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> |
2b188cc1 | 73 | |
c826bd7a DD |
74 | #define CREATE_TRACE_POINTS |
75 | #include <trace/events/io_uring.h> | |
76 | ||
2b188cc1 JA |
77 | #include <uapi/linux/io_uring.h> |
78 | ||
79 | #include "internal.h" | |
561fb04a | 80 | #include "io-wq.h" |
2b188cc1 | 81 | |
5277deaa | 82 | #define IORING_MAX_ENTRIES 32768 |
33a107f0 | 83 | #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES) |
65e19f54 JA |
84 | |
85 | /* | |
86 | * Shift of 9 is 512 entries, or exactly one page on 64-bit archs | |
87 | */ | |
88 | #define IORING_FILE_TABLE_SHIFT 9 | |
89 | #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT) | |
90 | #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1) | |
91 | #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE) | |
2b188cc1 JA |
92 | |
93 | struct io_uring { | |
94 | u32 head ____cacheline_aligned_in_smp; | |
95 | u32 tail ____cacheline_aligned_in_smp; | |
96 | }; | |
97 | ||
1e84b97b | 98 | /* |
75b28aff HV |
99 | * This data is shared with the application through the mmap at offsets |
100 | * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING. | |
1e84b97b SB |
101 | * |
102 | * The offsets to the member fields are published through struct | |
103 | * io_sqring_offsets when calling io_uring_setup. | |
104 | */ | |
75b28aff | 105 | struct io_rings { |
1e84b97b SB |
106 | /* |
107 | * Head and tail offsets into the ring; the offsets need to be | |
108 | * masked to get valid indices. | |
109 | * | |
75b28aff HV |
110 | * The kernel controls head of the sq ring and the tail of the cq ring, |
111 | * and the application controls tail of the sq ring and the head of the | |
112 | * cq ring. | |
1e84b97b | 113 | */ |
75b28aff | 114 | struct io_uring sq, cq; |
1e84b97b | 115 | /* |
75b28aff | 116 | * Bitmasks to apply to head and tail offsets (constant, equals |
1e84b97b SB |
117 | * ring_entries - 1) |
118 | */ | |
75b28aff HV |
119 | u32 sq_ring_mask, cq_ring_mask; |
120 | /* Ring sizes (constant, power of 2) */ | |
121 | u32 sq_ring_entries, cq_ring_entries; | |
1e84b97b SB |
122 | /* |
123 | * Number of invalid entries dropped by the kernel due to | |
124 | * invalid index stored in array | |
125 | * | |
126 | * Written by the kernel, shouldn't be modified by the | |
127 | * application (i.e. get number of "new events" by comparing to | |
128 | * cached value). | |
129 | * | |
130 | * After a new SQ head value was read by the application this | |
131 | * counter includes all submissions that were dropped reaching | |
132 | * the new SQ head (and possibly more). | |
133 | */ | |
75b28aff | 134 | u32 sq_dropped; |
1e84b97b SB |
135 | /* |
136 | * Runtime flags | |
137 | * | |
138 | * Written by the kernel, shouldn't be modified by the | |
139 | * application. | |
140 | * | |
141 | * The application needs a full memory barrier before checking | |
142 | * for IORING_SQ_NEED_WAKEUP after updating the sq tail. | |
143 | */ | |
75b28aff | 144 | u32 sq_flags; |
1e84b97b SB |
145 | /* |
146 | * Number of completion events lost because the queue was full; | |
147 | * this should be avoided by the application by making sure | |
0b4295b5 | 148 | * there are not more requests pending than there is space in |
1e84b97b SB |
149 | * the completion queue. |
150 | * | |
151 | * Written by the kernel, shouldn't be modified by the | |
152 | * application (i.e. get number of "new events" by comparing to | |
153 | * cached value). | |
154 | * | |
155 | * As completion events come in out of order this counter is not | |
156 | * ordered with any other data. | |
157 | */ | |
75b28aff | 158 | u32 cq_overflow; |
1e84b97b SB |
159 | /* |
160 | * Ring buffer of completion events. | |
161 | * | |
162 | * The kernel writes completion events fresh every time they are | |
163 | * produced, so the application is allowed to modify pending | |
164 | * entries. | |
165 | */ | |
75b28aff | 166 | struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp; |
2b188cc1 JA |
167 | }; |
168 | ||
edafccee JA |
169 | struct io_mapped_ubuf { |
170 | u64 ubuf; | |
171 | size_t len; | |
172 | struct bio_vec *bvec; | |
173 | unsigned int nr_bvecs; | |
174 | }; | |
175 | ||
65e19f54 JA |
176 | struct fixed_file_table { |
177 | struct file **files; | |
31b51510 JA |
178 | }; |
179 | ||
2b188cc1 JA |
180 | struct io_ring_ctx { |
181 | struct { | |
182 | struct percpu_ref refs; | |
183 | } ____cacheline_aligned_in_smp; | |
184 | ||
185 | struct { | |
186 | unsigned int flags; | |
187 | bool compat; | |
188 | bool account_mem; | |
1d7bb1d5 | 189 | bool cq_overflow_flushed; |
1b4a51b6 | 190 | bool drain_next; |
2b188cc1 | 191 | |
75b28aff HV |
192 | /* |
193 | * Ring buffer of indices into array of io_uring_sqe, which is | |
194 | * mmapped by the application using the IORING_OFF_SQES offset. | |
195 | * | |
196 | * This indirection could e.g. be used to assign fixed | |
197 | * io_uring_sqe entries to operations and only submit them to | |
198 | * the queue when needed. | |
199 | * | |
200 | * The kernel modifies neither the indices array nor the entries | |
201 | * array. | |
202 | */ | |
203 | u32 *sq_array; | |
2b188cc1 JA |
204 | unsigned cached_sq_head; |
205 | unsigned sq_entries; | |
206 | unsigned sq_mask; | |
6c271ce2 | 207 | unsigned sq_thread_idle; |
498ccd9e | 208 | unsigned cached_sq_dropped; |
206aefde | 209 | atomic_t cached_cq_overflow; |
2b188cc1 | 210 | struct io_uring_sqe *sq_sqes; |
de0617e4 JA |
211 | |
212 | struct list_head defer_list; | |
5262f567 | 213 | struct list_head timeout_list; |
1d7bb1d5 | 214 | struct list_head cq_overflow_list; |
fcb323cc JA |
215 | |
216 | wait_queue_head_t inflight_wait; | |
2b188cc1 JA |
217 | } ____cacheline_aligned_in_smp; |
218 | ||
206aefde JA |
219 | struct io_rings *rings; |
220 | ||
2b188cc1 | 221 | /* IO offload */ |
561fb04a | 222 | struct io_wq *io_wq; |
6c271ce2 | 223 | struct task_struct *sqo_thread; /* if using sq thread polling */ |
2b188cc1 | 224 | struct mm_struct *sqo_mm; |
6c271ce2 | 225 | wait_queue_head_t sqo_wait; |
75b28aff | 226 | |
6b06314c JA |
227 | /* |
228 | * If used, fixed file set. Writers must ensure that ->refs is dead, | |
229 | * readers must ensure that ->refs is alive as long as the file* is | |
230 | * used. Only updated through io_uring_register(2). | |
231 | */ | |
65e19f54 | 232 | struct fixed_file_table *file_table; |
6b06314c JA |
233 | unsigned nr_user_files; |
234 | ||
edafccee JA |
235 | /* if used, fixed mapped user buffers */ |
236 | unsigned nr_user_bufs; | |
237 | struct io_mapped_ubuf *user_bufs; | |
238 | ||
2b188cc1 JA |
239 | struct user_struct *user; |
240 | ||
0b8c0ec7 | 241 | const struct cred *creds; |
181e448d | 242 | |
206aefde JA |
243 | /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */ |
244 | struct completion *completions; | |
245 | ||
0ddf92e8 JA |
246 | /* if all else fails... */ |
247 | struct io_kiocb *fallback_req; | |
248 | ||
206aefde JA |
249 | #if defined(CONFIG_UNIX) |
250 | struct socket *ring_sock; | |
251 | #endif | |
252 | ||
253 | struct { | |
254 | unsigned cached_cq_tail; | |
255 | unsigned cq_entries; | |
256 | unsigned cq_mask; | |
257 | atomic_t cq_timeouts; | |
258 | struct wait_queue_head cq_wait; | |
259 | struct fasync_struct *cq_fasync; | |
260 | struct eventfd_ctx *cq_ev_fd; | |
261 | } ____cacheline_aligned_in_smp; | |
2b188cc1 JA |
262 | |
263 | struct { | |
264 | struct mutex uring_lock; | |
265 | wait_queue_head_t wait; | |
266 | } ____cacheline_aligned_in_smp; | |
267 | ||
268 | struct { | |
269 | spinlock_t completion_lock; | |
def596e9 JA |
270 | bool poll_multi_file; |
271 | /* | |
272 | * ->poll_list is protected by the ctx->uring_lock for | |
273 | * io_uring instances that don't use IORING_SETUP_SQPOLL. | |
274 | * For SQPOLL, only the single threaded io_sq_thread() will | |
275 | * manipulate the list, hence no extra locking is needed there. | |
276 | */ | |
277 | struct list_head poll_list; | |
78076bb6 JA |
278 | struct hlist_head *cancel_hash; |
279 | unsigned cancel_hash_bits; | |
31b51510 | 280 | |
fcb323cc JA |
281 | spinlock_t inflight_lock; |
282 | struct list_head inflight_list; | |
2b188cc1 | 283 | } ____cacheline_aligned_in_smp; |
2b188cc1 JA |
284 | }; |
285 | ||
09bb8394 JA |
286 | /* |
287 | * First field must be the file pointer in all the | |
288 | * iocb unions! See also 'struct kiocb' in <linux/fs.h> | |
289 | */ | |
221c5eb2 JA |
290 | struct io_poll_iocb { |
291 | struct file *file; | |
0969e783 JA |
292 | union { |
293 | struct wait_queue_head *head; | |
294 | u64 addr; | |
295 | }; | |
221c5eb2 | 296 | __poll_t events; |
8c838788 | 297 | bool done; |
221c5eb2 | 298 | bool canceled; |
392edb45 | 299 | struct wait_queue_entry wait; |
221c5eb2 JA |
300 | }; |
301 | ||
ad8a48ac JA |
302 | struct io_timeout_data { |
303 | struct io_kiocb *req; | |
304 | struct hrtimer timer; | |
305 | struct timespec64 ts; | |
306 | enum hrtimer_mode mode; | |
cc42e0ac | 307 | u32 seq_offset; |
ad8a48ac JA |
308 | }; |
309 | ||
8ed8d3c3 JA |
310 | struct io_accept { |
311 | struct file *file; | |
312 | struct sockaddr __user *addr; | |
313 | int __user *addr_len; | |
314 | int flags; | |
315 | }; | |
316 | ||
317 | struct io_sync { | |
318 | struct file *file; | |
319 | loff_t len; | |
320 | loff_t off; | |
321 | int flags; | |
322 | }; | |
323 | ||
fbf23849 JA |
324 | struct io_cancel { |
325 | struct file *file; | |
326 | u64 addr; | |
327 | }; | |
328 | ||
b29472ee JA |
329 | struct io_timeout { |
330 | struct file *file; | |
331 | u64 addr; | |
332 | int flags; | |
333 | }; | |
334 | ||
f499a021 JA |
335 | struct io_async_connect { |
336 | struct sockaddr_storage address; | |
337 | }; | |
338 | ||
03b1230c JA |
339 | struct io_async_msghdr { |
340 | struct iovec fast_iov[UIO_FASTIOV]; | |
341 | struct iovec *iov; | |
342 | struct sockaddr __user *uaddr; | |
343 | struct msghdr msg; | |
344 | }; | |
345 | ||
f67676d1 JA |
346 | struct io_async_rw { |
347 | struct iovec fast_iov[UIO_FASTIOV]; | |
348 | struct iovec *iov; | |
349 | ssize_t nr_segs; | |
350 | ssize_t size; | |
351 | }; | |
352 | ||
1a6b74fc JA |
353 | struct io_async_ctx { |
354 | struct io_uring_sqe sqe; | |
f67676d1 JA |
355 | union { |
356 | struct io_async_rw rw; | |
03b1230c | 357 | struct io_async_msghdr msg; |
f499a021 | 358 | struct io_async_connect connect; |
2d28390a | 359 | struct io_timeout_data timeout; |
f67676d1 | 360 | }; |
1a6b74fc JA |
361 | }; |
362 | ||
09bb8394 JA |
363 | /* |
364 | * NOTE! Each of the iocb union members has the file pointer | |
365 | * as the first entry in their struct definition. So you can | |
366 | * access the file pointer through any of the sub-structs, | |
367 | * or directly as just 'ki_filp' in this struct. | |
368 | */ | |
2b188cc1 | 369 | struct io_kiocb { |
221c5eb2 | 370 | union { |
09bb8394 | 371 | struct file *file; |
221c5eb2 JA |
372 | struct kiocb rw; |
373 | struct io_poll_iocb poll; | |
8ed8d3c3 JA |
374 | struct io_accept accept; |
375 | struct io_sync sync; | |
fbf23849 | 376 | struct io_cancel cancel; |
b29472ee | 377 | struct io_timeout timeout; |
221c5eb2 | 378 | }; |
2b188cc1 | 379 | |
cf6fd4bd | 380 | const struct io_uring_sqe *sqe; |
1a6b74fc | 381 | struct io_async_ctx *io; |
cf6fd4bd PB |
382 | struct file *ring_file; |
383 | int ring_fd; | |
384 | bool has_user; | |
385 | bool in_async; | |
386 | bool needs_fixed_file; | |
d625c6ee | 387 | u8 opcode; |
2b188cc1 JA |
388 | |
389 | struct io_ring_ctx *ctx; | |
eac406c6 JA |
390 | union { |
391 | struct list_head list; | |
78076bb6 | 392 | struct hlist_node hash_node; |
eac406c6 | 393 | }; |
9e645e11 | 394 | struct list_head link_list; |
2b188cc1 | 395 | unsigned int flags; |
c16361c1 | 396 | refcount_t refs; |
8449eeda | 397 | #define REQ_F_NOWAIT 1 /* must not punt to workers */ |
def596e9 | 398 | #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */ |
6b06314c | 399 | #define REQ_F_FIXED_FILE 4 /* ctx owns file */ |
4d7dd462 | 400 | #define REQ_F_LINK_NEXT 8 /* already grabbed next link */ |
e2033e33 SB |
401 | #define REQ_F_IO_DRAIN 16 /* drain existing IO first */ |
402 | #define REQ_F_IO_DRAINED 32 /* drain done */ | |
9e645e11 | 403 | #define REQ_F_LINK 64 /* linked sqes */ |
2665abfd | 404 | #define REQ_F_LINK_TIMEOUT 128 /* has linked timeout */ |
f7b76ac9 | 405 | #define REQ_F_FAIL_LINK 256 /* fail rest of links */ |
1b4a51b6 | 406 | #define REQ_F_DRAIN_LINK 512 /* link should be fully drained */ |
5262f567 | 407 | #define REQ_F_TIMEOUT 1024 /* timeout request */ |
491381ce JA |
408 | #define REQ_F_ISREG 2048 /* regular file */ |
409 | #define REQ_F_MUST_PUNT 4096 /* must be punted even for NONBLOCK */ | |
93bd25bb | 410 | #define REQ_F_TIMEOUT_NOSEQ 8192 /* no timeout sequence */ |
fb4b3d3f LT |
411 | #define REQ_F_INFLIGHT 16384 /* on inflight list */ |
412 | #define REQ_F_COMP_LOCKED 32768 /* completion under lock */ | |
4e88d6e7 | 413 | #define REQ_F_HARDLINK 65536 /* doesn't sever on completion < 0 */ |
8ed8d3c3 | 414 | #define REQ_F_PREPPED 131072 /* request already opcode prepared */ |
2b188cc1 | 415 | u64 user_data; |
9e645e11 | 416 | u32 result; |
de0617e4 | 417 | u32 sequence; |
2b188cc1 | 418 | |
fcb323cc JA |
419 | struct list_head inflight_entry; |
420 | ||
561fb04a | 421 | struct io_wq_work work; |
2b188cc1 JA |
422 | }; |
423 | ||
424 | #define IO_PLUG_THRESHOLD 2 | |
def596e9 | 425 | #define IO_IOPOLL_BATCH 8 |
2b188cc1 | 426 | |
9a56a232 JA |
427 | struct io_submit_state { |
428 | struct blk_plug plug; | |
429 | ||
2579f913 JA |
430 | /* |
431 | * io_kiocb alloc cache | |
432 | */ | |
433 | void *reqs[IO_IOPOLL_BATCH]; | |
434 | unsigned int free_reqs; | |
435 | unsigned int cur_req; | |
436 | ||
9a56a232 JA |
437 | /* |
438 | * File reference cache | |
439 | */ | |
440 | struct file *file; | |
441 | unsigned int fd; | |
442 | unsigned int has_refs; | |
443 | unsigned int used_refs; | |
444 | unsigned int ios_left; | |
445 | }; | |
446 | ||
561fb04a | 447 | static void io_wq_submit_work(struct io_wq_work **workptr); |
78e19bbe | 448 | static void io_cqring_fill_event(struct io_kiocb *req, long res); |
4fe2c963 | 449 | static void __io_free_req(struct io_kiocb *req); |
ec9c02ad | 450 | static void io_put_req(struct io_kiocb *req); |
78e19bbe | 451 | static void io_double_put_req(struct io_kiocb *req); |
978db57e | 452 | static void __io_double_put_req(struct io_kiocb *req); |
94ae5e77 JA |
453 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req); |
454 | static void io_queue_linked_timeout(struct io_kiocb *req); | |
de0617e4 | 455 | |
2b188cc1 JA |
456 | static struct kmem_cache *req_cachep; |
457 | ||
458 | static const struct file_operations io_uring_fops; | |
459 | ||
460 | struct sock *io_uring_get_socket(struct file *file) | |
461 | { | |
462 | #if defined(CONFIG_UNIX) | |
463 | if (file->f_op == &io_uring_fops) { | |
464 | struct io_ring_ctx *ctx = file->private_data; | |
465 | ||
466 | return ctx->ring_sock->sk; | |
467 | } | |
468 | #endif | |
469 | return NULL; | |
470 | } | |
471 | EXPORT_SYMBOL(io_uring_get_socket); | |
472 | ||
473 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) | |
474 | { | |
475 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
476 | ||
206aefde | 477 | complete(&ctx->completions[0]); |
2b188cc1 JA |
478 | } |
479 | ||
480 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) | |
481 | { | |
482 | struct io_ring_ctx *ctx; | |
78076bb6 | 483 | int hash_bits; |
2b188cc1 JA |
484 | |
485 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
486 | if (!ctx) | |
487 | return NULL; | |
488 | ||
0ddf92e8 JA |
489 | ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL); |
490 | if (!ctx->fallback_req) | |
491 | goto err; | |
492 | ||
206aefde JA |
493 | ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL); |
494 | if (!ctx->completions) | |
495 | goto err; | |
496 | ||
78076bb6 JA |
497 | /* |
498 | * Use 5 bits less than the max cq entries, that should give us around | |
499 | * 32 entries per hash list if totally full and uniformly spread. | |
500 | */ | |
501 | hash_bits = ilog2(p->cq_entries); | |
502 | hash_bits -= 5; | |
503 | if (hash_bits <= 0) | |
504 | hash_bits = 1; | |
505 | ctx->cancel_hash_bits = hash_bits; | |
506 | ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head), | |
507 | GFP_KERNEL); | |
508 | if (!ctx->cancel_hash) | |
509 | goto err; | |
510 | __hash_init(ctx->cancel_hash, 1U << hash_bits); | |
511 | ||
21482896 | 512 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
206aefde JA |
513 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) |
514 | goto err; | |
2b188cc1 JA |
515 | |
516 | ctx->flags = p->flags; | |
517 | init_waitqueue_head(&ctx->cq_wait); | |
1d7bb1d5 | 518 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
206aefde JA |
519 | init_completion(&ctx->completions[0]); |
520 | init_completion(&ctx->completions[1]); | |
2b188cc1 JA |
521 | mutex_init(&ctx->uring_lock); |
522 | init_waitqueue_head(&ctx->wait); | |
523 | spin_lock_init(&ctx->completion_lock); | |
def596e9 | 524 | INIT_LIST_HEAD(&ctx->poll_list); |
de0617e4 | 525 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 526 | INIT_LIST_HEAD(&ctx->timeout_list); |
fcb323cc JA |
527 | init_waitqueue_head(&ctx->inflight_wait); |
528 | spin_lock_init(&ctx->inflight_lock); | |
529 | INIT_LIST_HEAD(&ctx->inflight_list); | |
2b188cc1 | 530 | return ctx; |
206aefde | 531 | err: |
0ddf92e8 JA |
532 | if (ctx->fallback_req) |
533 | kmem_cache_free(req_cachep, ctx->fallback_req); | |
206aefde | 534 | kfree(ctx->completions); |
78076bb6 | 535 | kfree(ctx->cancel_hash); |
206aefde JA |
536 | kfree(ctx); |
537 | return NULL; | |
2b188cc1 JA |
538 | } |
539 | ||
9d858b21 | 540 | static inline bool __req_need_defer(struct io_kiocb *req) |
7adf4eaf | 541 | { |
a197f664 JL |
542 | struct io_ring_ctx *ctx = req->ctx; |
543 | ||
498ccd9e JA |
544 | return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped |
545 | + atomic_read(&ctx->cached_cq_overflow); | |
7adf4eaf JA |
546 | } |
547 | ||
9d858b21 | 548 | static inline bool req_need_defer(struct io_kiocb *req) |
de0617e4 | 549 | { |
9d858b21 BL |
550 | if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) == REQ_F_IO_DRAIN) |
551 | return __req_need_defer(req); | |
de0617e4 | 552 | |
9d858b21 | 553 | return false; |
de0617e4 JA |
554 | } |
555 | ||
7adf4eaf | 556 | static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx) |
de0617e4 JA |
557 | { |
558 | struct io_kiocb *req; | |
559 | ||
7adf4eaf | 560 | req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list); |
9d858b21 | 561 | if (req && !req_need_defer(req)) { |
de0617e4 JA |
562 | list_del_init(&req->list); |
563 | return req; | |
564 | } | |
565 | ||
566 | return NULL; | |
567 | } | |
568 | ||
5262f567 JA |
569 | static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx) |
570 | { | |
7adf4eaf JA |
571 | struct io_kiocb *req; |
572 | ||
573 | req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list); | |
93bd25bb JA |
574 | if (req) { |
575 | if (req->flags & REQ_F_TIMEOUT_NOSEQ) | |
576 | return NULL; | |
fb4b3d3f | 577 | if (!__req_need_defer(req)) { |
93bd25bb JA |
578 | list_del_init(&req->list); |
579 | return req; | |
580 | } | |
7adf4eaf JA |
581 | } |
582 | ||
583 | return NULL; | |
5262f567 JA |
584 | } |
585 | ||
de0617e4 | 586 | static void __io_commit_cqring(struct io_ring_ctx *ctx) |
2b188cc1 | 587 | { |
75b28aff | 588 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 589 | |
75b28aff | 590 | if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) { |
2b188cc1 | 591 | /* order cqe stores with ring update */ |
75b28aff | 592 | smp_store_release(&rings->cq.tail, ctx->cached_cq_tail); |
2b188cc1 | 593 | |
2b188cc1 JA |
594 | if (wq_has_sleeper(&ctx->cq_wait)) { |
595 | wake_up_interruptible(&ctx->cq_wait); | |
596 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
597 | } | |
598 | } | |
599 | } | |
600 | ||
d625c6ee | 601 | static inline bool io_req_needs_user(struct io_kiocb *req) |
18d9be1a | 602 | { |
d625c6ee JA |
603 | return !(req->opcode == IORING_OP_READ_FIXED || |
604 | req->opcode == IORING_OP_WRITE_FIXED); | |
561fb04a JA |
605 | } |
606 | ||
94ae5e77 JA |
607 | static inline bool io_prep_async_work(struct io_kiocb *req, |
608 | struct io_kiocb **link) | |
18d9be1a | 609 | { |
561fb04a | 610 | bool do_hashed = false; |
54a91f3b | 611 | |
cf6fd4bd | 612 | if (req->sqe) { |
d625c6ee | 613 | switch (req->opcode) { |
6cc47d1d JA |
614 | case IORING_OP_WRITEV: |
615 | case IORING_OP_WRITE_FIXED: | |
53108d47 JA |
616 | /* only regular files should be hashed for writes */ |
617 | if (req->flags & REQ_F_ISREG) | |
618 | do_hashed = true; | |
5f8fd2d3 JA |
619 | /* fall-through */ |
620 | case IORING_OP_READV: | |
621 | case IORING_OP_READ_FIXED: | |
622 | case IORING_OP_SENDMSG: | |
623 | case IORING_OP_RECVMSG: | |
624 | case IORING_OP_ACCEPT: | |
625 | case IORING_OP_POLL_ADD: | |
f8e85cf2 | 626 | case IORING_OP_CONNECT: |
5f8fd2d3 JA |
627 | /* |
628 | * We know REQ_F_ISREG is not set on some of these | |
629 | * opcodes, but this enables us to keep the check in | |
630 | * just one place. | |
631 | */ | |
632 | if (!(req->flags & REQ_F_ISREG)) | |
633 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
6cc47d1d JA |
634 | break; |
635 | } | |
d625c6ee | 636 | if (io_req_needs_user(req)) |
561fb04a | 637 | req->work.flags |= IO_WQ_WORK_NEEDS_USER; |
54a91f3b JA |
638 | } |
639 | ||
94ae5e77 | 640 | *link = io_prep_linked_timeout(req); |
561fb04a JA |
641 | return do_hashed; |
642 | } | |
643 | ||
a197f664 | 644 | static inline void io_queue_async_work(struct io_kiocb *req) |
561fb04a | 645 | { |
a197f664 | 646 | struct io_ring_ctx *ctx = req->ctx; |
94ae5e77 JA |
647 | struct io_kiocb *link; |
648 | bool do_hashed; | |
649 | ||
650 | do_hashed = io_prep_async_work(req, &link); | |
561fb04a JA |
651 | |
652 | trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work, | |
653 | req->flags); | |
654 | if (!do_hashed) { | |
655 | io_wq_enqueue(ctx->io_wq, &req->work); | |
656 | } else { | |
657 | io_wq_enqueue_hashed(ctx->io_wq, &req->work, | |
658 | file_inode(req->file)); | |
659 | } | |
94ae5e77 JA |
660 | |
661 | if (link) | |
662 | io_queue_linked_timeout(link); | |
18d9be1a JA |
663 | } |
664 | ||
5262f567 JA |
665 | static void io_kill_timeout(struct io_kiocb *req) |
666 | { | |
667 | int ret; | |
668 | ||
2d28390a | 669 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); |
5262f567 JA |
670 | if (ret != -1) { |
671 | atomic_inc(&req->ctx->cq_timeouts); | |
842f9612 | 672 | list_del_init(&req->list); |
78e19bbe | 673 | io_cqring_fill_event(req, 0); |
ec9c02ad | 674 | io_put_req(req); |
5262f567 JA |
675 | } |
676 | } | |
677 | ||
678 | static void io_kill_timeouts(struct io_ring_ctx *ctx) | |
679 | { | |
680 | struct io_kiocb *req, *tmp; | |
681 | ||
682 | spin_lock_irq(&ctx->completion_lock); | |
683 | list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list) | |
684 | io_kill_timeout(req); | |
685 | spin_unlock_irq(&ctx->completion_lock); | |
686 | } | |
687 | ||
de0617e4 JA |
688 | static void io_commit_cqring(struct io_ring_ctx *ctx) |
689 | { | |
690 | struct io_kiocb *req; | |
691 | ||
5262f567 JA |
692 | while ((req = io_get_timeout_req(ctx)) != NULL) |
693 | io_kill_timeout(req); | |
694 | ||
de0617e4 JA |
695 | __io_commit_cqring(ctx); |
696 | ||
697 | while ((req = io_get_deferred_req(ctx)) != NULL) { | |
698 | req->flags |= REQ_F_IO_DRAINED; | |
a197f664 | 699 | io_queue_async_work(req); |
de0617e4 JA |
700 | } |
701 | } | |
702 | ||
2b188cc1 JA |
703 | static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx) |
704 | { | |
75b28aff | 705 | struct io_rings *rings = ctx->rings; |
2b188cc1 JA |
706 | unsigned tail; |
707 | ||
708 | tail = ctx->cached_cq_tail; | |
115e12e5 SB |
709 | /* |
710 | * writes to the cq entry need to come after reading head; the | |
711 | * control dependency is enough as we're using WRITE_ONCE to | |
712 | * fill the cq entry | |
713 | */ | |
75b28aff | 714 | if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries) |
2b188cc1 JA |
715 | return NULL; |
716 | ||
717 | ctx->cached_cq_tail++; | |
75b28aff | 718 | return &rings->cqes[tail & ctx->cq_mask]; |
2b188cc1 JA |
719 | } |
720 | ||
1d7bb1d5 JA |
721 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
722 | { | |
723 | if (waitqueue_active(&ctx->wait)) | |
724 | wake_up(&ctx->wait); | |
725 | if (waitqueue_active(&ctx->sqo_wait)) | |
726 | wake_up(&ctx->sqo_wait); | |
727 | if (ctx->cq_ev_fd) | |
728 | eventfd_signal(ctx->cq_ev_fd, 1); | |
729 | } | |
730 | ||
c4a2ed72 JA |
731 | /* Returns true if there are no backlogged entries after the flush */ |
