) | |
3704 | break; | |
3705 | } | |
fc4df999 | 3706 | ret = io_poll_add(req, nxt); |
221c5eb2 JA |
3707 | break; |
3708 | case IORING_OP_POLL_REMOVE: | |
3529d8c2 JA |
3709 | if (sqe) { |
3710 | ret = io_poll_remove_prep(req, sqe); | |
3711 | if (ret < 0) | |
3712 | break; | |
3713 | } | |
fc4df999 | 3714 | ret = io_poll_remove(req); |
221c5eb2 | 3715 | break; |
5d17b4a4 | 3716 | case IORING_OP_SYNC_FILE_RANGE: |
3529d8c2 JA |
3717 | if (sqe) { |
3718 | ret = io_prep_sfr(req, sqe); | |
3719 | if (ret < 0) | |
3720 | break; | |
3721 | } | |
fc4df999 | 3722 | ret = io_sync_file_range(req, nxt, force_nonblock); |
5d17b4a4 | 3723 | break; |
0fa03c62 | 3724 | case IORING_OP_SENDMSG: |
3529d8c2 JA |
3725 | if (sqe) { |
3726 | ret = io_sendmsg_prep(req, sqe); | |
3727 | if (ret < 0) | |
3728 | break; | |
3729 | } | |
fc4df999 | 3730 | ret = io_sendmsg(req, nxt, force_nonblock); |
0fa03c62 | 3731 | break; |
aa1fa28f | 3732 | case IORING_OP_RECVMSG: |
3529d8c2 JA |
3733 | if (sqe) { |
3734 | ret = io_recvmsg_prep(req, sqe); | |
3735 | if (ret) | |
3736 | break; | |
3737 | } | |
fc4df999 | 3738 | ret = io_recvmsg(req, nxt, force_nonblock); |
aa1fa28f | 3739 | break; |
5262f567 | 3740 | case IORING_OP_TIMEOUT: |
3529d8c2 JA |
3741 | if (sqe) { |
3742 | ret = io_timeout_prep(req, sqe, false); | |
3743 | if (ret) | |
3744 | break; | |
3745 | } | |
fc4df999 | 3746 | ret = io_timeout(req); |
5262f567 | 3747 | break; |
11365043 | 3748 | case IORING_OP_TIMEOUT_REMOVE: |
3529d8c2 JA |
3749 | if (sqe) { |
3750 | ret = io_timeout_remove_prep(req, sqe); | |
3751 | if (ret) | |
3752 | break; | |
3753 | } | |
fc4df999 | 3754 | ret = io_timeout_remove(req); |
11365043 | 3755 | break; |
17f2fe35 | 3756 | case IORING_OP_ACCEPT: |
3529d8c2 JA |
3757 | if (sqe) { |
3758 | ret = io_accept_prep(req, sqe); | |
3759 | if (ret) | |
3760 | break; | |
3761 | } | |
fc4df999 | 3762 | ret = io_accept(req, nxt, force_nonblock); |
17f2fe35 | 3763 | break; |
f8e85cf2 | 3764 | case IORING_OP_CONNECT: |
3529d8c2 JA |
3765 | if (sqe) { |
3766 | ret = io_connect_prep(req, sqe); | |
3767 | if (ret) | |
3768 | break; | |
3769 | } | |
fc4df999 | 3770 | ret = io_connect(req, nxt, force_nonblock); |
f8e85cf2 | 3771 | break; |
62755e35 | 3772 | case IORING_OP_ASYNC_CANCEL: |
3529d8c2 JA |
3773 | if (sqe) { |
3774 | ret = io_async_cancel_prep(req, sqe); | |
3775 | if (ret) | |
3776 | break; | |
3777 | } | |
fc4df999 | 3778 | ret = io_async_cancel(req, nxt); |
62755e35 | 3779 | break; |
d63d1b5e JA |
3780 | case IORING_OP_FALLOCATE: |
3781 | if (sqe) { | |
3782 | ret = io_fallocate_prep(req, sqe); | |
3783 | if (ret) | |
3784 | break; | |
3785 | } | |
3786 | ret = io_fallocate(req, nxt, force_nonblock); | |
3787 | break; | |
15b71abe JA |
3788 | case IORING_OP_OPENAT: |
3789 | if (sqe) { | |
3790 | ret = io_openat_prep(req, sqe); | |
3791 | if (ret) | |
3792 | break; | |
3793 | } | |
3794 | ret = io_openat(req, nxt, force_nonblock); | |
3795 | break; | |
b5dba59e JA |
3796 | case IORING_OP_CLOSE: |
3797 | if (sqe) { | |
3798 | ret = io_close_prep(req, sqe); | |
3799 | if (ret) | |
3800 | break; | |
3801 | } | |
3802 | ret = io_close(req, nxt, force_nonblock); | |
3803 | break; | |
05f3fb3c JA |
3804 | case IORING_OP_FILES_UPDATE: |
3805 | if (sqe) { | |
3806 | ret = io_files_update_prep(req, sqe); | |
3807 | if (ret) | |
3808 | break; | |
3809 | } | |
3810 | ret = io_files_update(req, force_nonblock); | |
3811 | break; | |
eddc7ef5 JA |
3812 | case IORING_OP_STATX: |
3813 | if (sqe) { | |
3814 | ret = io_statx_prep(req, sqe); | |
3815 | if (ret) | |
3816 | break; | |
3817 | } | |
3818 | ret = io_statx(req, nxt, force_nonblock); | |
3819 | break; | |
2b188cc1 JA |
3820 | default: |
3821 | ret = -EINVAL; | |
3822 | break; | |
3823 | } | |
3824 | ||
def596e9 JA |
3825 | if (ret) |
3826 | return ret; | |
3827 | ||
3828 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
11ba820b JA |
3829 | const bool in_async = io_wq_current_is_worker(); |
3830 | ||
9e645e11 | 3831 | if (req->result == -EAGAIN) |
def596e9 JA |
3832 | return -EAGAIN; |
3833 | ||
11ba820b JA |
3834 | /* workqueue context doesn't hold uring_lock, grab it now */ |
3835 | if (in_async) | |
3836 | mutex_lock(&ctx->uring_lock); | |
3837 | ||
def596e9 | 3838 | io_iopoll_req_issued(req); |
11ba820b JA |
3839 | |
3840 | if (in_async) | |
3841 | mutex_unlock(&ctx->uring_lock); | |
def596e9 JA |
3842 | } |
3843 | ||
3844 | return 0; | |
2b188cc1 JA |
3845 | } |
3846 | ||
561fb04a | 3847 | static void io_wq_submit_work(struct io_wq_work **workptr) |
2b188cc1 | 3848 | { |
561fb04a | 3849 | struct io_wq_work *work = *workptr; |
2b188cc1 | 3850 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
561fb04a JA |
3851 | struct io_kiocb *nxt = NULL; |
3852 | int ret = 0; | |
2b188cc1 | 3853 | |
0c9d5ccd JA |
3854 | /* if NO_CANCEL is set, we must still run the work */ |
3855 | if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) == | |
3856 | IO_WQ_WORK_CANCEL) { | |
561fb04a | 3857 | ret = -ECANCELED; |
0c9d5ccd | 3858 | } |
31b51510 | 3859 | |
561fb04a | 3860 | if (!ret) { |
cf6fd4bd PB |
3861 | req->has_user = (work->flags & IO_WQ_WORK_HAS_MM) != 0; |
3862 | req->in_async = true; | |
561fb04a | 3863 | do { |
3529d8c2 | 3864 | ret = io_issue_sqe(req, NULL, &nxt, false); |
561fb04a JA |
3865 | /* |
3866 | * We can get EAGAIN for polled IO even though we're | |
3867 | * forcing a sync submission from here, since we can't | |
3868 | * wait for request slots on the block side. | |
3869 | */ | |
3870 | if (ret != -EAGAIN) | |
3871 | break; | |
3872 | cond_resched(); | |
3873 | } while (1); | |
3874 | } | |
31b51510 | 3875 | |
561fb04a | 3876 | /* drop submission reference */ |
ec9c02ad | 3877 | io_put_req(req); |
817869d2 | 3878 | |
561fb04a | 3879 | if (ret) { |
4e88d6e7 | 3880 | req_set_fail_links(req); |
78e19bbe | 3881 | io_cqring_add_event(req, ret); |
817869d2 | 3882 | io_put_req(req); |
edafccee | 3883 | } |
2b188cc1 | 3884 | |
561fb04a | 3885 | /* if a dependent link is ready, pass it back */ |
78912934 JA |
3886 | if (!ret && nxt) |
3887 | io_wq_assign_next(workptr, nxt); | |
2b188cc1 JA |
3888 | } |
3889 | ||
15b71abe | 3890 | static int io_req_needs_file(struct io_kiocb *req, int fd) |
09bb8394 | 3891 | { |
d3656344 | 3892 | if (!io_op_defs[req->opcode].needs_file) |
9e3aa61a | 3893 | return 0; |
d3656344 JA |
3894 | if (fd == -1 && io_op_defs[req->opcode].fd_non_neg) |
3895 | return 0; | |
3896 | return 1; | |
09bb8394 JA |
3897 | } |
3898 | ||
65e19f54 JA |
3899 | static inline struct file *io_file_from_index(struct io_ring_ctx *ctx, |
3900 | int index) | |
3901 | { | |
3902 | struct fixed_file_table *table; | |
3903 | ||
05f3fb3c JA |
3904 | table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT]; |
3905 | return table->files[index & IORING_FILE_TABLE_MASK];; | |
65e19f54 JA |
3906 | } |
3907 | ||
3529d8c2 JA |
3908 | static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req, |
3909 | const struct io_uring_sqe *sqe) | |
09bb8394 | 3910 | { |
a197f664 | 3911 | struct io_ring_ctx *ctx = req->ctx; |
09bb8394 | 3912 | unsigned flags; |
d3656344 | 3913 | int fd; |
09bb8394 | 3914 | |
3529d8c2 JA |
3915 | flags = READ_ONCE(sqe->flags); |
3916 | fd = READ_ONCE(sqe->fd); | |
09bb8394 | 3917 | |
4fe2c963 | 3918 | if (flags & IOSQE_IO_DRAIN) |
de0617e4 | 3919 | req->flags |= REQ_F_IO_DRAIN; |
de0617e4 | 3920 | |
d3656344 JA |
3921 | if (!io_req_needs_file(req, fd)) |
3922 | return 0; | |
09bb8394 JA |
3923 | |
3924 | if (flags & IOSQE_FIXED_FILE) { | |
05f3fb3c | 3925 | if (unlikely(!ctx->file_data || |
09bb8394 JA |
3926 | (unsigned) fd >= ctx->nr_user_files)) |
3927 | return -EBADF; | |
b7620121 | 3928 | fd = array_index_nospec(fd, ctx->nr_user_files); |
65e19f54 JA |
3929 | req->file = io_file_from_index(ctx, fd); |
3930 | if (!req->file) | |
08a45173 | 3931 | return -EBADF; |
09bb8394 | 3932 | req->flags |= REQ_F_FIXED_FILE; |
05f3fb3c | 3933 | percpu_ref_get(&ctx->file_data->refs); |
09bb8394 | 3934 | } else { |
cf6fd4bd | 3935 | if (req->needs_fixed_file) |
09bb8394 | 3936 | return -EBADF; |
c826bd7a | 3937 | trace_io_uring_file_get(ctx, fd); |
09bb8394 JA |
3938 | req->file = io_file_get(state, fd); |
3939 | if (unlikely(!req->file)) | |
3940 | return -EBADF; | |
3941 | } | |
3942 | ||
3943 | return 0; | |
3944 | } | |
3945 | ||
a197f664 | 3946 | static int io_grab_files(struct io_kiocb *req) |
fcb323cc JA |
3947 | { |
3948 | int ret = -EBADF; | |
a197f664 | 3949 | struct io_ring_ctx *ctx = req->ctx; |
fcb323cc | 3950 | |
b5dba59e JA |
3951 | if (!req->ring_file) |
3952 | return -EBADF; | |
3953 | ||
fcb323cc JA |
3954 | rcu_read_lock(); |
3955 | spin_lock_irq(&ctx->inflight_lock); | |
3956 | /* | |
3957 | * We use the f_ops->flush() handler to ensure that we can flush | |
3958 | * out work accessing these files if the fd is closed. Check if | |
3959 | * the fd has changed since we started down this path, and disallow | |
3960 | * this operation if it has. | |
3961 | */ | |
cf6fd4bd | 3962 | if (fcheck(req->ring_fd) == req->ring_file) { |
fcb323cc JA |
3963 | list_add(&req->inflight_entry, &ctx->inflight_list); |
3964 | req->flags |= REQ_F_INFLIGHT; | |
3965 | req->work.files = current->files; | |
3966 | ret = 0; | |
3967 | } | |
3968 | spin_unlock_irq(&ctx->inflight_lock); | |
3969 | rcu_read_unlock(); | |
3970 | ||
3971 | return ret; | |
3972 | } | |
3973 | ||
2665abfd | 3974 | static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer) |
2b188cc1 | 3975 | { |
ad8a48ac JA |
3976 | struct io_timeout_data *data = container_of(timer, |
3977 | struct io_timeout_data, timer); | |
3978 | struct io_kiocb *req = data->req; | |
2665abfd JA |
3979 | struct io_ring_ctx *ctx = req->ctx; |
3980 | struct io_kiocb *prev = NULL; | |
3981 | unsigned long flags; | |
2665abfd JA |
3982 | |
3983 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
3984 | ||
3985 | /* | |
3986 | * We don't expect the list to be empty, that will only happen if we | |
3987 | * race with the completion of the linked work. | |
3988 | */ | |
4493233e PB |
3989 | if (!list_empty(&req->link_list)) { |
3990 | prev = list_entry(req->link_list.prev, struct io_kiocb, | |
3991 | link_list); | |
5d960724 | 3992 | if (refcount_inc_not_zero(&prev->refs)) { |
4493233e | 3993 | list_del_init(&req->link_list); |
5d960724 JA |
3994 | prev->flags &= ~REQ_F_LINK_TIMEOUT; |
3995 | } else | |
76a46e06 | 3996 | prev = NULL; |
2665abfd JA |
3997 | } |
3998 | ||
3999 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
4000 | ||
4001 | if (prev) { | |
4e88d6e7 | 4002 | req_set_fail_links(prev); |
b0dd8a41 JA |
4003 | io_async_find_and_cancel(ctx, req, prev->user_data, NULL, |
4004 | -ETIME); | |
76a46e06 | 4005 | io_put_req(prev); |
47f46768 JA |
4006 | } else { |
4007 | io_cqring_add_event(req, -ETIME); | |
4008 | io_put_req(req); | |
2665abfd | 4009 | } |
2665abfd JA |
4010 | return HRTIMER_NORESTART; |
4011 | } | |
4012 | ||
ad8a48ac | 4013 | static void io_queue_linked_timeout(struct io_kiocb *req) |
2665abfd | 4014 | { |
76a46e06 | 4015 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd | 4016 | |
76a46e06 JA |
