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 | |
148 | * there are not more requests pending thatn there is space in | |
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; | |
292 | struct wait_queue_head *head; | |
293 | __poll_t events; | |
8c838788 | 294 | bool done; |
221c5eb2 | 295 | bool canceled; |
e944475e | 296 | struct wait_queue_entry *wait; |
221c5eb2 JA |
297 | }; |
298 | ||
ad8a48ac JA |
299 | struct io_timeout_data { |
300 | struct io_kiocb *req; | |
301 | struct hrtimer timer; | |
302 | struct timespec64 ts; | |
303 | enum hrtimer_mode mode; | |
cc42e0ac | 304 | u32 seq_offset; |
ad8a48ac JA |
305 | }; |
306 | ||
f499a021 JA |
307 | struct io_async_connect { |
308 | struct sockaddr_storage address; | |
309 | }; | |
310 | ||
03b1230c JA |
311 | struct io_async_msghdr { |
312 | struct iovec fast_iov[UIO_FASTIOV]; | |
313 | struct iovec *iov; | |
314 | struct sockaddr __user *uaddr; | |
315 | struct msghdr msg; | |
316 | }; | |
317 | ||
f67676d1 JA |
318 | struct io_async_rw { |
319 | struct iovec fast_iov[UIO_FASTIOV]; | |
320 | struct iovec *iov; | |
321 | ssize_t nr_segs; | |
322 | ssize_t size; | |
323 | }; | |
324 | ||
1a6b74fc JA |
325 | struct io_async_ctx { |
326 | struct io_uring_sqe sqe; | |
f67676d1 JA |
327 | union { |
328 | struct io_async_rw rw; | |
03b1230c | 329 | struct io_async_msghdr msg; |
f499a021 | 330 | struct io_async_connect connect; |
2d28390a | 331 | struct io_timeout_data timeout; |
f67676d1 | 332 | }; |
1a6b74fc JA |
333 | }; |
334 | ||
09bb8394 JA |
335 | /* |
336 | * NOTE! Each of the iocb union members has the file pointer | |
337 | * as the first entry in their struct definition. So you can | |
338 | * access the file pointer through any of the sub-structs, | |
339 | * or directly as just 'ki_filp' in this struct. | |
340 | */ | |
2b188cc1 | 341 | struct io_kiocb { |
221c5eb2 | 342 | union { |
09bb8394 | 343 | struct file *file; |
221c5eb2 JA |
344 | struct kiocb rw; |
345 | struct io_poll_iocb poll; | |
346 | }; | |
2b188cc1 | 347 | |
cf6fd4bd | 348 | const struct io_uring_sqe *sqe; |
1a6b74fc | 349 | struct io_async_ctx *io; |
cf6fd4bd PB |
350 | struct file *ring_file; |
351 | int ring_fd; | |
352 | bool has_user; | |
353 | bool in_async; | |
354 | bool needs_fixed_file; | |
2b188cc1 JA |
355 | |
356 | struct io_ring_ctx *ctx; | |
eac406c6 JA |
357 | union { |
358 | struct list_head list; | |
78076bb6 | 359 | struct hlist_node hash_node; |
eac406c6 | 360 | }; |
9e645e11 | 361 | struct list_head link_list; |
2b188cc1 | 362 | unsigned int flags; |
c16361c1 | 363 | refcount_t refs; |
8449eeda | 364 | #define REQ_F_NOWAIT 1 /* must not punt to workers */ |
def596e9 | 365 | #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */ |
6b06314c | 366 | #define REQ_F_FIXED_FILE 4 /* ctx owns file */ |
4d7dd462 | 367 | #define REQ_F_LINK_NEXT 8 /* already grabbed next link */ |
e2033e33 SB |
368 | #define REQ_F_IO_DRAIN 16 /* drain existing IO first */ |
369 | #define REQ_F_IO_DRAINED 32 /* drain done */ | |
9e645e11 | 370 | #define REQ_F_LINK 64 /* linked sqes */ |
2665abfd | 371 | #define REQ_F_LINK_TIMEOUT 128 /* has linked timeout */ |
f7b76ac9 | 372 | #define REQ_F_FAIL_LINK 256 /* fail rest of links */ |
1b4a51b6 | 373 | #define REQ_F_DRAIN_LINK 512 /* link should be fully drained */ |
5262f567 | 374 | #define REQ_F_TIMEOUT 1024 /* timeout request */ |
491381ce JA |
375 | #define REQ_F_ISREG 2048 /* regular file */ |
376 | #define REQ_F_MUST_PUNT 4096 /* must be punted even for NONBLOCK */ | |
93bd25bb | 377 | #define REQ_F_TIMEOUT_NOSEQ 8192 /* no timeout sequence */ |
fb4b3d3f LT |
378 | #define REQ_F_INFLIGHT 16384 /* on inflight list */ |
379 | #define REQ_F_COMP_LOCKED 32768 /* completion under lock */ | |
2b188cc1 | 380 | u64 user_data; |
9e645e11 | 381 | u32 result; |
de0617e4 | 382 | u32 sequence; |
2b188cc1 | 383 | |
fcb323cc JA |
384 | struct list_head inflight_entry; |
385 | ||
561fb04a | 386 | struct io_wq_work work; |
2b188cc1 JA |
387 | }; |
388 | ||