732 | static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force) | |
1d7bb1d5 JA |
733 | { |
734 | struct io_rings *rings = ctx->rings; | |
735 | struct io_uring_cqe *cqe; | |
736 | struct io_kiocb *req; | |
737 | unsigned long flags; | |
738 | LIST_HEAD(list); | |
739 | ||
740 | if (!force) { | |
741 | if (list_empty_careful(&ctx->cq_overflow_list)) | |
c4a2ed72 | 742 | return true; |
1d7bb1d5 JA |
743 | if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) == |
744 | rings->cq_ring_entries)) | |
c4a2ed72 | 745 | return false; |
1d7bb1d5 JA |
746 | } |
747 | ||
748 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
749 | ||
750 | /* if force is set, the ring is going away. always drop after that */ | |
751 | if (force) | |
752 | ctx->cq_overflow_flushed = true; | |
753 | ||
c4a2ed72 | 754 | cqe = NULL; |
1d7bb1d5 JA |
755 | while (!list_empty(&ctx->cq_overflow_list)) { |
756 | cqe = io_get_cqring(ctx); | |
757 | if (!cqe && !force) | |
758 | break; | |
759 | ||
760 | req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb, | |
761 | list); | |
762 | list_move(&req->list, &list); | |
763 | if (cqe) { | |
764 | WRITE_ONCE(cqe->user_data, req->user_data); | |
765 | WRITE_ONCE(cqe->res, req->result); | |
766 | WRITE_ONCE(cqe->flags, 0); | |
767 | } else { | |
768 | WRITE_ONCE(ctx->rings->cq_overflow, | |
769 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
770 | } | |
771 | } | |
772 | ||
773 | io_commit_cqring(ctx); | |
774 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
775 | io_cqring_ev_posted(ctx); | |
776 | ||
777 | while (!list_empty(&list)) { | |
778 | req = list_first_entry(&list, struct io_kiocb, list); | |
779 | list_del(&req->list); | |
ec9c02ad | 780 | io_put_req(req); |
1d7bb1d5 | 781 | } |
c4a2ed72 JA |
782 | |
783 | return cqe != NULL; | |
1d7bb1d5 JA |
784 | } |
785 | ||
78e19bbe | 786 | static void io_cqring_fill_event(struct io_kiocb *req, long res) |
2b188cc1 | 787 | { |
78e19bbe | 788 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
789 | struct io_uring_cqe *cqe; |
790 | ||
78e19bbe | 791 | trace_io_uring_complete(ctx, req->user_data, res); |
51c3ff62 | 792 | |
2b188cc1 JA |
793 | /* |
794 | * If we can't get a cq entry, userspace overflowed the | |
795 | * submission (by quite a lot). Increment the overflow count in | |
796 | * the ring. | |
797 | */ | |
798 | cqe = io_get_cqring(ctx); | |
1d7bb1d5 | 799 | if (likely(cqe)) { |
78e19bbe | 800 | WRITE_ONCE(cqe->user_data, req->user_data); |
2b188cc1 | 801 | WRITE_ONCE(cqe->res, res); |
c71ffb67 | 802 | WRITE_ONCE(cqe->flags, 0); |
1d7bb1d5 | 803 | } else if (ctx->cq_overflow_flushed) { |
498ccd9e JA |
804 | WRITE_ONCE(ctx->rings->cq_overflow, |
805 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
1d7bb1d5 JA |
806 | } else { |
807 | refcount_inc(&req->refs); | |
808 | req->result = res; | |
809 | list_add_tail(&req->list, &ctx->cq_overflow_list); | |
2b188cc1 JA |
810 | } |
811 | } | |
812 | ||
78e19bbe | 813 | static void io_cqring_add_event(struct io_kiocb *req, long res) |
2b188cc1 | 814 | { |
78e19bbe | 815 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
816 | unsigned long flags; |
817 | ||
818 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
78e19bbe | 819 | io_cqring_fill_event(req, res); |
2b188cc1 JA |
820 | io_commit_cqring(ctx); |
821 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
822 | ||
8c838788 | 823 | io_cqring_ev_posted(ctx); |
2b188cc1 JA |
824 | } |
825 | ||
0ddf92e8 JA |
826 | static inline bool io_is_fallback_req(struct io_kiocb *req) |
827 | { | |
828 | return req == (struct io_kiocb *) | |
829 | ((unsigned long) req->ctx->fallback_req & ~1UL); | |
830 | } | |
831 | ||
832 | static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx) | |
833 | { | |
834 | struct io_kiocb *req; | |
835 | ||
836 | req = ctx->fallback_req; | |
837 | if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req)) | |
838 | return req; | |
839 | ||
840 | return NULL; | |
841 | } | |
842 | ||
2579f913 JA |
843 | static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx, |
844 | struct io_submit_state *state) | |
2b188cc1 | 845 | { |
fd6fab2c | 846 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
2b188cc1 JA |
847 | struct io_kiocb *req; |
848 | ||
849 | if (!percpu_ref_tryget(&ctx->refs)) | |
850 | return NULL; | |
851 | ||
2579f913 | 852 | if (!state) { |
fd6fab2c | 853 | req = kmem_cache_alloc(req_cachep, gfp); |
2579f913 | 854 | if (unlikely(!req)) |
0ddf92e8 | 855 | goto fallback; |
2579f913 JA |
856 | } else if (!state->free_reqs) { |
857 | size_t sz; | |
858 | int ret; | |
859 | ||
860 | sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs)); | |
fd6fab2c JA |
861 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs); |
862 | ||
863 | /* | |
864 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
865 | * retry single alloc to be on the safe side. | |
866 | */ | |
867 | if (unlikely(ret <= 0)) { | |
868 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
869 | if (!state->reqs[0]) | |
0ddf92e8 | 870 | goto fallback; |
fd6fab2c JA |
871 | ret = 1; |
872 | } | |
2579f913 JA |
873 | state->free_reqs = ret - 1; |
874 | state->cur_req = 1; | |
875 | req = state->reqs[0]; | |
876 | } else { | |
877 | req = state->reqs[state->cur_req]; | |
878 | state->free_reqs--; | |
879 | state->cur_req++; | |
2b188cc1 JA |
880 | } |
881 | ||
0ddf92e8 | 882 | got_it: |
1a6b74fc | 883 | req->io = NULL; |
cf6fd4bd | 884 | req->ring_file = NULL; |
60c112b0 | 885 | req->file = NULL; |
2579f913 JA |
886 | req->ctx = ctx; |
887 | req->flags = 0; | |
e65ef56d JA |
888 | /* one is dropped after submission, the other at completion */ |
889 | refcount_set(&req->refs, 2); | |
9e645e11 | 890 | req->result = 0; |
561fb04a | 891 | INIT_IO_WORK(&req->work, io_wq_submit_work); |
2579f913 | 892 | return req; |
0ddf92e8 JA |
893 | fallback: |
894 | req = io_get_fallback_req(ctx); | |
895 | if (req) | |
896 | goto got_it; | |
6805b32e | 897 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
898 | return NULL; |
899 | } | |
900 | ||
def596e9 JA |
901 | static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr) |
902 | { | |
903 | if (*nr) { | |
904 | kmem_cache_free_bulk(req_cachep, *nr, reqs); | |
6805b32e | 905 | percpu_ref_put_many(&ctx->refs, *nr); |
def596e9 JA |
906 | *nr = 0; |
907 | } | |
908 | } | |
909 | ||
9e645e11 | 910 | static void __io_free_req(struct io_kiocb *req) |
2b188cc1 | 911 | { |
fcb323cc JA |
912 | struct io_ring_ctx *ctx = req->ctx; |
913 | ||
1a6b74fc JA |
914 | if (req->io) |
915 | kfree(req->io); | |
09bb8394 JA |
916 | if (req->file && !(req->flags & REQ_F_FIXED_FILE)) |
917 | fput(req->file); | |
fcb323cc JA |
918 | if (req->flags & REQ_F_INFLIGHT) { |
919 | unsigned long flags; | |
920 | ||
921 | spin_lock_irqsave(&ctx->inflight_lock, flags); | |
922 | list_del(&req->inflight_entry); | |
923 | if (waitqueue_active(&ctx->inflight_wait)) | |
924 | wake_up(&ctx->inflight_wait); | |
925 | spin_unlock_irqrestore(&ctx->inflight_lock, flags); | |
926 | } | |
927 | percpu_ref_put(&ctx->refs); | |
0ddf92e8 JA |
928 | if (likely(!io_is_fallback_req(req))) |
929 | kmem_cache_free(req_cachep, req); | |
930 | else | |
931 | clear_bit_unlock(0, (unsigned long *) ctx->fallback_req); | |
e65ef56d JA |
932 | } |
933 | ||
a197f664 | 934 | static bool io_link_cancel_timeout(struct io_kiocb *req) |
2665abfd | 935 | { |
a197f664 | 936 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd JA |
937 | int ret; |
938 | ||
2d28390a | 939 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); |
2665abfd | 940 | if (ret != -1) { |
78e19bbe | 941 | io_cqring_fill_event(req, -ECANCELED); |
2665abfd JA |
942 | io_commit_cqring(ctx); |
943 | req->flags &= ~REQ_F_LINK; | |
ec9c02ad | 944 | io_put_req(req); |
2665abfd JA |
945 | return true; |
946 | } | |
947 | ||
948 | return false; | |
e65ef56d JA |
949 | } |
950 | ||
ba816ad6 | 951 | static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr) |
9e645e11 | 952 | { |
2665abfd | 953 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd | 954 | bool wake_ev = false; |
9e645e11 | 955 | |
4d7dd462 JA |
956 | /* Already got next link */ |
957 | if (req->flags & REQ_F_LINK_NEXT) | |
958 | return; | |
959 | ||
9e645e11 JA |
960 | /* |
961 | * The list should never be empty when we are called here. But could | |
962 | * potentially happen if the chain is messed up, check to be on the | |
963 | * safe side. | |
964 | */ | |
4493233e PB |
965 | while (!list_empty(&req->link_list)) { |
966 | struct io_kiocb *nxt = list_first_entry(&req->link_list, | |
967 | struct io_kiocb, link_list); | |
94ae5e77 | 968 | |
4493233e PB |
969 | if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) && |
970 | (nxt->flags & REQ_F_TIMEOUT))) { | |
971 | list_del_init(&nxt->link_list); | |
94ae5e77 | 972 | wake_ev |= io_link_cancel_timeout(nxt); |
94ae5e77 JA |
973 | req->flags &= ~REQ_F_LINK_TIMEOUT; |
974 | continue; | |
975 | } | |
9e645e11 | 976 | |
4493233e PB |
977 | list_del_init(&req->link_list); |
978 | if (!list_empty(&nxt->link_list)) | |
979 | nxt->flags |= REQ_F_LINK; | |
b18fdf71 | 980 | *nxtptr = nxt; |
94ae5e77 | 981 | break; |
9e645e11 | 982 | } |
2665abfd | 983 | |
4d7dd462 | 984 | req->flags |= REQ_F_LINK_NEXT; |
2665abfd JA |
985 | if (wake_ev) |
986 | io_cqring_ev_posted(ctx); | |
9e645e11 JA |
987 | } |
988 | ||
989 | /* | |
990 | * Called if REQ_F_LINK is set, and we fail the head request | |
991 | */ | |
992 | static void io_fail_links(struct io_kiocb *req) | |
993 | { | |
2665abfd | 994 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd JA |
995 | unsigned long flags; |
996 | ||
997 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
9e645e11 JA |
998 | |
999 | while (!list_empty(&req->link_list)) { | |
4493233e PB |
1000 | struct io_kiocb *link = list_first_entry(&req->link_list, |
1001 | struct io_kiocb, link_list); | |
9e645e11 | 1002 | |
4493233e | 1003 | list_del_init(&link->link_list); |
c826bd7a | 1004 | trace_io_uring_fail_link(req, link); |
2665abfd JA |
1005 | |
1006 | if ((req->flags & REQ_F_LINK_TIMEOUT) && | |
d625c6ee | 1007 | link->opcode == IORING_OP_LINK_TIMEOUT) { |
a197f664 | 1008 | io_link_cancel_timeout(link); |
2665abfd | 1009 | } else { |
78e19bbe | 1010 | io_cqring_fill_event(link, -ECANCELED); |
978db57e | 1011 | __io_double_put_req(link); |
2665abfd | 1012 | } |
5d960724 | 1013 | req->flags &= ~REQ_F_LINK_TIMEOUT; |
9e645e11 | 1014 | } |
2665abfd JA |
1015 | |
1016 | io_commit_cqring(ctx); | |
1017 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1018 | io_cqring_ev_posted(ctx); | |
9e645e11 JA |
1019 | } |
1020 | ||
4d7dd462 | 1021 | static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt) |
9e645e11 | 1022 | { |
4d7dd462 | 1023 | if (likely(!(req->flags & REQ_F_LINK))) |
2665abfd | 1024 | return; |
2665abfd | 1025 | |
9e645e11 JA |
1026 | /* |
1027 | * If LINK is set, we have dependent requests in this chain. If we | |
1028 | * didn't fail this request, queue the first one up, moving any other | |
1029 | * dependencies to the next request. In case of failure, fail the rest | |
1030 | * of the chain. | |
1031 | */ | |
2665abfd JA |
1032 | if (req->flags & REQ_F_FAIL_LINK) { |
1033 | io_fail_links(req); | |
7c9e7f0f JA |
1034 | } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) == |
1035 | REQ_F_LINK_TIMEOUT) { | |
2665abfd JA |
1036 | struct io_ring_ctx *ctx = req->ctx; |
1037 | unsigned long flags; | |
1038 | ||
1039 | /* | |
1040 | * If this is a timeout link, we could be racing with the | |
1041 | * timeout timer. Grab the completion lock for this case to | |
7c9e7f0f | 1042 | * protect against that. |
2665abfd JA |
1043 | */ |
1044 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
1045 | io_req_link_next(req, nxt); | |
1046 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1047 | } else { | |
1048 | io_req_link_next(req, nxt); | |
9e645e11 | 1049 | } |
4d7dd462 | 1050 | } |
9e645e11 | 1051 | |
c69f8dbe JL |
1052 | static void io_free_req(struct io_kiocb *req) |
1053 | { | |
944e58bf PB |
1054 | struct io_kiocb *nxt = NULL; |
1055 | ||
1056 | io_req_find_next(req, &nxt); | |
70cf9f32 | 1057 | __io_free_req(req); |
944e58bf PB |
1058 | |
1059 | if (nxt) | |
1060 | io_queue_async_work(nxt); | |
c69f8dbe JL |
1061 | } |
1062 | ||
ba816ad6 JA |
1063 | /* |
1064 | * Drop reference to request, return next in chain (if there is one) if this | |
1065 | * was the last reference to this request. | |
1066 | */ | |
f9bd67f6 | 1067 | __attribute__((nonnull)) |
ec9c02ad | 1068 | static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr) |
e65ef56d | 1069 | { |
f9bd67f6 | 1070 | io_req_find_next(req, nxtptr); |
4d7dd462 | 1071 | |
e65ef56d | 1072 | if (refcount_dec_and_test(&req->refs)) |
4d7dd462 | 1073 | __io_free_req(req); |
2b188cc1 JA |
1074 | } |
1075 | ||
e65ef56d JA |
1076 | static void io_put_req(struct io_kiocb *req) |
1077 | { | |
1078 | if (refcount_dec_and_test(&req->refs)) | |
1079 | io_free_req(req); | |
2b188cc1 JA |
1080 | } |
1081 | ||
978db57e JA |
1082 | /* |
1083 | * Must only be used if we don't need to care about links, usually from | |
1084 | * within the completion handling itself. | |
1085 | */ | |
1086 | static void __io_double_put_req(struct io_kiocb *req) | |
78e19bbe JA |
1087 | { |
1088 | /* drop both submit and complete references */ | |
1089 | if (refcount_sub_and_test(2, &req->refs)) | |
1090 | __io_free_req(req); | |
1091 | } | |
1092 | ||
978db57e JA |
1093 | static void io_double_put_req(struct io_kiocb *req) |
1094 | { | |
1095 | /* drop both submit and complete references */ | |
1096 | if (refcount_sub_and_test(2, &req->refs)) | |
1097 | io_free_req(req); | |
1098 | } | |
1099 | ||
1d7bb1d5 | 1100 | static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush) |
a3a0e43f | 1101 | { |
84f97dc2 JA |
1102 | struct io_rings *rings = ctx->rings; |
1103 | ||
1d7bb1d5 JA |
1104 | /* |
1105 | * noflush == true is from the waitqueue handler, just ensure we wake | |
1106 | * up the task, and the next invocation will flush the entries. We | |
1107 | * cannot safely to it from here. | |
1108 | */ | |
1109 | if (noflush && !list_empty(&ctx->cq_overflow_list)) | |
1110 | return -1U; | |
1111 | ||
1112 | io_cqring_overflow_flush(ctx, false); | |
1113 | ||
a3a0e43f JA |
1114 | /* See comment at the top of this file */ |
1115 | smp_rmb(); | |
75b28aff | 1116 | return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head); |
a3a0e43f JA |
1117 | } |
1118 | ||
fb5ccc98 PB |
1119 | static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) |
1120 | { | |
1121 | struct io_rings *rings = ctx->rings; | |
1122 | ||
1123 | /* make sure SQ entry isn't read before tail */ | |
1124 | return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; | |
1125 | } | |
1126 | ||
def596e9 JA |
1127 | /* |
1128 | * Find and free completed poll iocbs | |
1129 | */ | |
1130 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1131 | struct list_head *done) | |
1132 | { | |
1133 | void *reqs[IO_IOPOLL_BATCH]; | |
1134 | struct io_kiocb *req; | |
09bb8394 | 1135 | int to_free; |
def596e9 | 1136 | |
09bb8394 | 1137 | to_free = 0; |
def596e9 JA |
1138 | while (!list_empty(done)) { |
1139 | req = list_first_entry(done, struct io_kiocb, list); | |
1140 | list_del(&req->list); | |
1141 | ||
78e19bbe | 1142 | io_cqring_fill_event(req, req->result); |
def596e9 JA |
1143 | (*nr_events)++; |
1144 | ||
09bb8394 JA |
1145 | if (refcount_dec_and_test(&req->refs)) { |
1146 | /* If we're not using fixed files, we have to pair the | |
1147 | * completion part with the file put. Use regular | |
1148 | * completions for those, only batch free for fixed | |
9e645e11 | 1149 | * file and non-linked commands. |
09bb8394 | 1150 | */ |
1a6b74fc JA |
1151 | if (((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) == |
1152 | REQ_F_FIXED_FILE) && !io_is_fallback_req(req) && | |
1153 | !req->io) { | |
09bb8394 JA |
1154 | reqs[to_free++] = req; |
1155 | if (to_free == ARRAY_SIZE(reqs)) | |
1156 | io_free_req_many(ctx, reqs, &to_free); | |
6b06314c | 1157 | } else { |
09bb8394 | 1158 | io_free_req(req); |
6b06314c | 1159 | } |
9a56a232 | 1160 | } |
def596e9 | 1161 | } |
def596e9 | 1162 | |
09bb8394 | 1163 | io_commit_cqring(ctx); |
def596e9 JA |
1164 | io_free_req_many(ctx, reqs, &to_free); |
1165 | } | |
1166 | ||
1167 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1168 | long min) | |
1169 | { | |
1170 | struct io_kiocb *req, *tmp; | |
1171 | LIST_HEAD(done); | |
1172 | bool spin; | |
1173 | int ret; | |
1174 | ||
1175 | /* | |
1176 | * Only spin for completions if we don't have multiple devices hanging | |
1177 | * off our complete list, and we're under the requested amount. | |
1178 | */ | |
1179 | spin = !ctx->poll_multi_file && *nr_events < min; | |
1180 | ||
1181 | ret = 0; | |
1182 | list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) { | |
1183 | struct kiocb *kiocb = &req->rw; | |
1184 | ||
1185 | /* | |
1186 | * Move completed entries to our local list. If we find a | |
1187 | * request that requires polling, break out and complete | |
1188 | * the done list first, if we have entries there. | |
1189 | */ | |
1190 | if (req->flags & REQ_F_IOPOLL_COMPLETED) { | |
1191 | list_move_tail(&req->list, &done); | |
1192 | continue; | |
1193 | } | |
1194 | if (!list_empty(&done)) | |
1195 | break; | |
1196 | ||
1197 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
1198 | if (ret < 0) | |
1199 | break; | |
1200 | ||
1201 | if (ret && spin) | |
1202 | spin = false; | |
1203 | ret = 0; | |
1204 | } | |
1205 | ||
1206 | if (!list_empty(&done)) | |
1207 | io_iopoll_complete(ctx, nr_events, &done); | |
1208 | ||
1209 | return ret; | |
1210 | } | |
1211 | ||
1212 | /* | |
d195a66e | 1213 | * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a |
def596e9 JA |
1214 | * non-spinning poll check - we'll still enter the driver poll loop, but only |
1215 | * as a non-spinning completion check. | |
1216 | */ | |
1217 | static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1218 | long min) | |
1219 | { | |
08f5439f | 1220 | while (!list_empty(&ctx->poll_list) && !need_resched()) { |
def596e9 JA |
1221 | int ret; |
1222 | ||
1223 | ret = io_do_iopoll(ctx, nr_events, min); | |
1224 | if (ret < 0) | |
1225 | return ret; | |
1226 | if (!min || *nr_events >= min) | |
1227 | return 0; | |
1228 | } | |
1229 | ||
1230 | return 1; | |
1231 | } | |
1232 | ||
1233 | /* | |
1234 | * We can't just wait for polled events to come to us, we have to actively | |
1235 | * find and complete them. | |
1236 | */ | |
1237 | static void io_iopoll_reap_events(struct io_ring_ctx *ctx) | |
1238 | { | |
1239 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
1240 | return; | |
1241 | ||
1242 | mutex_lock(&ctx->uring_lock); | |
1243 | while (!list_empty(&ctx->poll_list)) { | |
1244 | unsigned int nr_events = 0; | |
1245 | ||
1246 | io_iopoll_getevents(ctx, &nr_events, 1); | |
08f5439f JA |
1247 | |
1248 | /* | |
1249 | * Ensure we allow local-to-the-cpu processing to take place, | |
1250 | * in this case we need to ensure that we reap all events. | |
1251 | */ | |
1252 | cond_resched(); | |
def596e9 JA |
1253 | } |
1254 | mutex_unlock(&ctx->uring_lock); | |
1255 | } | |
1256 | ||
2b2ed975 JA |
1257 | static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, |
1258 | long min) | |
def596e9 | 1259 | { |
2b2ed975 | 1260 | int iters = 0, ret = 0; |
500f9fba | 1261 | |
def596e9 JA |
1262 | do { |
1263 | int tmin = 0; | |
1264 | ||
a3a0e43f JA |
1265 | /* |
1266 | * Don't enter poll loop if we already have events pending. | |
1267 | * If we do, we can potentially be spinning for commands that | |
1268 | * already triggered a CQE (eg in error). | |
1269 | */ | |
1d7bb1d5 | 1270 | if (io_cqring_events(ctx, false)) |
a3a0e43f JA |
1271 | break; |
1272 | ||
500f9fba JA |
1273 | /* |
1274 | * If a submit got punted to a workqueue, we can have the | |
1275 | * application entering polling for a command before it gets | |
1276 | * issued. That app will hold the uring_lock for the duration | |
1277 | * of the poll right here, so we need to take a breather every | |
1278 | * now and then to ensure that the issue has a chance to add | |
1279 | * the poll to the issued list. Otherwise we can spin here | |
1280 | * forever, while the workqueue is stuck trying to acquire the | |
1281 | * very same mutex. | |
1282 | */ | |
1283 | if (!(++iters & 7)) { | |
1284 | mutex_unlock(&ctx->uring_lock); | |
1285 | mutex_lock(&ctx->uring_lock); | |
1286 | } | |
1287 | ||
def596e9 JA |
1288 | if (*nr_events < min) |
1289 | tmin = min - *nr_events; | |
1290 | ||
1291 | ret = io_iopoll_getevents(ctx, nr_events, tmin); | |
1292 | if (ret <= 0) | |
1293 | break; | |
1294 | ret = 0; | |
1295 | } while (min && !*nr_events && !need_resched()); | |
1296 | ||
2b2ed975 JA |
1297 | return ret; |
1298 | } | |
1299 | ||
1300 | static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, | |
1301 | long min) | |
1302 | { | |
1303 | int ret; | |
1304 | ||
1305 | /* | |
1306 | * We disallow the app entering submit/complete with polling, but we | |
1307 | * still need to lock the ring to prevent racing with polled issue | |
1308 | * that got punted to a workqueue. | |
1309 | */ | |
1310 | mutex_lock(&ctx->uring_lock); | |
1311 | ret = __io_iopoll_check(ctx, nr_events, min); | |
500f9fba | 1312 | mutex_unlock(&ctx->uring_lock); |
def596e9 JA |
1313 | return ret; |
1314 | } | |
1315 | ||
491381ce | 1316 | static void kiocb_end_write(struct io_kiocb *req) |
2b188cc1 | 1317 | { |
491381ce JA |
1318 | /* |
1319 | * Tell lockdep we inherited freeze protection from submission | |
1320 | * thread. | |
1321 | */ | |
1322 | if (req->flags & REQ_F_ISREG) { | |
1323 | struct inode *inode = file_inode(req->file); | |
2b188cc1 | 1324 | |
491381ce | 1325 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); |
2b188cc1 | 1326 | } |
491381ce | 1327 | file_end_write(req->file); |
2b188cc1 JA |
1328 | } |
1329 | ||
4e88d6e7 JA |
1330 | static inline void req_set_fail_links(struct io_kiocb *req) |
1331 | { | |
1332 | if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK) | |
1333 | req->flags |= REQ_F_FAIL_LINK; | |
1334 | } | |
1335 | ||
ba816ad6 | 1336 | static void io_complete_rw_common(struct kiocb *kiocb, long res) |
2b188cc1 JA |
1337 | { |
1338 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
1339 | ||
491381ce JA |
1340 | if (kiocb->ki_flags & IOCB_WRITE) |
1341 | kiocb_end_write(req); | |
2b188cc1 | 1342 | |
4e88d6e7 JA |
1343 | if (res != req->result) |
1344 | req_set_fail_links(req); | |
78e19bbe | 1345 | io_cqring_add_event(req, res); |
ba816ad6 JA |
1346 | } |
1347 | ||
1348 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
1349 | { | |
1350 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
1351 | ||
1352 | io_complete_rw_common(kiocb, res); | |
e65ef56d | 1353 | io_put_req(req); |
2b188cc1 JA |
1354 | } |
1355 | ||
ba816ad6 JA |
1356 | static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res) |
1357 | { | |
1358 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
ec9c02ad | 1359 | struct io_kiocb *nxt = NULL; |
ba816ad6 JA |
1360 | |
1361 | io_complete_rw_common(kiocb, res); | |
ec9c02ad JL |
1362 | io_put_req_find_next(req, &nxt); |
1363 | ||
1364 | return nxt; | |
2b188cc1 JA |
1365 | } |
1366 | ||
def596e9 JA |
1367 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
1368 | { | |
1369 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
1370 | ||
491381ce JA |
1371 | if (kiocb->ki_flags & IOCB_WRITE) |
1372 | kiocb_end_write(req); | |
def596e9 | 1373 | |
4e88d6e7 JA |
1374 | if (res != req->result) |
1375 | req_set_fail_links(req); | |
9e645e11 | 1376 | req->result = res; |
def596e9 JA |
1377 | if (res != -EAGAIN) |
1378 | req->flags |= REQ_F_IOPOLL_COMPLETED; | |
1379 | } | |
1380 | ||
1381 | /* | |
1382 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1383 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
1384 | * find it from a io_iopoll_getevents() thread before the issuer is done | |
1385 | * accessing the kiocb cookie. | |
1386 | */ | |
1387 | static void io_iopoll_req_issued(struct io_kiocb *req) | |
1388 | { | |
1389 | struct io_ring_ctx *ctx = req->ctx; | |
1390 | ||
1391 | /* | |
1392 | * Track whether we have multiple files in our lists. This will impact | |
1393 | * how we do polling eventually, not spinning if we're on potentially | |
1394 | * different devices. | |
1395 | */ | |
1396 | if (list_empty(&ctx->poll_list)) { | |
1397 | ctx->poll_multi_file = false; | |
1398 | } else if (!ctx->poll_multi_file) { | |
1399 | struct io_kiocb *list_req; | |
1400 | ||