4017 | /* |
4018 | * If the list is now empty, then our linked request finished before | |
4019 | * we got a chance to setup the timer | |
4020 | */ | |
4021 | spin_lock_irq(&ctx->completion_lock); | |
4493233e | 4022 | if (!list_empty(&req->link_list)) { |
2d28390a | 4023 | struct io_timeout_data *data = &req->io->timeout; |
94ae5e77 | 4024 | |
ad8a48ac JA |
4025 | data->timer.function = io_link_timeout_fn; |
4026 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), | |
4027 | data->mode); | |
2665abfd | 4028 | } |
76a46e06 | 4029 | spin_unlock_irq(&ctx->completion_lock); |
2665abfd | 4030 | |
2665abfd | 4031 | /* drop submission reference */ |
76a46e06 JA |
4032 | io_put_req(req); |
4033 | } | |
2665abfd | 4034 | |
ad8a48ac | 4035 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) |
2665abfd JA |
4036 | { |
4037 | struct io_kiocb *nxt; | |
4038 | ||
4039 | if (!(req->flags & REQ_F_LINK)) | |
4040 | return NULL; | |
4041 | ||
4493233e PB |
4042 | nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, |
4043 | link_list); | |
d625c6ee | 4044 | if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT) |
76a46e06 | 4045 | return NULL; |
2665abfd | 4046 | |
76a46e06 | 4047 | req->flags |= REQ_F_LINK_TIMEOUT; |
76a46e06 | 4048 | return nxt; |
2665abfd JA |
4049 | } |
4050 | ||
3529d8c2 | 4051 | static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2b188cc1 | 4052 | { |
4a0a7a18 | 4053 | struct io_kiocb *linked_timeout; |
f9bd67f6 | 4054 | struct io_kiocb *nxt = NULL; |
e0c5c576 | 4055 | int ret; |
2b188cc1 | 4056 | |
4a0a7a18 JA |
4057 | again: |
4058 | linked_timeout = io_prep_linked_timeout(req); | |
4059 | ||
3529d8c2 | 4060 | ret = io_issue_sqe(req, sqe, &nxt, true); |
491381ce JA |
4061 | |
4062 | /* | |
4063 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
4064 | * doesn't support non-blocking read/write attempts | |
4065 | */ | |
4066 | if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) || | |
4067 | (req->flags & REQ_F_MUST_PUNT))) { | |
bbad27b2 PB |
4068 | if (req->work.flags & IO_WQ_WORK_NEEDS_FILES) { |
4069 | ret = io_grab_files(req); | |
4070 | if (ret) | |
4071 | goto err; | |
2b188cc1 | 4072 | } |
bbad27b2 PB |
4073 | |
4074 | /* | |
4075 | * Queued up for async execution, worker will release | |
4076 | * submit reference when the iocb is actually submitted. | |
4077 | */ | |
4078 | io_queue_async_work(req); | |
4a0a7a18 | 4079 | goto done_req; |
2b188cc1 | 4080 | } |
e65ef56d | 4081 | |
fcb323cc | 4082 | err: |
76a46e06 | 4083 | /* drop submission reference */ |
ec9c02ad | 4084 | io_put_req(req); |
e65ef56d | 4085 | |
f9bd67f6 | 4086 | if (linked_timeout) { |
76a46e06 | 4087 | if (!ret) |
f9bd67f6 | 4088 | io_queue_linked_timeout(linked_timeout); |
76a46e06 | 4089 | else |
f9bd67f6 | 4090 | io_put_req(linked_timeout); |
76a46e06 JA |
4091 | } |
4092 | ||
e65ef56d | 4093 | /* and drop final reference, if we failed */ |
9e645e11 | 4094 | if (ret) { |
78e19bbe | 4095 | io_cqring_add_event(req, ret); |
4e88d6e7 | 4096 | req_set_fail_links(req); |
e65ef56d | 4097 | io_put_req(req); |
9e645e11 | 4098 | } |
4a0a7a18 JA |
4099 | done_req: |
4100 | if (nxt) { | |
4101 | req = nxt; | |
4102 | nxt = NULL; | |
4103 | goto again; | |
4104 | } | |
2b188cc1 JA |
4105 | } |
4106 | ||
3529d8c2 | 4107 | static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4fe2c963 JL |
4108 | { |
4109 | int ret; | |
4110 | ||
1b4a51b6 PB |
4111 | if (unlikely(req->ctx->drain_next)) { |
4112 | req->flags |= REQ_F_IO_DRAIN; | |
4113 | req->ctx->drain_next = false; | |
4114 | } | |
4115 | req->ctx->drain_next = (req->flags & REQ_F_DRAIN_LINK); | |
4116 | ||
3529d8c2 | 4117 | ret = io_req_defer(req, sqe); |
4fe2c963 JL |
4118 | if (ret) { |
4119 | if (ret != -EIOCBQUEUED) { | |
78e19bbe | 4120 | io_cqring_add_event(req, ret); |
4e88d6e7 | 4121 | req_set_fail_links(req); |
78e19bbe | 4122 | io_double_put_req(req); |
4fe2c963 | 4123 | } |
ce35a47a JA |
4124 | } else if ((req->flags & REQ_F_FORCE_ASYNC) && |
4125 | !io_wq_current_is_worker()) { | |
4126 | /* | |
4127 | * Never try inline submit of IOSQE_ASYNC is set, go straight | |
4128 | * to async execution. | |
4129 | */ | |
4130 | req->work.flags |= IO_WQ_WORK_CONCURRENT; | |
4131 | io_queue_async_work(req); | |
4132 | } else { | |
3529d8c2 | 4133 | __io_queue_sqe(req, sqe); |
ce35a47a | 4134 | } |
4fe2c963 JL |
4135 | } |
4136 | ||
1b4a51b6 | 4137 | static inline void io_queue_link_head(struct io_kiocb *req) |
4fe2c963 | 4138 | { |
94ae5e77 | 4139 | if (unlikely(req->flags & REQ_F_FAIL_LINK)) { |
1b4a51b6 PB |
4140 | io_cqring_add_event(req, -ECANCELED); |
4141 | io_double_put_req(req); | |
4142 | } else | |
3529d8c2 | 4143 | io_queue_sqe(req, NULL); |
4fe2c963 JL |
4144 | } |
4145 | ||
4e88d6e7 | 4146 | #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \ |
ce35a47a | 4147 | IOSQE_IO_HARDLINK | IOSQE_ASYNC) |
9e645e11 | 4148 | |
3529d8c2 JA |
4149 | static bool io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
4150 | struct io_submit_state *state, struct io_kiocb **link) | |
9e645e11 | 4151 | { |
a197f664 | 4152 | struct io_ring_ctx *ctx = req->ctx; |
32fe525b | 4153 | unsigned int sqe_flags; |
9e645e11 JA |
4154 | int ret; |
4155 | ||
32fe525b PB |
4156 | sqe_flags = READ_ONCE(sqe->flags); |
4157 | ||
9e645e11 | 4158 | /* enforce forwards compatibility on users */ |
32fe525b | 4159 | if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) { |
9e645e11 | 4160 | ret = -EINVAL; |
196be95c | 4161 | goto err_req; |
9e645e11 | 4162 | } |
32fe525b | 4163 | if (sqe_flags & IOSQE_ASYNC) |
ce35a47a | 4164 | req->flags |= REQ_F_FORCE_ASYNC; |
9e645e11 | 4165 | |
3529d8c2 | 4166 | ret = io_req_set_file(state, req, sqe); |
9e645e11 JA |
4167 | if (unlikely(ret)) { |
4168 | err_req: | |
78e19bbe JA |
4169 | io_cqring_add_event(req, ret); |
4170 | io_double_put_req(req); | |
2e6e1fde | 4171 | return false; |
9e645e11 JA |
4172 | } |
4173 | ||
9e645e11 JA |
4174 | /* |
4175 | * If we already have a head request, queue this one for async | |
4176 | * submittal once the head completes. If we don't have a head but | |
4177 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
4178 | * submitted sync once the chain is complete. If none of those | |
4179 | * conditions are true (normal request), then just queue it. | |
4180 | */ | |
4181 | if (*link) { | |
9d76377f | 4182 | struct io_kiocb *head = *link; |
9e645e11 | 4183 | |
32fe525b | 4184 | if (sqe_flags & IOSQE_IO_DRAIN) |
9d76377f | 4185 | head->flags |= REQ_F_DRAIN_LINK | REQ_F_IO_DRAIN; |
1b4a51b6 | 4186 | |
32fe525b | 4187 | if (sqe_flags & IOSQE_IO_HARDLINK) |
4e88d6e7 JA |
4188 | req->flags |= REQ_F_HARDLINK; |
4189 | ||
b7bb4f7d | 4190 | if (io_alloc_async_ctx(req)) { |
9e645e11 JA |
4191 | ret = -EAGAIN; |
4192 | goto err_req; | |
4193 | } | |
4194 | ||
3529d8c2 | 4195 | ret = io_req_defer_prep(req, sqe); |
2d28390a | 4196 | if (ret) { |
4e88d6e7 | 4197 | /* fail even hard links since we don't submit */ |
9d76377f | 4198 | head->flags |= REQ_F_FAIL_LINK; |
f67676d1 | 4199 | goto err_req; |
2d28390a | 4200 | } |
9d76377f PB |
4201 | trace_io_uring_link(ctx, req, head); |
4202 | list_add_tail(&req->link_list, &head->link_list); | |
32fe525b PB |
4203 | |
4204 | /* last request of a link, enqueue the link */ | |
4205 | if (!(sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK))) { | |
4206 | io_queue_link_head(head); | |
4207 | *link = NULL; | |
4208 | } | |
4209 | } else if (sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) { | |
9e645e11 | 4210 | req->flags |= REQ_F_LINK; |
32fe525b | 4211 | if (sqe_flags & IOSQE_IO_HARDLINK) |
4e88d6e7 | 4212 | req->flags |= REQ_F_HARDLINK; |
9e645e11 | 4213 | |
9e645e11 | 4214 | INIT_LIST_HEAD(&req->link_list); |
3529d8c2 JA |
4215 | ret = io_req_defer_prep(req, sqe); |
4216 | if (ret) | |
4217 | req->flags |= REQ_F_FAIL_LINK; | |
9e645e11 JA |
4218 | *link = req; |
4219 | } else { | |
3529d8c2 | 4220 | io_queue_sqe(req, sqe); |
9e645e11 | 4221 | } |
2e6e1fde PB |
4222 | |
4223 | return true; | |
9e645e11 JA |
4224 | } |
4225 | ||
9a56a232 JA |
4226 | /* |
4227 | * Batched submission is done, ensure local IO is flushed out. | |
4228 | */ | |
4229 | static void io_submit_state_end(struct io_submit_state *state) | |
4230 | { | |
4231 | blk_finish_plug(&state->plug); | |
3d6770fb | 4232 | io_file_put(state); |
2579f913 JA |
4233 | if (state->free_reqs) |
4234 | kmem_cache_free_bulk(req_cachep, state->free_reqs, | |
4235 | &state->reqs[state->cur_req]); | |
9a56a232 JA |
4236 | } |
4237 | ||
4238 | /* | |
4239 | * Start submission side cache. | |
4240 | */ | |
4241 | static void io_submit_state_start(struct io_submit_state *state, | |
22efde59 | 4242 | unsigned int max_ios) |
9a56a232 JA |
4243 | { |
4244 | blk_start_plug(&state->plug); | |
2579f913 | 4245 | state->free_reqs = 0; |
9a56a232 JA |
4246 | state->file = NULL; |
4247 | state->ios_left = max_ios; | |
4248 | } | |
4249 | ||
2b188cc1 JA |
4250 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
4251 | { | |
75b28aff | 4252 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 4253 | |
75b28aff | 4254 | if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) { |
2b188cc1 JA |
4255 | /* |
4256 | * Ensure any loads from the SQEs are done at this point, | |
4257 | * since once we write the new head, the application could | |
4258 | * write new data to them. | |
4259 | */ | |
75b28aff | 4260 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); |
2b188cc1 JA |
4261 | } |
4262 | } | |
4263 | ||
2b188cc1 | 4264 | /* |
3529d8c2 | 4265 | * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory |
2b188cc1 JA |
4266 | * that is mapped by userspace. This means that care needs to be taken to |
4267 | * ensure that reads are stable, as we cannot rely on userspace always | |
4268 | * being a good citizen. If members of the sqe are validated and then later | |
4269 | * used, it's important that those reads are done through READ_ONCE() to | |
4270 | * prevent a re-load down the line. | |
4271 | */ | |
3529d8c2 JA |
4272 | static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req, |
4273 | const struct io_uring_sqe **sqe_ptr) | |
2b188cc1 | 4274 | { |
75b28aff HV |
4275 | struct io_rings *rings = ctx->rings; |
4276 | u32 *sq_array = ctx->sq_array; | |
2b188cc1 JA |
4277 | unsigned head; |
4278 | ||
4279 | /* | |
4280 | * The cached sq head (or cq tail) serves two purposes: | |
4281 | * | |
4282 | * 1) allows us to batch the cost of updating the user visible | |
4283 | * head updates. | |
4284 | * 2) allows the kernel side to track the head on its own, even | |
4285 | * though the application is the one updating it. | |
4286 | */ | |
4287 | head = ctx->cached_sq_head; | |
e523a29c | 4288 | /* make sure SQ entry isn't read before tail */ |
9835d6fa | 4289 | if (unlikely(head == smp_load_acquire(&rings->sq.tail))) |
2b188cc1 JA |
4290 | return false; |
4291 | ||
75b28aff | 4292 | head = READ_ONCE(sq_array[head & ctx->sq_mask]); |
9835d6fa | 4293 | if (likely(head < ctx->sq_entries)) { |
cf6fd4bd PB |
4294 | /* |
4295 | * All io need record the previous position, if LINK vs DARIN, | |