389 | #define IO_PLUG_THRESHOLD 2 | |
def596e9 | 390 | #define IO_IOPOLL_BATCH 8 |
2b188cc1 | 391 | |
9a56a232 JA |
392 | struct io_submit_state { |
393 | struct blk_plug plug; | |
394 | ||
2579f913 JA |
395 | /* |
396 | * io_kiocb alloc cache | |
397 | */ | |
398 | void *reqs[IO_IOPOLL_BATCH]; | |
399 | unsigned int free_reqs; | |
400 | unsigned int cur_req; | |
401 | ||
9a56a232 JA |
402 | /* |
403 | * File reference cache | |
404 | */ | |
405 | struct file *file; | |
406 | unsigned int fd; | |
407 | unsigned int has_refs; | |
408 | unsigned int used_refs; | |
409 | unsigned int ios_left; | |
410 | }; | |
411 | ||
561fb04a | 412 | static void io_wq_submit_work(struct io_wq_work **workptr); |
78e19bbe | 413 | static void io_cqring_fill_event(struct io_kiocb *req, long res); |
4fe2c963 | 414 | static void __io_free_req(struct io_kiocb *req); |
ec9c02ad | 415 | static void io_put_req(struct io_kiocb *req); |
78e19bbe | 416 | static void io_double_put_req(struct io_kiocb *req); |
978db57e | 417 | static void __io_double_put_req(struct io_kiocb *req); |
94ae5e77 JA |
418 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req); |
419 | static void io_queue_linked_timeout(struct io_kiocb *req); | |
de0617e4 | 420 | |
2b188cc1 JA |
421 | static struct kmem_cache *req_cachep; |
422 | ||
423 | static const struct file_operations io_uring_fops; | |
424 | ||
425 | struct sock *io_uring_get_socket(struct file *file) | |
426 | { | |
427 | #if defined(CONFIG_UNIX) | |
428 | if (file->f_op == &io_uring_fops) { | |
429 | struct io_ring_ctx *ctx = file->private_data; | |
430 | ||
431 | return ctx->ring_sock->sk; | |
432 | } | |
433 | #endif | |
434 | return NULL; | |
435 | } | |
436 | EXPORT_SYMBOL(io_uring_get_socket); | |
437 | ||
438 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) | |
439 | { | |
440 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
441 | ||
206aefde | 442 | complete(&ctx->completions[0]); |
2b188cc1 JA |
443 | } |
444 | ||
445 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) | |
446 | { | |
447 | struct io_ring_ctx *ctx; | |
78076bb6 | 448 | int hash_bits; |
2b188cc1 JA |
449 | |
450 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
451 | if (!ctx) | |
452 | return NULL; | |
453 | ||
0ddf92e8 JA |
454 | ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL); |
455 | if (!ctx->fallback_req) | |
456 | goto err; | |
457 | ||
206aefde JA |
458 | ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL); |
459 | if (!ctx->completions) | |
460 | goto err; | |
461 | ||
78076bb6 JA |
462 | /* |
463 | * Use 5 bits less than the max cq entries, that should give us around | |
464 | * 32 entries per hash list if totally full and uniformly spread. | |
465 | */ | |
466 | hash_bits = ilog2(p->cq_entries); | |
467 | hash_bits -= 5; | |
468 | if (hash_bits <= 0) | |
469 | hash_bits = 1; | |
470 | ctx->cancel_hash_bits = hash_bits; | |
471 | ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head), | |
472 | GFP_KERNEL); | |
473 | if (!ctx->cancel_hash) | |
474 | goto err; | |
475 | __hash_init(ctx->cancel_hash, 1U << hash_bits); | |
476 | ||
21482896 | 477 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
206aefde JA |
478 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) |
479 | goto err; | |
2b188cc1 JA |
480 | |
481 | ctx->flags = p->flags; | |
482 | init_waitqueue_head(&ctx->cq_wait); | |
1d7bb1d5 | 483 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
206aefde JA |
484 | init_completion(&ctx->completions[0]); |
485 | init_completion(&ctx->completions[1]); | |
2b188cc1 JA |
486 | mutex_init(&ctx->uring_lock); |
487 | init_waitqueue_head(&ctx->wait); | |
488 | spin_lock_init(&ctx->completion_lock); | |
def596e9 | 489 | INIT_LIST_HEAD(&ctx->poll_list); |
de0617e4 | 490 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 491 | INIT_LIST_HEAD(&ctx->timeout_list); |
fcb323cc JA |
492 | init_waitqueue_head(&ctx->inflight_wait); |
493 | spin_lock_init(&ctx->inflight_lock); | |
494 | INIT_LIST_HEAD(&ctx->inflight_list); | |