1401 | list_req = list_first_entry(&ctx->poll_list, struct io_kiocb, | |
1402 | list); | |
1403 | if (list_req->rw.ki_filp != req->rw.ki_filp) | |
1404 | ctx->poll_multi_file = true; | |
1405 | } | |
1406 | ||
1407 | /* | |
1408 | * For fast devices, IO may have already completed. If it has, add | |
1409 | * it to the front so we find it first. | |
1410 | */ | |
1411 | if (req->flags & REQ_F_IOPOLL_COMPLETED) | |
1412 | list_add(&req->list, &ctx->poll_list); | |
1413 | else | |
1414 | list_add_tail(&req->list, &ctx->poll_list); | |
1415 | } | |
1416 | ||
3d6770fb | 1417 | static void io_file_put(struct io_submit_state *state) |
9a56a232 | 1418 | { |
3d6770fb | 1419 | if (state->file) { |
9a56a232 JA |
1420 | int diff = state->has_refs - state->used_refs; |
1421 | ||
1422 | if (diff) | |
1423 | fput_many(state->file, diff); | |
1424 | state->file = NULL; | |
1425 | } | |
1426 | } | |
1427 | ||
1428 | /* | |
1429 | * Get as many references to a file as we have IOs left in this submission, | |
1430 | * assuming most submissions are for one file, or at least that each file | |
1431 | * has more than one submission. | |
1432 | */ | |
1433 | static struct file *io_file_get(struct io_submit_state *state, int fd) | |
1434 | { | |
1435 | if (!state) | |
1436 | return fget(fd); | |
1437 | ||
1438 | if (state->file) { | |
1439 | if (state->fd == fd) { | |
1440 | state->used_refs++; | |
1441 | state->ios_left--; | |
1442 | return state->file; | |
1443 | } | |
3d6770fb | 1444 | io_file_put(state); |
9a56a232 JA |
1445 | } |
1446 | state->file = fget_many(fd, state->ios_left); | |
1447 | if (!state->file) | |
1448 | return NULL; | |
1449 | ||
1450 | state->fd = fd; | |
1451 | state->has_refs = state->ios_left; | |
1452 | state->used_refs = 1; | |
1453 | state->ios_left--; | |
1454 | return state->file; | |
1455 | } | |
1456 | ||
2b188cc1 JA |
1457 | /* |
1458 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1459 | * any file. For now, just ensure that anything potentially problematic is done | |
1460 | * inline. | |
1461 | */ | |
1462 | static bool io_file_supports_async(struct file *file) | |
1463 | { | |
1464 | umode_t mode = file_inode(file)->i_mode; | |
1465 | ||
10d59345 | 1466 | if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode)) |
2b188cc1 JA |
1467 | return true; |
1468 | if (S_ISREG(mode) && file->f_op != &io_uring_fops) | |
1469 | return true; | |
1470 | ||
1471 | return false; | |
1472 | } | |
1473 | ||
267bc904 | 1474 | static int io_prep_rw(struct io_kiocb *req, bool force_nonblock) |
2b188cc1 | 1475 | { |
cf6fd4bd | 1476 | const struct io_uring_sqe *sqe = req->sqe; |
def596e9 | 1477 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 | 1478 | struct kiocb *kiocb = &req->rw; |
09bb8394 JA |
1479 | unsigned ioprio; |
1480 | int ret; | |
2b188cc1 | 1481 | |
09bb8394 JA |
1482 | if (!req->file) |
1483 | return -EBADF; | |
2b188cc1 | 1484 | |
491381ce JA |
1485 | if (S_ISREG(file_inode(req->file)->i_mode)) |
1486 | req->flags |= REQ_F_ISREG; | |
1487 | ||
2b188cc1 JA |
1488 | kiocb->ki_pos = READ_ONCE(sqe->off); |
1489 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); | |
1490 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); | |
1491 | ||
1492 | ioprio = READ_ONCE(sqe->ioprio); | |
1493 | if (ioprio) { | |
1494 | ret = ioprio_check_cap(ioprio); | |
1495 | if (ret) | |
09bb8394 | 1496 | return ret; |
2b188cc1 JA |
1497 | |
1498 | kiocb->ki_ioprio = ioprio; | |
1499 | } else | |
1500 | kiocb->ki_ioprio = get_current_ioprio(); | |
1501 | ||
1502 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
1503 | if (unlikely(ret)) | |
09bb8394 | 1504 | return ret; |
8449eeda SB |
1505 | |
1506 | /* don't allow async punt if RWF_NOWAIT was requested */ | |
491381ce JA |
1507 | if ((kiocb->ki_flags & IOCB_NOWAIT) || |
1508 | (req->file->f_flags & O_NONBLOCK)) | |
8449eeda SB |
1509 | req->flags |= REQ_F_NOWAIT; |
1510 | ||
1511 | if (force_nonblock) | |
2b188cc1 | 1512 | kiocb->ki_flags |= IOCB_NOWAIT; |
8449eeda | 1513 | |
def596e9 | 1514 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
1515 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
1516 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 1517 | return -EOPNOTSUPP; |
2b188cc1 | 1518 | |
def596e9 JA |
1519 | kiocb->ki_flags |= IOCB_HIPRI; |
1520 | kiocb->ki_complete = io_complete_rw_iopoll; | |
6873e0bd | 1521 | req->result = 0; |
def596e9 | 1522 | } else { |
09bb8394 JA |
1523 | if (kiocb->ki_flags & IOCB_HIPRI) |
1524 | return -EINVAL; | |
def596e9 JA |
1525 | kiocb->ki_complete = io_complete_rw; |
1526 | } | |
2b188cc1 | 1527 | return 0; |
2b188cc1 JA |
1528 | } |
1529 | ||
1530 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
1531 | { | |
1532 | switch (ret) { | |
1533 | case -EIOCBQUEUED: | |
1534 | break; | |
1535 | case -ERESTARTSYS: | |
1536 | case -ERESTARTNOINTR: | |
1537 | case -ERESTARTNOHAND: | |
1538 | case -ERESTART_RESTARTBLOCK: | |
1539 | /* | |
1540 | * We can't just restart the syscall, since previously | |
1541 | * submitted sqes may already be in progress. Just fail this | |
1542 | * IO with EINTR. | |
1543 | */ | |
1544 | ret = -EINTR; | |
1545 | /* fall through */ | |
1546 | default: | |
1547 | kiocb->ki_complete(kiocb, ret, 0); | |
1548 | } | |
1549 | } | |
1550 | ||
ba816ad6 JA |
1551 | static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt, |
1552 | bool in_async) | |
1553 | { | |
f9bd67f6 | 1554 | if (in_async && ret >= 0 && kiocb->ki_complete == io_complete_rw) |
ba816ad6 JA |
1555 | *nxt = __io_complete_rw(kiocb, ret); |
1556 | else | |
1557 | io_rw_done(kiocb, ret); | |
1558 | } | |
1559 | ||
7d009165 PB |
1560 | static ssize_t io_import_fixed(struct io_ring_ctx *ctx, int rw, |
1561 | const struct io_uring_sqe *sqe, | |
1562 | struct iov_iter *iter) | |
edafccee JA |
1563 | { |
1564 | size_t len = READ_ONCE(sqe->len); | |
1565 | struct io_mapped_ubuf *imu; | |
1566 | unsigned index, buf_index; | |
1567 | size_t offset; | |
1568 | u64 buf_addr; | |
1569 | ||
1570 | /* attempt to use fixed buffers without having provided iovecs */ | |
1571 | if (unlikely(!ctx->user_bufs)) | |
1572 | return -EFAULT; | |
1573 | ||
1574 | buf_index = READ_ONCE(sqe->buf_index); | |
1575 | if (unlikely(buf_index >= ctx->nr_user_bufs)) | |
1576 | return -EFAULT; | |
1577 | ||
1578 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); | |
1579 | imu = &ctx->user_bufs[index]; | |
1580 | buf_addr = READ_ONCE(sqe->addr); | |
1581 | ||
1582 | /* overflow */ | |
1583 | if (buf_addr + len < buf_addr) | |
1584 | return -EFAULT; | |
1585 | /* not inside the mapped region */ | |
1586 | if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len) | |
1587 | return -EFAULT; | |
1588 | ||
1589 | /* | |
1590 | * May not be a start of buffer, set size appropriately | |
1591 | * and advance us to the beginning. | |
1592 | */ | |
1593 | offset = buf_addr - imu->ubuf; | |
1594 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
bd11b3a3 JA |
1595 | |
1596 | if (offset) { | |
1597 | /* | |
1598 | * Don't use iov_iter_advance() here, as it's really slow for | |
1599 | * using the latter parts of a big fixed buffer - it iterates | |
1600 | * over each segment manually. We can cheat a bit here, because | |
1601 | * we know that: | |
1602 | * | |
1603 | * 1) it's a BVEC iter, we set it up | |
1604 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
1605 | * first and last bvec | |
1606 | * | |
1607 | * So just find our index, and adjust the iterator afterwards. | |
1608 | * If the offset is within the first bvec (or the whole first | |
1609 | * bvec, just use iov_iter_advance(). This makes it easier | |
1610 | * since we can just skip the first segment, which may not | |
1611 | * be PAGE_SIZE aligned. | |
1612 | */ | |
1613 | const struct bio_vec *bvec = imu->bvec; | |
1614 | ||
1615 | if (offset <= bvec->bv_len) { | |
1616 | iov_iter_advance(iter, offset); | |
1617 | } else { | |
1618 | unsigned long seg_skip; | |
1619 | ||
1620 | /* skip first vec */ | |
1621 | offset -= bvec->bv_len; | |
1622 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
1623 | ||
1624 | iter->bvec = bvec + seg_skip; | |
1625 | iter->nr_segs -= seg_skip; | |
99c79f66 | 1626 | iter->count -= bvec->bv_len + offset; |
bd11b3a3 | 1627 | iter->iov_offset = offset & ~PAGE_MASK; |
bd11b3a3 JA |
1628 | } |
1629 | } | |
1630 | ||
5e559561 | 1631 | return len; |
edafccee JA |
1632 | } |
1633 | ||
cf6fd4bd PB |
1634 | static ssize_t io_import_iovec(int rw, struct io_kiocb *req, |
1635 | struct iovec **iovec, struct iov_iter *iter) | |
2b188cc1 | 1636 | { |
cf6fd4bd | 1637 | const struct io_uring_sqe *sqe = req->sqe; |
2b188cc1 JA |
1638 | void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
1639 | size_t sqe_len = READ_ONCE(sqe->len); | |
edafccee JA |
1640 | u8 opcode; |
1641 | ||
1642 | /* | |
1643 | * We're reading ->opcode for the second time, but the first read | |
1644 | * doesn't care whether it's _FIXED or not, so it doesn't matter | |
1645 | * whether ->opcode changes concurrently. The first read does care | |
1646 | * about whether it is a READ or a WRITE, so we don't trust this read | |
1647 | * for that purpose and instead let the caller pass in the read/write | |
1648 | * flag. | |
1649 | */ | |
d625c6ee | 1650 | opcode = req->opcode; |
7d009165 | 1651 | if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) { |
edafccee | 1652 | *iovec = NULL; |
7d009165 | 1653 | return io_import_fixed(req->ctx, rw, sqe, iter); |
edafccee | 1654 | } |
2b188cc1 | 1655 | |
f67676d1 JA |
1656 | if (req->io) { |
1657 | struct io_async_rw *iorw = &req->io->rw; | |
1658 | ||
1659 | *iovec = iorw->iov; | |
1660 | iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size); | |
1661 | if (iorw->iov == iorw->fast_iov) | |
1662 | *iovec = NULL; | |
1663 | return iorw->size; | |
1664 | } | |
1665 | ||
cf6fd4bd | 1666 | if (!req->has_user) |
2b188cc1 JA |
1667 | return -EFAULT; |
1668 | ||
1669 | #ifdef CONFIG_COMPAT | |
cf6fd4bd | 1670 | if (req->ctx->compat) |
2b188cc1 JA |
1671 | return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV, |
1672 | iovec, iter); | |
1673 | #endif | |
1674 | ||
1675 | return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter); | |
1676 | } | |
1677 | ||
31b51510 | 1678 | /* |
32960613 JA |
1679 | * For files that don't have ->read_iter() and ->write_iter(), handle them |
1680 | * by looping over ->read() or ->write() manually. | |
31b51510 | 1681 | */ |
32960613 JA |
1682 | static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb, |
1683 | struct iov_iter *iter) | |
1684 | { | |
1685 | ssize_t ret = 0; | |
1686 | ||
1687 | /* | |
1688 | * Don't support polled IO through this interface, and we can't | |
1689 | * support non-blocking either. For the latter, this just causes | |
1690 | * the kiocb to be handled from an async context. | |
1691 | */ | |
1692 | if (kiocb->ki_flags & IOCB_HIPRI) | |
1693 | return -EOPNOTSUPP; | |
1694 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
1695 | return -EAGAIN; | |
1696 | ||
1697 | while (iov_iter_count(iter)) { | |
311ae9e1 | 1698 | struct iovec iovec; |
32960613 JA |
1699 | ssize_t nr; |
1700 | ||
311ae9e1 PB |
1701 | if (!iov_iter_is_bvec(iter)) { |
1702 | iovec = iov_iter_iovec(iter); | |
1703 | } else { | |
1704 | /* fixed buffers import bvec */ | |
1705 | iovec.iov_base = kmap(iter->bvec->bv_page) | |
1706 | + iter->iov_offset; | |
1707 | iovec.iov_len = min(iter->count, | |
1708 | iter->bvec->bv_len - iter->iov_offset); | |
1709 | } | |
1710 | ||
32960613 JA |
1711 | if (rw == READ) { |
1712 | nr = file->f_op->read(file, iovec.iov_base, | |
1713 | iovec.iov_len, &kiocb->ki_pos); | |
1714 | } else { | |
1715 | nr = file->f_op->write(file, iovec.iov_base, | |
1716 | iovec.iov_len, &kiocb->ki_pos); | |
1717 | } | |
1718 | ||
311ae9e1 PB |
1719 | if (iov_iter_is_bvec(iter)) |
1720 | kunmap(iter->bvec->bv_page); | |
1721 | ||
32960613 JA |
1722 | if (nr < 0) { |
1723 | if (!ret) | |
1724 | ret = nr; | |
1725 | break; | |
1726 | } | |
1727 | ret += nr; | |
1728 | if (nr != iovec.iov_len) | |
1729 | break; | |
1730 | iov_iter_advance(iter, nr); | |
1731 | } | |
1732 | ||
1733 | return ret; | |
1734 | } | |
1735 | ||
b7bb4f7d | 1736 | static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size, |
f67676d1 JA |
1737 | struct iovec *iovec, struct iovec *fast_iov, |
1738 | struct iov_iter *iter) | |
1739 | { | |
1740 | req->io->rw.nr_segs = iter->nr_segs; | |
1741 | req->io->rw.size = io_size; | |
1742 | req->io->rw.iov = iovec; | |
1743 | if (!req->io->rw.iov) { | |
1744 | req->io->rw.iov = req->io->rw.fast_iov; | |
1745 | memcpy(req->io->rw.iov, fast_iov, | |
1746 | sizeof(struct iovec) * iter->nr_segs); | |
1747 | } | |
1748 | } | |
1749 | ||
b7bb4f7d | 1750 | static int io_alloc_async_ctx(struct io_kiocb *req) |
f67676d1 JA |
1751 | { |
1752 | req->io = kmalloc(sizeof(*req->io), GFP_KERNEL); | |
1753 | if (req->io) { | |
f67676d1 JA |
1754 | memcpy(&req->io->sqe, req->sqe, sizeof(req->io->sqe)); |
1755 | req->sqe = &req->io->sqe; | |
1756 | return 0; | |
1757 | } | |
1758 | ||
b7bb4f7d JA |
1759 | return 1; |
1760 | } | |
1761 | ||
1762 | static void io_rw_async(struct io_wq_work **workptr) | |
1763 | { | |
1764 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
1765 | struct iovec *iov = NULL; | |
1766 | ||
1767 | if (req->io->rw.iov != req->io->rw.fast_iov) | |
1768 | iov = req->io->rw.iov; | |
1769 | io_wq_submit_work(workptr); | |
1770 | kfree(iov); | |
1771 | } | |
1772 | ||
1773 | static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size, | |
1774 | struct iovec *iovec, struct iovec *fast_iov, | |
1775 | struct iov_iter *iter) | |
1776 | { | |
1777 | if (!req->io && io_alloc_async_ctx(req)) | |
1778 | return -ENOMEM; | |
1779 | ||
1780 | io_req_map_rw(req, io_size, iovec, fast_iov, iter); | |
1781 | req->work.func = io_rw_async; | |
1782 | return 0; | |
f67676d1 JA |
1783 | } |
1784 | ||
1785 | static int io_read_prep(struct io_kiocb *req, struct iovec **iovec, | |
1786 | struct iov_iter *iter, bool force_nonblock) | |
1787 | { | |
1788 | ssize_t ret; | |
1789 | ||
1790 | ret = io_prep_rw(req, force_nonblock); | |
1791 | if (ret) | |
1792 | return ret; | |
1793 | ||
1794 | if (unlikely(!(req->file->f_mode & FMODE_READ))) | |
1795 | return -EBADF; | |
1796 | ||
1797 | return io_import_iovec(READ, req, iovec, iter); | |
1798 | } | |
1799 | ||
267bc904 | 1800 | static int io_read(struct io_kiocb *req, struct io_kiocb **nxt, |
8358e3a8 | 1801 | bool force_nonblock) |
2b188cc1 JA |
1802 | { |
1803 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1804 | struct kiocb *kiocb = &req->rw; | |
1805 | struct iov_iter iter; | |
1806 | struct file *file; | |
31b51510 | 1807 | size_t iov_count; |
f67676d1 | 1808 | ssize_t io_size, ret; |
2b188cc1 | 1809 | |
f67676d1 JA |
1810 | if (!req->io) { |
1811 | ret = io_read_prep(req, &iovec, &iter, force_nonblock); | |
1812 | if (ret < 0) | |
1813 | return ret; | |
1814 | } else { | |
1815 | ret = io_import_iovec(READ, req, &iovec, &iter); | |
1816 | if (ret < 0) | |
1817 | return ret; | |
1818 | } | |
2b188cc1 | 1819 | |
f67676d1 JA |
1820 | file = req->file; |
1821 | io_size = ret; | |
9e645e11 | 1822 | if (req->flags & REQ_F_LINK) |
f67676d1 JA |
1823 | req->result = io_size; |
1824 | ||
1825 | /* | |
1826 | * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so | |
1827 | * we know to async punt it even if it was opened O_NONBLOCK | |
1828 | */ | |
1829 | if (force_nonblock && !io_file_supports_async(file)) { | |
1830 | req->flags |= REQ_F_MUST_PUNT; | |
1831 | goto copy_iov; | |
1832 | } | |
9e645e11 | 1833 | |
31b51510 JA |
1834 | iov_count = iov_iter_count(&iter); |
1835 | ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 JA |
1836 | if (!ret) { |
1837 | ssize_t ret2; | |
1838 | ||
32960613 JA |
1839 | if (file->f_op->read_iter) |
1840 | ret2 = call_read_iter(file, kiocb, &iter); | |
1841 | else | |
1842 | ret2 = loop_rw_iter(READ, file, kiocb, &iter); | |
1843 | ||
9d93a3f5 JA |
1844 | /* |
1845 | * In case of a short read, punt to async. This can happen | |
1846 | * if we have data partially cached. Alternatively we can | |
1847 | * return the short read, in which case the application will | |
1848 | * need to issue another SQE and wait for it. That SQE will | |
1849 | * need async punt anyway, so it's more efficient to do it | |
1850 | * here. | |
1851 | */ | |
491381ce JA |
1852 | if (force_nonblock && !(req->flags & REQ_F_NOWAIT) && |
1853 | (req->flags & REQ_F_ISREG) && | |
f67676d1 | 1854 | ret2 > 0 && ret2 < io_size) |
9d93a3f5 JA |
1855 | ret2 = -EAGAIN; |
1856 | /* Catch -EAGAIN return for forced non-blocking submission */ | |
f67676d1 | 1857 | if (!force_nonblock || ret2 != -EAGAIN) { |
cf6fd4bd | 1858 | kiocb_done(kiocb, ret2, nxt, req->in_async); |
f67676d1 JA |
1859 | } else { |
1860 | copy_iov: | |
b7bb4f7d | 1861 | ret = io_setup_async_rw(req, io_size, iovec, |
f67676d1 JA |
1862 | inline_vecs, &iter); |
1863 | if (ret) | |
1864 | goto out_free; | |
1865 | return -EAGAIN; | |
1866 | } | |
2b188cc1 | 1867 | } |
f67676d1 | 1868 | out_free: |
b7bb4f7d JA |
1869 | if (!io_wq_current_is_worker()) |
1870 | kfree(iovec); | |
2b188cc1 JA |
1871 | return ret; |
1872 | } | |
1873 | ||
f67676d1 JA |
1874 | static int io_write_prep(struct io_kiocb *req, struct iovec **iovec, |
1875 | struct iov_iter *iter, bool force_nonblock) | |
1876 | { | |
1877 | ssize_t ret; | |
1878 | ||
1879 | ret = io_prep_rw(req, force_nonblock); | |
1880 | if (ret) | |
1881 | return ret; | |
1882 | ||
1883 | if (unlikely(!(req->file->f_mode & FMODE_WRITE))) | |
1884 | return -EBADF; | |
1885 | ||
1886 | return io_import_iovec(WRITE, req, iovec, iter); | |
1887 | } | |
1888 | ||
267bc904 | 1889 | static int io_write(struct io_kiocb *req, struct io_kiocb **nxt, |
8358e3a8 | 1890 | bool force_nonblock) |
2b188cc1 JA |
1891 | { |
1892 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1893 | struct kiocb *kiocb = &req->rw; | |
1894 | struct iov_iter iter; | |
1895 | struct file *file; | |
31b51510 | 1896 | size_t iov_count; |
f67676d1 | 1897 | ssize_t ret, io_size; |
2b188cc1 | 1898 | |
f67676d1 JA |
1899 | if (!req->io) { |
1900 | ret = io_write_prep(req, &iovec, &iter, force_nonblock); | |
1901 | if (ret < 0) | |
1902 | return ret; | |
1903 | } else { | |
1904 | ret = io_import_iovec(WRITE, req, &iovec, &iter); | |
1905 | if (ret < 0) | |
1906 | return ret; | |
1907 | } | |
2b188cc1 | 1908 | |
2b188cc1 | 1909 | file = kiocb->ki_filp; |
f67676d1 | 1910 | io_size = ret; |
9e645e11 | 1911 | if (req->flags & REQ_F_LINK) |
f67676d1 | 1912 | req->result = io_size; |
9e645e11 | 1913 | |
f67676d1 JA |
1914 | /* |
1915 | * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so | |
1916 | * we know to async punt it even if it was opened O_NONBLOCK | |
1917 | */ | |
1918 | if (force_nonblock && !io_file_supports_async(req->file)) { | |
1919 | req->flags |= REQ_F_MUST_PUNT; | |
1920 | goto copy_iov; | |
1921 | } | |
31b51510 | 1922 | |
10d59345 JA |
1923 | /* file path doesn't support NOWAIT for non-direct_IO */ |
1924 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) && | |
1925 | (req->flags & REQ_F_ISREG)) | |
f67676d1 | 1926 | goto copy_iov; |
31b51510 | 1927 | |
f67676d1 | 1928 | iov_count = iov_iter_count(&iter); |
31b51510 | 1929 | ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count); |
2b188cc1 | 1930 | if (!ret) { |
9bf7933f RP |
1931 | ssize_t ret2; |
1932 | ||
2b188cc1 JA |
1933 | /* |
1934 | * Open-code file_start_write here to grab freeze protection, | |
1935 | * which will be released by another thread in | |
1936 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
1937 | * released so that it doesn't complain about the held lock when | |
1938 | * we return to userspace. | |
1939 | */ | |
491381ce | 1940 | if (req->flags & REQ_F_ISREG) { |
2b188cc1 JA |
1941 | __sb_start_write(file_inode(file)->i_sb, |
1942 | SB_FREEZE_WRITE, true); | |
1943 | __sb_writers_release(file_inode(file)->i_sb, | |
1944 | SB_FREEZE_WRITE); | |
1945 | } | |
1946 | kiocb->ki_flags |= IOCB_WRITE; | |
9bf7933f | 1947 | |
32960613 JA |
1948 | if (file->f_op->write_iter) |
1949 | ret2 = call_write_iter(file, kiocb, &iter); | |
1950 | else | |
1951 | ret2 = loop_rw_iter(WRITE, file, kiocb, &iter); | |
f67676d1 | 1952 | if (!force_nonblock || ret2 != -EAGAIN) { |
cf6fd4bd | 1953 | kiocb_done(kiocb, ret2, nxt, req->in_async); |
f67676d1 JA |
1954 | } else { |
1955 | copy_iov: | |
b7bb4f7d | 1956 | ret = io_setup_async_rw(req, io_size, iovec, |
f67676d1 JA |
1957 | inline_vecs, &iter); |
1958 | if (ret) | |
1959 | goto out_free; | |
1960 | return -EAGAIN; | |
1961 | } | |
2b188cc1 | 1962 | } |
31b51510 | 1963 | out_free: |
b7bb4f7d JA |
1964 | if (!io_wq_current_is_worker()) |
1965 | kfree(iovec); | |
2b188cc1 JA |
1966 | return ret; |
1967 | } | |
1968 | ||
1969 | /* | |
1970 | * IORING_OP_NOP just posts a completion event, nothing else. | |
1971 | */ | |
78e19bbe | 1972 | static int io_nop(struct io_kiocb *req) |
2b188cc1 JA |
1973 | { |
1974 | struct io_ring_ctx *ctx = req->ctx; | |
2b188cc1 | 1975 | |
def596e9 JA |
1976 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
1977 | return -EINVAL; | |
1978 | ||
78e19bbe | 1979 | io_cqring_add_event(req, 0); |
e65ef56d | 1980 | io_put_req(req); |
2b188cc1 JA |
1981 | return 0; |
1982 | } | |
1983 | ||
fc4df999 | 1984 | static int io_prep_fsync(struct io_kiocb *req) |
c992fe29 | 1985 | { |
fc4df999 | 1986 | const struct io_uring_sqe *sqe = req->sqe; |
6b06314c | 1987 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 1988 | |
8ed8d3c3 JA |
1989 | if (req->flags & REQ_F_PREPPED) |
1990 | return 0; | |
09bb8394 JA |
1991 | if (!req->file) |
1992 | return -EBADF; | |
c992fe29 | 1993 | |
6b06314c | 1994 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 1995 | return -EINVAL; |
edafccee | 1996 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
1997 | return -EINVAL; |
1998 | ||
8ed8d3c3 JA |
1999 | req->sync.flags = READ_ONCE(sqe->fsync_flags); |
2000 | if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC)) | |
2001 | return -EINVAL; | |
2002 | ||
2003 | req->sync.off = READ_ONCE(sqe->off); | |
2004 | req->sync.len = READ_ONCE(sqe->len); | |
2005 | req->flags |= REQ_F_PREPPED; | |
c992fe29 CH |
2006 | return 0; |
2007 | } | |
2008 | ||
8ed8d3c3 JA |
2009 | static bool io_req_cancelled(struct io_kiocb *req) |
2010 | { | |
2011 | if (req->work.flags & IO_WQ_WORK_CANCEL) { | |
2012 | req_set_fail_links(req); | |
2013 | io_cqring_add_event(req, -ECANCELED); | |
2014 | io_put_req(req); | |
2015 | return true; | |
2016 | } | |
2017 | ||
2018 | return false; | |
2019 | } | |
2020 | ||
2021 | static void io_fsync_finish(struct io_wq_work **workptr) | |
2022 | { | |
2023 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
2024 | loff_t end = req->sync.off + req->sync.len; | |
2025 | struct io_kiocb *nxt = NULL; | |
2026 | int ret; | |
2027 | ||
2028 | if (io_req_cancelled(req)) | |
2029 | return; | |
2030 | ||
2031 | ret = vfs_fsync_range(req->rw.ki_filp, req->sync.off, | |
2032 | end > 0 ? end : LLONG_MAX, | |
2033 | req->sync.flags & IORING_FSYNC_DATASYNC); | |
2034 | if (ret < 0) | |
2035 | req_set_fail_links(req); | |
2036 | io_cqring_add_event(req, ret); | |
2037 | io_put_req_find_next(req, &nxt); | |
2038 | if (nxt) | |
2039 | *workptr = &nxt->work; | |
2040 | } | |
2041 | ||
fc4df999 JA |
2042 | static int io_fsync(struct io_kiocb *req, struct io_kiocb **nxt, |
2043 | bool force_nonblock) | |
c992fe29 | 2044 | { |
8ed8d3c3 | 2045 | struct io_wq_work *work, *old_work; |
c992fe29 CH |
2046 | int ret; |
2047 | ||
fc4df999 | 2048 | ret = io_prep_fsync(req); |
c992fe29 CH |
2049 | if (ret) |
2050 | return ret; | |
2051 | ||
2052 | /* fsync always requires a blocking context */ | |
8ed8d3c3 JA |
2053 | if (force_nonblock) { |
2054 | io_put_req(req); | |
2055 | req->work.func = io_fsync_finish; | |
c992fe29 | 2056 | return -EAGAIN; |
8ed8d3c3 | 2057 | } |
c992fe29 | 2058 | |
8ed8d3c3 JA |
2059 | work = old_work = &req->work; |
2060 | io_fsync_finish(&work); | |
2061 | if (work && work != old_work) | |
2062 | *nxt = container_of(work, struct io_kiocb, work); | |
c992fe29 CH |
2063 | return 0; |
2064 | } | |
2065 | ||
fc4df999 | 2066 | static int io_prep_sfr(struct io_kiocb *req) |
5d17b4a4 | 2067 | { |
fc4df999 | 2068 | const struct io_uring_sqe *sqe = req->sqe; |
5d17b4a4 | 2069 | struct io_ring_ctx *ctx = req->ctx; |
5d17b4a4 | 2070 | |
8ed8d3c3 JA |
2071 | if (req->flags & REQ_F_PREPPED) |
2072 | return 0; | |
5d17b4a4 JA |
2073 | if (!req->file) |
2074 | return -EBADF; | |
5d17b4a4 JA |
2075 | |
2076 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