4296 | * it can be used to mark the position of the first IO in the | |
4297 | * link list. | |
4298 | */ | |
4299 | req->sequence = ctx->cached_sq_head; | |
3529d8c2 JA |
4300 | *sqe_ptr = &ctx->sq_sqes[head]; |
4301 | req->opcode = READ_ONCE((*sqe_ptr)->opcode); | |
4302 | req->user_data = READ_ONCE((*sqe_ptr)->user_data); | |
2b188cc1 JA |
4303 | ctx->cached_sq_head++; |
4304 | return true; | |
4305 | } | |
4306 | ||
4307 | /* drop invalid entries */ | |
4308 | ctx->cached_sq_head++; | |
498ccd9e JA |
4309 | ctx->cached_sq_dropped++; |
4310 | WRITE_ONCE(rings->sq_dropped, ctx->cached_sq_dropped); | |
2b188cc1 JA |
4311 | return false; |
4312 | } | |
4313 | ||
fb5ccc98 | 4314 | static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr, |
ae9428ca PB |
4315 | struct file *ring_file, int ring_fd, |
4316 | struct mm_struct **mm, bool async) | |
6c271ce2 JA |
4317 | { |
4318 | struct io_submit_state state, *statep = NULL; | |
9e645e11 | 4319 | struct io_kiocb *link = NULL; |
9e645e11 | 4320 | int i, submitted = 0; |
95a1b3ff | 4321 | bool mm_fault = false; |
6c271ce2 | 4322 | |
c4a2ed72 | 4323 | /* if we have a backlog and couldn't flush it all, return BUSY */ |
ad3eb2c8 JA |
4324 | if (test_bit(0, &ctx->sq_check_overflow)) { |
4325 | if (!list_empty(&ctx->cq_overflow_list) && | |
4326 | !io_cqring_overflow_flush(ctx, false)) | |
4327 | return -EBUSY; | |
4328 | } | |
6c271ce2 JA |
4329 | |
4330 | if (nr > IO_PLUG_THRESHOLD) { | |
22efde59 | 4331 | io_submit_state_start(&state, nr); |
6c271ce2 JA |
4332 | statep = &state; |
4333 | } | |
4334 | ||
4335 | for (i = 0; i < nr; i++) { | |
3529d8c2 | 4336 | const struct io_uring_sqe *sqe; |
196be95c | 4337 | struct io_kiocb *req; |
fb5ccc98 | 4338 | |
196be95c PB |
4339 | req = io_get_req(ctx, statep); |
4340 | if (unlikely(!req)) { | |
4341 | if (!submitted) | |
4342 | submitted = -EAGAIN; | |
fb5ccc98 | 4343 | break; |
196be95c | 4344 | } |
3529d8c2 | 4345 | if (!io_get_sqring(ctx, req, &sqe)) { |
196be95c PB |
4346 | __io_free_req(req); |
4347 | break; | |
4348 | } | |
fb5ccc98 | 4349 | |
d3656344 JA |
4350 | /* will complete beyond this point, count as submitted */ |
4351 | submitted++; | |
4352 | ||
4353 | if (unlikely(req->opcode >= IORING_OP_LAST)) { | |
4354 | io_cqring_add_event(req, -EINVAL); | |
4355 | io_double_put_req(req); | |
4356 | break; | |
4357 | } | |
4358 | ||
4359 | if (io_op_defs[req->opcode].needs_mm && !*mm) { | |
95a1b3ff PB |
4360 | mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm); |
4361 | if (!mm_fault) { | |
4362 | use_mm(ctx->sqo_mm); | |
4363 | *mm = ctx->sqo_mm; | |
4364 | } | |
9e645e11 | 4365 | } |
9e645e11 | 4366 | |
cf6fd4bd PB |
4367 | req->ring_file = ring_file; |
4368 | req->ring_fd = ring_fd; | |
4369 | req->has_user = *mm != NULL; | |
4370 | req->in_async = async; | |
4371 | req->needs_fixed_file = async; | |
d625c6ee | 4372 | trace_io_uring_submit_sqe(ctx, req->user_data, true, async); |
3529d8c2 | 4373 | if (!io_submit_sqe(req, sqe, statep, &link)) |
2e6e1fde | 4374 | break; |
6c271ce2 JA |
4375 | } |
4376 | ||
9e645e11 | 4377 | if (link) |
1b4a51b6 | 4378 | io_queue_link_head(link); |
6c271ce2 JA |
4379 | if (statep) |
4380 | io_submit_state_end(&state); | |
4381 | ||
ae9428ca PB |
4382 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
4383 | io_commit_sqring(ctx); | |
4384 | ||
6c271ce2 JA |
4385 | return submitted; |
4386 | } | |
4387 | ||
4388 | static int io_sq_thread(void *data) | |
4389 | { | |
6c271ce2 JA |
4390 | struct io_ring_ctx *ctx = data; |
4391 | struct mm_struct *cur_mm = NULL; | |
181e448d | 4392 | const struct cred *old_cred; |
6c271ce2 JA |
4393 | mm_segment_t old_fs; |
4394 | DEFINE_WAIT(wait); | |
4395 | unsigned inflight; | |
4396 | unsigned long timeout; | |
c1edbf5f | 4397 | int ret; |
6c271ce2 | 4398 | |
206aefde | 4399 | complete(&ctx->completions[1]); |
a4c0b3de | 4400 | |
6c271ce2 JA |
4401 | old_fs = get_fs(); |
4402 | set_fs(USER_DS); | |
181e448d | 4403 | old_cred = override_creds(ctx->creds); |
6c271ce2 | 4404 | |
c1edbf5f | 4405 | ret = timeout = inflight = 0; |
2bbcd6d3 | 4406 | while (!kthread_should_park()) { |
fb5ccc98 | 4407 | unsigned int to_submit; |
6c271ce2 JA |
4408 | |
4409 | if (inflight) { | |
4410 | unsigned nr_events = 0; | |
4411 | ||
4412 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
2b2ed975 JA |
4413 | /* |
4414 | * inflight is the count of the maximum possible | |
4415 | * entries we submitted, but it can be smaller | |
4416 | * if we dropped some of them. If we don't have | |
4417 | * poll entries available, then we know that we | |
4418 | * have nothing left to poll for. Reset the | |
4419 | * inflight count to zero in that case. | |
4420 | */ | |
4421 | mutex_lock(&ctx->uring_lock); | |
4422 | if (!list_empty(&ctx->poll_list)) | |
4423 | __io_iopoll_check(ctx, &nr_events, 0); | |
4424 | else | |
4425 | inflight = 0; | |
4426 | mutex_unlock(&ctx->uring_lock); | |
6c271ce2 JA |
4427 | } else { |
4428 | /* | |
4429 | * Normal IO, just pretend everything completed. | |
4430 | * We don't have to poll completions for that. | |
4431 | */ | |
4432 | nr_events = inflight; | |
4433 | } | |
4434 | ||
4435 | inflight -= nr_events; | |
4436 | if (!inflight) | |
4437 | timeout = jiffies + ctx->sq_thread_idle; | |
4438 | } | |
4439 | ||
fb5ccc98 | 4440 | to_submit = io_sqring_entries(ctx); |
c1edbf5f JA |
4441 | |
4442 | /* | |
4443 | * If submit got -EBUSY, flag us as needing the application | |
4444 | * to enter the kernel to reap and flush events. | |
4445 | */ | |
4446 | if (!to_submit || ret == -EBUSY) { | |
6c271ce2 JA |
4447 | /* |
4448 | * We're polling. If we're within the defined idle | |
4449 | * period, then let us spin without work before going | |
c1edbf5f JA |
4450 | * to sleep. The exception is if we got EBUSY doing |
4451 | * more IO, we should wait for the application to | |
4452 | * reap events and wake us up. | |
6c271ce2 | 4453 | */ |
c1edbf5f JA |
4454 | if (inflight || |
4455 | (!time_after(jiffies, timeout) && ret != -EBUSY)) { | |
9831a90c | 4456 | cond_resched(); |
6c271ce2 JA |
4457 | continue; |
4458 | } | |
4459 | ||
4460 | /* | |
4461 | * Drop cur_mm before scheduling, we can't hold it for | |
4462 | * long periods (or over schedule()). Do this before | |
4463 | * adding ourselves to the waitqueue, as the unuse/drop | |
4464 | * may sleep. | |
4465 | */ | |
4466 | if (cur_mm) { | |
4467 | unuse_mm(cur_mm); | |
4468 | mmput(cur_mm); | |
4469 | cur_mm = NULL; | |
4470 | } | |
4471 | ||
4472 | prepare_to_wait(&ctx->sqo_wait, &wait, | |
4473 | TASK_INTERRUPTIBLE); | |
4474 | ||
4475 | /* Tell userspace we may need a wakeup call */ | |
75b28aff | 4476 | ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP; |
0d7bae69 SB |
4477 | /* make sure to read SQ tail after writing flags */ |
4478 | smp_mb(); | |
6c271ce2 | 4479 | |
fb5ccc98 | 4480 | to_submit = io_sqring_entries(ctx); |
c1edbf5f | 4481 | if (!to_submit || ret == -EBUSY) { |
2bbcd6d3 | 4482 | if (kthread_should_park()) { |
6c271ce2 JA |
4483 | finish_wait(&ctx->sqo_wait, &wait); |
4484 | break; | |
4485 | } | |
4486 | if (signal_pending(current)) | |
4487 | flush_signals(current); | |
4488 | schedule(); | |
4489 | finish_wait(&ctx->sqo_wait, &wait); | |
4490 | ||
75b28aff | 4491 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
4492 | continue; |
4493 | } | |
4494 | finish_wait(&ctx->sqo_wait, &wait); | |
4495 | ||
75b28aff | 4496 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
4497 | } |
4498 | ||
fb5ccc98 | 4499 | to_submit = min(to_submit, ctx->sq_entries); |
8a4955ff | 4500 | mutex_lock(&ctx->uring_lock); |
1d7bb1d5 | 4501 | ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true); |
8a4955ff | 4502 | mutex_unlock(&ctx->uring_lock); |
1d7bb1d5 JA |
4503 | if (ret > 0) |
4504 | inflight += ret; | |
6c271ce2 JA |
4505 | } |
4506 | ||
4507 | set_fs(old_fs); | |
4508 | if (cur_mm) { | |
4509 | unuse_mm(cur_mm); | |
4510 | mmput(cur_mm); | |
4511 | } | |
181e448d | 4512 | revert_creds(old_cred); |
06058632 | 4513 | |
2bbcd6d3 | 4514 | kthread_parkme(); |
06058632 | 4515 | |
6c271ce2 JA |
4516 | return 0; |
4517 | } | |
4518 | ||
bda52162 JA |
4519 | struct io_wait_queue { |
4520 | struct wait_queue_entry wq; | |
4521 | struct io_ring_ctx *ctx; | |
4522 | unsigned to_wait; | |
4523 | unsigned nr_timeouts; | |
4524 | }; | |
4525 | ||
1d7bb1d5 | 4526 | static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush) |
bda52162 JA |
4527 | { |
4528 | struct io_ring_ctx *ctx = iowq->ctx; | |
4529 | ||
4530 | /* | |
d195a66e | 4531 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
4532 | * started waiting. For timeouts, we always want to return to userspace, |
4533 | * regardless of event count. | |
4534 | */ | |
1d7bb1d5 | 4535 | return io_cqring_events(ctx, noflush) >= iowq->to_wait || |
bda52162 JA |
4536 | atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
4537 | } | |
4538 | ||
4539 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
4540 | int wake_flags, void *key) | |
4541 | { | |
4542 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
4543 | wq); | |
4544 | ||
1d7bb1d5 JA |
4545 | /* use noflush == true, as we can't safely rely on locking context */ |
4546 | if (!io_should_wake(iowq, true)) | |
bda52162 JA |
4547 | return -1; |
4548 | ||
4549 | return autoremove_wake_function(curr, mode, wake_flags, key); | |
4550 | } | |
4551 | ||
2b188cc1 JA |
4552 | /* |
4553 | * Wait until events become available, if we don't already have some. The | |
4554 | * application must reap them itself, as they reside on the shared cq ring. | |
4555 | */ | |
4556 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
4557 | const sigset_t __user *sig, size_t sigsz) | |
4558 | { | |
bda52162 JA |
4559 | struct io_wait_queue iowq = { |
4560 | .wq = { | |
4561 | .private = current, | |
4562 | .func = io_wake_function, | |
4563 | .entry = LIST_HEAD_INIT(iowq.wq.entry), | |
4564 | }, | |
4565 | .ctx = ctx, | |
4566 | .to_wait = min_events, | |
4567 | }; | |
75b28aff | 4568 | struct io_rings *rings = ctx->rings; |
e9ffa5c2 | 4569 | int ret = 0; |
2b188cc1 | 4570 | |
1d7bb1d5 | 4571 | if (io_cqring_events(ctx, false) >= min_events) |
2b188cc1 JA |
4572 | return 0; |
4573 | ||
4574 | if (sig) { | |
9e75ad5d AB |
4575 | #ifdef CONFIG_COMPAT |
4576 | if (in_compat_syscall()) | |
4577 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 4578 | sigsz); |
9e75ad5d AB |
4579 | else |
4580 | #endif | |
b772434b | 4581 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 4582 | |
2b188cc1 JA |
4583 | if (ret) |
4584 | return ret; | |
4585 | } | |
4586 | ||
bda52162 | 4587 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
c826bd7a | 4588 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 JA |
4589 | do { |
4590 | prepare_to_wait_exclusive(&ctx->wait, &iowq.wq, | |
4591 | TASK_INTERRUPTIBLE); | |
1d7bb1d5 | 4592 | if (io_should_wake(&iowq, false)) |
bda52162 JA |
4593 | break; |
4594 | schedule(); | |
4595 | if (signal_pending(current)) { | |
e9ffa5c2 | 4596 | ret = -EINTR; |
bda52162 JA |
4597 | break; |
4598 | } | |
4599 | } while (1); | |
4600 | finish_wait(&ctx->wait, &iowq.wq); | |
4601 | ||
e9ffa5c2 | 4602 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 4603 | |
75b28aff | 4604 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
4605 | } |
4606 | ||
6b06314c JA |
4607 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
4608 | { | |
4609 | #if defined(CONFIG_UNIX) | |
4610 | if (ctx->ring_sock) { | |
4611 | struct sock *sock = ctx->ring_sock->sk; | |
4612 | struct sk_buff *skb; | |
4613 | ||
4614 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
4615 | kfree_skb(skb); | |
4616 | } | |
4617 | #else | |
4618 | int i; | |
4619 | ||
65e19f54 JA |
4620 | for (i = 0; i < ctx->nr_user_files; i++) { |
4621 | struct file *file; | |
4622 | ||
4623 | file = io_file_from_index(ctx, i); | |
4624 | if (file) | |
4625 | fput(file); | |
4626 | } | |
6b06314c JA |
4627 | #endif |
4628 | } | |
4629 | ||
05f3fb3c JA |
4630 | static void io_file_ref_kill(struct percpu_ref *ref) |
4631 | { | |
4632 | struct fixed_file_data *data; | |
4633 | ||
4634 | data = container_of(ref, struct fixed_file_data, refs); | |
4635 | complete(&data->done); | |
4636 | } | |
4637 | ||
6b06314c JA |
4638 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) |
4639 | { | |
05f3fb3c | 4640 | struct fixed_file_data *data = ctx->file_data; |
65e19f54 JA |
4641 | unsigned nr_tables, i; |
4642 | ||
05f3fb3c | 4643 | if (!data) |
6b06314c JA |
4644 | return -ENXIO; |
4645 | ||
05f3fb3c JA |
4646 | /* protect against inflight atomic switch, which drops the ref */ |
4647 | flush_work(&data->ref_work); | |
4648 | percpu_ref_get(&data->refs); | |
4649 | percpu_ref_kill_and_confirm(&data->refs, io_file_ref_kill); | |
4650 | wait_for_completion(&data->done); | |
4651 | percpu_ref_put(&data->refs); | |
4652 | percpu_ref_exit(&data->refs); | |
4653 | ||
6b06314c | 4654 | __io_sqe_files_unregister(ctx); |
65e19f54 JA |
4655 | nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE); |
4656 | for (i = 0; i < nr_tables; i++) | |
05f3fb3c JA |
4657 | kfree(data->table[i].files); |
4658 | kfree(data->table); | |
4659 | kfree(data); | |
4660 | ctx->file_data = NULL; | |
6b06314c JA |
4661 | ctx->nr_user_files = 0; |
4662 | return 0; | |
4663 | } | |
4664 | ||
6c271ce2 JA |
4665 | static void io_sq_thread_stop(struct io_ring_ctx *ctx) |
4666 | { | |
4667 | if (ctx->sqo_thread) { | |
206aefde | 4668 | wait_for_completion(&ctx->completions[1]); |
2bbcd6d3 RP |
4669 | /* |
4670 | * The park is a bit of a work-around, without it we get | |
4671 | * warning spews on shutdown with SQPOLL set and affinity | |
4672 | * set to a single CPU. | |
4673 | */ | |
06058632 | 4674 | kthread_park(ctx->sqo_thread); |
6c271ce2 JA |
4675 | kthread_stop(ctx->sqo_thread); |
4676 | ctx->sqo_thread = NULL; | |
4677 | } | |
4678 | } | |
4679 | ||
6b06314c JA |
4680 | static void io_finish_async(struct io_ring_ctx *ctx) |
4681 | { | |
6c271ce2 JA |
4682 | io_sq_thread_stop(ctx); |
4683 | ||
561fb04a JA |
4684 | if (ctx->io_wq) { |
4685 | io_wq_destroy(ctx->io_wq); | |
4686 | ctx->io_wq = NULL; | |
6b06314c JA |
4687 | } |
4688 | } | |
4689 | ||
4690 | #if defined(CONFIG_UNIX) | |
6b06314c JA |
4691 | /* |
4692 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
4693 | * the io_uring can be safely unregistered on process exit, even if we have | |
4694 | * loops in the file referencing. | |
4695 | */ | |
4696 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
4697 | { | |
4698 | struct sock *sk = ctx->ring_sock->sk; | |
4699 | struct scm_fp_list *fpl; | |
4700 | struct sk_buff *skb; | |
08a45173 | 4701 | int i, nr_files; |
6b06314c JA |
4702 | |
4703 | if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) { | |
4704 | unsigned long inflight = ctx->user->unix_inflight + nr; | |
4705 | ||
4706 | if (inflight > task_rlimit(current, RLIMIT_NOFILE)) | |
4707 | return -EMFILE; | |
4708 | } | |
4709 | ||
4710 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); | |
4711 | if (!fpl) | |
4712 | return -ENOMEM; | |
4713 | ||
4714 | skb = alloc_skb(0, GFP_KERNEL); | |
4715 | if (!skb) { | |
4716 | kfree(fpl); | |
4717 | return -ENOMEM; | |
4718 | } | |
4719 | ||
4720 | skb->sk = sk; | |
6b06314c | 4721 | |
08a45173 | 4722 | nr_files = 0; |
6b06314c JA |
4723 | fpl->user = get_uid(ctx->user); |
4724 | for (i = 0; i < nr; i++) { | |
65e19f54 JA |
4725 | struct file *file = io_file_from_index(ctx, i + offset); |
4726 | ||
4727 | if (!file) | |
08a45173 | 4728 | continue; |
65e19f54 | 4729 | fpl->fp[nr_files] = get_file(file); |
08a45173 JA |
4730 | unix_inflight(fpl->user, fpl->fp[nr_files]); |
4731 | nr_files++; | |
6b06314c JA |
4732 | } |
4733 | ||
08a45173 JA |
4734 | if (nr_files) { |
4735 | fpl->max = SCM_MAX_FD; | |
4736 | fpl->count = nr_files; | |
4737 | UNIXCB(skb).fp = fpl; | |
05f3fb3c | 4738 | skb->destructor = unix_destruct_scm; |
08a45173 JA |
4739 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); |
4740 | skb_queue_head(&sk->sk_receive_queue, skb); | |
6b06314c | 4741 | |
08a45173 JA |
4742 | for (i = 0; i < nr_files; i++) |
4743 | fput(fpl->fp[i]); | |
4744 | } else { | |
4745 | kfree_skb(skb); | |
4746 | kfree(fpl); | |
4747 | } | |
6b06314c JA |
4748 | |
4749 | return 0; | |
4750 | } | |
4751 | ||
4752 | /* | |
4753 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
4754 | * causes regular reference counting to break down. We rely on the UNIX | |
4755 | * garbage collection to take care of this problem for us. | |
4756 | */ | |
4757 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
4758 | { | |
4759 | unsigned left, total; | |
4760 | int ret = 0; | |
4761 | ||
4762 | total = 0; | |
4763 | left = ctx->nr_user_files; | |
4764 | while (left) { | |
4765 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
4766 | |
4767 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
4768 | if (ret) | |
4769 | break; | |
4770 | left -= this_files; | |
4771 | total += this_files; | |
4772 | } | |
4773 | ||
4774 | if (!ret) | |
4775 | return 0; | |
4776 | ||
4777 | while (total < ctx->nr_user_files) { | |
65e19f54 JA |
4778 | struct file *file = io_file_from_index(ctx, total); |
4779 | ||
4780 | if (file) | |
4781 | fput(file); | |
6b06314c JA |
4782 | total++; |
4783 | } | |
4784 | ||
4785 | return ret; | |
4786 | } | |
4787 | #else | |
4788 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
4789 | { | |
4790 | return 0; | |
4791 | } | |
4792 | #endif | |
4793 | ||
65e19f54 JA |
4794 | static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables, |
4795 | unsigned nr_files) | |
4796 | { | |
4797 | int i; | |
4798 | ||
4799 | for (i = 0; i < nr_tables; i++) { | |
05f3fb3c | 4800 | struct fixed_file_table *table = &ctx->file_data->table[i]; |
65e19f54 JA |
4801 | unsigned this_files; |
4802 | ||
4803 | this_files = min(nr_files, IORING_MAX_FILES_TABLE); | |
4804 | table->files = kcalloc(this_files, sizeof(struct file *), | |
4805 | GFP_KERNEL); | |
4806 | if (!table->files) | |
4807 | break; | |
4808 | nr_files -= this_files; | |
4809 | } | |
4810 | ||
4811 | if (i == nr_tables) | |
4812 | return 0; | |
4813 | ||
4814 | for (i = 0; i < nr_tables; i++) { | |
05f3fb3c | 4815 | struct fixed_file_table *table = &ctx->file_data->table[i]; |
65e19f54 JA |
4816 | kfree(table->files); |
4817 | } | |
4818 | return 1; | |
4819 | } | |
4820 | ||
05f3fb3c JA |
4821 | static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file) |
4822 | { | |
4823 | #if defined(CONFIG_UNIX) | |
4824 | struct sock *sock = ctx->ring_sock->sk; | |
4825 | struct sk_buff_head list, *head = &sock->sk_receive_queue; | |
4826 | struct sk_buff *skb; | |
4827 | int i; | |
4828 | ||
4829 | __skb_queue_head_init(&list); | |
4830 | ||
4831 | /* | |
4832 | * Find the skb that holds this file in its SCM_RIGHTS. When found, | |
4833 | * remove this entry and rearrange the file array. | |
4834 | */ | |
4835 | skb = skb_dequeue(head); | |
4836 | while (skb) { | |
4837 | struct scm_fp_list *fp; | |
4838 | ||
4839 | fp = UNIXCB(skb).fp; | |
4840 | for (i = 0; i < fp->count; i++) { | |
4841 | int left; | |
4842 | ||
4843 | if (fp->fp[i] != file) | |
4844 | continue; | |
4845 | ||
4846 | unix_notinflight(fp->user, fp->fp[i]); | |
4847 | left = fp->count - 1 - i; | |
4848 | if (left) { | |
4849 | memmove(&fp->fp[i], &fp->fp[i + 1], | |
4850 | left * sizeof(struct file *)); | |
4851 | } | |
4852 | fp->count--; | |
4853 | if (!fp->count) { | |
4854 | kfree_skb(skb); | |
4855 | skb = NULL; | |
4856 | } else { | |
4857 | __skb_queue_tail(&list, skb); | |
4858 | } | |
4859 | fput(file); | |
4860 | file = NULL; | |
4861 | break; | |
4862 | } | |
4863 | ||
4864 | if (!file) | |
4865 | break; | |
4866 | ||
4867 | __skb_queue_tail(&list, skb); | |
4868 | ||
4869 | skb = skb_dequeue(head); | |
4870 | } | |
4871 | ||
4872 | if (skb_peek(&list)) { | |
4873 | spin_lock_irq(&head->lock); | |
4874 | while ((skb = __skb_dequeue(&list)) != NULL) | |
4875 | __skb_queue_tail(head, skb); | |
4876 | spin_unlock_irq(&head->lock); | |
4877 | } | |
4878 | #else | |
4879 | fput(file); | |
4880 | #endif | |
4881 | } | |
4882 | ||
4883 | struct io_file_put { | |
4884 | struct llist_node llist; | |
4885 | struct file *file; | |
4886 | struct completion *done; | |
4887 | }; | |
4888 | ||
4889 | static void io_ring_file_ref_switch(struct work_struct *work) | |
4890 | { | |
4891 | struct io_file_put *pfile, *tmp; | |
4892 | struct fixed_file_data *data; | |
4893 | struct llist_node *node; | |
4894 | ||
4895 | data = container_of(work, struct fixed_file_data, ref_work); | |
4896 | ||
4897 | while ((node = llist_del_all(&data->put_llist)) != NULL) { | |
4898 | llist_for_each_entry_safe(pfile, tmp, node, llist) { | |
4899 | io_ring_file_put(data->ctx, pfile->file); | |
4900 | if (pfile->done) | |
4901 | complete(pfile->done); | |
4902 | else | |
4903 | kfree(pfile); | |
4904 | } | |
4905 | } | |
4906 | ||
4907 | percpu_ref_get(&data->refs); | |
4908 | percpu_ref_switch_to_percpu(&data->refs); | |
4909 | } | |
4910 | ||
4911 | static void io_file_data_ref_zero(struct percpu_ref *ref) | |
4912 | { | |
4913 | struct fixed_file_data *data; | |
4914 | ||
4915 | data = container_of(ref, struct fixed_file_data, refs); | |
4916 | ||
4917 | /* we can't safely switch from inside this context, punt to wq */ | |
4918 | queue_work(system_wq, &data->ref_work); | |
4919 | } | |
4920 | ||
6b06314c JA |
4921 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, |
4922 | unsigned nr_args) | |
4923 | { | |
4924 | __s32 __user *fds = (__s32 __user *) arg; | |
65e19f54 | 4925 | unsigned nr_tables; |
05f3fb3c | 4926 | struct file *file; |
6b06314c JA |
4927 | int fd, ret = 0; |
4928 | unsigned i; | |
4929 | ||
05f3fb3c | 4930 | if (ctx->file_data) |
6b06314c JA |
4931 | return -EBUSY; |
4932 | if (!nr_args) | |
4933 | return -EINVAL; | |
4934 | if (nr_args > IORING_MAX_FIXED_FILES) | |
4935 | return -EMFILE; | |
4936 | ||
05f3fb3c JA |
4937 | ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL); |
4938 | if (!ctx->file_data) | |
4939 | return -ENOMEM; | |
4940 | ctx->file_data->ctx = ctx; | |
4941 | init_completion(&ctx->file_data->done); | |
4942 | ||