2077 | return -EINVAL; | |
2078 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) | |
2079 | return -EINVAL; | |
2080 | ||
8ed8d3c3 JA |
2081 | req->sync.off = READ_ONCE(sqe->off); |
2082 | req->sync.len = READ_ONCE(sqe->len); | |
2083 | req->sync.flags = READ_ONCE(sqe->sync_range_flags); | |
2084 | req->flags |= REQ_F_PREPPED; | |
2085 | return 0; | |
2086 | } | |
2087 | ||
2088 | static void io_sync_file_range_finish(struct io_wq_work **workptr) | |
2089 | { | |
2090 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
2091 | struct io_kiocb *nxt = NULL; | |
2092 | int ret; | |
2093 | ||
2094 | if (io_req_cancelled(req)) | |
2095 | return; | |
2096 | ||
2097 | ret = sync_file_range(req->rw.ki_filp, req->sync.off, req->sync.len, | |
2098 | req->sync.flags); | |
2099 | if (ret < 0) | |
2100 | req_set_fail_links(req); | |
2101 | io_cqring_add_event(req, ret); | |
2102 | io_put_req_find_next(req, &nxt); | |
2103 | if (nxt) | |
2104 | *workptr = &nxt->work; | |
5d17b4a4 JA |
2105 | } |
2106 | ||
fc4df999 | 2107 | static int io_sync_file_range(struct io_kiocb *req, struct io_kiocb **nxt, |
5d17b4a4 JA |
2108 | bool force_nonblock) |
2109 | { | |
8ed8d3c3 | 2110 | struct io_wq_work *work, *old_work; |
5d17b4a4 JA |
2111 | int ret; |
2112 | ||
fc4df999 | 2113 | ret = io_prep_sfr(req); |
5d17b4a4 JA |
2114 | if (ret) |
2115 | return ret; | |
2116 | ||
2117 | /* sync_file_range always requires a blocking context */ | |
8ed8d3c3 JA |
2118 | if (force_nonblock) { |
2119 | io_put_req(req); | |
2120 | req->work.func = io_sync_file_range_finish; | |
5d17b4a4 | 2121 | return -EAGAIN; |
8ed8d3c3 | 2122 | } |
5d17b4a4 | 2123 | |
8ed8d3c3 JA |
2124 | work = old_work = &req->work; |
2125 | io_sync_file_range_finish(&work); | |
2126 | if (work && work != old_work) | |
2127 | *nxt = container_of(work, struct io_kiocb, work); | |
5d17b4a4 JA |
2128 | return 0; |
2129 | } | |
2130 | ||
b7bb4f7d JA |
2131 | #if defined(CONFIG_NET) |
2132 | static void io_sendrecv_async(struct io_wq_work **workptr) | |
2133 | { | |
2134 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
2135 | struct iovec *iov = NULL; | |
2136 | ||
2137 | if (req->io->rw.iov != req->io->rw.fast_iov) | |
2138 | iov = req->io->msg.iov; | |
2139 | io_wq_submit_work(workptr); | |
2140 | kfree(iov); | |
2141 | } | |
2142 | #endif | |
2143 | ||
03b1230c JA |
2144 | static int io_sendmsg_prep(struct io_kiocb *req, struct io_async_ctx *io) |
2145 | { | |
0fa03c62 | 2146 | #if defined(CONFIG_NET) |
03b1230c JA |
2147 | const struct io_uring_sqe *sqe = req->sqe; |
2148 | struct user_msghdr __user *msg; | |
2149 | unsigned flags; | |
2150 | ||
2151 | flags = READ_ONCE(sqe->msg_flags); | |
2152 | msg = (struct user_msghdr __user *)(unsigned long) READ_ONCE(sqe->addr); | |
d9688565 | 2153 | io->msg.iov = io->msg.fast_iov; |
03b1230c JA |
2154 | return sendmsg_copy_msghdr(&io->msg.msg, msg, flags, &io->msg.iov); |
2155 | #else | |
2156 | return 0; | |
2157 | #endif | |
2158 | } | |
2159 | ||
fc4df999 JA |
2160 | static int io_sendmsg(struct io_kiocb *req, struct io_kiocb **nxt, |
2161 | bool force_nonblock) | |
aa1fa28f | 2162 | { |
03b1230c | 2163 | #if defined(CONFIG_NET) |
fc4df999 | 2164 | const struct io_uring_sqe *sqe = req->sqe; |
0b416c3e | 2165 | struct io_async_msghdr *kmsg = NULL; |
0fa03c62 JA |
2166 | struct socket *sock; |
2167 | int ret; | |
2168 | ||
2169 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
2170 | return -EINVAL; | |
2171 | ||
2172 | sock = sock_from_file(req->file, &ret); | |
2173 | if (sock) { | |
b7bb4f7d | 2174 | struct io_async_ctx io; |
03b1230c | 2175 | struct sockaddr_storage addr; |
0fa03c62 JA |
2176 | unsigned flags; |
2177 | ||
2178 | flags = READ_ONCE(sqe->msg_flags); | |
2179 | if (flags & MSG_DONTWAIT) | |
2180 | req->flags |= REQ_F_NOWAIT; | |
2181 | else if (force_nonblock) | |
2182 | flags |= MSG_DONTWAIT; | |
2183 | ||
03b1230c | 2184 | if (req->io) { |
0b416c3e JA |
2185 | kmsg = &req->io->msg; |
2186 | kmsg->msg.msg_name = &addr; | |
2187 | /* if iov is set, it's allocated already */ | |
2188 | if (!kmsg->iov) | |
2189 | kmsg->iov = kmsg->fast_iov; | |
2190 | kmsg->msg.msg_iter.iov = kmsg->iov; | |
03b1230c | 2191 | } else { |
0b416c3e JA |
2192 | kmsg = &io.msg; |
2193 | kmsg->msg.msg_name = &addr; | |
03b1230c JA |
2194 | ret = io_sendmsg_prep(req, &io); |
2195 | if (ret) | |
2196 | goto out; | |
2197 | } | |
0fa03c62 | 2198 | |
0b416c3e | 2199 | ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags); |
03b1230c | 2200 | if (force_nonblock && ret == -EAGAIN) { |
b7bb4f7d JA |
2201 | if (req->io) |
2202 | return -EAGAIN; | |
2203 | if (io_alloc_async_ctx(req)) | |
2204 | return -ENOMEM; | |
2205 | memcpy(&req->io->msg, &io.msg, sizeof(io.msg)); | |
2206 | req->work.func = io_sendrecv_async; | |
0b416c3e | 2207 | return -EAGAIN; |
03b1230c | 2208 | } |
441cdbd5 JA |
2209 | if (ret == -ERESTARTSYS) |
2210 | ret = -EINTR; | |
0fa03c62 JA |
2211 | } |
2212 | ||
03b1230c | 2213 | out: |
b7bb4f7d | 2214 | if (!io_wq_current_is_worker() && kmsg && kmsg->iov != kmsg->fast_iov) |
0b416c3e | 2215 | kfree(kmsg->iov); |
78e19bbe | 2216 | io_cqring_add_event(req, ret); |
4e88d6e7 JA |
2217 | if (ret < 0) |
2218 | req_set_fail_links(req); | |
ec9c02ad | 2219 | io_put_req_find_next(req, nxt); |
5d17b4a4 | 2220 | return 0; |
03b1230c JA |
2221 | #else |
2222 | return -EOPNOTSUPP; | |
aa1fa28f | 2223 | #endif |
03b1230c | 2224 | } |
aa1fa28f | 2225 | |
03b1230c | 2226 | static int io_recvmsg_prep(struct io_kiocb *req, struct io_async_ctx *io) |
aa1fa28f JA |
2227 | { |
2228 | #if defined(CONFIG_NET) | |
03b1230c JA |
2229 | const struct io_uring_sqe *sqe = req->sqe; |
2230 | struct user_msghdr __user *msg; | |
2231 | unsigned flags; | |
2232 | ||
2233 | flags = READ_ONCE(sqe->msg_flags); | |
2234 | msg = (struct user_msghdr __user *)(unsigned long) READ_ONCE(sqe->addr); | |
d9688565 | 2235 | io->msg.iov = io->msg.fast_iov; |
03b1230c JA |
2236 | return recvmsg_copy_msghdr(&io->msg.msg, msg, flags, &io->msg.uaddr, |
2237 | &io->msg.iov); | |
aa1fa28f | 2238 | #else |
03b1230c | 2239 | return 0; |
aa1fa28f JA |
2240 | #endif |
2241 | } | |
2242 | ||
fc4df999 JA |
2243 | static int io_recvmsg(struct io_kiocb *req, struct io_kiocb **nxt, |
2244 | bool force_nonblock) | |
aa1fa28f JA |
2245 | { |
2246 | #if defined(CONFIG_NET) | |
fc4df999 | 2247 | const struct io_uring_sqe *sqe = req->sqe; |
0b416c3e | 2248 | struct io_async_msghdr *kmsg = NULL; |
03b1230c JA |
2249 | struct socket *sock; |
2250 | int ret; | |
2251 | ||
2252 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
2253 | return -EINVAL; | |
2254 | ||
2255 | sock = sock_from_file(req->file, &ret); | |
2256 | if (sock) { | |
2257 | struct user_msghdr __user *msg; | |
b7bb4f7d | 2258 | struct io_async_ctx io; |
03b1230c | 2259 | struct sockaddr_storage addr; |
03b1230c JA |
2260 | unsigned flags; |
2261 | ||
2262 | flags = READ_ONCE(sqe->msg_flags); | |
2263 | if (flags & MSG_DONTWAIT) | |
2264 | req->flags |= REQ_F_NOWAIT; | |
2265 | else if (force_nonblock) | |
2266 | flags |= MSG_DONTWAIT; | |
2267 | ||
2268 | msg = (struct user_msghdr __user *) (unsigned long) | |
2269 | READ_ONCE(sqe->addr); | |
2270 | if (req->io) { | |
0b416c3e JA |
2271 | kmsg = &req->io->msg; |
2272 | kmsg->msg.msg_name = &addr; | |
2273 | /* if iov is set, it's allocated already */ | |
2274 | if (!kmsg->iov) | |
2275 | kmsg->iov = kmsg->fast_iov; | |
2276 | kmsg->msg.msg_iter.iov = kmsg->iov; | |
03b1230c | 2277 | } else { |
0b416c3e JA |
2278 | kmsg = &io.msg; |
2279 | kmsg->msg.msg_name = &addr; | |
03b1230c JA |
2280 | ret = io_recvmsg_prep(req, &io); |
2281 | if (ret) | |
2282 | goto out; | |
2283 | } | |
2284 | ||
0b416c3e | 2285 | ret = __sys_recvmsg_sock(sock, &kmsg->msg, msg, kmsg->uaddr, flags); |
03b1230c | 2286 | if (force_nonblock && ret == -EAGAIN) { |
b7bb4f7d JA |
2287 | if (req->io) |
2288 | return -EAGAIN; | |
2289 | if (io_alloc_async_ctx(req)) | |
2290 | return -ENOMEM; | |
2291 | memcpy(&req->io->msg, &io.msg, sizeof(io.msg)); | |
2292 | req->work.func = io_sendrecv_async; | |
0b416c3e | 2293 | return -EAGAIN; |
03b1230c JA |
2294 | } |
2295 | if (ret == -ERESTARTSYS) | |
2296 | ret = -EINTR; | |
2297 | } | |
2298 | ||
2299 | out: | |
b7bb4f7d | 2300 | if (!io_wq_current_is_worker() && kmsg && kmsg->iov != kmsg->fast_iov) |
0b416c3e | 2301 | kfree(kmsg->iov); |
03b1230c | 2302 | io_cqring_add_event(req, ret); |
4e88d6e7 JA |
2303 | if (ret < 0) |
2304 | req_set_fail_links(req); | |
03b1230c JA |
2305 | io_put_req_find_next(req, nxt); |
2306 | return 0; | |
0fa03c62 JA |
2307 | #else |
2308 | return -EOPNOTSUPP; | |
2309 | #endif | |
2310 | } | |
5d17b4a4 | 2311 | |
8ed8d3c3 | 2312 | static int io_accept_prep(struct io_kiocb *req) |
17f2fe35 JA |
2313 | { |
2314 | #if defined(CONFIG_NET) | |
fc4df999 | 2315 | const struct io_uring_sqe *sqe = req->sqe; |
8ed8d3c3 JA |
2316 | struct io_accept *accept = &req->accept; |
2317 | ||
2318 | if (req->flags & REQ_F_PREPPED) | |
2319 | return 0; | |
17f2fe35 JA |
2320 | |
2321 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) | |
2322 | return -EINVAL; | |
8042d6ce | 2323 | if (sqe->ioprio || sqe->len || sqe->buf_index) |
17f2fe35 JA |
2324 | return -EINVAL; |
2325 | ||
8ed8d3c3 JA |
2326 | accept->addr = (struct sockaddr __user *) |
2327 | (unsigned long) READ_ONCE(sqe->addr); | |
2328 | accept->addr_len = (int __user *) (unsigned long) READ_ONCE(sqe->addr2); | |
2329 | accept->flags = READ_ONCE(sqe->accept_flags); | |
2330 | req->flags |= REQ_F_PREPPED; | |
2331 | return 0; | |
2332 | #else | |
2333 | return -EOPNOTSUPP; | |
2334 | #endif | |
2335 | } | |
17f2fe35 | 2336 | |
8ed8d3c3 JA |
2337 | #if defined(CONFIG_NET) |
2338 | static int __io_accept(struct io_kiocb *req, struct io_kiocb **nxt, | |
2339 | bool force_nonblock) | |
2340 | { | |
2341 | struct io_accept *accept = &req->accept; | |
2342 | unsigned file_flags; | |
2343 | int ret; | |
2344 | ||
2345 | file_flags = force_nonblock ? O_NONBLOCK : 0; | |
2346 | ret = __sys_accept4_file(req->file, file_flags, accept->addr, | |
2347 | accept->addr_len, accept->flags); | |
2348 | if (ret == -EAGAIN && force_nonblock) | |
17f2fe35 | 2349 | return -EAGAIN; |
8e3cca12 JA |
2350 | if (ret == -ERESTARTSYS) |
2351 | ret = -EINTR; | |
4e88d6e7 JA |
2352 | if (ret < 0) |
2353 | req_set_fail_links(req); | |
78e19bbe | 2354 | io_cqring_add_event(req, ret); |
ec9c02ad | 2355 | io_put_req_find_next(req, nxt); |
17f2fe35 | 2356 | return 0; |
8ed8d3c3 JA |
2357 | } |
2358 | ||
2359 | static void io_accept_finish(struct io_wq_work **workptr) | |
2360 | { | |
2361 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
2362 | struct io_kiocb *nxt = NULL; | |
2363 | ||
2364 | if (io_req_cancelled(req)) | |
2365 | return; | |
2366 | __io_accept(req, &nxt, false); | |
2367 | if (nxt) | |
2368 | *workptr = &nxt->work; | |
2369 | } | |
2370 | #endif | |
2371 | ||
2372 | static int io_accept(struct io_kiocb *req, struct io_kiocb **nxt, | |
2373 | bool force_nonblock) | |
2374 | { | |
2375 | #if defined(CONFIG_NET) | |
2376 | int ret; | |
2377 | ||
2378 | ret = io_accept_prep(req); | |
2379 | if (ret) | |
2380 | return ret; | |
2381 | ||
2382 | ret = __io_accept(req, nxt, force_nonblock); | |
2383 | if (ret == -EAGAIN && force_nonblock) { | |
2384 | req->work.func = io_accept_finish; | |
2385 | req->work.flags |= IO_WQ_WORK_NEEDS_FILES; | |
2386 | io_put_req(req); | |
2387 | return -EAGAIN; | |
2388 | } | |
2389 | return 0; | |
0fa03c62 JA |
2390 | #else |
2391 | return -EOPNOTSUPP; | |
2392 | #endif | |
2393 | } | |
5d17b4a4 | 2394 | |
f499a021 JA |
2395 | static int io_connect_prep(struct io_kiocb *req, struct io_async_ctx *io) |
2396 | { | |
2397 | #if defined(CONFIG_NET) | |
2398 | const struct io_uring_sqe *sqe = req->sqe; | |
2399 | struct sockaddr __user *addr; | |
2400 | int addr_len; | |
2401 | ||
2402 | addr = (struct sockaddr __user *) (unsigned long) READ_ONCE(sqe->addr); | |
2403 | addr_len = READ_ONCE(sqe->addr2); | |
2404 | return move_addr_to_kernel(addr, addr_len, &io->connect.address); | |
2405 | #else | |
2406 | return 0; | |
2407 | #endif | |
2408 | } | |
2409 | ||
fc4df999 JA |
2410 | static int io_connect(struct io_kiocb *req, struct io_kiocb **nxt, |
2411 | bool force_nonblock) | |
f8e85cf2 JA |
2412 | { |
2413 | #if defined(CONFIG_NET) | |
fc4df999 | 2414 | const struct io_uring_sqe *sqe = req->sqe; |
f499a021 | 2415 | struct io_async_ctx __io, *io; |
f8e85cf2 JA |
2416 | unsigned file_flags; |
2417 | int addr_len, ret; | |
2418 | ||
2419 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) | |
2420 | return -EINVAL; | |
2421 | if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags) | |
2422 | return -EINVAL; | |
2423 | ||
f8e85cf2 JA |
2424 | addr_len = READ_ONCE(sqe->addr2); |
2425 | file_flags = force_nonblock ? O_NONBLOCK : 0; | |
2426 | ||
f499a021 JA |
2427 | if (req->io) { |
2428 | io = req->io; | |
2429 | } else { | |
2430 | ret = io_connect_prep(req, &__io); | |
2431 | if (ret) | |
2432 | goto out; | |
2433 | io = &__io; | |
2434 | } | |
2435 | ||
2436 | ret = __sys_connect_file(req->file, &io->connect.address, addr_len, | |
2437 | file_flags); | |
87f80d62 | 2438 | if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) { |
b7bb4f7d JA |
2439 | if (req->io) |
2440 | return -EAGAIN; | |
2441 | if (io_alloc_async_ctx(req)) { | |
f499a021 JA |
2442 | ret = -ENOMEM; |
2443 | goto out; | |
2444 | } | |
b7bb4f7d | 2445 | memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect)); |
f8e85cf2 | 2446 | return -EAGAIN; |
f499a021 | 2447 | } |
f8e85cf2 JA |
2448 | if (ret == -ERESTARTSYS) |
2449 | ret = -EINTR; | |
f499a021 | 2450 | out: |
4e88d6e7 JA |
2451 | if (ret < 0) |
2452 | req_set_fail_links(req); | |
f8e85cf2 JA |
2453 | io_cqring_add_event(req, ret); |
2454 | io_put_req_find_next(req, nxt); | |
2455 | return 0; | |
2456 | #else | |
2457 | return -EOPNOTSUPP; | |
2458 | #endif | |
2459 | } | |
2460 | ||
221c5eb2 JA |
2461 | static void io_poll_remove_one(struct io_kiocb *req) |
2462 | { | |
2463 | struct io_poll_iocb *poll = &req->poll; | |
2464 | ||
2465 | spin_lock(&poll->head->lock); | |
2466 | WRITE_ONCE(poll->canceled, true); | |
392edb45 JA |
2467 | if (!list_empty(&poll->wait.entry)) { |
2468 | list_del_init(&poll->wait.entry); | |
a197f664 | 2469 | io_queue_async_work(req); |
221c5eb2 JA |
2470 | } |
2471 | spin_unlock(&poll->head->lock); | |
78076bb6 | 2472 | hash_del(&req->hash_node); |
221c5eb2 JA |
2473 | } |
2474 | ||
2475 | static void io_poll_remove_all(struct io_ring_ctx *ctx) | |
2476 | { | |
78076bb6 | 2477 | struct hlist_node *tmp; |
221c5eb2 | 2478 | struct io_kiocb *req; |
78076bb6 | 2479 | int i; |
221c5eb2 JA |
2480 | |
2481 | spin_lock_irq(&ctx->completion_lock); | |
78076bb6 JA |
2482 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { |
2483 | struct hlist_head *list; | |
2484 | ||
2485 | list = &ctx->cancel_hash[i]; | |
2486 | hlist_for_each_entry_safe(req, tmp, list, hash_node) | |
2487 | io_poll_remove_one(req); | |
221c5eb2 JA |
2488 | } |
2489 | spin_unlock_irq(&ctx->completion_lock); | |
2490 | } | |
2491 | ||
47f46768 JA |
2492 | static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr) |
2493 | { | |
78076bb6 | 2494 | struct hlist_head *list; |
47f46768 JA |
2495 | struct io_kiocb *req; |
2496 | ||
78076bb6 JA |
2497 | list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)]; |
2498 | hlist_for_each_entry(req, list, hash_node) { | |
2499 | if (sqe_addr == req->user_data) { | |
eac406c6 JA |
2500 | io_poll_remove_one(req); |
2501 | return 0; | |
2502 | } | |
47f46768 JA |
2503 | } |
2504 | ||
2505 | return -ENOENT; | |
2506 | } | |
2507 | ||
0969e783 JA |
2508 | static int io_poll_remove_prep(struct io_kiocb *req) |
2509 | { | |
2510 | const struct io_uring_sqe *sqe = req->sqe; | |
2511 | ||
2512 | if (req->flags & REQ_F_PREPPED) | |
2513 | return 0; | |
2514 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
2515 | return -EINVAL; | |
2516 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index || | |
2517 | sqe->poll_events) | |
2518 | return -EINVAL; | |
2519 | ||
2520 | req->poll.addr = READ_ONCE(sqe->addr); | |
2521 | req->flags |= REQ_F_PREPPED; | |
2522 | return 0; | |
2523 | } | |
2524 | ||
221c5eb2 JA |
2525 | /* |
2526 | * Find a running poll command that matches one specified in sqe->addr, | |
2527 | * and remove it if found. | |
2528 | */ | |
fc4df999 | 2529 | static int io_poll_remove(struct io_kiocb *req) |
221c5eb2 JA |
2530 | { |
2531 | struct io_ring_ctx *ctx = req->ctx; | |
0969e783 | 2532 | u64 addr; |
47f46768 | 2533 | int ret; |
221c5eb2 | 2534 | |
0969e783 JA |
2535 | ret = io_poll_remove_prep(req); |
2536 | if (ret) | |
2537 | return ret; | |
221c5eb2 | 2538 | |
0969e783 | 2539 | addr = req->poll.addr; |
221c5eb2 | 2540 | spin_lock_irq(&ctx->completion_lock); |
0969e783 | 2541 | ret = io_poll_cancel(ctx, addr); |
221c5eb2 JA |
2542 | spin_unlock_irq(&ctx->completion_lock); |
2543 | ||
78e19bbe | 2544 | io_cqring_add_event(req, ret); |
4e88d6e7 JA |
2545 | if (ret < 0) |
2546 | req_set_fail_links(req); | |
e65ef56d | 2547 | io_put_req(req); |
221c5eb2 JA |
2548 | return 0; |
2549 | } | |
2550 | ||
b0dd8a41 | 2551 | static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error) |
221c5eb2 | 2552 | { |
a197f664 JL |
2553 | struct io_ring_ctx *ctx = req->ctx; |
2554 | ||
8c838788 | 2555 | req->poll.done = true; |
b0dd8a41 JA |
2556 | if (error) |
2557 | io_cqring_fill_event(req, error); | |
2558 | else | |
2559 | io_cqring_fill_event(req, mangle_poll(mask)); | |
8c838788 | 2560 | io_commit_cqring(ctx); |
221c5eb2 JA |
2561 | } |
2562 | ||
561fb04a | 2563 | static void io_poll_complete_work(struct io_wq_work **workptr) |
221c5eb2 | 2564 | { |
561fb04a | 2565 | struct io_wq_work *work = *workptr; |
221c5eb2 JA |
2566 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
2567 | struct io_poll_iocb *poll = &req->poll; | |
2568 | struct poll_table_struct pt = { ._key = poll->events }; | |
2569 | struct io_ring_ctx *ctx = req->ctx; | |
89723d0b | 2570 | struct io_kiocb *nxt = NULL; |
221c5eb2 | 2571 | __poll_t mask = 0; |
b0dd8a41 | 2572 | int ret = 0; |
221c5eb2 | 2573 | |
b0dd8a41 | 2574 | if (work->flags & IO_WQ_WORK_CANCEL) { |
561fb04a | 2575 | WRITE_ONCE(poll->canceled, true); |
b0dd8a41 JA |
2576 | ret = -ECANCELED; |
2577 | } else if (READ_ONCE(poll->canceled)) { | |
2578 | ret = -ECANCELED; | |
2579 | } | |
561fb04a | 2580 | |
b0dd8a41 | 2581 | if (ret != -ECANCELED) |
221c5eb2 JA |
2582 | mask = vfs_poll(poll->file, &pt) & poll->events; |
2583 | ||
2584 | /* | |
2585 | * Note that ->ki_cancel callers also delete iocb from active_reqs after | |
2586 | * calling ->ki_cancel. We need the ctx_lock roundtrip here to | |
2587 | * synchronize with them. In the cancellation case the list_del_init | |
2588 | * itself is not actually needed, but harmless so we keep it in to | |
2589 | * avoid further branches in the fast path. | |
2590 | */ | |
2591 | spin_lock_irq(&ctx->completion_lock); | |
b0dd8a41 | 2592 | if (!mask && ret != -ECANCELED) { |
392edb45 | 2593 | add_wait_queue(poll->head, &poll->wait); |
221c5eb2 JA |
2594 | spin_unlock_irq(&ctx->completion_lock); |
2595 | return; | |
2596 | } | |
78076bb6 | 2597 | hash_del(&req->hash_node); |
b0dd8a41 | 2598 | io_poll_complete(req, mask, ret); |
221c5eb2 JA |
2599 | spin_unlock_irq(&ctx->completion_lock); |
2600 | ||
8c838788 | 2601 | io_cqring_ev_posted(ctx); |
89723d0b | 2602 | |
4e88d6e7 JA |
2603 | if (ret < 0) |
2604 | req_set_fail_links(req); | |
ec9c02ad | 2605 | io_put_req_find_next(req, &nxt); |
89723d0b JA |
2606 | if (nxt) |
2607 | *workptr = &nxt->work; | |
221c5eb2 JA |
2608 | } |
2609 | ||
2610 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
2611 | void *key) | |
2612 | { | |
e944475e | 2613 | struct io_poll_iocb *poll = wait->private; |
221c5eb2 JA |
2614 | struct io_kiocb *req = container_of(poll, struct io_kiocb, poll); |
2615 | struct io_ring_ctx *ctx = req->ctx; | |
2616 | __poll_t mask = key_to_poll(key); | |
8c838788 | 2617 | unsigned long flags; |
221c5eb2 JA |
2618 | |
2619 | /* for instances that support it check for an event match first: */ | |
8c838788 JA |
2620 | if (mask && !(mask & poll->events)) |
2621 | return 0; | |
221c5eb2 | 2622 | |
392edb45 | 2623 | list_del_init(&poll->wait.entry); |
221c5eb2 | 2624 | |
7c9e7f0f JA |
2625 | /* |
2626 | * Run completion inline if we can. We're using trylock here because | |
2627 | * we are violating the completion_lock -> poll wq lock ordering. | |
2628 | * If we have a link timeout we're going to need the completion_lock | |
2629 | * for finalizing the request, mark us as having grabbed that already. | |
2630 | */ | |
8c838788 | 2631 | if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) { |
78076bb6 | 2632 | hash_del(&req->hash_node); |
b0dd8a41 | 2633 | io_poll_complete(req, mask, 0); |
7c9e7f0f JA |
2634 | req->flags |= REQ_F_COMP_LOCKED; |
2635 | io_put_req(req); | |
8c838788 | 2636 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
221c5eb2 | 2637 | |
8c838788 | 2638 | io_cqring_ev_posted(ctx); |
8c838788 | 2639 | } else { |
a197f664 | 2640 | io_queue_async_work(req); |
221c5eb2 JA |
2641 | } |
2642 | ||
221c5eb2 JA |
2643 | return 1; |
2644 | } | |
2645 | ||
2646 | struct io_poll_table { | |
2647 | struct poll_table_struct pt; | |
2648 | struct io_kiocb *req; | |
2649 | int error; | |
2650 | }; | |
2651 | ||
2652 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
2653 | struct poll_table_struct *p) | |
2654 | { | |
2655 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
2656 | ||
2657 | if (unlikely(pt->req->poll.head)) { | |
2658 | pt->error = -EINVAL; | |
2659 | return; | |
2660 | } | |
2661 | ||
2662 | pt->error = 0; | |
2663 | pt->req->poll.head = head; | |
392edb45 | 2664 | add_wait_queue(head, &pt->req->poll.wait); |
221c5eb2 JA |
2665 | } |
2666 | ||
eac406c6 JA |
2667 | static void io_poll_req_insert(struct io_kiocb *req) |
2668 | { | |
2669 | struct io_ring_ctx *ctx = req->ctx; | |
78076bb6 JA |
2670 | struct hlist_head *list; |
2671 | ||
2672 | list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)]; | |
2673 | hlist_add_head(&req->hash_node, list); | |
eac406c6 JA |
2674 | } |
2675 | ||
0969e783 | 2676 | static int io_poll_add_prep(struct io_kiocb *req) |
221c5eb2 | 2677 | { |
fc4df999 | 2678 | const struct io_uring_sqe *sqe = req->sqe; |
221c5eb2 | 2679 | struct io_poll_iocb *poll = &req->poll; |
221c5eb2 | 2680 | u16 events; |
221c5eb2 | 2681 | |
0969e783 JA |
2682 | if (req->flags & REQ_F_PREPPED) |
2683 | return 0; | |
221c5eb2 JA |
2684 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
2685 | return -EINVAL; | |
2686 | if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
2687 | return -EINVAL; | |
09bb8394 JA |
2688 | if (!poll->file) |
2689 | return -EBADF; | |
221c5eb2 | 2690 | |
0969e783 | 2691 | req->flags |= REQ_F_PREPPED; |
221c5eb2 JA |
2692 | events = READ_ONCE(sqe->poll_events); |
2693 | poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP; | |
0969e783 JA |
2694 | return 0; |
2695 | } | |
2696 | ||
2697 | static int io_poll_add(struct io_kiocb *req, struct io_kiocb **nxt) | |
2698 | { | |
2699 | struct io_poll_iocb *poll = &req->poll; | |
2700 | struct io_ring_ctx *ctx = req->ctx; | |
2701 | struct io_poll_table ipt; | |
2702 | bool cancel = false; | |
2703 | __poll_t mask; | |
2704 | int ret; | |
2705 | ||
2706 | ret = io_poll_add_prep(req); | |
2707 | if (ret) | |
2708 | return ret; | |
2709 | ||
2710 | INIT_IO_WORK(&req->work, io_poll_complete_work); | |
78076bb6 | 2711 | INIT_HLIST_NODE(&req->hash_node); |
221c5eb2 | 2712 | |
221c5eb2 | 2713 | poll->head = NULL; |
8c838788 | 2714 | poll->done = false; |
221c5eb2 JA |
2715 | poll->canceled = false; |
2716 | ||
2717 | ipt.pt._qproc = io_poll_queue_proc; | |
2718 | ipt.pt._key = poll->events; | |
2719 | ipt.req = req; | |
2720 | ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */ | |
2721 | ||
2722 | /* initialized the list so that we can do list_empty checks */ | |
392edb45 JA |
2723 | INIT_LIST_HEAD(&poll->wait.entry); |
2724 | init_waitqueue_func_entry(&poll->wait, io_poll_wake); | |
2725 | poll->wait.private = poll; | |
221c5eb2 | 2726 | |
36703247 JA |
2727 | INIT_LIST_HEAD(&req->list); |
2728 | ||
221c5eb2 | 2729 | mask = vfs_poll(poll->file, &ipt.pt) & poll->events; |
221c5eb2 JA |
2730 | |
2731 | spin_lock_irq(&ctx->completion_lock); | |
8c838788 JA |
2732 | if (likely(poll->head)) { |
2733 | spin_lock(&poll->head->lock); | |
392edb45 | 2734 | if (unlikely(list_empty(&poll->wait.entry))) { |
8c838788 JA |
2735 | if (ipt.error) |
2736 | cancel = true; | |
2737 | ipt.error = 0; | |