65e19f54 | 4943 | nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE); |
05f3fb3c JA |
4944 | ctx->file_data->table = kcalloc(nr_tables, |
4945 | sizeof(struct fixed_file_table), | |
65e19f54 | 4946 | GFP_KERNEL); |
05f3fb3c JA |
4947 | if (!ctx->file_data->table) { |
4948 | kfree(ctx->file_data); | |
4949 | ctx->file_data = NULL; | |
6b06314c | 4950 | return -ENOMEM; |
05f3fb3c JA |
4951 | } |
4952 | ||
4953 | if (percpu_ref_init(&ctx->file_data->refs, io_file_data_ref_zero, | |
4954 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) { | |
4955 | kfree(ctx->file_data->table); | |
4956 | kfree(ctx->file_data); | |
4957 | ctx->file_data = NULL; | |
4958 | return -ENOMEM; | |
4959 | } | |
4960 | ctx->file_data->put_llist.first = NULL; | |
4961 | INIT_WORK(&ctx->file_data->ref_work, io_ring_file_ref_switch); | |
6b06314c | 4962 | |
65e19f54 | 4963 | if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) { |
05f3fb3c JA |
4964 | percpu_ref_exit(&ctx->file_data->refs); |
4965 | kfree(ctx->file_data->table); | |
4966 | kfree(ctx->file_data); | |
4967 | ctx->file_data = NULL; | |
65e19f54 JA |
4968 | return -ENOMEM; |
4969 | } | |
4970 | ||
08a45173 | 4971 | for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { |
65e19f54 JA |
4972 | struct fixed_file_table *table; |
4973 | unsigned index; | |
4974 | ||
6b06314c JA |
4975 | ret = -EFAULT; |
4976 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) | |
4977 | break; | |
08a45173 JA |
4978 | /* allow sparse sets */ |
4979 | if (fd == -1) { | |
4980 | ret = 0; | |
4981 | continue; | |
4982 | } | |
6b06314c | 4983 | |
05f3fb3c | 4984 | table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT]; |
65e19f54 | 4985 | index = i & IORING_FILE_TABLE_MASK; |
05f3fb3c | 4986 | file = fget(fd); |
6b06314c JA |
4987 | |
4988 | ret = -EBADF; | |
05f3fb3c | 4989 | if (!file) |
6b06314c | 4990 | break; |
05f3fb3c | 4991 | |
6b06314c JA |
4992 | /* |
4993 | * Don't allow io_uring instances to be registered. If UNIX | |
4994 | * isn't enabled, then this causes a reference cycle and this | |
4995 | * instance can never get freed. If UNIX is enabled we'll | |
4996 | * handle it just fine, but there's still no point in allowing | |
4997 | * a ring fd as it doesn't support regular read/write anyway. | |
4998 | */ | |
05f3fb3c JA |
4999 | if (file->f_op == &io_uring_fops) { |
5000 | fput(file); | |
6b06314c JA |
5001 | break; |
5002 | } | |
6b06314c | 5003 | ret = 0; |
05f3fb3c | 5004 | table->files[index] = file; |
6b06314c JA |
5005 | } |
5006 | ||
5007 | if (ret) { | |
65e19f54 | 5008 | for (i = 0; i < ctx->nr_user_files; i++) { |
65e19f54 JA |
5009 | file = io_file_from_index(ctx, i); |
5010 | if (file) | |
5011 | fput(file); | |
5012 | } | |
5013 | for (i = 0; i < nr_tables; i++) | |
05f3fb3c | 5014 | kfree(ctx->file_data->table[i].files); |
6b06314c | 5015 | |
05f3fb3c JA |
5016 | kfree(ctx->file_data->table); |
5017 | kfree(ctx->file_data); | |
5018 | ctx->file_data = NULL; | |
6b06314c JA |
5019 | ctx->nr_user_files = 0; |
5020 | return ret; | |
5021 | } | |
5022 | ||
5023 | ret = io_sqe_files_scm(ctx); | |
5024 | if (ret) | |
5025 | io_sqe_files_unregister(ctx); | |
5026 | ||
5027 | return ret; | |
5028 | } | |
5029 | ||
c3a31e60 JA |
5030 | static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file, |
5031 | int index) | |
5032 | { | |
5033 | #if defined(CONFIG_UNIX) | |
5034 | struct sock *sock = ctx->ring_sock->sk; | |
5035 | struct sk_buff_head *head = &sock->sk_receive_queue; | |
5036 | struct sk_buff *skb; | |
5037 | ||
5038 | /* | |
5039 | * See if we can merge this file into an existing skb SCM_RIGHTS | |
5040 | * file set. If there's no room, fall back to allocating a new skb | |
5041 | * and filling it in. | |
5042 | */ | |
5043 | spin_lock_irq(&head->lock); | |
5044 | skb = skb_peek(head); | |
5045 | if (skb) { | |
5046 | struct scm_fp_list *fpl = UNIXCB(skb).fp; | |
5047 | ||
5048 | if (fpl->count < SCM_MAX_FD) { | |
5049 | __skb_unlink(skb, head); | |
5050 | spin_unlock_irq(&head->lock); | |
5051 | fpl->fp[fpl->count] = get_file(file); | |
5052 | unix_inflight(fpl->user, fpl->fp[fpl->count]); | |
5053 | fpl->count++; | |
5054 | spin_lock_irq(&head->lock); | |
5055 | __skb_queue_head(head, skb); | |
5056 | } else { | |
5057 | skb = NULL; | |
5058 | } | |
5059 | } | |
5060 | spin_unlock_irq(&head->lock); | |
5061 | ||
5062 | if (skb) { | |
5063 | fput(file); | |
5064 | return 0; | |
5065 | } | |
5066 | ||
5067 | return __io_sqe_files_scm(ctx, 1, index); | |
5068 | #else | |
5069 | return 0; | |
5070 | #endif | |
5071 | } | |
5072 | ||
05f3fb3c | 5073 | static void io_atomic_switch(struct percpu_ref *ref) |
c3a31e60 | 5074 | { |
05f3fb3c JA |
5075 | struct fixed_file_data *data; |
5076 | ||
5077 | data = container_of(ref, struct fixed_file_data, refs); | |
5078 | clear_bit(FFD_F_ATOMIC, &data->state); | |
5079 | } | |
5080 | ||
5081 | static bool io_queue_file_removal(struct fixed_file_data *data, | |
5082 | struct file *file) | |
5083 | { | |
5084 | struct io_file_put *pfile, pfile_stack; | |
5085 | DECLARE_COMPLETION_ONSTACK(done); | |
5086 | ||
5087 | /* | |
5088 | * If we fail allocating the struct we need for doing async reomval | |
5089 | * of this file, just punt to sync and wait for it. | |
5090 | */ | |
5091 | pfile = kzalloc(sizeof(*pfile), GFP_KERNEL); | |
5092 | if (!pfile) { | |
5093 | pfile = &pfile_stack; | |
5094 | pfile->done = &done; | |
5095 | } | |
5096 | ||
5097 | pfile->file = file; | |
5098 | llist_add(&pfile->llist, &data->put_llist); | |
5099 | ||
5100 | if (pfile == &pfile_stack) { | |
5101 | if (!test_and_set_bit(FFD_F_ATOMIC, &data->state)) { | |
5102 | percpu_ref_put(&data->refs); | |
5103 | percpu_ref_switch_to_atomic(&data->refs, | |
5104 | io_atomic_switch); | |
5105 | } | |
5106 | wait_for_completion(&done); | |
5107 | flush_work(&data->ref_work); | |
5108 | return false; | |
5109 | } | |
5110 | ||
5111 | return true; | |
5112 | } | |
5113 | ||
5114 | static int __io_sqe_files_update(struct io_ring_ctx *ctx, | |
5115 | struct io_uring_files_update *up, | |
5116 | unsigned nr_args) | |
5117 | { | |
5118 | struct fixed_file_data *data = ctx->file_data; | |
5119 | bool ref_switch = false; | |
5120 | struct file *file; | |
c3a31e60 JA |
5121 | __s32 __user *fds; |
5122 | int fd, i, err; | |
5123 | __u32 done; | |
5124 | ||
05f3fb3c | 5125 | if (check_add_overflow(up->offset, nr_args, &done)) |
c3a31e60 JA |
5126 | return -EOVERFLOW; |
5127 | if (done > ctx->nr_user_files) | |
5128 | return -EINVAL; | |
5129 | ||
5130 | done = 0; | |
05f3fb3c | 5131 | fds = u64_to_user_ptr(up->fds); |
c3a31e60 | 5132 | while (nr_args) { |
65e19f54 JA |
5133 | struct fixed_file_table *table; |
5134 | unsigned index; | |
5135 | ||
c3a31e60 JA |
5136 | err = 0; |
5137 | if (copy_from_user(&fd, &fds[done], sizeof(fd))) { | |
5138 | err = -EFAULT; | |
5139 | break; | |
5140 | } | |
05f3fb3c JA |
5141 | i = array_index_nospec(up->offset, ctx->nr_user_files); |
5142 | table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT]; | |
65e19f54 JA |
5143 | index = i & IORING_FILE_TABLE_MASK; |
5144 | if (table->files[index]) { | |
05f3fb3c | 5145 | file = io_file_from_index(ctx, index); |
65e19f54 | 5146 | table->files[index] = NULL; |
05f3fb3c JA |
5147 | if (io_queue_file_removal(data, file)) |
5148 | ref_switch = true; | |
c3a31e60 JA |
5149 | } |
5150 | if (fd != -1) { | |
c3a31e60 JA |
5151 | file = fget(fd); |
5152 | if (!file) { | |
5153 | err = -EBADF; | |
5154 | break; | |
5155 | } | |
5156 | /* | |
5157 | * Don't allow io_uring instances to be registered. If | |
5158 | * UNIX isn't enabled, then this causes a reference | |
5159 | * cycle and this instance can never get freed. If UNIX | |
5160 | * is enabled we'll handle it just fine, but there's | |
5161 | * still no point in allowing a ring fd as it doesn't | |
5162 | * support regular read/write anyway. | |
5163 | */ | |
5164 | if (file->f_op == &io_uring_fops) { | |
5165 | fput(file); | |
5166 | err = -EBADF; | |
5167 | break; | |
5168 | } | |
65e19f54 | 5169 | table->files[index] = file; |
c3a31e60 JA |
5170 | err = io_sqe_file_register(ctx, file, i); |
5171 | if (err) | |
5172 | break; | |
5173 | } | |
5174 | nr_args--; | |
5175 | done++; | |
05f3fb3c JA |
5176 | up->offset++; |
5177 | } | |
5178 | ||
5179 | if (ref_switch && !test_and_set_bit(FFD_F_ATOMIC, &data->state)) { | |
5180 | percpu_ref_put(&data->refs); | |
5181 | percpu_ref_switch_to_atomic(&data->refs, io_atomic_switch); | |
c3a31e60 JA |
5182 | } |
5183 | ||
5184 | return done ? done : err; | |
5185 | } | |
05f3fb3c JA |
5186 | static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg, |
5187 | unsigned nr_args) | |
5188 | { | |
5189 | struct io_uring_files_update up; | |
5190 | ||
5191 | if (!ctx->file_data) | |
5192 | return -ENXIO; | |
5193 | if (!nr_args) | |
5194 | return -EINVAL; | |
5195 | if (copy_from_user(&up, arg, sizeof(up))) | |
5196 | return -EFAULT; | |
5197 | if (up.resv) | |
5198 | return -EINVAL; | |
5199 | ||
5200 | return __io_sqe_files_update(ctx, &up, nr_args); | |
5201 | } | |
c3a31e60 | 5202 | |
7d723065 JA |
5203 | static void io_put_work(struct io_wq_work *work) |
5204 | { | |
5205 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
5206 | ||
5207 | io_put_req(req); | |
5208 | } | |
5209 | ||
5210 | static void io_get_work(struct io_wq_work *work) | |
5211 | { | |
5212 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
5213 | ||
5214 | refcount_inc(&req->refs); | |
5215 | } | |
5216 | ||
6c271ce2 JA |
5217 | static int io_sq_offload_start(struct io_ring_ctx *ctx, |
5218 | struct io_uring_params *p) | |
2b188cc1 | 5219 | { |
576a347b | 5220 | struct io_wq_data data; |
561fb04a | 5221 | unsigned concurrency; |
2b188cc1 JA |
5222 | int ret; |
5223 | ||
6c271ce2 | 5224 | init_waitqueue_head(&ctx->sqo_wait); |
2b188cc1 JA |
5225 | mmgrab(current->mm); |
5226 | ctx->sqo_mm = current->mm; | |
5227 | ||
6c271ce2 | 5228 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3ec482d1 JA |
5229 | ret = -EPERM; |
5230 | if (!capable(CAP_SYS_ADMIN)) | |
5231 | goto err; | |
5232 | ||
917257da JA |
5233 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
5234 | if (!ctx->sq_thread_idle) | |
5235 | ctx->sq_thread_idle = HZ; | |
5236 | ||
6c271ce2 | 5237 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 5238 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 5239 | |
917257da | 5240 | ret = -EINVAL; |
44a9bd18 JA |
5241 | if (cpu >= nr_cpu_ids) |
5242 | goto err; | |
7889f44d | 5243 | if (!cpu_online(cpu)) |
917257da JA |
5244 | goto err; |
5245 | ||
6c271ce2 JA |
5246 | ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread, |
5247 | ctx, cpu, | |
5248 | "io_uring-sq"); | |
5249 | } else { | |
5250 | ctx->sqo_thread = kthread_create(io_sq_thread, ctx, | |
5251 | "io_uring-sq"); | |
5252 | } | |
5253 | if (IS_ERR(ctx->sqo_thread)) { | |
5254 | ret = PTR_ERR(ctx->sqo_thread); | |
5255 | ctx->sqo_thread = NULL; | |
5256 | goto err; | |
5257 | } | |
5258 | wake_up_process(ctx->sqo_thread); | |
5259 | } else if (p->flags & IORING_SETUP_SQ_AFF) { | |
5260 | /* Can't have SQ_AFF without SQPOLL */ | |