2738 | mask = 0; | |
2739 | } | |
2740 | if (mask || ipt.error) | |
392edb45 | 2741 | list_del_init(&poll->wait.entry); |
8c838788 JA |
2742 | else if (cancel) |
2743 | WRITE_ONCE(poll->canceled, true); | |
2744 | else if (!poll->done) /* actually waiting for an event */ | |
eac406c6 | 2745 | io_poll_req_insert(req); |
8c838788 JA |
2746 | spin_unlock(&poll->head->lock); |
2747 | } | |
2748 | if (mask) { /* no async, we'd stolen it */ | |
221c5eb2 | 2749 | ipt.error = 0; |
b0dd8a41 | 2750 | io_poll_complete(req, mask, 0); |
221c5eb2 | 2751 | } |
221c5eb2 JA |
2752 | spin_unlock_irq(&ctx->completion_lock); |
2753 | ||
8c838788 JA |
2754 | if (mask) { |
2755 | io_cqring_ev_posted(ctx); | |
ec9c02ad | 2756 | io_put_req_find_next(req, nxt); |
221c5eb2 | 2757 | } |
8c838788 | 2758 | return ipt.error; |
221c5eb2 JA |
2759 | } |
2760 | ||
5262f567 JA |
2761 | static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer) |
2762 | { | |
ad8a48ac JA |
2763 | struct io_timeout_data *data = container_of(timer, |
2764 | struct io_timeout_data, timer); | |
2765 | struct io_kiocb *req = data->req; | |
2766 | struct io_ring_ctx *ctx = req->ctx; | |
5262f567 JA |
2767 | unsigned long flags; |
2768 | ||
5262f567 JA |
2769 | atomic_inc(&ctx->cq_timeouts); |
2770 | ||
2771 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
ef03681a | 2772 | /* |
11365043 JA |
2773 | * We could be racing with timeout deletion. If the list is empty, |
2774 | * then timeout lookup already found it and will be handling it. | |
ef03681a | 2775 | */ |
842f9612 | 2776 | if (!list_empty(&req->list)) { |
11365043 | 2777 | struct io_kiocb *prev; |
5262f567 | 2778 | |
11365043 JA |
2779 | /* |
2780 | * Adjust the reqs sequence before the current one because it | |
d195a66e | 2781 | * will consume a slot in the cq_ring and the cq_tail |
11365043 JA |
2782 | * pointer will be increased, otherwise other timeout reqs may |
2783 | * return in advance without waiting for enough wait_nr. | |
2784 | */ | |
2785 | prev = req; | |
2786 | list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list) | |
2787 | prev->sequence++; | |
11365043 | 2788 | list_del_init(&req->list); |
11365043 | 2789 | } |
5262f567 | 2790 | |
78e19bbe | 2791 | io_cqring_fill_event(req, -ETIME); |
5262f567 JA |
2792 | io_commit_cqring(ctx); |
2793 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
2794 | ||
2795 | io_cqring_ev_posted(ctx); | |
4e88d6e7 | 2796 | req_set_fail_links(req); |
5262f567 JA |
2797 | io_put_req(req); |
2798 | return HRTIMER_NORESTART; | |
2799 | } | |
2800 | ||
47f46768 JA |
2801 | static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data) |
2802 | { | |
2803 | struct io_kiocb *req; | |
2804 | int ret = -ENOENT; | |
2805 | ||
2806 | list_for_each_entry(req, &ctx->timeout_list, list) { | |
2807 | if (user_data == req->user_data) { | |
2808 | list_del_init(&req->list); | |
2809 | ret = 0; | |
2810 | break; | |
2811 | } | |
2812 | } | |
2813 | ||
2814 | if (ret == -ENOENT) | |
2815 | return ret; | |
2816 | ||
2d28390a | 2817 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); |
47f46768 JA |
2818 | if (ret == -1) |
2819 | return -EALREADY; | |
2820 | ||
4e88d6e7 | 2821 | req_set_fail_links(req); |
47f46768 JA |
2822 | io_cqring_fill_event(req, -ECANCELED); |
2823 | io_put_req(req); | |
2824 | return 0; | |
2825 | } | |
2826 | ||
b29472ee JA |
2827 | static int io_timeout_remove_prep(struct io_kiocb *req) |
2828 | { | |
2829 | const struct io_uring_sqe *sqe = req->sqe; | |
2830 | ||
2831 | if (req->flags & REQ_F_PREPPED) | |
2832 | return 0; | |
2833 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
2834 | return -EINVAL; | |
2835 | if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len) | |
2836 | return -EINVAL; | |
2837 | ||
2838 | req->timeout.addr = READ_ONCE(sqe->addr); | |
2839 | req->timeout.flags = READ_ONCE(sqe->timeout_flags); | |
2840 | if (req->timeout.flags) | |
2841 | return -EINVAL; | |
2842 | ||
2843 | req->flags |= REQ_F_PREPPED; | |
2844 | return 0; | |
2845 | } | |
2846 | ||
11365043 JA |
2847 | /* |
2848 | * Remove or update an existing timeout command | |
2849 | */ | |
fc4df999 | 2850 | static int io_timeout_remove(struct io_kiocb *req) |
11365043 JA |
2851 | { |
2852 | struct io_ring_ctx *ctx = req->ctx; | |
47f46768 | 2853 | int ret; |
11365043 | 2854 | |
b29472ee JA |
2855 | ret = io_timeout_remove_prep(req); |
2856 | if (ret) | |
2857 | return ret; | |
11365043 | 2858 | |
11365043 | 2859 | spin_lock_irq(&ctx->completion_lock); |
b29472ee | 2860 | ret = io_timeout_cancel(ctx, req->timeout.addr); |
11365043 | 2861 | |
47f46768 | 2862 | io_cqring_fill_event(req, ret); |
11365043 JA |
2863 | io_commit_cqring(ctx); |
2864 | spin_unlock_irq(&ctx->completion_lock); | |
5262f567 | 2865 | io_cqring_ev_posted(ctx); |
4e88d6e7 JA |
2866 | if (ret < 0) |
2867 | req_set_fail_links(req); | |
ec9c02ad | 2868 | io_put_req(req); |
11365043 | 2869 | return 0; |
5262f567 JA |
2870 | } |
2871 | ||
2d28390a JA |
2872 | static int io_timeout_prep(struct io_kiocb *req, struct io_async_ctx *io, |
2873 | bool is_timeout_link) | |
5262f567 | 2874 | { |
cf6fd4bd | 2875 | const struct io_uring_sqe *sqe = req->sqe; |
ad8a48ac | 2876 | struct io_timeout_data *data; |
a41525ab | 2877 | unsigned flags; |
5262f567 | 2878 | |
ad8a48ac | 2879 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5262f567 | 2880 | return -EINVAL; |
ad8a48ac | 2881 | if (sqe->ioprio || sqe->buf_index || sqe->len != 1) |
a41525ab | 2882 | return -EINVAL; |
2d28390a JA |
2883 | if (sqe->off && is_timeout_link) |
2884 | return -EINVAL; | |
a41525ab JA |
2885 | flags = READ_ONCE(sqe->timeout_flags); |
2886 | if (flags & ~IORING_TIMEOUT_ABS) | |
5262f567 | 2887 | return -EINVAL; |
bdf20073 | 2888 | |
2d28390a | 2889 | data = &io->timeout; |
ad8a48ac | 2890 | data->req = req; |
ad8a48ac JA |
2891 | req->flags |= REQ_F_TIMEOUT; |
2892 | ||
2893 | if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr))) | |
5262f567 JA |
2894 | return -EFAULT; |
2895 | ||
11365043 | 2896 | if (flags & IORING_TIMEOUT_ABS) |
ad8a48ac | 2897 | data->mode = HRTIMER_MODE_ABS; |
11365043 | 2898 | else |
ad8a48ac | 2899 | data->mode = HRTIMER_MODE_REL; |
11365043 | 2900 | |
ad8a48ac JA |
2901 | hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode); |
2902 | return 0; | |
2903 | } | |
2904 | ||
fc4df999 | 2905 | static int io_timeout(struct io_kiocb *req) |
ad8a48ac | 2906 | { |
fc4df999 | 2907 | const struct io_uring_sqe *sqe = req->sqe; |
ad8a48ac JA |
2908 | unsigned count; |
2909 | struct io_ring_ctx *ctx = req->ctx; | |
2910 | struct io_timeout_data *data; | |
2911 | struct list_head *entry; | |
2912 | unsigned span = 0; | |
b7bb4f7d | 2913 | int ret; |
ad8a48ac | 2914 | |
b7bb4f7d JA |
2915 | if (!req->io) { |
2916 | if (io_alloc_async_ctx(req)) | |
2d28390a | 2917 | return -ENOMEM; |
b7bb4f7d JA |
2918 | ret = io_timeout_prep(req, req->io, false); |
2919 | if (ret) | |
2d28390a | 2920 | return ret; |
2d28390a JA |
2921 | } |
2922 | data = &req->io->timeout; | |
93bd25bb | 2923 | |
5262f567 JA |
2924 | /* |
2925 | * sqe->off holds how many events that need to occur for this | |
93bd25bb JA |
2926 | * timeout event to be satisfied. If it isn't set, then this is |
2927 | * a pure timeout request, sequence isn't used. | |
5262f567 JA |
2928 | */ |
2929 | count = READ_ONCE(sqe->off); | |
93bd25bb JA |
2930 | if (!count) { |
2931 | req->flags |= REQ_F_TIMEOUT_NOSEQ; | |
2932 | spin_lock_irq(&ctx->completion_lock); | |
2933 | entry = ctx->timeout_list.prev; | |
2934 | goto add; | |
2935 | } | |
5262f567 JA |
2936 | |
2937 | req->sequence = ctx->cached_sq_head + count - 1; | |
2d28390a | 2938 | data->seq_offset = count; |
5262f567 JA |
2939 | |
2940 | /* | |
2941 | * Insertion sort, ensuring the first entry in the list is always | |
2942 | * the one we need first. | |
2943 | */ | |
5262f567 JA |
2944 | spin_lock_irq(&ctx->completion_lock); |
2945 | list_for_each_prev(entry, &ctx->timeout_list) { | |
2946 | struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list); | |
5da0fb1a | 2947 | unsigned nxt_sq_head; |
2948 | long long tmp, tmp_nxt; | |
2d28390a | 2949 | u32 nxt_offset = nxt->io->timeout.seq_offset; |
5262f567 | 2950 | |
93bd25bb JA |
2951 | if (nxt->flags & REQ_F_TIMEOUT_NOSEQ) |
2952 | continue; | |
2953 | ||
5da0fb1a | 2954 | /* |
2955 | * Since cached_sq_head + count - 1 can overflow, use type long | |
2956 | * long to store it. | |
2957 | */ | |
2958 | tmp = (long long)ctx->cached_sq_head + count - 1; | |
cc42e0ac PB |
2959 | nxt_sq_head = nxt->sequence - nxt_offset + 1; |
2960 | tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1; | |
5da0fb1a | 2961 | |
2962 | /* | |
2963 | * cached_sq_head may overflow, and it will never overflow twice | |
2964 | * once there is some timeout req still be valid. | |
2965 | */ | |
2966 | if (ctx->cached_sq_head < nxt_sq_head) | |
8b07a65a | 2967 | tmp += UINT_MAX; |
5da0fb1a | 2968 | |
a1f58ba4 | 2969 | if (tmp > tmp_nxt) |
5262f567 | 2970 | break; |
a1f58ba4 | 2971 | |
2972 | /* | |
2973 | * Sequence of reqs after the insert one and itself should | |
2974 | * be adjusted because each timeout req consumes a slot. | |
2975 | */ | |
2976 | span++; | |
2977 | nxt->sequence++; | |
5262f567 | 2978 | } |
a1f58ba4 | 2979 | req->sequence -= span; |
93bd25bb | 2980 | add: |
5262f567 | 2981 | list_add(&req->list, entry); |
ad8a48ac JA |
2982 | data->timer.function = io_timeout_fn; |
2983 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode); | |
5262f567 | 2984 | spin_unlock_irq(&ctx->completion_lock); |
5262f567 JA |
2985 | return 0; |
2986 | } | |
5262f567 | 2987 | |
62755e35 JA |
2988 | static bool io_cancel_cb(struct io_wq_work *work, void *data) |
2989 | { | |
2990 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
2991 | ||
2992 | return req->user_data == (unsigned long) data; | |
2993 | } | |
2994 | ||
e977d6d3 | 2995 | static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr) |
62755e35 | 2996 | { |
62755e35 | 2997 | enum io_wq_cancel cancel_ret; |
62755e35 JA |
2998 | int ret = 0; |
2999 | ||
62755e35 JA |
3000 | cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr); |
3001 | switch (cancel_ret) { | |
3002 | case IO_WQ_CANCEL_OK: | |
3003 | ret = 0; | |
3004 | break; | |
3005 | case IO_WQ_CANCEL_RUNNING: | |
3006 | ret = -EALREADY; | |
3007 | break; | |
3008 | case IO_WQ_CANCEL_NOTFOUND: | |
3009 | ret = -ENOENT; | |
3010 | break; | |
3011 | } | |
3012 | ||
e977d6d3 JA |
3013 | return ret; |
3014 | } | |
3015 | ||
47f46768 JA |
3016 | static void io_async_find_and_cancel(struct io_ring_ctx *ctx, |
3017 | struct io_kiocb *req, __u64 sqe_addr, | |
b0dd8a41 | 3018 | struct io_kiocb **nxt, int success_ret) |
47f46768 JA |
3019 | { |
3020 | unsigned long flags; | |
3021 | int ret; | |
3022 | ||
3023 | ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr); | |
3024 | if (ret != -ENOENT) { | |
3025 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
3026 | goto done; | |
3027 | } | |
3028 | ||
3029 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
3030 | ret = io_timeout_cancel(ctx, sqe_addr); | |
3031 | if (ret != -ENOENT) | |
3032 | goto done; | |
3033 | ret = io_poll_cancel(ctx, sqe_addr); | |
3034 | done: | |
b0dd8a41 JA |
3035 | if (!ret) |
3036 | ret = success_ret; | |
47f46768 JA |
3037 | io_cqring_fill_event(req, ret); |
3038 | io_commit_cqring(ctx); | |
3039 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
3040 | io_cqring_ev_posted(ctx); | |
3041 | ||
4e88d6e7 JA |
3042 | if (ret < 0) |
3043 | req_set_fail_links(req); | |
47f46768 JA |
3044 | io_put_req_find_next(req, nxt); |
3045 | } | |
3046 | ||
fbf23849 | 3047 | static int io_async_cancel_prep(struct io_kiocb *req) |
e977d6d3 | 3048 | { |
fc4df999 | 3049 | const struct io_uring_sqe *sqe = req->sqe; |
e977d6d3 | 3050 | |
fbf23849 JA |
3051 | if (req->flags & REQ_F_PREPPED) |
3052 | return 0; | |
3053 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
e977d6d3 JA |
3054 | return -EINVAL; |
3055 | if (sqe->flags || sqe->ioprio || sqe->off || sqe->len || | |
3056 | sqe->cancel_flags) | |
3057 | return -EINVAL; | |
3058 | ||
fbf23849 JA |
3059 | req->flags |= REQ_F_PREPPED; |
3060 | req->cancel.addr = READ_ONCE(sqe->addr); | |
3061 | return 0; | |
3062 | } | |
3063 | ||
3064 | static int io_async_cancel(struct io_kiocb *req, struct io_kiocb **nxt) | |
3065 | { | |
3066 | struct io_ring_ctx *ctx = req->ctx; | |
3067 | int ret; | |
3068 | ||
3069 | ret = io_async_cancel_prep(req); | |
3070 | if (ret) | |
3071 | return ret; | |
3072 | ||
3073 | io_async_find_and_cancel(ctx, req, req->cancel.addr, nxt, 0); | |
5262f567 JA |
3074 | return 0; |
3075 | } | |
3076 | ||
b7bb4f7d | 3077 | static int io_req_defer_prep(struct io_kiocb *req) |
f67676d1 JA |
3078 | { |
3079 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
b7bb4f7d | 3080 | struct io_async_ctx *io = req->io; |
f67676d1 | 3081 | struct iov_iter iter; |
e781573e | 3082 | ssize_t ret = 0; |
f67676d1 | 3083 | |
d625c6ee | 3084 | switch (req->opcode) { |
e781573e JA |
3085 | case IORING_OP_NOP: |
3086 | break; | |
f67676d1 JA |
3087 | case IORING_OP_READV: |
3088 | case IORING_OP_READ_FIXED: | |
b7bb4f7d JA |
3089 | /* ensure prep does right import */ |
3090 | req->io = NULL; | |
f67676d1 | 3091 | ret = io_read_prep(req, &iovec, &iter, true); |
b7bb4f7d JA |
3092 | req->io = io; |
3093 | if (ret < 0) | |
3094 | break; | |
3095 | io_req_map_rw(req, ret, iovec, inline_vecs, &iter); | |
3096 | ret = 0; | |
f67676d1 JA |
3097 | break; |
3098 | case IORING_OP_WRITEV: | |
3099 | case IORING_OP_WRITE_FIXED: | |
b7bb4f7d JA |
3100 | /* ensure prep does right import */ |
3101 | req->io = NULL; | |
f67676d1 | 3102 | ret = io_write_prep(req, &iovec, &iter, true); |
b7bb4f7d JA |
3103 | req->io = io; |
3104 | if (ret < 0) | |
3105 | break; | |
3106 | io_req_map_rw(req, ret, iovec, inline_vecs, &iter); | |
3107 | ret = 0; | |
f67676d1 | 3108 | break; |
0969e783 JA |
3109 | case IORING_OP_POLL_ADD: |
3110 | ret = io_poll_add_prep(req); | |
3111 | break; | |
3112 | case IORING_OP_POLL_REMOVE: | |
3113 | ret = io_poll_remove_prep(req); | |
3114 | break; | |
8ed8d3c3 JA |
3115 | case IORING_OP_FSYNC: |
3116 | ret = io_prep_fsync(req); | |
3117 | break; | |
3118 | case IORING_OP_SYNC_FILE_RANGE: | |
3119 | ret = io_prep_sfr(req); | |
3120 | break; | |
03b1230c JA |
3121 | case IORING_OP_SENDMSG: |
3122 | ret = io_sendmsg_prep(req, io); | |
3123 | break; | |
3124 | case IORING_OP_RECVMSG: | |
3125 | ret = io_recvmsg_prep(req, io); | |
3126 | break; | |
f499a021 JA |
3127 | case IORING_OP_CONNECT: |
3128 | ret = io_connect_prep(req, io); | |
3129 | break; | |
2d28390a | 3130 | case IORING_OP_TIMEOUT: |
b7bb4f7d JA |
3131 | ret = io_timeout_prep(req, io, false); |
3132 | break; | |
b29472ee JA |
3133 | case IORING_OP_TIMEOUT_REMOVE: |
3134 | ret = io_timeout_remove_prep(req); | |
3135 | break; | |
fbf23849 JA |
3136 | case IORING_OP_ASYNC_CANCEL: |
3137 | ret = io_async_cancel_prep(req); | |
3138 | break; | |
2d28390a | 3139 | case IORING_OP_LINK_TIMEOUT: |
b7bb4f7d JA |
3140 | ret = io_timeout_prep(req, io, true); |
3141 | break; | |
8ed8d3c3 JA |
3142 | case IORING_OP_ACCEPT: |
3143 | ret = io_accept_prep(req); | |
3144 | break; | |
f67676d1 | 3145 | default: |
e781573e JA |
3146 | printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n", |
3147 | req->opcode); | |
3148 | ret = -EINVAL; | |
b7bb4f7d | 3149 | break; |
f67676d1 JA |
3150 | } |
3151 | ||
b7bb4f7d | 3152 | return ret; |
f67676d1 JA |
3153 | } |
3154 | ||
a197f664 | 3155 | static int io_req_defer(struct io_kiocb *req) |
de0617e4 | 3156 | { |
a197f664 | 3157 | struct io_ring_ctx *ctx = req->ctx; |
f67676d1 | 3158 | int ret; |
de0617e4 | 3159 | |
9d858b21 BL |
3160 | /* Still need defer if there is pending req in defer list. */ |
3161 | if (!req_need_defer(req) && list_empty(&ctx->defer_list)) | |
de0617e4 JA |
3162 | return 0; |
3163 | ||
b7bb4f7d | 3164 | if (io_alloc_async_ctx(req)) |
de0617e4 JA |
3165 | return -EAGAIN; |
3166 | ||
b7bb4f7d JA |
3167 | ret = io_req_defer_prep(req); |
3168 | if (ret < 0) | |
2d28390a | 3169 | return ret; |
2d28390a | 3170 | |
de0617e4 | 3171 | spin_lock_irq(&ctx->completion_lock); |
9d858b21 | 3172 | if (!req_need_defer(req) && list_empty(&ctx->defer_list)) { |
de0617e4 | 3173 | spin_unlock_irq(&ctx->completion_lock); |
de0617e4 JA |
3174 | return 0; |
3175 | } | |
3176 | ||
915967f6 | 3177 | trace_io_uring_defer(ctx, req, req->user_data); |
de0617e4 JA |
3178 | list_add_tail(&req->list, &ctx->defer_list); |
3179 | spin_unlock_irq(&ctx->completion_lock); | |
3180 | return -EIOCBQUEUED; | |
3181 | } | |
3182 | ||
f9bd67f6 | 3183 | __attribute__((nonnull)) |
d732447f PB |
3184 | static int io_issue_sqe(struct io_kiocb *req, struct io_kiocb **nxt, |
3185 | bool force_nonblock) | |
2b188cc1 | 3186 | { |
a197f664 | 3187 | struct io_ring_ctx *ctx = req->ctx; |
d625c6ee | 3188 | int ret; |
2b188cc1 | 3189 | |
d625c6ee | 3190 | switch (req->opcode) { |
2b188cc1 | 3191 | case IORING_OP_NOP: |
78e19bbe | 3192 | ret = io_nop(req); |
2b188cc1 JA |
3193 | break; |
3194 | case IORING_OP_READV: | |
cf6fd4bd | 3195 | if (unlikely(req->sqe->buf_index)) |
edafccee | 3196 | return -EINVAL; |
267bc904 | 3197 | ret = io_read(req, nxt, force_nonblock); |
2b188cc1 JA |
3198 | break; |
3199 | case IORING_OP_WRITEV: | |
cf6fd4bd | 3200 | if (unlikely(req->sqe->buf_index)) |
edafccee | 3201 | return -EINVAL; |
267bc904 | 3202 | ret = io_write(req, nxt, force_nonblock); |
edafccee JA |
3203 | break; |
3204 | case IORING_OP_READ_FIXED: | |
267bc904 | 3205 | ret = io_read(req, nxt, force_nonblock); |
edafccee JA |
3206 | break; |
3207 | case IORING_OP_WRITE_FIXED: | |
267bc904 | 3208 | ret = io_write(req, nxt, force_nonblock); |
2b188cc1 | 3209 | break; |
c992fe29 | 3210 | case IORING_OP_FSYNC: |
fc4df999 | 3211 | ret = io_fsync(req, nxt, force_nonblock); |
c992fe29 | 3212 | break; |
221c5eb2 | 3213 | case IORING_OP_POLL_ADD: |
fc4df999 | 3214 | ret = io_poll_add(req, nxt); |
221c5eb2 JA |
3215 | break; |
3216 | case IORING_OP_POLL_REMOVE: | |
fc4df999 | 3217 | ret = io_poll_remove(req); |
221c5eb2 | 3218 | break; |
5d17b4a4 | 3219 | case IORING_OP_SYNC_FILE_RANGE: |
fc4df999 | 3220 | ret = io_sync_file_range(req, nxt, force_nonblock); |
5d17b4a4 | 3221 | break; |
0fa03c62 | 3222 | case IORING_OP_SENDMSG: |
fc4df999 | 3223 | ret = io_sendmsg(req, nxt, force_nonblock); |
0fa03c62 | 3224 | break; |
aa1fa28f | 3225 | case IORING_OP_RECVMSG: |
fc4df999 | 3226 | ret = io_recvmsg(req, nxt, force_nonblock); |
aa1fa28f | 3227 | break; |
5262f567 | 3228 | case IORING_OP_TIMEOUT: |
fc4df999 | 3229 | ret = io_timeout(req); |
5262f567 | 3230 | break; |
11365043 | 3231 | case IORING_OP_TIMEOUT_REMOVE: |
fc4df999 | 3232 | ret = io_timeout_remove(req); |
11365043 | 3233 | break; |
17f2fe35 | 3234 | case IORING_OP_ACCEPT: |
fc4df999 | 3235 | ret = io_accept(req, nxt, force_nonblock); |
17f2fe35 | 3236 | break; |
f8e85cf2 | 3237 | case IORING_OP_CONNECT: |
fc4df999 | 3238 | ret = io_connect(req, nxt, force_nonblock); |
f8e85cf2 | 3239 | break; |
62755e35 | 3240 | case IORING_OP_ASYNC_CANCEL: |
fc4df999 | 3241 | ret = io_async_cancel(req, nxt); |
62755e35 | 3242 | break; |
2b188cc1 JA |
3243 | default: |
3244 | ret = -EINVAL; | |
3245 | break; | |
3246 | } | |
3247 | ||
def596e9 JA |
3248 | if (ret) |
3249 | return ret; | |
3250 | ||
3251 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
9e645e11 | 3252 | if (req->result == -EAGAIN) |
def596e9 JA |
3253 | return -EAGAIN; |
3254 | ||
def596e9 | 3255 | io_iopoll_req_issued(req); |
def596e9 JA |
3256 | } |
3257 | ||
3258 | return 0; | |
2b188cc1 JA |
3259 | } |
3260 | ||
b76da70f JA |
3261 | static void io_link_work_cb(struct io_wq_work **workptr) |
3262 | { | |
3263 | struct io_wq_work *work = *workptr; | |
3264 | struct io_kiocb *link = work->data; | |
3265 | ||
3266 | io_queue_linked_timeout(link); | |
3267 | work->func = io_wq_submit_work; | |
3268 | } | |
3269 | ||
561fb04a | 3270 | static void io_wq_submit_work(struct io_wq_work **workptr) |
2b188cc1 | 3271 | { |
561fb04a | 3272 | struct io_wq_work *work = *workptr; |
2b188cc1 | 3273 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
561fb04a JA |
3274 | struct io_kiocb *nxt = NULL; |
3275 | int ret = 0; | |
2b188cc1 | 3276 | |
561fb04a JA |
3277 | /* Ensure we clear previously set non-block flag */ |
3278 | req->rw.ki_flags &= ~IOCB_NOWAIT; | |
2b188cc1 | 3279 | |
561fb04a JA |
3280 | if (work->flags & IO_WQ_WORK_CANCEL) |
3281 | ret = -ECANCELED; | |
31b51510 | 3282 | |
561fb04a | 3283 | if (!ret) { |
cf6fd4bd PB |
3284 | req->has_user = (work->flags & IO_WQ_WORK_HAS_MM) != 0; |
3285 | req->in_async = true; | |
561fb04a | 3286 | do { |
d732447f | 3287 | ret = io_issue_sqe(req, &nxt, false); |
561fb04a JA |
3288 | /* |
3289 | * We can get EAGAIN for polled IO even though we're | |
3290 | * forcing a sync submission from here, since we can't | |
3291 | * wait for request slots on the block side. | |
3292 | */ | |
3293 | if (ret != -EAGAIN) | |
3294 | break; | |
3295 | cond_resched(); | |
3296 | } while (1); | |
3297 | } | |
31b51510 | 3298 | |
561fb04a | 3299 | /* drop submission reference */ |
ec9c02ad | 3300 | io_put_req(req); |
817869d2 | 3301 | |
561fb04a | 3302 | if (ret) { |
4e88d6e7 | 3303 | req_set_fail_links(req); |
78e19bbe | 3304 | io_cqring_add_event(req, ret); |
817869d2 | 3305 | io_put_req(req); |
edafccee | 3306 | } |
2b188cc1 | 3307 | |
561fb04a JA |
3308 | /* if a dependent link is ready, pass it back */ |
3309 | if (!ret && nxt) { | |
94ae5e77 JA |
3310 | struct io_kiocb *link; |
3311 | ||
3312 | io_prep_async_work(nxt, &link); | |
561fb04a | 3313 | *workptr = &nxt->work; |
b76da70f JA |
3314 | if (link) { |
3315 | nxt->work.flags |= IO_WQ_WORK_CB; | |
3316 | nxt->work.func = io_link_work_cb; | |
3317 | nxt->work.data = link; | |
3318 | } | |
31b51510 | 3319 | } |
2b188cc1 JA |
3320 | } |
3321 | ||
9e3aa61a JA |
3322 | static bool io_req_op_valid(int op) |
3323 | { | |
3324 | return op >= IORING_OP_NOP && op < IORING_OP_LAST; | |
3325 | } | |
3326 | ||
d625c6ee | 3327 | static int io_req_needs_file(struct io_kiocb *req) |
09bb8394 | 3328 | { |
d625c6ee | 3329 | switch (req->opcode) { |
09bb8394 JA |
3330 | case IORING_OP_NOP: |
3331 | case IORING_OP_POLL_REMOVE: | |
5683e540 | 3332 | case IORING_OP_TIMEOUT: |
a320e9fa PB |
3333 | case IORING_OP_TIMEOUT_REMOVE: |
3334 | case IORING_OP_ASYNC_CANCEL: | |
3335 | case IORING_OP_LINK_TIMEOUT: | |
9e3aa61a | 3336 | return 0; |
09bb8394 | 3337 | default: |
d625c6ee | 3338 | if (io_req_op_valid(req->opcode)) |
9e3aa61a JA |
3339 | return 1; |
3340 | return -EINVAL; | |
09bb8394 JA |
3341 | } |
3342 | } | |
3343 | ||
65e19f54 JA |
3344 | static inline struct file *io_file_from_index(struct io_ring_ctx *ctx, |
3345 | int index) | |
3346 | { | |
3347 | struct fixed_file_table *table; | |
3348 | ||
3349 | table = &ctx->file_table[index >> IORING_FILE_TABLE_SHIFT]; | |
3350 | return table->files[index & IORING_FILE_TABLE_MASK]; | |
3351 | } | |
3352 | ||
a197f664 | 3353 | static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req) |
09bb8394 | 3354 | { |
a197f664 | 3355 | struct io_ring_ctx *ctx = req->ctx; |
09bb8394 | 3356 | unsigned flags; |
9e3aa61a | 3357 | int fd, ret; |
09bb8394 | 3358 | |
cf6fd4bd PB |
3359 | flags = READ_ONCE(req->sqe->flags); |
3360 | fd = READ_ONCE(req->sqe->fd); | |
09bb8394 | 3361 | |
4fe2c963 | 3362 | if (flags & IOSQE_IO_DRAIN) |
de0617e4 | 3363 | req->flags |= REQ_F_IO_DRAIN; |
de0617e4 | 3364 | |
d625c6ee | 3365 | ret = io_req_needs_file(req); |
9e3aa61a JA |
3366 | if (ret <= 0) |
3367 | return ret; | |
09bb8394 JA |
3368 | |
3369 | if (flags & IOSQE_FIXED_FILE) { | |
65e19f54 | 3370 | if (unlikely(!ctx->file_table || |
09bb8394 JA |
3371 | (unsigned) fd >= ctx->nr_user_files)) |
3372 | return -EBADF; | |
b7620121 | 3373 | fd = array_index_nospec(fd, ctx->nr_user_files); |
65e19f54 JA |
3374 | req->file = io_file_from_index(ctx, fd); |
3375 | if (!req->file) | |
08a45173 | 3376 | return -EBADF; |
09bb8394 JA |
3377 | req->flags |= REQ_F_FIXED_FILE; |
3378 | } else { | |
cf6fd4bd | 3379 | if (req->needs_fixed_file) |
09bb8394 | 3380 | return -EBADF; |
c826bd7a | 3381 | trace_io_uring_file_get(ctx, fd); |
09bb8394 JA |
3382 | req->file = io_file_get(state, fd); |
3383 | if (unlikely(!req->file)) | |
3384 | return -EBADF; | |
3385 | } | |
3386 | ||
3387 | return 0; | |
3388 | } | |
3389 | ||
a197f664 | 3390 | static int io_grab_files(struct io_kiocb *req) |
fcb323cc JA |
3391 | { |
3392 | int ret = -EBADF; | |
a197f664 | 3393 | struct io_ring_ctx *ctx = req->ctx; |
fcb323cc JA |
3394 | |
3395 | rcu_read_lock(); | |
3396 | spin_lock_irq(&ctx->inflight_lock); | |
3397 | /* | |
3398 | * We use the f_ops->flush() handler to ensure that we can flush | |
3399 | * out work accessing these files if the fd is closed. Check if | |
3400 | * the fd has changed since we started down this path, and disallow | |
3401 | * this operation if it has. | |
3402 | */ | |
cf6fd4bd | 3403 | if (fcheck(req->ring_fd) == req->ring_file) { |
fcb323cc JA |
3404 | list_add(&req->inflight_entry, &ctx->inflight_list); |
3405 | req->flags |= REQ_F_INFLIGHT; | |
3406 | req->work.files = current->files; | |
3407 | ret = 0; | |
3408 | } | |
3409 | spin_unlock_irq(&ctx->inflight_lock); | |
3410 | rcu_read_unlock(); | |
3411 | ||
3412 | return ret; | |
3413 | } | |
3414 | ||
2665abfd | 3415 | static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer) |
2b188cc1 | 3416 | { |
ad8a48ac JA |
3417 | struct io_timeout_data *data = container_of(timer, |
3418 | struct io_timeout_data, timer); | |
3419 | struct io_kiocb *req = data->req; | |
2665abfd JA |
3420 | struct io_ring_ctx *ctx = req->ctx; |
3421 | struct io_kiocb *prev = NULL; | |
3422 | unsigned long flags; | |
2665abfd JA |
3423 | |
3424 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
3425 | ||
3426 | /* | |
3427 | * We don't expect the list to be empty, that will only happen if we | |
3428 | * race with the completion of the linked work. | |
3429 | */ | |
4493233e PB |
3430 | if (!list_empty(&req->link_list)) { |
3431 | prev = list_entry(req->link_list.prev, struct io_kiocb, | |
3432 | link_list); | |
5d960724 | 3433 | if (refcount_inc_not_zero(&prev->refs)) { |
4493233e | 3434 | list_del_init(&req->link_list); |
5d960724 JA |
3435 | prev->flags &= ~REQ_F_LINK_TIMEOUT; |
3436 | } else | |
76a46e06 | 3437 | prev = NULL; |
2665abfd JA |
3438 | } |
3439 | ||
3440 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
3441 | ||
3442 | if (prev) { | |
4e88d6e7 | 3443 | req_set_fail_links(prev); |
b0dd8a41 JA |
3444 | io_async_find_and_cancel(ctx, req, prev->user_data, NULL, |
3445 | -ETIME); | |
76a46e06 | 3446 | io_put_req(prev); |
47f46768 JA |
3447 | } else { |
3448 | io_cqring_add_event(req, -ETIME); | |
3449 | io_put_req(req); | |
2665abfd | 3450 | } |
2665abfd JA |
3451 | return HRTIMER_NORESTART; |
3452 | } | |
3453 | ||
ad8a48ac | 3454 | static void io_queue_linked_timeout(struct io_kiocb *req) |
2665abfd | 3455 | { |
76a46e06 | 3456 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd | 3457 | |
76a46e06 JA |
3458 | /* |
3459 | * If the list is now empty, then our linked request finished before | |
3460 | * we got a chance to setup the timer | |
3461 | */ | |
3462 | spin_lock_irq(&ctx->completion_lock); | |
4493233e | 3463 | if (!list_empty(&req->link_list)) { |
2d28390a | 3464 | struct io_timeout_data *data = &req->io->timeout; |
94ae5e77 | 3465 | |
ad8a48ac JA |
3466 | data->timer.function = io_link_timeout_fn; |
3467 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), | |
3468 | data->mode); | |
2665abfd | 3469 | } |
76a46e06 | 3470 | spin_unlock_irq(&ctx->completion_lock); |
2665abfd | 3471 | |
2665abfd | 3472 | /* drop submission reference */ |
76a46e06 JA |
3473 | io_put_req(req); |
3474 | } | |
2665abfd | 3475 | |
ad8a48ac | 3476 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) |
2665abfd JA |
3477 | { |
3478 | struct io_kiocb *nxt; | |
3479 | ||
3480 | if (!(req->flags & REQ_F_LINK)) | |
3481 | return NULL; | |
3482 | ||
4493233e PB |
3483 | nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, |
3484 | link_list); | |
d625c6ee | 3485 | if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT) |
76a46e06 | 3486 | return NULL; |
2665abfd | 3487 | |
76a46e06 | 3488 | req->flags |= REQ_F_LINK_TIMEOUT; |
76a46e06 | 3489 | return nxt; |
2665abfd JA |
3490 | } |
3491 | ||
0e0702da | 3492 | static void __io_queue_sqe(struct io_kiocb *req) |
2b188cc1 | 3493 | { |
4a0a7a18 | 3494 | struct io_kiocb *linked_timeout; |
f9bd67f6 | 3495 | struct io_kiocb *nxt = NULL; |
e0c5c576 | 3496 | int ret; |
2b188cc1 | 3497 | |
4a0a7a18 JA |
3498 | again: |
3499 | linked_timeout = io_prep_linked_timeout(req); | |
3500 | ||
f9bd67f6 | 3501 | ret = io_issue_sqe(req, &nxt, true); |
491381ce JA |
3502 | |
3503 | /* | |
3504 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
3505 | * doesn't support non-blocking read/write attempts | |
3506 | */ | |
3507 | if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) || | |
3508 | (req->flags & REQ_F_MUST_PUNT))) { | |
bbad27b2 PB |
3509 | if (req->work.flags & IO_WQ_WORK_NEEDS_FILES) { |
3510 | ret = io_grab_files(req); | |
3511 | if (ret) | |
3512 | goto err; | |
2b188cc1 | 3513 | } |
bbad27b2 PB |
3514 | |
3515 | /* | |
3516 | * Queued up for async execution, worker will release | |
3517 | * submit reference when the iocb is actually submitted. | |
3518 | */ | |
3519 | io_queue_async_work(req); | |
4a0a7a18 | 3520 | goto done_req; |
2b188cc1 | 3521 | } |
e65ef56d | 3522 | |
fcb323cc | 3523 | err: |
76a46e06 | 3524 | /* drop submission reference */ |
ec9c02ad | 3525 | io_put_req(req); |
e65ef56d | 3526 | |
f9bd67f6 | 3527 | if (linked_timeout) { |
76a46e06 | 3528 | if (!ret) |
f9bd67f6 | 3529 | io_queue_linked_timeout(linked_timeout); |
76a46e06 | 3530 | else |
f9bd67f6 | 3531 | io_put_req(linked_timeout); |
76a46e06 JA |
3532 | } |
3533 | ||
e65ef56d | 3534 | /* and drop final reference, if we failed */ |
9e645e11 | 3535 | if (ret) { |
78e19bbe | 3536 | io_cqring_add_event(req, ret); |
4e88d6e7 | 3537 | req_set_fail_links(req); |
e65ef56d | 3538 | io_put_req(req); |
9e645e11 | 3539 | } |
4a0a7a18 JA |
3540 | done_req: |
3541 | if (nxt) { | |
3542 | req = nxt; | |
3543 | nxt = NULL; | |
3544 | goto again; | |
3545 | } | |
2b188cc1 JA |
3546 | } |
3547 | ||
0e0702da | 3548 | static void io_queue_sqe(struct io_kiocb *req) |
4fe2c963 JL |
3549 | { |
3550 | int ret; | |
3551 | ||
1b4a51b6 PB |
3552 | if (unlikely(req->ctx->drain_next)) { |
3553 | req->flags |= REQ_F_IO_DRAIN; | |
3554 | req->ctx->drain_next = false; | |
3555 | } | |
3556 | req->ctx->drain_next = (req->flags & REQ_F_DRAIN_LINK); | |
3557 | ||
a197f664 | 3558 | ret = io_req_defer(req); |
4fe2c963 JL |
3559 | if (ret) { |
3560 | if (ret != -EIOCBQUEUED) { | |
78e19bbe | 3561 | io_cqring_add_event(req, ret); |
4e88d6e7 | 3562 | req_set_fail_links(req); |
78e19bbe | 3563 | io_double_put_req(req); |
4fe2c963 | 3564 | } |
0e0702da JA |
3565 | } else |
3566 | __io_queue_sqe(req); | |
4fe2c963 JL |
3567 | } |
3568 | ||
1b4a51b6 | 3569 | static inline void io_queue_link_head(struct io_kiocb *req) |
4fe2c963 | 3570 | { |
94ae5e77 | 3571 | if (unlikely(req->flags & REQ_F_FAIL_LINK)) { |
1b4a51b6 PB |
3572 | io_cqring_add_event(req, -ECANCELED); |
3573 | io_double_put_req(req); | |
3574 | } else | |
0e0702da | 3575 | io_queue_sqe(req); |
4fe2c963 JL |
3576 | } |
3577 | ||
4e88d6e7 JA |
3578 | #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \ |
3579 | IOSQE_IO_HARDLINK) | |
9e645e11 | 3580 | |
2e6e1fde | 3581 | static bool io_submit_sqe(struct io_kiocb *req, struct io_submit_state *state, |
a197f664 | 3582 | struct io_kiocb **link) |
9e645e11 | 3583 | { |
a197f664 | 3584 | struct io_ring_ctx *ctx = req->ctx; |
9e645e11 JA |
3585 | int ret; |
3586 | ||
3587 | /* enforce forwards compatibility on users */ | |
cf6fd4bd | 3588 | if (unlikely(req->sqe->flags & ~SQE_VALID_FLAGS)) { |
9e645e11 | 3589 | ret = -EINVAL; |
196be95c | 3590 | goto err_req; |
9e645e11 JA |
3591 | } |
3592 | ||
a197f664 | 3593 | ret = io_req_set_file(state, req); |
9e645e11 JA |
3594 | if (unlikely(ret)) { |
3595 | err_req: | |
78e19bbe JA |
3596 | io_cqring_add_event(req, ret); |
3597 | io_double_put_req(req); | |
2e6e1fde | 3598 | return false; |
9e645e11 JA |
3599 | } |
3600 | ||
9e645e11 JA |
3601 | /* |
3602 | * If we already have a head request, queue this one for async | |
3603 | * submittal once the head completes. If we don't have a head but | |
3604 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
3605 | * submitted sync once the chain is complete. If none of those | |
3606 | * conditions are true (normal request), then just queue it. | |
3607 | */ | |
3608 | if (*link) { | |
3609 | struct io_kiocb *prev = *link; | |
3610 | ||
cf6fd4bd | 3611 | if (req->sqe->flags & IOSQE_IO_DRAIN) |
1b4a51b6 PB |
3612 | (*link)->flags |= REQ_F_DRAIN_LINK | REQ_F_IO_DRAIN; |
3613 | ||
4e88d6e7 JA |
3614 | if (req->sqe->flags & IOSQE_IO_HARDLINK) |
3615 | req->flags |= REQ_F_HARDLINK; | |
3616 | ||
b7bb4f7d | 3617 | if (io_alloc_async_ctx(req)) { |
9e645e11 JA |
3618 | ret = -EAGAIN; |
3619 | goto err_req; | |
3620 | } | |
3621 | ||
b7bb4f7d | 3622 | ret = io_req_defer_prep(req); |
2d28390a | 3623 | if (ret) { |
4e88d6e7 | 3624 | /* fail even hard links since we don't submit */ |
2d28390a | 3625 | prev->flags |= REQ_F_FAIL_LINK; |
f67676d1 | 3626 | goto err_req; |
2d28390a | 3627 | } |
c826bd7a | 3628 | trace_io_uring_link(ctx, req, prev); |
4493233e | 3629 | list_add_tail(&req->link_list, &prev->link_list); |
4e88d6e7 | 3630 | } else if (req->sqe->flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) { |
9e645e11 | 3631 | req->flags |= REQ_F_LINK; |
4e88d6e7 JA |
3632 | if (req->sqe->flags & IOSQE_IO_HARDLINK) |
3633 | req->flags |= REQ_F_HARDLINK; | |
9e645e11 | 3634 | |
9e645e11 JA |
3635 | INIT_LIST_HEAD(&req->link_list); |
3636 | *link = req; | |
3637 | } else { | |
a197f664 | 3638 | io_queue_sqe(req); |
9e645e11 | 3639 | } |
2e6e1fde PB |
3640 | |
3641 | return true; | |
9e645e11 JA |
3642 | } |
3643 | ||
9a56a232 JA |
3644 | /* |
3645 | * Batched submission is done, ensure local IO is flushed out. | |
3646 | */ | |
3647 | static void io_submit_state_end(struct io_submit_state *state) | |
3648 | { | |
3649 | blk_finish_plug(&state->plug); | |
3d6770fb | 3650 | io_file_put(state); |
2579f913 JA |
3651 | if (state->free_reqs) |
3652 | kmem_cache_free_bulk(req_cachep, state->free_reqs, | |
3653 | &state->reqs[state->cur_req]); | |
9a56a232 JA |
3654 | } |
3655 | ||
3656 | /* | |
3657 | * Start submission side cache. | |
3658 | */ | |
3659 | static void io_submit_state_start(struct io_submit_state *state, | |
22efde59 | 3660 | unsigned int max_ios) |
9a56a232 JA |
3661 | { |
3662 | blk_start_plug(&state->plug); | |
2579f913 | 3663 | state->free_reqs = 0; |
9a56a232 JA |
3664 | state->file = NULL; |
3665 | state->ios_left = max_ios; | |
3666 | } | |
3667 | ||
2b188cc1 JA |
3668 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
3669 | { | |
75b28aff | 3670 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 3671 | |
75b28aff | 3672 | if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) { |
2b188cc1 JA |
3673 | /* |
3674 | * Ensure any loads from the SQEs are done at this point, | |
3675 | * since once we write the new head, the application could | |
3676 | * write new data to them. | |
3677 | */ | |
75b28aff | 3678 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); |
2b188cc1 JA |
3679 | } |
3680 | } | |
3681 | ||
2b188cc1 | 3682 | /* |
d195a66e | 3683 | * Fetch an sqe, if one is available. Note that req->sqe will point to memory |
2b188cc1 JA |
3684 | * that is mapped by userspace. This means that care needs to be taken to |
3685 | * ensure that reads are stable, as we cannot rely on userspace always | |
3686 | * being a good citizen. If members of the sqe are validated and then later | |
3687 | * used, it's important that those reads are done through READ_ONCE() to | |
3688 | * prevent a re-load down the line. | |
3689 | */ | |
cf6fd4bd | 3690 | static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req) |
2b188cc1 | 3691 | { |
75b28aff HV |
3692 | struct io_rings *rings = ctx->rings; |
3693 | u32 *sq_array = ctx->sq_array; | |
2b188cc1 JA |
3694 | unsigned head; |
3695 | ||
3696 | /* | |
3697 | * The cached sq head (or cq tail) serves two purposes: | |
3698 | * | |
3699 | * 1) allows us to batch the cost of updating the user visible | |
3700 | * head updates. | |
3701 | * 2) allows the kernel side to track the head on its own, even | |
3702 | * though the application is the one updating it. | |
3703 | */ | |
3704 | head = ctx->cached_sq_head; | |
e523a29c | 3705 | /* make sure SQ entry isn't read before tail */ |
9835d6fa | 3706 | if (unlikely(head == smp_load_acquire(&rings->sq.tail))) |
2b188cc1 JA |
3707 | return false; |
3708 | ||
75b28aff | 3709 | head = READ_ONCE(sq_array[head & ctx->sq_mask]); |
9835d6fa | 3710 | if (likely(head < ctx->sq_entries)) { |
cf6fd4bd PB |
3711 | /* |
3712 | * All io need record the previous position, if LINK vs DARIN, | |
3713 | * it can be used to mark the position of the first IO in the | |
3714 | * link list. | |
3715 | */ | |
3716 | req->sequence = ctx->cached_sq_head; | |
3717 | req->sqe = &ctx->sq_sqes[head]; | |
d625c6ee JA |
3718 | req->opcode = READ_ONCE(req->sqe->opcode); |
3719 | req->user_data = READ_ONCE(req->sqe->user_data); | |
2b188cc1 JA |
3720 | ctx->cached_sq_head++; |
3721 | return true; | |
3722 | } | |
3723 | ||
3724 | /* drop invalid entries */ | |
3725 | ctx->cached_sq_head++; | |
498ccd9e JA |
3726 | ctx->cached_sq_dropped++; |
3727 | WRITE_ONCE(rings->sq_dropped, ctx->cached_sq_dropped); | |
2b188cc1 JA |
3728 | return false; |
3729 | } | |
3730 | ||
fb5ccc98 | 3731 | static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr, |
ae9428ca PB |
3732 | struct file *ring_file, int ring_fd, |
3733 | struct mm_struct **mm, bool async) | |
6c271ce2 JA |
3734 | { |
3735 | struct io_submit_state state, *statep = NULL; | |
9e645e11 | 3736 | struct io_kiocb *link = NULL; |
9e645e11 | 3737 | int i, submitted = 0; |
95a1b3ff | 3738 | bool mm_fault = false; |
6c271ce2 | 3739 | |
c4a2ed72 JA |
3740 | /* if we have a backlog and couldn't flush it all, return BUSY */ |
3741 | if (!list_empty(&ctx->cq_overflow_list) && | |
3742 | !io_cqring_overflow_flush(ctx, false)) | |
1d7bb1d5 | 3743 | return -EBUSY; |
6c271ce2 JA |
3744 | |
3745 | if (nr > IO_PLUG_THRESHOLD) { | |
22efde59 | 3746 | io_submit_state_start(&state, nr); |
6c271ce2 JA |
3747 | statep = &state; |
3748 | } | |
3749 | ||
3750 | for (i = 0; i < nr; i++) { | |
196be95c | 3751 | struct io_kiocb *req; |
50585b9a | 3752 | unsigned int sqe_flags; |
fb5ccc98 | 3753 | |
196be95c PB |
3754 | req = io_get_req(ctx, statep); |
3755 | if (unlikely(!req)) { | |
3756 | if (!submitted) | |
3757 | submitted = -EAGAIN; | |
fb5ccc98 | 3758 | break; |
196be95c | 3759 | } |
cf6fd4bd | 3760 | if (!io_get_sqring(ctx, req)) { |
196be95c PB |
3761 | __io_free_req(req); |
3762 | break; | |
3763 | } | |
fb5ccc98 | 3764 | |
d625c6ee | 3765 | if (io_req_needs_user(req) && !*mm) { |
95a1b3ff PB |
3766 | mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm); |
3767 | if (!mm_fault) { | |
3768 | use_mm(ctx->sqo_mm); | |
3769 | *mm = ctx->sqo_mm; | |
3770 | } | |
9e645e11 | 3771 | } |
9e645e11 | 3772 | |
2e6e1fde | 3773 | submitted++; |
cf6fd4bd | 3774 | sqe_flags = req->sqe->flags; |
50585b9a | 3775 | |
cf6fd4bd PB |
3776 | req->ring_file = ring_file; |
3777 | req->ring_fd = ring_fd; | |
3778 | req->has_user = *mm != NULL; | |
3779 | req->in_async = async; | |
3780 | req->needs_fixed_file = async; | |
d625c6ee | 3781 | trace_io_uring_submit_sqe(ctx, req->user_data, true, async); |
2e6e1fde PB |
3782 | if (!io_submit_sqe(req, statep, &link)) |
3783 | break; | |
e5eb6366 PB |
3784 | /* |
3785 | * If previous wasn't linked and we have a linked command, | |
3786 | * that's the end of the chain. Submit the previous link. | |
3787 | */ | |
ffbb8d6b | 3788 | if (!(sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) && link) { |
1b4a51b6 | 3789 | io_queue_link_head(link); |
e5eb6366 | 3790 | link = NULL; |
6c271ce2 | 3791 | } |
6c271ce2 JA |
3792 | } |
3793 | ||
9e645e11 | 3794 | if (link) |
1b4a51b6 | 3795 | io_queue_link_head(link); |
6c271ce2 JA |
3796 | if (statep) |
3797 | io_submit_state_end(&state); | |
3798 | ||
ae9428ca PB |
3799 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
3800 | io_commit_sqring(ctx); | |
3801 | ||
6c271ce2 JA |
3802 | return submitted; |
3803 | } | |
3804 | ||
3805 | static int io_sq_thread(void *data) | |
3806 | { | |
6c271ce2 JA |
3807 | struct io_ring_ctx *ctx = data; |
3808 | struct mm_struct *cur_mm = NULL; | |
181e448d | 3809 | const struct cred *old_cred; |
6c271ce2 JA |
3810 | mm_segment_t old_fs; |
3811 | DEFINE_WAIT(wait); | |
3812 | unsigned inflight; | |
3813 | unsigned long timeout; | |
c1edbf5f | 3814 | int ret; |
6c271ce2 | 3815 | |
206aefde | 3816 | complete(&ctx->completions[1]); |
a4c0b3de | 3817 | |
6c271ce2 JA |
3818 | old_fs = get_fs(); |
3819 | set_fs(USER_DS); | |
181e448d | 3820 | old_cred = override_creds(ctx->creds); |
6c271ce2 | 3821 | |
c1edbf5f | 3822 | ret = timeout = inflight = 0; |
2bbcd6d3 | 3823 | while (!kthread_should_park()) { |
fb5ccc98 | 3824 | unsigned int to_submit; |
6c271ce2 JA |
3825 | |
3826 | if (inflight) { | |
3827 | unsigned nr_events = 0; | |
3828 | ||
3829 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
2b2ed975 JA |
3830 | /* |
3831 | * inflight is the count of the maximum possible | |
3832 | * entries we submitted, but it can be smaller | |
3833 | * if we dropped some of them. If we don't have | |
3834 | * poll entries available, then we know that we | |
3835 | * have nothing left to poll for. Reset the | |
3836 | * inflight count to zero in that case. | |
3837 | */ | |
3838 | mutex_lock(&ctx->uring_lock); | |
3839 | if (!list_empty(&ctx->poll_list)) | |
3840 | __io_iopoll_check(ctx, &nr_events, 0); | |
3841 | else | |
3842 | inflight = 0; | |
3843 | mutex_unlock(&ctx->uring_lock); | |
6c271ce2 JA |
3844 | } else { |
3845 | /* | |
3846 | * Normal IO, just pretend everything completed. | |
3847 | * We don't have to poll completions for that. | |
3848 | */ | |
3849 | nr_events = inflight; | |
3850 | } | |
3851 | ||
3852 | inflight -= nr_events; | |
3853 | if (!inflight) | |
3854 | timeout = jiffies + ctx->sq_thread_idle; | |
3855 | } | |
3856 | ||
fb5ccc98 | 3857 | to_submit = io_sqring_entries(ctx); |
c1edbf5f JA |
3858 | |
3859 | /* | |
3860 | * If submit got -EBUSY, flag us as needing the application | |
3861 | * to enter the kernel to reap and flush events. | |
3862 | */ | |
3863 | if (!to_submit || ret == -EBUSY) { | |
6c271ce2 JA |
3864 | /* |
3865 | * We're polling. If we're within the defined idle | |
3866 | * period, then let us spin without work before going | |
c1edbf5f JA |
3867 | * to sleep. The exception is if we got EBUSY doing |
3868 | * more IO, we should wait for the application to | |
3869 | * reap events and wake us up. | |
6c271ce2 | 3870 | */ |
c1edbf5f JA |
3871 | if (inflight || |
3872 | (!time_after(jiffies, timeout) && ret != -EBUSY)) { | |
9831a90c | 3873 | cond_resched(); |
6c271ce2 JA |
3874 | continue; |
3875 | } | |
3876 | ||
3877 | /* | |
3878 | * Drop cur_mm before scheduling, we can't hold it for | |
3879 | * long periods (or over schedule()). Do this before | |
3880 | * adding ourselves to the waitqueue, as the unuse/drop | |
3881 | * may sleep. | |
3882 | */ | |
3883 | if (cur_mm) { | |
3884 | unuse_mm(cur_mm); | |
3885 | mmput(cur_mm); | |
3886 | cur_mm = NULL; | |
3887 | } | |
3888 | ||
3889 | prepare_to_wait(&ctx->sqo_wait, &wait, | |
3890 | TASK_INTERRUPTIBLE); | |
3891 | ||
3892 | /* Tell userspace we may need a wakeup call */ | |
75b28aff | 3893 | ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP; |
0d7bae69 SB |
3894 | /* make sure to read SQ tail after writing flags */ |
3895 | smp_mb(); | |
6c271ce2 | 3896 | |
fb5ccc98 | 3897 | to_submit = io_sqring_entries(ctx); |
c1edbf5f | 3898 | if (!to_submit || ret == -EBUSY) { |
2bbcd6d3 | 3899 | if (kthread_should_park()) { |
6c271ce2 JA |
3900 | finish_wait(&ctx->sqo_wait, &wait); |
3901 | break; | |
3902 | } | |
3903 | if (signal_pending(current)) | |
3904 | flush_signals(current); | |
3905 | schedule(); | |
3906 | finish_wait(&ctx->sqo_wait, &wait); | |
3907 | ||
75b28aff | 3908 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
3909 | continue; |
3910 | } | |
3911 | finish_wait(&ctx->sqo_wait, &wait); | |
3912 | ||
75b28aff | 3913 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
3914 | } |
3915 | ||
fb5ccc98 | 3916 | to_submit = min(to_submit, ctx->sq_entries); |
8a4955ff | 3917 | mutex_lock(&ctx->uring_lock); |
1d7bb1d5 | 3918 | ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true); |
8a4955ff | 3919 | mutex_unlock(&ctx->uring_lock); |
1d7bb1d5 JA |
3920 | if (ret > 0) |
3921 | inflight += ret; | |
6c271ce2 JA |
3922 | } |
3923 | ||
3924 | set_fs(old_fs); | |
3925 | if (cur_mm) { | |
3926 | unuse_mm(cur_mm); | |
3927 | mmput(cur_mm); | |
3928 | } | |
181e448d | 3929 | revert_creds(old_cred); |
06058632 | 3930 | |
2bbcd6d3 | 3931 | kthread_parkme(); |
06058632 | 3932 | |
6c271ce2 JA |
3933 | return 0; |
3934 | } | |
3935 | ||
bda52162 JA |
3936 | struct io_wait_queue { |
3937 | struct wait_queue_entry wq; | |
3938 | struct io_ring_ctx *ctx; | |
3939 | unsigned to_wait; | |
3940 | unsigned nr_timeouts; | |
3941 | }; | |
3942 | ||
1d7bb1d5 | 3943 | static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush) |
bda52162 JA |
3944 | { |
3945 | struct io_ring_ctx *ctx = iowq->ctx; | |
3946 | ||
3947 | /* | |
d195a66e | 3948 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
3949 | * started waiting. For timeouts, we always want to return to userspace, |
3950 | * regardless of event count. | |
3951 | */ | |
1d7bb1d5 | 3952 | return io_cqring_events(ctx, noflush) >= iowq->to_wait || |
bda52162 JA |
3953 | atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
3954 | } | |
3955 | ||
3956 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
3957 | int wake_flags, void *key) | |
3958 | { | |
3959 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
3960 | wq); | |
3961 | ||
1d7bb1d5 JA |
3962 | /* use noflush == true, as we can't safely rely on locking context */ |
3963 | if (!io_should_wake(iowq, true)) | |
bda52162 JA |
3964 | return -1; |
3965 | ||
3966 | return autoremove_wake_function(curr, mode, wake_flags, key); | |
3967 | } | |
3968 | ||
2b188cc1 JA |
3969 | /* |
3970 | * Wait until events become available, if we don't already have some. The | |
3971 | * application must reap them itself, as they reside on the shared cq ring. | |
3972 | */ | |
3973 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
3974 | const sigset_t __user *sig, size_t sigsz) | |
3975 | { | |
bda52162 JA |
3976 | struct io_wait_queue iowq = { |
3977 | .wq = { | |
3978 | .private = current, | |
3979 | .func = io_wake_function, | |
3980 | .entry = LIST_HEAD_INIT(iowq.wq.entry), | |
3981 | }, | |
3982 | .ctx = ctx, | |
3983 | .to_wait = min_events, | |
3984 | }; | |
75b28aff | 3985 | struct io_rings *rings = ctx->rings; |
e9ffa5c2 | 3986 | int ret = 0; |
2b188cc1 | 3987 | |
1d7bb1d5 | 3988 | if (io_cqring_events(ctx, false) >= min_events) |
2b188cc1 JA |
3989 | return 0; |
3990 | ||
3991 | if (sig) { | |
9e75ad5d AB |
3992 | #ifdef CONFIG_COMPAT |
3993 | if (in_compat_syscall()) | |
3994 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 3995 | sigsz); |
9e75ad5d AB |
3996 | else |
3997 | #endif | |
b772434b | 3998 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 3999 | |
2b188cc1 JA |
4000 | if (ret) |
4001 | return ret; | |
4002 | } | |
4003 | ||
bda52162 | 4004 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
c826bd7a | 4005 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 JA |
4006 | do { |
4007 | prepare_to_wait_exclusive(&ctx->wait, &iowq.wq, | |
4008 | TASK_INTERRUPTIBLE); | |
1d7bb1d5 | 4009 | if (io_should_wake(&iowq, false)) |
bda52162 JA |
4010 | break; |
4011 | schedule(); | |
4012 | if (signal_pending(current)) { | |
e9ffa5c2 | 4013 | ret = -EINTR; |
bda52162 JA |
4014 | break; |
4015 | } | |
4016 | } while (1); | |
4017 | finish_wait(&ctx->wait, &iowq.wq); | |
4018 | ||
e9ffa5c2 | 4019 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 4020 | |
75b28aff | 4021 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
4022 | } |
4023 | ||
6b06314c JA |
4024 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
4025 | { | |
4026 | #if defined(CONFIG_UNIX) | |
4027 | if (ctx->ring_sock) { | |
4028 | struct sock *sock = ctx->ring_sock->sk; | |
4029 | struct sk_buff *skb; | |
4030 | ||
4031 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
4032 | kfree_skb(skb); | |
4033 | } | |
4034 | #else | |
4035 | int i; | |
4036 | ||
65e19f54 JA |
4037 | for (i = 0; i < ctx->nr_user_files; i++) { |
4038 | struct file *file; | |
4039 | ||
4040 | file = io_file_from_index(ctx, i); | |
4041 | if (file) | |
4042 | fput(file); | |
4043 | } | |
6b06314c JA |
4044 | #endif |
4045 | } | |
4046 | ||
4047 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) | |