5261 | ret = -EINVAL; | |
5262 | goto err; | |
5263 | } | |
5264 | ||
576a347b JA |
5265 | data.mm = ctx->sqo_mm; |
5266 | data.user = ctx->user; | |
181e448d | 5267 | data.creds = ctx->creds; |
576a347b JA |
5268 | data.get_work = io_get_work; |
5269 | data.put_work = io_put_work; | |
5270 | ||
561fb04a JA |
5271 | /* Do QD, or 4 * CPUS, whatever is smallest */ |
5272 | concurrency = min(ctx->sq_entries, 4 * num_online_cpus()); | |
576a347b | 5273 | ctx->io_wq = io_wq_create(concurrency, &data); |
975c99a5 JA |
5274 | if (IS_ERR(ctx->io_wq)) { |
5275 | ret = PTR_ERR(ctx->io_wq); | |
5276 | ctx->io_wq = NULL; | |
2b188cc1 JA |
5277 | goto err; |
5278 | } | |
5279 | ||
5280 | return 0; | |
5281 | err: | |
54a91f3b | 5282 | io_finish_async(ctx); |
2b188cc1 JA |
5283 | mmdrop(ctx->sqo_mm); |
5284 | ctx->sqo_mm = NULL; | |
5285 | return ret; | |
5286 | } | |
5287 | ||
5288 | static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages) | |
5289 | { | |
5290 | atomic_long_sub(nr_pages, &user->locked_vm); | |
5291 | } | |
5292 | ||
5293 | static int io_account_mem(struct user_struct *user, unsigned long nr_pages) | |
5294 | { | |
5295 | unsigned long page_limit, cur_pages, new_pages; | |
5296 | ||
5297 | /* Don't allow more pages than we can safely lock */ | |
5298 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
5299 | ||
5300 | do { | |
5301 | cur_pages = atomic_long_read(&user->locked_vm); | |
5302 | new_pages = cur_pages + nr_pages; | |
5303 | if (new_pages > page_limit) | |
5304 | return -ENOMEM; | |
5305 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
5306 | new_pages) != cur_pages); | |
5307 | ||
5308 | return 0; | |
5309 | } | |
5310 | ||
5311 | static void io_mem_free(void *ptr) | |
5312 | { | |
52e04ef4 MR |
5313 | struct page *page; |
5314 | ||
5315 | if (!ptr) | |
5316 | return; | |
2b188cc1 | 5317 | |
52e04ef4 | 5318 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
5319 | if (put_page_testzero(page)) |
5320 | free_compound_page(page); | |
5321 | } | |
5322 | ||
5323 | static void *io_mem_alloc(size_t size) | |
5324 | { | |
5325 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
5326 | __GFP_NORETRY; | |
5327 | ||
5328 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
5329 | } | |
5330 | ||
75b28aff HV |
5331 | static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries, |
5332 | size_t *sq_offset) | |
5333 | { | |
5334 | struct io_rings *rings; | |
5335 | size_t off, sq_array_size; | |
5336 | ||
5337 | off = struct_size(rings, cqes, cq_entries); | |
5338 | if (off == SIZE_MAX) | |
5339 | return SIZE_MAX; | |
5340 | ||
5341 | #ifdef CONFIG_SMP | |
5342 | off = ALIGN(off, SMP_CACHE_BYTES); | |
5343 | if (off == 0) | |
5344 | return SIZE_MAX; | |
5345 | #endif | |
5346 | ||
5347 | sq_array_size = array_size(sizeof(u32), sq_entries); | |
5348 | if (sq_array_size == SIZE_MAX) | |
5349 | return SIZE_MAX; | |
5350 | ||
5351 | if (check_add_overflow(off, sq_array_size, &off)) | |
5352 | return SIZE_MAX; | |
5353 | ||
5354 | if (sq_offset) | |
5355 | *sq_offset = off; | |
5356 | ||
5357 | return off; | |
5358 | } | |
5359 | ||
2b188cc1 JA |
5360 | static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries) |
5361 | { | |
75b28aff | 5362 | size_t pages; |
2b188cc1 | 5363 | |
75b28aff HV |
5364 | pages = (size_t)1 << get_order( |
5365 | rings_size(sq_entries, cq_entries, NULL)); | |
5366 | pages += (size_t)1 << get_order( | |
5367 | array_size(sizeof(struct io_uring_sqe), sq_entries)); | |
2b188cc1 | 5368 | |
75b28aff | 5369 | return pages; |
2b188cc1 JA |
5370 | } |
5371 | ||
edafccee JA |
5372 | static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx) |
5373 | { | |
5374 | int i, j; | |
5375 | ||
5376 | if (!ctx->user_bufs) | |
5377 | return -ENXIO; | |
5378 | ||
5379 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
5380 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
5381 | ||
5382 | for (j = 0; j < imu->nr_bvecs; j++) | |
27c4d3a3 | 5383 | put_user_page(imu->bvec[j].bv_page); |
edafccee JA |
5384 | |
5385 | if (ctx->account_mem) | |
5386 | io_unaccount_mem(ctx->user, imu->nr_bvecs); | |
d4ef6475 | 5387 | kvfree(imu->bvec); |
edafccee JA |
5388 | imu->nr_bvecs = 0; |
5389 | } | |
5390 | ||
5391 | kfree(ctx->user_bufs); | |
5392 | ctx->user_bufs = NULL; | |
5393 | ctx->nr_user_bufs = 0; | |
5394 | return 0; | |
5395 | } | |
5396 | ||
5397 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
5398 | void __user *arg, unsigned index) | |
5399 | { | |
5400 | struct iovec __user *src; | |
5401 | ||
5402 | #ifdef CONFIG_COMPAT | |
5403 | if (ctx->compat) { | |
5404 | struct compat_iovec __user *ciovs; | |
5405 | struct compat_iovec ciov; | |
5406 | ||
5407 | ciovs = (struct compat_iovec __user *) arg; | |
5408 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
5409 | return -EFAULT; | |
5410 | ||
d55e5f5b | 5411 | dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base); |
edafccee JA |
5412 | dst->iov_len = ciov.iov_len; |
5413 | return 0; | |
5414 | } | |
5415 | #endif | |
5416 | src = (struct iovec __user *) arg; | |
5417 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
5418 | return -EFAULT; | |
5419 | return 0; | |
5420 | } | |
5421 | ||
5422 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, | |
5423 | unsigned nr_args) | |
5424 | { | |
5425 | struct vm_area_struct **vmas = NULL; | |
5426 | struct page **pages = NULL; | |
5427 | int i, j, got_pages = 0; | |
5428 | int ret = -EINVAL; | |
5429 | ||
5430 | if (ctx->user_bufs) | |
5431 | return -EBUSY; | |
5432 | if (!nr_args || nr_args > UIO_MAXIOV) | |
5433 | return -EINVAL; | |
5434 | ||
5435 | ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf), | |
5436 | GFP_KERNEL); | |
5437 | if (!ctx->user_bufs) | |
5438 | return -ENOMEM; | |
5439 | ||
5440 | for (i = 0; i < nr_args; i++) { | |
5441 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
5442 | unsigned long off, start, end, ubuf; | |
5443 | int pret, nr_pages; | |
5444 | struct iovec iov; | |
5445 | size_t size; | |
5446 | ||
5447 | ret = io_copy_iov(ctx, &iov, arg, i); | |
5448 | if (ret) | |
a278682d | 5449 | goto err; |
edafccee JA |
5450 | |
5451 | /* | |
5452 | * Don't impose further limits on the size and buffer | |
5453 | * constraints here, we'll -EINVAL later when IO is | |
5454 | * submitted if they are wrong. | |
5455 | */ | |
5456 | ret = -EFAULT; | |
5457 | if (!iov.iov_base || !iov.iov_len) | |
5458 | goto err; | |
5459 | ||
5460 | /* arbitrary limit, but we need something */ | |
5461 | if (iov.iov_len > SZ_1G) | |
5462 | goto err; | |
5463 | ||
5464 | ubuf = (unsigned long) iov.iov_base; | |
5465 | end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
5466 | start = ubuf >> PAGE_SHIFT; | |
5467 | nr_pages = end - start; | |
5468 | ||
5469 | if (ctx->account_mem) { | |
5470 | ret = io_account_mem(ctx->user, nr_pages); | |
5471 | if (ret) | |
5472 | goto err; | |
5473 | } | |
5474 | ||
5475 | ret = 0; | |
5476 | if (!pages || nr_pages > got_pages) { | |
5477 | kfree(vmas); | |
5478 | kfree(pages); | |
d4ef6475 | 5479 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), |
edafccee | 5480 | GFP_KERNEL); |
d4ef6475 | 5481 | vmas = kvmalloc_array(nr_pages, |
edafccee JA |
5482 | sizeof(struct vm_area_struct *), |
5483 | GFP_KERNEL); | |
5484 | if (!pages || !vmas) { | |
5485 | ret = -ENOMEM; | |
5486 | if (ctx->account_mem) | |
5487 | io_unaccount_mem(ctx->user, nr_pages); | |
5488 | goto err; | |
5489 | } | |
5490 | got_pages = nr_pages; | |
5491 | } | |
5492 | ||
d4ef6475 | 5493 | imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec), |
edafccee JA |
5494 | GFP_KERNEL); |
5495 | ret = -ENOMEM; | |
5496 | if (!imu->bvec) { | |
5497 | if (ctx->account_mem) | |
5498 | io_unaccount_mem(ctx->user, nr_pages); | |
5499 | goto err; | |
5500 | } | |
5501 | ||
5502 | ret = 0; | |
5503 | down_read(¤t->mm->mmap_sem); | |
932f4a63 IW |
5504 | pret = get_user_pages(ubuf, nr_pages, |
5505 | FOLL_WRITE | FOLL_LONGTERM, | |
5506 | pages, vmas); | |
edafccee JA |
5507 | if (pret == nr_pages) { |
5508 | /* don't support file backed memory */ | |
5509 | for (j = 0; j < nr_pages; j++) { | |
5510 | struct vm_area_struct *vma = vmas[j]; | |
5511 | ||
5512 | if (vma->vm_file && | |
5513 | !is_file_hugepages(vma->vm_file)) { | |
5514 | ret = -EOPNOTSUPP; | |
5515 | break; | |
5516 | } | |
5517 | } | |
5518 | } else { | |
5519 | ret = pret < 0 ? pret : -EFAULT; | |
5520 | } | |
5521 | up_read(¤t->mm->mmap_sem); | |
5522 | if (ret) { | |
5523 | /* | |
5524 | * if we did partial map, or found file backed vmas, | |
5525 | * release any pages we did get | |
5526 | */ | |
27c4d3a3 JH |
5527 | if (pret > 0) |
5528 | put_user_pages(pages, pret); | |
edafccee JA |
5529 | if (ctx->account_mem) |
5530 | io_unaccount_mem(ctx->user, nr_pages); | |
d4ef6475 | 5531 | kvfree(imu->bvec); |
edafccee JA |
5532 | goto err; |
5533 | } | |
5534 | ||
5535 | off = ubuf & ~PAGE_MASK; | |
5536 | size = iov.iov_len; | |
5537 | for (j = 0; j < nr_pages; j++) { | |
5538 | size_t vec_len; | |
5539 | ||
5540 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
5541 | imu->bvec[j].bv_page = pages[j]; | |
5542 | imu->bvec[j].bv_len = vec_len; | |
5543 | imu->bvec[j].bv_offset = off; | |
5544 | off = 0; | |
5545 | size -= vec_len; | |
5546 | } | |
5547 | /* store original address for later verification */ | |
5548 | imu->ubuf = ubuf; | |
5549 | imu->len = iov.iov_len; | |
5550 | imu->nr_bvecs = nr_pages; | |
5551 | ||
5552 | ctx->nr_user_bufs++; | |
5553 | } | |
d4ef6475 MR |
5554 | kvfree(pages); |
5555 | kvfree(vmas); | |
edafccee JA |
5556 | return 0; |
5557 | err: | |
d4ef6475 MR |
5558 | kvfree(pages); |
5559 | kvfree(vmas); | |
edafccee JA |
5560 | io_sqe_buffer_unregister(ctx); |
5561 | return ret; | |
5562 | } | |
5563 | ||
9b402849 JA |
5564 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
5565 | { | |
5566 | __s32 __user *fds = arg; | |
5567 | int fd; | |
5568 | ||
5569 | if (ctx->cq_ev_fd) | |
5570 | return -EBUSY; | |
5571 | ||
5572 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
5573 | return -EFAULT; | |
5574 | ||
5575 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
5576 | if (IS_ERR(ctx->cq_ev_fd)) { | |
5577 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
5578 | ctx->cq_ev_fd = NULL; | |
5579 | return ret; | |
5580 | } | |
5581 | ||
5582 | return 0; | |
5583 | } | |
5584 | ||
5585 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
5586 | { | |
5587 | if (ctx->cq_ev_fd) { | |
5588 | eventfd_ctx_put(ctx->cq_ev_fd); | |
5589 | ctx->cq_ev_fd = NULL; | |
5590 | return 0; | |
5591 | } | |
5592 | ||
5593 | return -ENXIO; | |
5594 | } | |
5595 | ||
2b188cc1 JA |
5596 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
5597 | { | |
6b06314c | 5598 | io_finish_async(ctx); |
2b188cc1 JA |
5599 | if (ctx->sqo_mm) |
5600 | mmdrop(ctx->sqo_mm); | |
def596e9 JA |
5601 | |
5602 | io_iopoll_reap_events(ctx); | |
edafccee | 5603 | io_sqe_buffer_unregister(ctx); |
6b06314c | 5604 | io_sqe_files_unregister(ctx); |
9b402849 | 5605 | io_eventfd_unregister(ctx); |
def596e9 | 5606 | |
2b188cc1 | 5607 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
5608 | if (ctx->ring_sock) { |
5609 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 5610 | sock_release(ctx->ring_sock); |
355e8d26 | 5611 | } |
2b188cc1 JA |
5612 | #endif |
5613 | ||
75b28aff | 5614 | io_mem_free(ctx->rings); |
2b188cc1 | 5615 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
5616 | |
5617 | percpu_ref_exit(&ctx->refs); | |
5618 | if (ctx->account_mem) | |
5619 | io_unaccount_mem(ctx->user, | |
5620 | ring_pages(ctx->sq_entries, ctx->cq_entries)); | |
5621 | free_uid(ctx->user); | |
181e448d | 5622 | put_cred(ctx->creds); |
206aefde | 5623 | kfree(ctx->completions); |
78076bb6 | 5624 | kfree(ctx->cancel_hash); |
0ddf92e8 | 5625 | kmem_cache_free(req_cachep, ctx->fallback_req); |
2b188cc1 JA |
5626 | kfree(ctx); |
5627 | } | |
5628 | ||
5629 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
5630 | { | |
5631 | struct io_ring_ctx *ctx = file->private_data; | |
5632 | __poll_t mask = 0; | |
5633 | ||
5634 | poll_wait(file, &ctx->cq_wait, wait); | |
4f7067c3 SB |
5635 | /* |
5636 | * synchronizes with barrier from wq_has_sleeper call in | |
5637 | * io_commit_cqring | |
5638 | */ | |
2b188cc1 | 5639 | smp_rmb(); |
75b28aff HV |
5640 | if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head != |
5641 | ctx->rings->sq_ring_entries) | |
2b188cc1 | 5642 | mask |= EPOLLOUT | EPOLLWRNORM; |
daa5de54 | 5643 | if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail) |
2b188cc1 JA |
5644 | mask |= EPOLLIN | EPOLLRDNORM; |
5645 | ||
5646 | return mask; | |
5647 | } | |
5648 | ||
5649 | static int io_uring_fasync(int fd, struct file *file, int on) | |
5650 | { | |
5651 | struct io_ring_ctx *ctx = file->private_data; | |
5652 | ||
5653 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
5654 | } | |
5655 | ||
5656 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) | |
5657 | { | |
5658 | mutex_lock(&ctx->uring_lock); | |
5659 | percpu_ref_kill(&ctx->refs); | |
5660 | mutex_unlock(&ctx->uring_lock); | |
5661 | ||
5262f567 | 5662 | io_kill_timeouts(ctx); |
221c5eb2 | 5663 | io_poll_remove_all(ctx); |
561fb04a JA |
5664 | |
5665 | if (ctx->io_wq) | |
5666 | io_wq_cancel_all(ctx->io_wq); | |
5667 | ||
def596e9 | 5668 | io_iopoll_reap_events(ctx); |
15dff286 JA |
5669 | /* if we failed setting up the ctx, we might not have any rings */ |
5670 | if (ctx->rings) | |
5671 | io_cqring_overflow_flush(ctx, true); | |
206aefde | 5672 | wait_for_completion(&ctx->completions[0]); |
2b188cc1 JA |
5673 | io_ring_ctx_free(ctx); |
5674 | } | |
5675 | ||
5676 | static int io_uring_release(struct inode *inode, struct file *file) | |
5677 | { | |
5678 | struct io_ring_ctx *ctx = file->private_data; | |
5679 | ||
5680 | file->private_data = NULL; | |
5681 | io_ring_ctx_wait_and_kill(ctx); | |
5682 | return 0; | |
5683 | } | |
5684 | ||
fcb323cc JA |
5685 | static void io_uring_cancel_files(struct io_ring_ctx *ctx, |
5686 | struct files_struct *files) | |
5687 | { | |
5688 | struct io_kiocb *req; | |
5689 | DEFINE_WAIT(wait); | |
5690 | ||
5691 | while (!list_empty_careful(&ctx->inflight_list)) { | |
768134d4 | 5692 | struct io_kiocb *cancel_req = NULL; |
fcb323cc JA |
5693 | |
5694 | spin_lock_irq(&ctx->inflight_lock); | |
5695 | list_for_each_entry(req, &ctx->inflight_list, inflight_entry) { | |
768134d4 JA |
5696 | if (req->work.files != files) |
5697 | continue; | |
5698 | /* req is being completed, ignore */ | |
5699 | if (!refcount_inc_not_zero(&req->refs)) | |
5700 | continue; | |
5701 | cancel_req = req; | |
5702 | break; | |
fcb323cc | 5703 | } |
768134d4 | 5704 | if (cancel_req) |
fcb323cc | 5705 | prepare_to_wait(&ctx->inflight_wait, &wait, |
768134d4 | 5706 | TASK_UNINTERRUPTIBLE); |
fcb323cc JA |
5707 | spin_unlock_irq(&ctx->inflight_lock); |
5708 | ||
768134d4 JA |
5709 | /* We need to keep going until we don't find a matching req */ |
5710 | if (!cancel_req) | |
fcb323cc | 5711 | break; |
2f6d9b9d BL |
5712 | |
5713 | io_wq_cancel_work(ctx->io_wq, &cancel_req->work); | |
5714 | io_put_req(cancel_req); | |
fcb323cc JA |
5715 | schedule(); |
5716 | } | |
768134d4 | 5717 | finish_wait(&ctx->inflight_wait, &wait); |
fcb323cc JA |
5718 | } |
5719 | ||
5720 | static int io_uring_flush(struct file *file, void *data) | |
5721 | { | |
5722 | struct io_ring_ctx *ctx = file->private_data; | |
5723 | ||
5724 | io_uring_cancel_files(ctx, data); | |
1d7bb1d5 JA |
5725 | if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) { |
5726 | io_cqring_overflow_flush(ctx, true); | |
fcb323cc | 5727 | io_wq_cancel_all(ctx->io_wq); |
1d7bb1d5 | 5728 | } |
fcb323cc JA |
5729 | return 0; |
5730 | } | |
5731 | ||
6c5c240e RP |
5732 | static void *io_uring_validate_mmap_request(struct file *file, |
5733 | loff_t pgoff, size_t sz) | |
2b188cc1 | 5734 | { |
2b188cc1 | 5735 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 5736 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
5737 | struct page *page; |
5738 | void *ptr; | |
5739 | ||
5740 | switch (offset) { | |
5741 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
5742 | case IORING_OFF_CQ_RING: |
5743 | ptr = ctx->rings; | |
2b188cc1 JA |
5744 | break; |
5745 | case IORING_OFF_SQES: | |
5746 | ptr = ctx->sq_sqes; | |
5747 | break; | |
2b188cc1 | 5748 | default: |
6c5c240e | 5749 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
5750 | } |
5751 | ||
5752 | page = virt_to_head_page(ptr); | |
a50b854e | 5753 | if (sz > page_size(page)) |
6c5c240e RP |
5754 | return ERR_PTR(-EINVAL); |
5755 | ||
5756 | return ptr; | |
5757 | } | |
5758 | ||
5759 | #ifdef CONFIG_MMU | |
5760 | ||
5761 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
5762 | { | |
5763 | size_t sz = vma->vm_end - vma->vm_start; | |
5764 | unsigned long pfn; | |
5765 | void *ptr; | |
5766 | ||
5767 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
5768 | if (IS_ERR(ptr)) | |
5769 | return PTR_ERR(ptr); | |
2b188cc1 JA |
5770 | |
5771 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
5772 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
5773 | } | |
5774 | ||
6c5c240e RP |
5775 | #else /* !CONFIG_MMU */ |
5776 | ||
5777 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
5778 | { | |
5779 | return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL; | |
5780 | } | |
5781 | ||
5782 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
5783 | { | |
5784 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
5785 | } | |
5786 | ||
5787 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
5788 | unsigned long addr, unsigned long len, | |
5789 | unsigned long pgoff, unsigned long flags) | |
5790 | { | |
5791 | void *ptr; | |
5792 | ||
5793 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
5794 | if (IS_ERR(ptr)) | |
5795 | return PTR_ERR(ptr); | |
5796 | ||
5797 | return (unsigned long) ptr; | |
5798 | } | |
5799 | ||
5800 | #endif /* !CONFIG_MMU */ | |
5801 | ||
2b188cc1 JA |
5802 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
5803 | u32, min_complete, u32, flags, const sigset_t __user *, sig, | |
5804 | size_t, sigsz) | |
5805 | { | |
5806 | struct io_ring_ctx *ctx; | |
5807 | long ret = -EBADF; | |
5808 | int submitted = 0; | |
5809 | struct fd f; | |
5810 | ||
6c271ce2 | 5811 | if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP)) |
2b188cc1 JA |
5812 | return -EINVAL; |
5813 | ||
5814 | f = fdget(fd); | |
5815 | if (!f.file) | |
5816 | return -EBADF; | |
5817 | ||
5818 | ret = -EOPNOTSUPP; | |
5819 | if (f.file->f_op != &io_uring_fops) | |
5820 | goto out_fput; | |
5821 | ||
5822 | ret = -ENXIO; | |
5823 | ctx = f.file->private_data; | |
5824 | if (!percpu_ref_tryget(&ctx->refs)) | |
5825 | goto out_fput; | |
5826 | ||
6c271ce2 JA |
5827 | /* |
5828 | * For SQ polling, the thread will do all submissions and completions. | |
5829 | * Just return the requested submit count, and wake the thread if | |
5830 | * we were asked to. | |
5831 | */ | |
b2a9eada | 5832 | ret = 0; |
6c271ce2 | 5833 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
c1edbf5f JA |
5834 | if (!list_empty_careful(&ctx->cq_overflow_list)) |
5835 | io_cqring_overflow_flush(ctx, false); | |
6c271ce2 JA |
5836 | if (flags & IORING_ENTER_SQ_WAKEUP) |
5837 | wake_up(&ctx->sqo_wait); | |
5838 | submitted = to_submit; | |
b2a9eada | 5839 | } else if (to_submit) { |
ae9428ca | 5840 | struct mm_struct *cur_mm; |
2b188cc1 | 5841 | |
44d28279 JA |
5842 | if (current->mm != ctx->sqo_mm || |
5843 | current_cred() != ctx->creds) { | |
5844 | ret = -EPERM; | |
5845 | goto out; | |
5846 | } | |
5847 | ||
ae9428ca | 5848 | to_submit = min(to_submit, ctx->sq_entries); |
2b188cc1 | 5849 | mutex_lock(&ctx->uring_lock); |
ae9428ca PB |
5850 | /* already have mm, so io_submit_sqes() won't try to grab it */ |
5851 | cur_mm = ctx->sqo_mm; | |
5852 | submitted = io_submit_sqes(ctx, to_submit, f.file, fd, | |
5853 | &cur_mm, false); | |
2b188cc1 | 5854 | mutex_unlock(&ctx->uring_lock); |
7c504e65 PB |
5855 | |
5856 | if (submitted != to_submit) | |
5857 | goto out; | |
2b188cc1 JA |
5858 | } |
5859 | if (flags & IORING_ENTER_GETEVENTS) { | |
def596e9 JA |
5860 | unsigned nr_events = 0; |
5861 | ||
2b188cc1 JA |
5862 | min_complete = min(min_complete, ctx->cq_entries); |
5863 | ||
def596e9 | 5864 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 | 5865 | ret = io_iopoll_check(ctx, &nr_events, min_complete); |
def596e9 JA |
5866 | } else { |
5867 | ret = io_cqring_wait(ctx, min_complete, sig, sigsz); | |
5868 | } | |
2b188cc1 JA |
5869 | } |
5870 | ||
7c504e65 | 5871 | out: |
6805b32e | 5872 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
5873 | out_fput: |
5874 | fdput(f); | |
5875 | return submitted ? submitted : ret; | |
5876 | } | |
5877 | ||
5878 | static const struct file_operations io_uring_fops = { | |
5879 | .release = io_uring_release, | |
fcb323cc | 5880 | .flush = io_uring_flush, |
2b188cc1 | 5881 | .mmap = io_uring_mmap, |
6c5c240e RP |
5882 | #ifndef CONFIG_MMU |
5883 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
5884 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
5885 | #endif | |
2b188cc1 JA |
5886 | .poll = io_uring_poll, |
5887 | .fasync = io_uring_fasync, | |
5888 | }; | |
5889 | ||
5890 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
5891 | struct io_uring_params *p) | |
5892 | { | |
75b28aff HV |
5893 | struct io_rings *rings; |
5894 | size_t size, sq_array_offset; | |
2b188cc1 | 5895 | |
75b28aff HV |
5896 | size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset); |
5897 | if (size == SIZE_MAX) | |
5898 | return -EOVERFLOW; | |
5899 | ||
5900 | rings = io_mem_alloc(size); | |
5901 | if (!rings) | |
2b188cc1 JA |
5902 | return -ENOMEM; |
5903 | ||
75b28aff HV |
5904 | ctx->rings = rings; |
5905 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
5906 | rings->sq_ring_mask = p->sq_entries - 1; | |
5907 | rings->cq_ring_mask = p->cq_entries - 1; | |
5908 | rings->sq_ring_entries = p->sq_entries; | |
5909 | rings->cq_ring_entries = p->cq_entries; | |
5910 | ctx->sq_mask = rings->sq_ring_mask; | |
5911 | ctx->cq_mask = rings->cq_ring_mask; | |
5912 | ctx->sq_entries = rings->sq_ring_entries; | |