4048 | { | |
65e19f54 JA |
4049 | unsigned nr_tables, i; |
4050 | ||
4051 | if (!ctx->file_table) | |
6b06314c JA |
4052 | return -ENXIO; |
4053 | ||
4054 | __io_sqe_files_unregister(ctx); | |
65e19f54 JA |
4055 | nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE); |
4056 | for (i = 0; i < nr_tables; i++) | |
4057 | kfree(ctx->file_table[i].files); | |
4058 | kfree(ctx->file_table); | |
4059 | ctx->file_table = NULL; | |
6b06314c JA |
4060 | ctx->nr_user_files = 0; |
4061 | return 0; | |
4062 | } | |
4063 | ||
6c271ce2 JA |
4064 | static void io_sq_thread_stop(struct io_ring_ctx *ctx) |
4065 | { | |
4066 | if (ctx->sqo_thread) { | |
206aefde | 4067 | wait_for_completion(&ctx->completions[1]); |
2bbcd6d3 RP |
4068 | /* |
4069 | * The park is a bit of a work-around, without it we get | |
4070 | * warning spews on shutdown with SQPOLL set and affinity | |
4071 | * set to a single CPU. | |
4072 | */ | |
06058632 | 4073 | kthread_park(ctx->sqo_thread); |
6c271ce2 JA |
4074 | kthread_stop(ctx->sqo_thread); |
4075 | ctx->sqo_thread = NULL; | |
4076 | } | |
4077 | } | |
4078 | ||
6b06314c JA |
4079 | static void io_finish_async(struct io_ring_ctx *ctx) |
4080 | { | |
6c271ce2 JA |
4081 | io_sq_thread_stop(ctx); |
4082 | ||
561fb04a JA |
4083 | if (ctx->io_wq) { |
4084 | io_wq_destroy(ctx->io_wq); | |
4085 | ctx->io_wq = NULL; | |
6b06314c JA |
4086 | } |
4087 | } | |
4088 | ||
4089 | #if defined(CONFIG_UNIX) | |
4090 | static void io_destruct_skb(struct sk_buff *skb) | |
4091 | { | |
4092 | struct io_ring_ctx *ctx = skb->sk->sk_user_data; | |
8a997340 | 4093 | |
561fb04a JA |
4094 | if (ctx->io_wq) |
4095 | io_wq_flush(ctx->io_wq); | |
6b06314c | 4096 | |
6b06314c JA |
4097 | unix_destruct_scm(skb); |
4098 | } | |
4099 | ||
4100 | /* | |
4101 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
4102 | * the io_uring can be safely unregistered on process exit, even if we have | |
4103 | * loops in the file referencing. | |
4104 | */ | |
4105 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
4106 | { | |
4107 | struct sock *sk = ctx->ring_sock->sk; | |
4108 | struct scm_fp_list *fpl; | |
4109 | struct sk_buff *skb; | |
08a45173 | 4110 | int i, nr_files; |
6b06314c JA |
4111 | |
4112 | if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) { | |
4113 | unsigned long inflight = ctx->user->unix_inflight + nr; | |
4114 | ||
4115 | if (inflight > task_rlimit(current, RLIMIT_NOFILE)) | |
4116 | return -EMFILE; | |
4117 | } | |
4118 | ||
4119 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); | |
4120 | if (!fpl) | |
4121 | return -ENOMEM; | |
4122 | ||
4123 | skb = alloc_skb(0, GFP_KERNEL); | |
4124 | if (!skb) { | |
4125 | kfree(fpl); | |
4126 | return -ENOMEM; | |
4127 | } | |
4128 | ||
4129 | skb->sk = sk; | |
6b06314c | 4130 | |
08a45173 | 4131 | nr_files = 0; |
6b06314c JA |
4132 | fpl->user = get_uid(ctx->user); |
4133 | for (i = 0; i < nr; i++) { | |
65e19f54 JA |
4134 | struct file *file = io_file_from_index(ctx, i + offset); |
4135 | ||
4136 | if (!file) | |
08a45173 | 4137 | continue; |
65e19f54 | 4138 | fpl->fp[nr_files] = get_file(file); |
08a45173 JA |
4139 | unix_inflight(fpl->user, fpl->fp[nr_files]); |
4140 | nr_files++; | |
6b06314c JA |
4141 | } |
4142 | ||
08a45173 JA |
4143 | if (nr_files) { |
4144 | fpl->max = SCM_MAX_FD; | |
4145 | fpl->count = nr_files; | |
4146 | UNIXCB(skb).fp = fpl; | |
4147 | skb->destructor = io_destruct_skb; | |
4148 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); | |
4149 | skb_queue_head(&sk->sk_receive_queue, skb); | |
6b06314c | 4150 | |
08a45173 JA |
4151 | for (i = 0; i < nr_files; i++) |
4152 | fput(fpl->fp[i]); | |
4153 | } else { | |
4154 | kfree_skb(skb); | |
4155 | kfree(fpl); | |
4156 | } | |
6b06314c JA |
4157 | |
4158 | return 0; | |
4159 | } | |
4160 | ||
4161 | /* | |
4162 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
4163 | * causes regular reference counting to break down. We rely on the UNIX | |
4164 | * garbage collection to take care of this problem for us. | |
4165 | */ | |
4166 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
4167 | { | |
4168 | unsigned left, total; | |
4169 | int ret = 0; | |
4170 | ||
4171 | total = 0; | |
4172 | left = ctx->nr_user_files; | |
4173 | while (left) { | |
4174 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
4175 | |
4176 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
4177 | if (ret) | |
4178 | break; | |
4179 | left -= this_files; | |
4180 | total += this_files; | |
4181 | } | |
4182 | ||
4183 | if (!ret) | |
4184 | return 0; | |
4185 | ||
4186 | while (total < ctx->nr_user_files) { | |
65e19f54 JA |
4187 | struct file *file = io_file_from_index(ctx, total); |
4188 | ||
4189 | if (file) | |
4190 | fput(file); | |
6b06314c JA |
4191 | total++; |
4192 | } | |
4193 | ||
4194 | return ret; | |
4195 | } | |
4196 | #else | |
4197 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
4198 | { | |
4199 | return 0; | |
4200 | } | |
4201 | #endif | |
4202 | ||
65e19f54 JA |
4203 | static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables, |
4204 | unsigned nr_files) | |
4205 | { | |
4206 | int i; | |
4207 | ||
4208 | for (i = 0; i < nr_tables; i++) { | |
4209 | struct fixed_file_table *table = &ctx->file_table[i]; | |
4210 | unsigned this_files; | |
4211 | ||
4212 | this_files = min(nr_files, IORING_MAX_FILES_TABLE); | |
4213 | table->files = kcalloc(this_files, sizeof(struct file *), | |
4214 | GFP_KERNEL); | |
4215 | if (!table->files) | |
4216 | break; | |
4217 | nr_files -= this_files; | |
4218 | } | |
4219 | ||
4220 | if (i == nr_tables) | |
4221 | return 0; | |
4222 | ||
4223 | for (i = 0; i < nr_tables; i++) { | |
4224 | struct fixed_file_table *table = &ctx->file_table[i]; | |
4225 | kfree(table->files); | |
4226 | } | |
4227 | return 1; | |
4228 | } | |
4229 | ||
6b06314c JA |
4230 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, |
4231 | unsigned nr_args) | |
4232 | { | |
4233 | __s32 __user *fds = (__s32 __user *) arg; | |
65e19f54 | 4234 | unsigned nr_tables; |
6b06314c JA |
4235 | int fd, ret = 0; |
4236 | unsigned i; | |
4237 | ||
65e19f54 | 4238 | if (ctx->file_table) |
6b06314c JA |
4239 | return -EBUSY; |
4240 | if (!nr_args) | |
4241 | return -EINVAL; | |
4242 | if (nr_args > IORING_MAX_FIXED_FILES) | |
4243 | return -EMFILE; | |
4244 | ||
65e19f54 JA |
4245 | nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE); |
4246 | ctx->file_table = kcalloc(nr_tables, sizeof(struct fixed_file_table), | |
4247 | GFP_KERNEL); | |
4248 | if (!ctx->file_table) | |
6b06314c JA |
4249 | return -ENOMEM; |
4250 | ||
65e19f54 JA |
4251 | if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) { |
4252 | kfree(ctx->file_table); | |
46568e9b | 4253 | ctx->file_table = NULL; |
65e19f54 JA |
4254 | return -ENOMEM; |
4255 | } | |
4256 | ||
08a45173 | 4257 | for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { |
65e19f54 JA |
4258 | struct fixed_file_table *table; |
4259 | unsigned index; | |
4260 | ||
6b06314c JA |
4261 | ret = -EFAULT; |
4262 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) | |
4263 | break; | |
08a45173 JA |
4264 | /* allow sparse sets */ |
4265 | if (fd == -1) { | |
4266 | ret = 0; | |
4267 | continue; | |
4268 | } | |
6b06314c | 4269 | |
65e19f54 JA |
4270 | table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT]; |
4271 | index = i & IORING_FILE_TABLE_MASK; | |
4272 | table->files[index] = fget(fd); | |
6b06314c JA |
4273 | |
4274 | ret = -EBADF; | |
65e19f54 | 4275 | if (!table->files[index]) |
6b06314c JA |
4276 | break; |
4277 | /* | |
4278 | * Don't allow io_uring instances to be registered. If UNIX | |
4279 | * isn't enabled, then this causes a reference cycle and this | |
4280 | * instance can never get freed. If UNIX is enabled we'll | |
4281 | * handle it just fine, but there's still no point in allowing | |
4282 | * a ring fd as it doesn't support regular read/write anyway. | |
4283 | */ | |
65e19f54 JA |
4284 | if (table->files[index]->f_op == &io_uring_fops) { |
4285 | fput(table->files[index]); | |
6b06314c JA |
4286 | break; |
4287 | } | |
6b06314c JA |
4288 | ret = 0; |
4289 | } | |
4290 | ||
4291 | if (ret) { | |
65e19f54 JA |
4292 | for (i = 0; i < ctx->nr_user_files; i++) { |
4293 | struct file *file; | |
6b06314c | 4294 | |
65e19f54 JA |
4295 | file = io_file_from_index(ctx, i); |
4296 | if (file) | |
4297 | fput(file); | |
4298 | } | |
4299 | for (i = 0; i < nr_tables; i++) | |
4300 | kfree(ctx->file_table[i].files); | |
6b06314c | 4301 | |
65e19f54 JA |
4302 | kfree(ctx->file_table); |
4303 | ctx->file_table = NULL; | |
6b06314c JA |
4304 | ctx->nr_user_files = 0; |
4305 | return ret; | |
4306 | } | |
4307 | ||
4308 | ret = io_sqe_files_scm(ctx); | |
4309 | if (ret) | |
4310 | io_sqe_files_unregister(ctx); | |
4311 | ||
4312 | return ret; | |
4313 | } | |
4314 | ||
c3a31e60 JA |
4315 | static void io_sqe_file_unregister(struct io_ring_ctx *ctx, int index) |
4316 | { | |
4317 | #if defined(CONFIG_UNIX) | |
65e19f54 | 4318 | struct file *file = io_file_from_index(ctx, index); |
c3a31e60 JA |
4319 | struct sock *sock = ctx->ring_sock->sk; |
4320 | struct sk_buff_head list, *head = &sock->sk_receive_queue; | |
4321 | struct sk_buff *skb; | |
4322 | int i; | |
4323 | ||
4324 | __skb_queue_head_init(&list); | |
4325 | ||
4326 | /* | |
4327 | * Find the skb that holds this file in its SCM_RIGHTS. When found, | |
4328 | * remove this entry and rearrange the file array. | |
4329 | */ | |
4330 | skb = skb_dequeue(head); | |
4331 | while (skb) { | |
4332 | struct scm_fp_list *fp; | |
4333 | ||
4334 | fp = UNIXCB(skb).fp; | |
4335 | for (i = 0; i < fp->count; i++) { | |
4336 | int left; | |
4337 | ||
4338 | if (fp->fp[i] != file) | |
4339 | continue; | |
4340 | ||
4341 | unix_notinflight(fp->user, fp->fp[i]); | |
4342 | left = fp->count - 1 - i; | |
4343 | if (left) { | |
4344 | memmove(&fp->fp[i], &fp->fp[i + 1], | |
4345 | left * sizeof(struct file *)); | |
4346 | } | |
4347 | fp->count--; | |
4348 | if (!fp->count) { | |
4349 | kfree_skb(skb); | |
4350 | skb = NULL; | |
4351 | } else { | |
4352 | __skb_queue_tail(&list, skb); | |
4353 | } | |
4354 | fput(file); | |
4355 | file = NULL; | |
4356 | break; | |
4357 | } | |
4358 | ||
4359 | if (!file) | |
4360 | break; | |
4361 | ||
4362 | __skb_queue_tail(&list, skb); | |
4363 | ||
4364 | skb = skb_dequeue(head); | |
4365 | } | |
4366 | ||
4367 | if (skb_peek(&list)) { | |
4368 | spin_lock_irq(&head->lock); | |
4369 | while ((skb = __skb_dequeue(&list)) != NULL) | |
4370 | __skb_queue_tail(head, skb); | |
4371 | spin_unlock_irq(&head->lock); | |
4372 | } | |
4373 | #else | |
65e19f54 | 4374 | fput(io_file_from_index(ctx, index)); |
c3a31e60 JA |
4375 | #endif |
4376 | } | |
4377 | ||
4378 | static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file, | |
4379 | int index) | |
4380 | { | |
4381 | #if defined(CONFIG_UNIX) | |
4382 | struct sock *sock = ctx->ring_sock->sk; | |
4383 | struct sk_buff_head *head = &sock->sk_receive_queue; | |
4384 | struct sk_buff *skb; | |
4385 | ||
4386 | /* | |
4387 | * See if we can merge this file into an existing skb SCM_RIGHTS | |
4388 | * file set. If there's no room, fall back to allocating a new skb | |
4389 | * and filling it in. | |
4390 | */ | |
4391 | spin_lock_irq(&head->lock); | |
4392 | skb = skb_peek(head); | |
4393 | if (skb) { | |
4394 | struct scm_fp_list *fpl = UNIXCB(skb).fp; | |
4395 | ||
4396 | if (fpl->count < SCM_MAX_FD) { | |
4397 | __skb_unlink(skb, head); | |
4398 | spin_unlock_irq(&head->lock); | |
4399 | fpl->fp[fpl->count] = get_file(file); | |
4400 | unix_inflight(fpl->user, fpl->fp[fpl->count]); | |
4401 | fpl->count++; | |
4402 | spin_lock_irq(&head->lock); | |
4403 | __skb_queue_head(head, skb); | |
4404 | } else { | |
4405 | skb = NULL; | |
4406 | } | |
4407 | } | |
4408 | spin_unlock_irq(&head->lock); | |
4409 | ||
4410 | if (skb) { | |
4411 | fput(file); | |
4412 | return 0; | |
4413 | } | |
4414 | ||
4415 | return __io_sqe_files_scm(ctx, 1, index); | |
4416 | #else | |
4417 | return 0; | |
4418 | #endif | |
4419 | } | |
4420 | ||
4421 | static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg, | |
4422 | unsigned nr_args) | |
4423 | { | |
4424 | struct io_uring_files_update up; | |
4425 | __s32 __user *fds; | |
4426 | int fd, i, err; | |
4427 | __u32 done; | |
4428 | ||
65e19f54 | 4429 | if (!ctx->file_table) |
c3a31e60 JA |
4430 | return -ENXIO; |
4431 | if (!nr_args) | |
4432 | return -EINVAL; | |
4433 | if (copy_from_user(&up, arg, sizeof(up))) | |
4434 | return -EFAULT; | |
4435 | if (check_add_overflow(up.offset, nr_args, &done)) | |
4436 | return -EOVERFLOW; | |
4437 | if (done > ctx->nr_user_files) | |
4438 | return -EINVAL; | |
4439 | ||
4440 | done = 0; | |
4441 | fds = (__s32 __user *) up.fds; | |
4442 | while (nr_args) { | |
65e19f54 JA |
4443 | struct fixed_file_table *table; |
4444 | unsigned index; | |
4445 | ||
c3a31e60 JA |
4446 | err = 0; |
4447 | if (copy_from_user(&fd, &fds[done], sizeof(fd))) { | |
4448 | err = -EFAULT; | |
4449 | break; | |
4450 | } | |
4451 | i = array_index_nospec(up.offset, ctx->nr_user_files); | |
65e19f54 JA |
4452 | table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT]; |
4453 | index = i & IORING_FILE_TABLE_MASK; | |
4454 | if (table->files[index]) { | |
c3a31e60 | 4455 | io_sqe_file_unregister(ctx, i); |
65e19f54 | 4456 | table->files[index] = NULL; |
c3a31e60 JA |
4457 | } |
4458 | if (fd != -1) { | |
4459 | struct file *file; | |
4460 | ||
4461 | file = fget(fd); | |
4462 | if (!file) { | |
4463 | err = -EBADF; | |
4464 | break; | |
4465 | } | |
4466 | /* | |
4467 | * Don't allow io_uring instances to be registered. If | |
4468 | * UNIX isn't enabled, then this causes a reference | |
4469 | * cycle and this instance can never get freed. If UNIX | |
4470 | * is enabled we'll handle it just fine, but there's | |
4471 | * still no point in allowing a ring fd as it doesn't | |
4472 | * support regular read/write anyway. | |
4473 | */ | |
4474 | if (file->f_op == &io_uring_fops) { | |
4475 | fput(file); | |
4476 | err = -EBADF; | |
4477 | break; | |
4478 | } | |
65e19f54 | 4479 | table->files[index] = file; |
c3a31e60 JA |
4480 | err = io_sqe_file_register(ctx, file, i); |
4481 | if (err) | |
4482 | break; | |
4483 | } | |
4484 | nr_args--; | |
4485 | done++; | |
4486 | up.offset++; | |
4487 | } | |
4488 | ||
4489 | return done ? done : err; | |
4490 | } | |
4491 | ||
7d723065 JA |
4492 | static void io_put_work(struct io_wq_work *work) |
4493 | { | |
4494 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
4495 | ||
4496 | io_put_req(req); | |
4497 | } | |
4498 | ||
4499 | static void io_get_work(struct io_wq_work *work) | |
4500 | { | |
4501 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
4502 | ||
4503 | refcount_inc(&req->refs); | |
4504 | } | |
4505 | ||
6c271ce2 JA |
4506 | static int io_sq_offload_start(struct io_ring_ctx *ctx, |
4507 | struct io_uring_params *p) | |
2b188cc1 | 4508 | { |
576a347b | 4509 | struct io_wq_data data; |
561fb04a | 4510 | unsigned concurrency; |
2b188cc1 JA |
4511 | int ret; |
4512 | ||
6c271ce2 | 4513 | init_waitqueue_head(&ctx->sqo_wait); |
2b188cc1 JA |
4514 | mmgrab(current->mm); |
4515 | ctx->sqo_mm = current->mm; | |
4516 | ||
6c271ce2 | 4517 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3ec482d1 JA |
4518 | ret = -EPERM; |
4519 | if (!capable(CAP_SYS_ADMIN)) | |
4520 | goto err; | |
4521 | ||
917257da JA |
4522 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
4523 | if (!ctx->sq_thread_idle) | |
4524 | ctx->sq_thread_idle = HZ; | |
4525 | ||
6c271ce2 | 4526 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 4527 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 4528 | |
917257da | 4529 | ret = -EINVAL; |
44a9bd18 JA |
4530 | if (cpu >= nr_cpu_ids) |
4531 | goto err; | |
7889f44d | 4532 | if (!cpu_online(cpu)) |
917257da JA |
4533 | goto err; |
4534 | ||
6c271ce2 JA |
4535 | ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread, |
4536 | ctx, cpu, | |
4537 | "io_uring-sq"); | |
4538 | } else { | |
4539 | ctx->sqo_thread = kthread_create(io_sq_thread, ctx, | |
4540 | "io_uring-sq"); | |
4541 | } | |
4542 | if (IS_ERR(ctx->sqo_thread)) { | |
4543 | ret = PTR_ERR(ctx->sqo_thread); | |
4544 | ctx->sqo_thread = NULL; | |
4545 | goto err; | |
4546 | } | |
4547 | wake_up_process(ctx->sqo_thread); | |
4548 | } else if (p->flags & IORING_SETUP_SQ_AFF) { | |
4549 | /* Can't have SQ_AFF without SQPOLL */ | |
4550 | ret = -EINVAL; | |
4551 | goto err; | |
4552 | } | |
4553 | ||
576a347b JA |
4554 | data.mm = ctx->sqo_mm; |
4555 | data.user = ctx->user; | |
181e448d | 4556 | data.creds = ctx->creds; |
576a347b JA |
4557 | data.get_work = io_get_work; |
4558 | data.put_work = io_put_work; | |
4559 | ||
561fb04a JA |
4560 | /* Do QD, or 4 * CPUS, whatever is smallest */ |
4561 | concurrency = min(ctx->sq_entries, 4 * num_online_cpus()); | |
576a347b | 4562 | ctx->io_wq = io_wq_create(concurrency, &data); |
975c99a5 JA |
4563 | if (IS_ERR(ctx->io_wq)) { |
4564 | ret = PTR_ERR(ctx->io_wq); | |
4565 | ctx->io_wq = NULL; | |
2b188cc1 JA |
4566 | goto err; |
4567 | } | |
4568 | ||
4569 | return 0; | |
4570 | err: | |
54a91f3b | 4571 | io_finish_async(ctx); |
2b188cc1 JA |
4572 | mmdrop(ctx->sqo_mm); |
4573 | ctx->sqo_mm = NULL; | |
4574 | return ret; | |
4575 | } | |
4576 | ||
4577 | static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages) | |
4578 | { | |
4579 | atomic_long_sub(nr_pages, &user->locked_vm); | |
4580 | } | |
4581 | ||
4582 | static int io_account_mem(struct user_struct *user, unsigned long nr_pages) | |
4583 | { | |
4584 | unsigned long page_limit, cur_pages, new_pages; | |
4585 | ||
4586 | /* Don't allow more pages than we can safely lock */ | |
4587 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
4588 | ||
4589 | do { | |
4590 | cur_pages = atomic_long_read(&user->locked_vm); | |
4591 | new_pages = cur_pages + nr_pages; | |
4592 | if (new_pages > page_limit) | |
4593 | return -ENOMEM; | |
4594 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
4595 | new_pages) != cur_pages); | |
4596 | ||
4597 | return 0; | |
4598 | } | |
4599 | ||
4600 | static void io_mem_free(void *ptr) | |
4601 | { | |
52e04ef4 MR |
4602 | struct page *page; |
4603 | ||
4604 | if (!ptr) | |
4605 | return; | |
2b188cc1 | 4606 | |
52e04ef4 | 4607 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
4608 | if (put_page_testzero(page)) |
4609 | free_compound_page(page); | |
4610 | } | |
4611 | ||
4612 | static void *io_mem_alloc(size_t size) | |
4613 | { | |
4614 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
4615 | __GFP_NORETRY; | |
4616 | ||
4617 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
4618 | } | |
4619 | ||
75b28aff HV |
4620 | static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries, |
4621 | size_t *sq_offset) | |
4622 | { | |
4623 | struct io_rings *rings; | |
4624 | size_t off, sq_array_size; | |
4625 | ||
4626 | off = struct_size(rings, cqes, cq_entries); | |
4627 | if (off == SIZE_MAX) | |
4628 | return SIZE_MAX; | |
4629 | ||
4630 | #ifdef CONFIG_SMP | |
4631 | off = ALIGN(off, SMP_CACHE_BYTES); | |
4632 | if (off == 0) | |
4633 | return SIZE_MAX; | |
4634 | #endif | |
4635 | ||
4636 | sq_array_size = array_size(sizeof(u32), sq_entries); | |
4637 | if (sq_array_size == SIZE_MAX) | |
4638 | return SIZE_MAX; | |
4639 | ||
4640 | if (check_add_overflow(off, sq_array_size, &off)) | |
4641 | return SIZE_MAX; | |
4642 | ||
4643 | if (sq_offset) | |
4644 | *sq_offset = off; | |
4645 | ||
4646 | return off; | |
4647 | } | |
4648 | ||
2b188cc1 JA |
4649 | static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries) |
4650 | { | |
75b28aff | 4651 | size_t pages; |
2b188cc1 | 4652 | |
75b28aff HV |
4653 | pages = (size_t)1 << get_order( |
4654 | rings_size(sq_entries, cq_entries, NULL)); | |
4655 | pages += (size_t)1 << get_order( | |
4656 | array_size(sizeof(struct io_uring_sqe), sq_entries)); | |
2b188cc1 | 4657 | |
75b28aff | 4658 | return pages; |
2b188cc1 JA |
4659 | } |
4660 | ||
edafccee JA |
4661 | static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx) |
4662 | { | |
4663 | int i, j; | |
4664 | ||
4665 | if (!ctx->user_bufs) | |
4666 | return -ENXIO; | |
4667 | ||
4668 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
4669 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
4670 | ||
4671 | for (j = 0; j < imu->nr_bvecs; j++) | |
27c4d3a3 | 4672 | put_user_page(imu->bvec[j].bv_page); |
edafccee JA |
4673 | |
4674 | if (ctx->account_mem) | |
4675 | io_unaccount_mem(ctx->user, imu->nr_bvecs); | |
d4ef6475 | 4676 | kvfree(imu->bvec); |
edafccee JA |
4677 | imu->nr_bvecs = 0; |
4678 | } | |
4679 | ||
4680 | kfree(ctx->user_bufs); | |
4681 | ctx->user_bufs = NULL; | |
4682 | ctx->nr_user_bufs = 0; | |
4683 | return 0; | |
4684 | } | |
4685 | ||
4686 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
4687 | void __user *arg, unsigned index) | |
4688 | { | |
4689 | struct iovec __user *src; | |
4690 | ||
4691 | #ifdef CONFIG_COMPAT | |
4692 | if (ctx->compat) { | |
4693 | struct compat_iovec __user *ciovs; | |
4694 | struct compat_iovec ciov; | |
4695 | ||
4696 | ciovs = (struct compat_iovec __user *) arg; | |
4697 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
4698 | return -EFAULT; | |
4699 | ||
4700 | dst->iov_base = (void __user *) (unsigned long) ciov.iov_base; | |
4701 | dst->iov_len = ciov.iov_len; | |
4702 | return 0; | |
4703 | } | |
4704 | #endif | |
4705 | src = (struct iovec __user *) arg; | |
4706 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
4707 | return -EFAULT; | |
4708 | return 0; | |
4709 | } | |
4710 | ||
4711 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, | |
4712 | unsigned nr_args) | |
4713 | { | |
4714 | struct vm_area_struct **vmas = NULL; | |
4715 | struct page **pages = NULL; | |
4716 | int i, j, got_pages = 0; | |
4717 | int ret = -EINVAL; | |
4718 | ||
4719 | if (ctx->user_bufs) | |
4720 | return -EBUSY; | |
4721 | if (!nr_args || nr_args > UIO_MAXIOV) | |
4722 | return -EINVAL; | |
4723 | ||
4724 | ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf), | |
4725 | GFP_KERNEL); | |
4726 | if (!ctx->user_bufs) | |
4727 | return -ENOMEM; | |
4728 | ||
4729 | for (i = 0; i < nr_args; i++) { | |
4730 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
4731 | unsigned long off, start, end, ubuf; | |
4732 | int pret, nr_pages; | |
4733 | struct iovec iov; | |
4734 | size_t size; | |
4735 | ||
4736 | ret = io_copy_iov(ctx, &iov, arg, i); | |
4737 | if (ret) | |
a278682d | 4738 | goto err; |
edafccee JA |
4739 | |
4740 | /* | |
4741 | * Don't impose further limits on the size and buffer | |
4742 | * constraints here, we'll -EINVAL later when IO is | |
4743 | * submitted if they are wrong. | |
4744 | */ | |
4745 | ret = -EFAULT; | |
4746 | if (!iov.iov_base || !iov.iov_len) | |
4747 | goto err; | |
4748 | ||
4749 | /* arbitrary limit, but we need something */ | |
4750 | if (iov.iov_len > SZ_1G) | |
4751 | goto err; | |
4752 | ||
4753 | ubuf = (unsigned long) iov.iov_base; | |
4754 | end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
4755 | start = ubuf >> PAGE_SHIFT; | |
4756 | nr_pages = end - start; | |
4757 | ||
4758 | if (ctx->account_mem) { | |
4759 | ret = io_account_mem(ctx->user, nr_pages); | |
4760 | if (ret) | |
4761 | goto err; | |
4762 | } | |
4763 | ||
4764 | ret = 0; | |
4765 | if (!pages || nr_pages > got_pages) { | |
4766 | kfree(vmas); | |
4767 | kfree(pages); | |
d4ef6475 | 4768 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), |
edafccee | 4769 | GFP_KERNEL); |
d4ef6475 | 4770 | vmas = kvmalloc_array(nr_pages, |
edafccee JA |
4771 | sizeof(struct vm_area_struct *), |
4772 | GFP_KERNEL); | |
4773 | if (!pages || !vmas) { | |
4774 | ret = -ENOMEM; | |
4775 | if (ctx->account_mem) | |
4776 | io_unaccount_mem(ctx->user, nr_pages); | |
4777 | goto err; | |
4778 | } | |
4779 | got_pages = nr_pages; | |
4780 | } | |
4781 | ||
d4ef6475 | 4782 | imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec), |
edafccee JA |
4783 | GFP_KERNEL); |
4784 | ret = -ENOMEM; | |
4785 | if (!imu->bvec) { | |
4786 | if (ctx->account_mem) | |
4787 | io_unaccount_mem(ctx->user, nr_pages); | |
4788 | goto err; | |
4789 | } | |
4790 | ||
4791 | ret = 0; | |
4792 | down_read(¤t->mm->mmap_sem); | |
932f4a63 IW |
4793 | pret = get_user_pages(ubuf, nr_pages, |
4794 | FOLL_WRITE | FOLL_LONGTERM, | |
4795 | pages, vmas); | |
edafccee JA |
4796 | if (pret == nr_pages) { |
4797 | /* don't support file backed memory */ | |
4798 | for (j = 0; j < nr_pages; j++) { | |
4799 | struct vm_area_struct *vma = vmas[j]; | |
4800 | ||
4801 | if (vma->vm_file && | |
4802 | !is_file_hugepages(vma->vm_file)) { | |
4803 | ret = -EOPNOTSUPP; | |
4804 | break; | |
4805 | } | |
4806 | } | |
4807 | } else { | |
4808 | ret = pret < 0 ? pret : -EFAULT; | |
4809 | } | |
4810 | up_read(¤t->mm->mmap_sem); | |
4811 | if (ret) { | |
4812 | /* | |
4813 | * if we did partial map, or found file backed vmas, | |
4814 | * release any pages we did get | |
4815 | */ | |
27c4d3a3 JH |
4816 | if (pret > 0) |
4817 | put_user_pages(pages, pret); | |
edafccee JA |
4818 | if (ctx->account_mem) |
4819 | io_unaccount_mem(ctx->user, nr_pages); | |
d4ef6475 | 4820 | kvfree(imu->bvec); |
edafccee JA |
4821 | goto err; |
4822 | } | |
4823 | ||
4824 | off = ubuf & ~PAGE_MASK; | |
4825 | size = iov.iov_len; | |
4826 | for (j = 0; j < nr_pages; j++) { | |
4827 | size_t vec_len; | |
4828 | ||
4829 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
4830 | imu->bvec[j].bv_page = pages[j]; | |
4831 | imu->bvec[j].bv_len = vec_len; | |
4832 | imu->bvec[j].bv_offset = off; | |
4833 | off = 0; | |
4834 | size -= vec_len; | |
4835 | } | |
4836 | /* store original address for later verification */ | |
4837 | imu->ubuf = ubuf; | |
4838 | imu->len = iov.iov_len; | |
4839 | imu->nr_bvecs = nr_pages; | |
4840 | ||
4841 | ctx->nr_user_bufs++; | |
4842 | } | |
d4ef6475 MR |
4843 | kvfree(pages); |
4844 | kvfree(vmas); | |
edafccee JA |
4845 | return 0; |
4846 | err: | |
d4ef6475 MR |
4847 | kvfree(pages); |
4848 | kvfree(vmas); | |
edafccee JA |
4849 | io_sqe_buffer_unregister(ctx); |
4850 | return ret; | |
4851 | } | |
4852 | ||
9b402849 JA |
4853 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
4854 | { | |
4855 | __s32 __user *fds = arg; | |
4856 | int fd; | |
4857 | ||
4858 | if (ctx->cq_ev_fd) | |
4859 | return -EBUSY; | |
4860 | ||
4861 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
4862 | return -EFAULT; | |
4863 | ||
4864 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
4865 | if (IS_ERR(ctx->cq_ev_fd)) { | |
4866 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
4867 | ctx->cq_ev_fd = NULL; | |
4868 | return ret; | |
4869 | } | |
4870 | ||
4871 | return 0; | |
4872 | } | |
4873 | ||
4874 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
4875 | { | |
4876 | if (ctx->cq_ev_fd) { | |
4877 | eventfd_ctx_put(ctx->cq_ev_fd); | |
4878 | ctx->cq_ev_fd = NULL; | |
4879 | return 0; | |
4880 | } | |
4881 | ||
4882 | return -ENXIO; | |
4883 | } | |
4884 | ||
2b188cc1 JA |
4885 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
4886 | { | |
6b06314c | 4887 | io_finish_async(ctx); |
2b188cc1 JA |
4888 | if (ctx->sqo_mm) |
4889 | mmdrop(ctx->sqo_mm); | |
def596e9 JA |
4890 | |
4891 | io_iopoll_reap_events(ctx); | |
edafccee | 4892 | io_sqe_buffer_unregister(ctx); |
6b06314c | 4893 | io_sqe_files_unregister(ctx); |
9b402849 | 4894 | io_eventfd_unregister(ctx); |
def596e9 | 4895 | |
2b188cc1 | 4896 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
4897 | if (ctx->ring_sock) { |
4898 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 4899 | sock_release(ctx->ring_sock); |
355e8d26 | 4900 | } |
2b188cc1 JA |
4901 | #endif |
4902 | ||
75b28aff | 4903 | io_mem_free(ctx->rings); |
2b188cc1 | 4904 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
4905 | |
4906 | percpu_ref_exit(&ctx->refs); | |
4907 | if (ctx->account_mem) | |
4908 | io_unaccount_mem(ctx->user, | |
4909 | ring_pages(ctx->sq_entries, ctx->cq_entries)); | |
4910 | free_uid(ctx->user); | |
181e448d | 4911 | put_cred(ctx->creds); |
206aefde | 4912 | kfree(ctx->completions); |
78076bb6 | 4913 | kfree(ctx->cancel_hash); |
0ddf92e8 | 4914 | kmem_cache_free(req_cachep, ctx->fallback_req); |
2b188cc1 JA |
4915 | kfree(ctx); |
4916 | } | |
4917 | ||
4918 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
4919 | { | |
4920 | struct io_ring_ctx *ctx = file->private_data; | |
4921 | __poll_t mask = 0; | |
4922 | ||
4923 | poll_wait(file, &ctx->cq_wait, wait); | |
4f7067c3 SB |
4924 | /* |
4925 | * synchronizes with barrier from wq_has_sleeper call in | |
4926 | * io_commit_cqring | |
4927 | */ | |
2b188cc1 | 4928 | smp_rmb(); |
75b28aff HV |
4929 | if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head != |
4930 | ctx->rings->sq_ring_entries) | |
2b188cc1 | 4931 | mask |= EPOLLOUT | EPOLLWRNORM; |
daa5de54 | 4932 | if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail) |
2b188cc1 JA |
4933 | mask |= EPOLLIN | EPOLLRDNORM; |
4934 | ||
4935 | return mask; | |
4936 | } | |
4937 | ||
4938 | static int io_uring_fasync(int fd, struct file *file, int on) | |
4939 | { | |
4940 | struct io_ring_ctx *ctx = file->private_data; | |
4941 | ||
4942 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
4943 | } | |
4944 | ||
4945 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) | |
4946 | { | |
4947 | mutex_lock(&ctx->uring_lock); | |
4948 | percpu_ref_kill(&ctx->refs); | |
4949 | mutex_unlock(&ctx->uring_lock); | |
4950 | ||
5262f567 | 4951 | io_kill_timeouts(ctx); |
221c5eb2 | 4952 | io_poll_remove_all(ctx); |
561fb04a JA |
4953 | |
4954 | if (ctx->io_wq) | |
4955 | io_wq_cancel_all(ctx->io_wq); | |
4956 | ||
def596e9 | 4957 | io_iopoll_reap_events(ctx); |
15dff286 JA |
4958 | /* if we failed setting up the ctx, we might not have any rings */ |
4959 | if (ctx->rings) | |
4960 | io_cqring_overflow_flush(ctx, true); | |
206aefde | 4961 | wait_for_completion(&ctx->completions[0]); |
2b188cc1 JA |
4962 | io_ring_ctx_free(ctx); |
4963 | } | |
4964 | ||
4965 | static int io_uring_release(struct inode *inode, struct file *file) | |
4966 | { | |
4967 | struct io_ring_ctx *ctx = file->private_data; | |
4968 | ||
4969 | file->private_data = NULL; | |
4970 | io_ring_ctx_wait_and_kill(ctx); | |
4971 | return 0; | |
4972 | } | |
4973 | ||
fcb323cc JA |
4974 | static void io_uring_cancel_files(struct io_ring_ctx *ctx, |
4975 | struct files_struct *files) | |
4976 | { | |
4977 | struct io_kiocb *req; | |
4978 | DEFINE_WAIT(wait); | |
4979 | ||
4980 | while (!list_empty_careful(&ctx->inflight_list)) { | |
768134d4 | 4981 | struct io_kiocb *cancel_req = NULL; |
fcb323cc JA |
4982 | |
4983 | spin_lock_irq(&ctx->inflight_lock); | |
4984 | list_for_each_entry(req, &ctx->inflight_list, inflight_entry) { | |
768134d4 JA |
4985 | if (req->work.files != files) |
4986 | continue; | |
4987 | /* req is being completed, ignore */ | |
4988 | if (!refcount_inc_not_zero(&req->refs)) | |
4989 | continue; | |
4990 | cancel_req = req; | |
4991 | break; | |
fcb323cc | 4992 | } |
768134d4 | 4993 | if (cancel_req) |
fcb323cc | 4994 | prepare_to_wait(&ctx->inflight_wait, &wait, |
768134d4 | 4995 | TASK_UNINTERRUPTIBLE); |
fcb323cc JA |
4996 | spin_unlock_irq(&ctx->inflight_lock); |
4997 | ||
768134d4 JA |
4998 | /* We need to keep going until we don't find a matching req */ |
4999 | if (!cancel_req) | |
fcb323cc | 5000 | break; |
2f6d9b9d BL |
5001 | |
5002 | io_wq_cancel_work(ctx->io_wq, &cancel_req->work); | |
5003 | io_put_req(cancel_req); | |
fcb323cc JA |
5004 | schedule(); |
5005 | } | |
768134d4 | 5006 | finish_wait(&ctx->inflight_wait, &wait); |
fcb323cc JA |
5007 | } |
5008 | ||
5009 | static int io_uring_flush(struct file *file, void *data) | |
5010 | { | |
5011 | struct io_ring_ctx *ctx = file->private_data; | |
5012 | ||
5013 | io_uring_cancel_files(ctx, data); | |
1d7bb1d5 JA |
5014 | if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) { |
5015 | io_cqring_overflow_flush(ctx, true); | |
fcb323cc | 5016 | io_wq_cancel_all(ctx->io_wq); |
1d7bb1d5 | 5017 | } |
fcb323cc JA |
5018 | return 0; |
5019 | } | |
5020 | ||
6c5c240e RP |
5021 | static void *io_uring_validate_mmap_request(struct file *file, |
5022 | loff_t pgoff, size_t sz) | |
2b188cc1 | 5023 | { |
2b188cc1 | 5024 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 5025 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
5026 | struct page *page; |
5027 | void *ptr; | |
5028 | ||
5029 | switch (offset) { | |
5030 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
5031 | case IORING_OFF_CQ_RING: |
5032 | ptr = ctx->rings; | |
2b188cc1 JA |
5033 | break; |
5034 | case IORING_OFF_SQES: | |
5035 | ptr = ctx->sq_sqes; | |
5036 | break; | |
2b188cc1 | 5037 | default: |
6c5c240e | 5038 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
5039 | } |
5040 | ||
5041 | page = virt_to_head_page(ptr); | |
a50b854e | 5042 | if (sz > page_size(page)) |
6c5c240e RP |
5043 | return ERR_PTR(-EINVAL); |
5044 | ||
5045 | return ptr; | |
5046 | } | |
5047 | ||
5048 | #ifdef CONFIG_MMU | |
5049 | ||
5050 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
5051 | { | |
5052 | size_t sz = vma->vm_end - vma->vm_start; | |
5053 | unsigned long pfn; | |
5054 | void *ptr; | |
5055 | ||
5056 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
5057 | if (IS_ERR(ptr)) | |
5058 | return PTR_ERR(ptr); | |
2b188cc1 JA |
5059 | |
5060 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
5061 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
5062 | } | |
5063 | ||
6c5c240e RP |
5064 | #else /* !CONFIG_MMU */ |
5065 | ||
5066 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
5067 | { | |
5068 | return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL; | |
5069 | } | |
5070 | ||
5071 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
5072 | { | |
5073 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
5074 | } | |
5075 | ||
5076 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
5077 | unsigned long addr, unsigned long len, | |
5078 | unsigned long pgoff, unsigned long flags) | |
5079 | { | |
5080 | void *ptr; | |
5081 | ||
5082 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
5083 | if (IS_ERR(ptr)) | |
5084 | return PTR_ERR(ptr); | |
5085 | ||
5086 | return (unsigned long) ptr; | |
5087 | } | |
5088 | ||
5089 | #endif /* !CONFIG_MMU */ | |
5090 | ||
2b188cc1 JA |
5091 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
5092 | u32, min_complete, u32, flags, const sigset_t __user *, sig, | |
5093 | size_t, sigsz) | |
5094 | { | |
5095 | struct io_ring_ctx *ctx; | |
5096 | long ret = -EBADF; | |
5097 | int submitted = 0; | |
5098 | struct fd f; | |
5099 | ||
6c271ce2 | 5100 | if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP)) |
2b188cc1 JA |
5101 | return -EINVAL; |
5102 | ||
5103 | f = fdget(fd); | |
5104 | if (!f.file) | |
5105 | return -EBADF; | |
5106 | ||
5107 | ret = -EOPNOTSUPP; | |
5108 | if (f.file->f_op != &io_uring_fops) | |
5109 | goto out_fput; | |
5110 | ||
5111 | ret = -ENXIO; | |
5112 | ctx = f.file->private_data; | |
5113 | if (!percpu_ref_tryget(&ctx->refs)) | |
5114 | goto out_fput; | |
5115 | ||
6c271ce2 JA |
5116 | /* |
5117 | * For SQ polling, the thread will do all submissions and completions. | |
5118 | * Just return the requested submit count, and wake the thread if | |
5119 | * we were asked to. | |
5120 | */ | |
b2a9eada | 5121 | ret = 0; |
6c271ce2 | 5122 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
c1edbf5f JA |
5123 | if (!list_empty_careful(&ctx->cq_overflow_list)) |
5124 | io_cqring_overflow_flush(ctx, false); | |
6c271ce2 JA |
5125 | if (flags & IORING_ENTER_SQ_WAKEUP) |
5126 | wake_up(&ctx->sqo_wait); | |
5127 | submitted = to_submit; | |
b2a9eada | 5128 | } else if (to_submit) { |
ae9428ca | 5129 | struct mm_struct *cur_mm; |
2b188cc1 | 5130 | |
ae9428ca | 5131 | to_submit = min(to_submit, ctx->sq_entries); |
2b188cc1 | 5132 | mutex_lock(&ctx->uring_lock); |
ae9428ca PB |
5133 | /* already have mm, so io_submit_sqes() won't try to grab it */ |
5134 | cur_mm = ctx->sqo_mm; | |
5135 | submitted = io_submit_sqes(ctx, to_submit, f.file, fd, | |
5136 | &cur_mm, false); | |
2b188cc1 | 5137 | mutex_unlock(&ctx->uring_lock); |
7c504e65 PB |
5138 | |
5139 | if (submitted != to_submit) | |
5140 | goto out; | |
2b188cc1 JA |
5141 | } |
5142 | if (flags & IORING_ENTER_GETEVENTS) { | |
def596e9 JA |
5143 | unsigned nr_events = 0; |
5144 | ||
2b188cc1 JA |
5145 | min_complete = min(min_complete, ctx->cq_entries); |
5146 | ||
def596e9 | 5147 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 | 5148 | ret = io_iopoll_check(ctx, &nr_events, min_complete); |
def596e9 JA |
5149 | } else { |
5150 | ret = io_cqring_wait(ctx, min_complete, sig, sigsz); | |
5151 | } | |
2b188cc1 JA |
5152 | } |
5153 | ||
7c504e65 | 5154 | out: |
6805b32e | 5155 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
5156 | out_fput: |
5157 | fdput(f); | |
5158 | return submitted ? submitted : ret; | |
5159 | } | |
5160 | ||
5161 | static const struct file_operations io_uring_fops = { | |
5162 | .release = io_uring_release, | |
fcb323cc | 5163 | .flush = io_uring_flush, |
2b188cc1 | 5164 | .mmap = io_uring_mmap, |
6c5c240e RP |
5165 | #ifndef CONFIG_MMU |
5166 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
5167 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
5168 | #endif | |
2b188cc1 JA |
5169 | .poll = io_uring_poll, |
5170 | .fasync = io_uring_fasync, | |
5171 | }; | |
5172 | ||
5173 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
5174 | struct io_uring_params *p) | |
5175 | { | |
75b28aff HV |
5176 | struct io_rings *rings; |
5177 | size_t size, sq_array_offset; | |
2b188cc1 | 5178 | |
75b28aff HV |
5179 | size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset); |
5180 | if (size == SIZE_MAX) | |
5181 | return -EOVERFLOW; | |
5182 | ||
5183 | rings = io_mem_alloc(size); | |
5184 | if (!rings) | |
2b188cc1 JA |
5185 | return -ENOMEM; |
5186 | ||
75b28aff HV |
5187 | ctx->rings = rings; |
5188 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
5189 | rings->sq_ring_mask = p->sq_entries - 1; | |
5190 | rings->cq_ring_mask = p->cq_entries - 1; | |
5191 | rings->sq_ring_entries = p->sq_entries; | |
5192 | rings->cq_ring_entries = p->cq_entries; | |
5193 | ctx->sq_mask = rings->sq_ring_mask; | |
5194 | ctx->cq_mask = rings->cq_ring_mask; | |
5195 | ctx->sq_entries = rings->sq_ring_entries; | |
5196 | ctx->cq_entries = rings->cq_ring_entries; | |
2b188cc1 JA |
5197 | |
5198 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 JA |
5199 | if (size == SIZE_MAX) { |
5200 | io_mem_free(ctx->rings); | |
5201 | ctx->rings = NULL; | |
2b188cc1 | 5202 | return -EOVERFLOW; |
eb065d30 | 5203 | } |
2b188cc1 JA |
5204 | |
5205 | ctx->sq_sqes = io_mem_alloc(size); | |
eb065d30 JA |
5206 | if (!ctx->sq_sqes) { |
5207 | io_mem_free(ctx->rings); | |
5208 | ctx->rings = NULL; | |
2b188cc1 | 5209 | return -ENOMEM; |
eb065d30 | 5210 | } |
2b188cc1 | 5211 | |
2b188cc1 JA |
5212 | return 0; |
5213 | } | |
5214 | ||
5215 | /* | |
5216 | * Allocate an anonymous fd, this is what constitutes the application | |
5217 | * visible backing of an io_uring instance. The application mmaps this | |
5218 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
5219 | * we have to tie this fd to a socket for file garbage collection purposes. | |
5220 | */ | |
5221 | static int io_uring_get_fd(struct io_ring_ctx *ctx) | |
5222 | { | |
5223 | struct file *file; | |
5224 | int ret; | |
5225 | ||
5226 | #if defined(CONFIG_UNIX) | |
5227 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, | |
5228 | &ctx->ring_sock); | |
5229 | if (ret) | |
5230 | return ret; | |
5231 | #endif | |
5232 | ||
5233 | ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
5234 | if (ret < 0) | |
5235 | goto err; | |
5236 | ||
5237 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, | |
5238 | O_RDWR | O_CLOEXEC); | |
5239 | if (IS_ERR(file)) { | |
5240 | put_unused_fd(ret); | |
5241 | ret = PTR_ERR(file); | |
5242 | goto err; | |
5243 | } | |
5244 | ||
5245 | #if defined(CONFIG_UNIX) | |
5246 | ctx->ring_sock->file = file; | |
6b06314c | 5247 | ctx->ring_sock->sk->sk_user_data = ctx; |
2b188cc1 JA |
5248 | #endif |
5249 | fd_install(ret, file); | |
5250 | return ret; | |
5251 | err: | |
5252 | #if defined(CONFIG_UNIX) | |
5253 | sock_release(ctx->ring_sock); | |
5254 | ctx->ring_sock = NULL; | |
5255 | #endif | |
5256 | return ret; | |
5257 | } | |
5258 | ||
5259 | static int io_uring_create(unsigned entries, struct io_uring_params *p) | |
5260 | { | |
5261 | struct user_struct *user = NULL; | |
5262 | struct io_ring_ctx *ctx; | |
5263 | bool account_mem; | |
5264 | int ret; | |
5265 | ||
5266 | if (!entries || entries > IORING_MAX_ENTRIES) | |
5267 | return -EINVAL; | |
5268 | ||
5269 | /* | |
5270 | * Use twice as many entries for the CQ ring. It's possible for the | |
5271 | * application to drive a higher depth than the size of the SQ ring, | |
5272 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
5273 | * some flexibility in overcommitting a bit. If the application has |
5274 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
5275 | * of CQ ring entries manually. | |
2b188cc1 JA |
5276 | */ |
5277 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
5278 | if (p->flags & IORING_SETUP_CQSIZE) { |
5279 | /* | |
5280 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
5281 | * to a power-of-two, if it isn't already. We do NOT impose | |
5282 | * any cq vs sq ring sizing. | |
5283 | */ | |
5284 | if (p->cq_entries < p->sq_entries || p->cq_entries > IORING_MAX_CQ_ENTRIES) | |
5285 | return -EINVAL; | |
5286 | p->cq_entries = roundup_pow_of_two(p->cq_entries); | |
5287 | } else { | |
5288 | p->cq_entries = 2 * p->sq_entries; | |
5289 | } | |
2b188cc1 JA |
5290 | |
5291 | user = get_uid(current_user()); | |
5292 | account_mem = !capable(CAP_IPC_LOCK); | |
5293 | ||
5294 | if (account_mem) { | |
5295 | ret = io_account_mem(user, | |
5296 | ring_pages(p->sq_entries, p->cq_entries)); | |
5297 | if (ret) { | |
5298 | free_uid(user); | |
5299 | return ret; | |
5300 | } | |
5301 | } | |
5302 | ||
5303 | ctx = io_ring_ctx_alloc(p); | |
5304 | if (!ctx) { | |
5305 | if (account_mem) | |
5306 | io_unaccount_mem(user, ring_pages(p->sq_entries, | |
5307 | p->cq_entries)); | |
5308 | free_uid(user); | |
5309 | return -ENOMEM; | |
5310 | } | |
5311 | ctx->compat = in_compat_syscall(); | |
5312 | ctx->account_mem = account_mem; | |
5313 | ctx->user = user; | |
0b8c0ec7 | 5314 | ctx->creds = get_current_cred(); |
2b188cc1 JA |
5315 | |
5316 | ret = io_allocate_scq_urings(ctx, p); | |
5317 | if (ret) | |
5318 | goto err; | |
5319 | ||
6c271ce2 | 5320 | ret = io_sq_offload_start(ctx, p); |
2b188cc1 JA |
5321 | if (ret) |
5322 | goto err; | |
5323 | ||
2b188cc1 | 5324 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
5325 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
5326 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
5327 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
5328 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
5329 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
5330 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
5331 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
5332 | |
5333 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
5334 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
5335 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
5336 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
5337 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
5338 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
5339 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
ac90f249 | 5340 | |
044c1ab3 JA |
5341 | /* |
5342 | * Install ring fd as the very last thing, so we don't risk someone | |
5343 | * having closed it before we finish setup | |
5344 | */ | |
5345 | ret = io_uring_get_fd(ctx); | |
5346 | if (ret < 0) | |
5347 | goto err; | |
5348 | ||
da8c9690 JA |
5349 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
5350 | IORING_FEAT_SUBMIT_STABLE; | |
c826bd7a | 5351 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
5352 | return ret; |
5353 | err: | |
5354 | io_ring_ctx_wait_and_kill(ctx); | |
5355 | return ret; | |
5356 | } | |
5357 | ||
5358 | /* | |
5359 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
5360 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
5361 | * params structure passed in. | |
5362 | */ | |
5363 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
5364 | { | |
5365 | struct io_uring_params p; | |
5366 | long ret; | |
5367 | int i; | |
5368 | ||
5369 | if (copy_from_user(&p, params, sizeof(p))) | |
5370 | return -EFAULT; | |
5371 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
5372 | if (p.resv[i]) | |
5373 | return -EINVAL; | |
5374 | } | |
5375 | ||
6c271ce2 | 5376 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
33a107f0 | 5377 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE)) |
2b188cc1 JA |
5378 | return -EINVAL; |
5379 | ||
5380 | ret = io_uring_create(entries, &p); | |
5381 | if (ret < 0) | |
5382 | return ret; | |
5383 | ||
5384 | if (copy_to_user(params, &p, sizeof(p))) | |
5385 | return -EFAULT; | |
5386 | ||
5387 | return ret; | |
5388 | } | |
5389 | ||
5390 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
5391 | struct io_uring_params __user *, params) | |
5392 | { | |
5393 | return io_uring_setup(entries, params); | |
5394 | } | |
5395 | ||
edafccee JA |
5396 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
5397 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
5398 | __releases(ctx->uring_lock) |
5399 | __acquires(ctx->uring_lock) | |
edafccee JA |
5400 | { |
5401 | int ret; | |
5402 | ||
35fa71a0 JA |
5403 | /* |
5404 | * We're inside the ring mutex, if the ref is already dying, then | |
5405 | * someone else killed the ctx or is already going through | |
5406 | * io_uring_register(). | |
5407 | */ | |
5408 | if (percpu_ref_is_dying(&ctx->refs)) | |
5409 | return -ENXIO; | |
5410 | ||
edafccee | 5411 | percpu_ref_kill(&ctx->refs); |
b19062a5 JA |
5412 | |
5413 | /* | |
5414 | * Drop uring mutex before waiting for references to exit. If another | |
5415 | * thread is currently inside io_uring_enter() it might need to grab | |
5416 | * the uring_lock to make progress. If we hold it here across the drain | |
5417 | * wait, then we can deadlock. It's safe to drop the mutex here, since | |
5418 | * no new references will come in after we've killed the percpu ref. | |
5419 | */ | |
5420 | mutex_unlock(&ctx->uring_lock); | |
206aefde | 5421 | wait_for_completion(&ctx->completions[0]); |
b19062a5 | 5422 | mutex_lock(&ctx->uring_lock); |
edafccee JA |
5423 | |
5424 | switch (opcode) { | |
5425 | case IORING_REGISTER_BUFFERS: | |
5426 | ret = io_sqe_buffer_register(ctx, arg, nr_args); | |
5427 | break; | |
5428 | case IORING_UNREGISTER_BUFFERS: | |
5429 | ret = -EINVAL; | |
5430 | if (arg || nr_args) | |
5431 | break; | |
5432 | ret = io_sqe_buffer_unregister(ctx); | |
5433 | break; | |
6b06314c JA |
5434 | case IORING_REGISTER_FILES: |
5435 | ret = io_sqe_files_register(ctx, arg, nr_args); | |
5436 | break; | |
5437 | case IORING_UNREGISTER_FILES: | |
5438 | ret = -EINVAL; | |
5439 | if (arg || nr_args) | |
5440 | break; | |
5441 | ret = io_sqe_files_unregister(ctx); | |
5442 | break; | |
c3a31e60 JA |
5443 | case IORING_REGISTER_FILES_UPDATE: |
5444 | ret = io_sqe_files_update(ctx, arg, nr_args); | |
5445 | break; | |
9b402849 JA |
5446 | case IORING_REGISTER_EVENTFD: |
5447 | ret = -EINVAL; | |
5448 | if (nr_args != 1) | |
5449 | break; | |
5450 | ret = io_eventfd_register(ctx, arg); | |
5451 | break; | |
5452 | case IORING_UNREGISTER_EVENTFD: | |
5453 | ret = -EINVAL; | |
5454 | if (arg || nr_args) | |
5455 | break; | |
5456 | ret = io_eventfd_unregister(ctx); | |
5457 | break; | |
edafccee JA |
5458 | default: |
5459 | ret = -EINVAL; | |
5460 | break; | |
5461 | } | |
5462 | ||
5463 | /* bring the ctx back to life */ | |
206aefde | 5464 | reinit_completion(&ctx->completions[0]); |
edafccee JA |
5465 | percpu_ref_reinit(&ctx->refs); |
5466 | return ret; | |
5467 | } | |
5468 | ||
5469 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
5470 | void __user *, arg, unsigned int, nr_args) | |
5471 | { | |
5472 | struct io_ring_ctx *ctx; | |
5473 | long ret = -EBADF; | |
5474 | struct fd f; | |
5475 | ||
5476 | f = fdget(fd); | |
5477 | if (!f.file) | |
5478 | return -EBADF; | |
5479 | ||
5480 | ret = -EOPNOTSUPP; | |
5481 | if (f.file->f_op != &io_uring_fops) | |
5482 | goto out_fput; | |
5483 | ||
5484 | ctx = f.file->private_data; | |
5485 | ||
5486 | mutex_lock(&ctx->uring_lock); | |
5487 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
5488 | mutex_unlock(&ctx->uring_lock); | |
c826bd7a DD |
5489 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, |
5490 | ctx->cq_ev_fd != NULL, ret); | |
edafccee JA |
5491 | out_fput: |
5492 | fdput(f); | |
5493 | return ret; | |
5494 | } | |
5495 | ||
2b188cc1 JA |
5496 | static int __init io_uring_init(void) |
5497 | { | |
5498 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC); | |
5499 | return 0; | |
5500 | }; | |
5501 | __initcall(io_uring_init); |