5913 | ctx->cq_entries = rings->cq_ring_entries; | |
2b188cc1 JA |
5914 | |
5915 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 JA |
5916 | if (size == SIZE_MAX) { |
5917 | io_mem_free(ctx->rings); | |
5918 | ctx->rings = NULL; | |
2b188cc1 | 5919 | return -EOVERFLOW; |
eb065d30 | 5920 | } |
2b188cc1 JA |
5921 | |
5922 | ctx->sq_sqes = io_mem_alloc(size); | |
eb065d30 JA |
5923 | if (!ctx->sq_sqes) { |
5924 | io_mem_free(ctx->rings); | |
5925 | ctx->rings = NULL; | |
2b188cc1 | 5926 | return -ENOMEM; |
eb065d30 | 5927 | } |
2b188cc1 | 5928 | |
2b188cc1 JA |
5929 | return 0; |
5930 | } | |
5931 | ||
5932 | /* | |
5933 | * Allocate an anonymous fd, this is what constitutes the application | |
5934 | * visible backing of an io_uring instance. The application mmaps this | |
5935 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
5936 | * we have to tie this fd to a socket for file garbage collection purposes. | |
5937 | */ | |
5938 | static int io_uring_get_fd(struct io_ring_ctx *ctx) | |
5939 | { | |
5940 | struct file *file; | |
5941 | int ret; | |
5942 | ||
5943 | #if defined(CONFIG_UNIX) | |
5944 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, | |
5945 | &ctx->ring_sock); | |
5946 | if (ret) | |
5947 | return ret; | |
5948 | #endif | |
5949 | ||
5950 | ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
5951 | if (ret < 0) | |
5952 | goto err; | |
5953 | ||
5954 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, | |
5955 | O_RDWR | O_CLOEXEC); | |
5956 | if (IS_ERR(file)) { | |
5957 | put_unused_fd(ret); | |
5958 | ret = PTR_ERR(file); | |
5959 | goto err; | |
5960 | } | |
5961 | ||
5962 | #if defined(CONFIG_UNIX) | |
5963 | ctx->ring_sock->file = file; | |
5964 | #endif | |
5965 | fd_install(ret, file); | |
5966 | return ret; | |
5967 | err: | |
5968 | #if defined(CONFIG_UNIX) | |
5969 | sock_release(ctx->ring_sock); | |
5970 | ctx->ring_sock = NULL; | |
5971 | #endif | |
5972 | return ret; | |
5973 | } | |
5974 | ||
5975 | static int io_uring_create(unsigned entries, struct io_uring_params *p) | |
5976 | { | |
5977 | struct user_struct *user = NULL; | |
5978 | struct io_ring_ctx *ctx; | |
5979 | bool account_mem; | |
5980 | int ret; | |
5981 | ||
5982 | if (!entries || entries > IORING_MAX_ENTRIES) | |
5983 | return -EINVAL; | |
5984 | ||
5985 | /* | |
5986 | * Use twice as many entries for the CQ ring. It's possible for the | |
5987 | * application to drive a higher depth than the size of the SQ ring, | |
5988 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
5989 | * some flexibility in overcommitting a bit. If the application has |
5990 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
5991 | * of CQ ring entries manually. | |
2b188cc1 JA |
5992 | */ |
5993 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
5994 | if (p->flags & IORING_SETUP_CQSIZE) { |
5995 | /* | |
5996 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
5997 | * to a power-of-two, if it isn't already. We do NOT impose | |
5998 | * any cq vs sq ring sizing. | |
5999 | */ | |
6000 | if (p->cq_entries < p->sq_entries || p->cq_entries > IORING_MAX_CQ_ENTRIES) | |
6001 | return -EINVAL; | |
6002 | p->cq_entries = roundup_pow_of_two(p->cq_entries); | |
6003 | } else { | |
6004 | p->cq_entries = 2 * p->sq_entries; | |
6005 | } | |
2b188cc1 JA |
6006 | |
6007 | user = get_uid(current_user()); | |
6008 | account_mem = !capable(CAP_IPC_LOCK); | |
6009 | ||
6010 | if (account_mem) { | |
6011 | ret = io_account_mem(user, | |
6012 | ring_pages(p->sq_entries, p->cq_entries)); | |
6013 | if (ret) { | |
6014 | free_uid(user); | |
6015 | return ret; | |
6016 | } | |
6017 | } | |
6018 | ||
6019 | ctx = io_ring_ctx_alloc(p); | |
6020 | if (!ctx) { | |
6021 | if (account_mem) | |
6022 | io_unaccount_mem(user, ring_pages(p->sq_entries, | |
6023 | p->cq_entries)); | |
6024 | free_uid(user); | |
6025 | return -ENOMEM; | |
6026 | } | |
6027 | ctx->compat = in_compat_syscall(); | |
6028 | ctx->account_mem = account_mem; | |
6029 | ctx->user = user; | |
0b8c0ec7 | 6030 | ctx->creds = get_current_cred(); |
2b188cc1 JA |
6031 | |
6032 | ret = io_allocate_scq_urings(ctx, p); | |
6033 | if (ret) | |
6034 | goto err; | |
6035 | ||
6c271ce2 | 6036 | ret = io_sq_offload_start(ctx, p); |
2b188cc1 JA |
6037 | if (ret) |
6038 | goto err; | |
6039 | ||
2b188cc1 | 6040 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
6041 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
6042 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
6043 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
6044 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
6045 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
6046 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
6047 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
6048 | |
6049 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
6050 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
6051 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
6052 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
6053 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
6054 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
6055 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
ac90f249 | 6056 | |
044c1ab3 JA |
6057 | /* |
6058 | * Install ring fd as the very last thing, so we don't risk someone | |
6059 | * having closed it before we finish setup | |
6060 | */ | |
6061 | ret = io_uring_get_fd(ctx); | |
6062 | if (ret < 0) | |
6063 | goto err; | |
6064 | ||
da8c9690 JA |
6065 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
6066 | IORING_FEAT_SUBMIT_STABLE; | |
c826bd7a | 6067 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
6068 | return ret; |
6069 | err: | |
6070 | io_ring_ctx_wait_and_kill(ctx); | |
6071 | return ret; | |
6072 | } | |
6073 | ||
6074 | /* | |
6075 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
6076 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
6077 | * params structure passed in. | |
6078 | */ | |
6079 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
6080 | { | |
6081 | struct io_uring_params p; | |
6082 | long ret; | |
6083 | int i; | |
6084 | ||
6085 | if (copy_from_user(&p, params, sizeof(p))) | |
6086 | return -EFAULT; | |
6087 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
6088 | if (p.resv[i]) | |
6089 | return -EINVAL; | |
6090 | } | |
6091 | ||
6c271ce2 | 6092 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
33a107f0 | 6093 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE)) |
2b188cc1 JA |
6094 | return -EINVAL; |
6095 | ||
6096 | ret = io_uring_create(entries, &p); | |
6097 | if (ret < 0) | |
6098 | return ret; | |
6099 | ||
6100 | if (copy_to_user(params, &p, sizeof(p))) | |
6101 | return -EFAULT; | |
6102 | ||
6103 | return ret; | |
6104 | } | |
6105 | ||
6106 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
6107 | struct io_uring_params __user *, params) | |
6108 | { | |
6109 | return io_uring_setup(entries, params); | |
6110 | } | |
6111 | ||
edafccee JA |
6112 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
6113 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
6114 | __releases(ctx->uring_lock) |
6115 | __acquires(ctx->uring_lock) | |
edafccee JA |
6116 | { |
6117 | int ret; | |
6118 | ||
35fa71a0 JA |
6119 | /* |
6120 | * We're inside the ring mutex, if the ref is already dying, then | |
6121 | * someone else killed the ctx or is already going through | |
6122 | * io_uring_register(). | |
6123 | */ | |
6124 | if (percpu_ref_is_dying(&ctx->refs)) | |
6125 | return -ENXIO; | |
6126 | ||
05f3fb3c JA |
6127 | if (opcode != IORING_UNREGISTER_FILES && |
6128 | opcode != IORING_REGISTER_FILES_UPDATE) { | |
6129 | percpu_ref_kill(&ctx->refs); | |
b19062a5 | 6130 | |
05f3fb3c JA |
6131 | /* |
6132 | * Drop uring mutex before waiting for references to exit. If | |
6133 | * another thread is currently inside io_uring_enter() it might | |
6134 | * need to grab the uring_lock to make progress. If we hold it | |
6135 | * here across the drain wait, then we can deadlock. It's safe | |
6136 | * to drop the mutex here, since no new references will come in | |
6137 | * after we've killed the percpu ref. | |
6138 | */ | |
6139 | mutex_unlock(&ctx->uring_lock); | |
6140 | wait_for_completion(&ctx->completions[0]); | |
6141 | mutex_lock(&ctx->uring_lock); | |
6142 | } | |
edafccee JA |
6143 | |
6144 | switch (opcode) { | |
6145 | case IORING_REGISTER_BUFFERS: | |
6146 | ret = io_sqe_buffer_register(ctx, arg, nr_args); | |
6147 | break; | |
6148 | case IORING_UNREGISTER_BUFFERS: | |
6149 | ret = -EINVAL; | |
6150 | if (arg || nr_args) | |
6151 | break; | |
6152 | ret = io_sqe_buffer_unregister(ctx); | |
6153 | break; | |
6b06314c JA |
6154 | case IORING_REGISTER_FILES: |
6155 | ret = io_sqe_files_register(ctx, arg, nr_args); | |
6156 | break; | |
6157 | case IORING_UNREGISTER_FILES: | |
6158 | ret = -EINVAL; | |
6159 | if (arg || nr_args) | |
6160 | break; | |
6161 | ret = io_sqe_files_unregister(ctx); | |
6162 | break; | |
c3a31e60 JA |
6163 | case IORING_REGISTER_FILES_UPDATE: |
6164 | ret = io_sqe_files_update(ctx, arg, nr_args); | |
6165 | break; | |
9b402849 JA |
6166 | case IORING_REGISTER_EVENTFD: |
6167 | ret = -EINVAL; | |
6168 | if (nr_args != 1) | |
6169 | break; | |
6170 | ret = io_eventfd_register(ctx, arg); | |
6171 | break; | |
6172 | case IORING_UNREGISTER_EVENTFD: | |
6173 | ret = -EINVAL; | |
6174 | if (arg || nr_args) | |
6175 | break; | |
6176 | ret = io_eventfd_unregister(ctx); | |
6177 | break; | |
edafccee JA |
6178 | default: |
6179 | ret = -EINVAL; | |
6180 | break; | |
6181 | } | |
6182 | ||
05f3fb3c JA |
6183 | |
6184 | if (opcode != IORING_UNREGISTER_FILES && | |
6185 | opcode != IORING_REGISTER_FILES_UPDATE) { | |
6186 | /* bring the ctx back to life */ | |
6187 | reinit_completion(&ctx->completions[0]); | |
6188 | percpu_ref_reinit(&ctx->refs); | |
6189 | } | |
edafccee JA |
6190 | return ret; |
6191 | } | |
6192 | ||
6193 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
6194 | void __user *, arg, unsigned int, nr_args) | |
6195 | { | |
6196 | struct io_ring_ctx *ctx; | |
6197 | long ret = -EBADF; | |
6198 | struct fd f; | |
6199 | ||
6200 | f = fdget(fd); | |
6201 | if (!f.file) | |
6202 | return -EBADF; | |
6203 | ||
6204 | ret = -EOPNOTSUPP; | |
6205 | if (f.file->f_op != &io_uring_fops) | |
6206 | goto out_fput; | |
6207 | ||
6208 | ctx = f.file->private_data; | |
6209 | ||
6210 | mutex_lock(&ctx->uring_lock); | |
6211 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
6212 | mutex_unlock(&ctx->uring_lock); | |
c826bd7a DD |
6213 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, |
6214 | ctx->cq_ev_fd != NULL, ret); | |
edafccee JA |
6215 | out_fput: |
6216 | fdput(f); | |
6217 | return ret; | |
6218 | } | |
6219 | ||
2b188cc1 JA |
6220 | static int __init io_uring_init(void) |
6221 | { | |
d3656344 | 6222 | BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST); |
2b188cc1 JA |
6223 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC); |
6224 | return 0; | |
6225 | }; | |
6226 | __initcall(io_uring_init); |