2b188cc1 | 495 | return ctx; |
206aefde | 496 | err: |
0ddf92e8 JA |
497 | if (ctx->fallback_req) |
498 | kmem_cache_free(req_cachep, ctx->fallback_req); | |
206aefde | 499 | kfree(ctx->completions); |
78076bb6 | 500 | kfree(ctx->cancel_hash); |
206aefde JA |
501 | kfree(ctx); |
502 | return NULL; | |
2b188cc1 JA |
503 | } |
504 | ||
9d858b21 | 505 | static inline bool __req_need_defer(struct io_kiocb *req) |
7adf4eaf | 506 | { |
a197f664 JL |
507 | struct io_ring_ctx *ctx = req->ctx; |
508 | ||
498ccd9e JA |
509 | return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped |
510 | + atomic_read(&ctx->cached_cq_overflow); | |
7adf4eaf JA |
511 | } |
512 | ||
9d858b21 | 513 | static inline bool req_need_defer(struct io_kiocb *req) |
de0617e4 | 514 | { |
9d858b21 BL |
515 | if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) == REQ_F_IO_DRAIN) |
516 | return __req_need_defer(req); | |
de0617e4 | 517 | |
9d858b21 | 518 | return false; |
de0617e4 JA |
519 | } |
520 | ||
7adf4eaf | 521 | static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx) |
de0617e4 JA |
522 | { |
523 | struct io_kiocb *req; | |
524 | ||
7adf4eaf | 525 | req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list); |
9d858b21 | 526 | if (req && !req_need_defer(req)) { |
de0617e4 JA |
527 | list_del_init(&req->list); |
528 | return req; | |
529 | } | |
530 | ||
531 | return NULL; | |
532 | } | |
533 | ||
5262f567 JA |
534 | static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx) |
535 | { | |
7adf4eaf JA |
536 | struct io_kiocb *req; |
537 | ||
538 | req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list); | |
93bd25bb JA |
539 | if (req) { |
540 | if (req->flags & REQ_F_TIMEOUT_NOSEQ) | |
541 | return NULL; | |
fb4b3d3f | 542 | if (!__req_need_defer(req)) { |
93bd25bb JA |
543 | list_del_init(&req->list); |
544 | return req; | |
545 | } | |
7adf4eaf JA |
546 | } |
547 | ||
548 | return NULL; | |
5262f567 JA |
549 | } |
550 | ||
de0617e4 | 551 | static void __io_commit_cqring(struct io_ring_ctx *ctx) |
2b188cc1 | 552 | { |
75b28aff | 553 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 554 | |
75b28aff | 555 | if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) { |
2b188cc1 | 556 | /* order cqe stores with ring update */ |
75b28aff | 557 | smp_store_release(&rings->cq.tail, ctx->cached_cq_tail); |
2b188cc1 | 558 | |
2b188cc1 JA |
559 | if (wq_has_sleeper(&ctx->cq_wait)) { |
560 | wake_up_interruptible(&ctx->cq_wait); | |
561 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
562 | } | |
563 | } | |
564 | } | |
565 | ||
561fb04a | 566 | static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe) |
18d9be1a | 567 | { |
561fb04a JA |
568 | u8 opcode = READ_ONCE(sqe->opcode); |
569 | ||
570 | return !(opcode == IORING_OP_READ_FIXED || | |
571 | opcode == IORING_OP_WRITE_FIXED); | |
572 | } | |
573 | ||
94ae5e77 JA |
574 | static inline bool io_prep_async_work(struct io_kiocb *req, |
575 | struct io_kiocb **link) | |
18d9be1a | 576 | { |
561fb04a | 577 | bool do_hashed = false; |
54a91f3b | 578 | |
cf6fd4bd PB |
579 | if (req->sqe) { |
580 | switch (req->sqe->opcode) { | |
6cc47d1d JA |
581 | case IORING_OP_WRITEV: |
582 | case IORING_OP_WRITE_FIXED: | |
561fb04a | 583 | do_hashed = true; |
5f8fd2d3 JA |
584 | /* fall-through */ |
585 | case IORING_OP_READV: | |
586 | case IORING_OP_READ_FIXED: | |
587 | case IORING_OP_SENDMSG: | |
588 | case IORING_OP_RECVMSG: | |
589 | case IORING_OP_ACCEPT: | |
590 | case IORING_OP_POLL_ADD: | |
f8e85cf2 | 591 | case IORING_OP_CONNECT: |
5f8fd2d3 JA |
592 | /* |
593 | * We know REQ_F_ISREG is not set on some of these | |
594 | * opcodes, but this enables us to keep the check in | |
595 | * just one place. | |
596 | */ | |
597 | if (!(req->flags & REQ_F_ISREG)) | |
598 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
6cc47d1d JA |
599 | break; |
600 | } | |
cf6fd4bd | 601 | if (io_sqe_needs_user(req->sqe)) |
561fb04a | 602 | req->work.flags |= IO_WQ_WORK_NEEDS_USER; |
54a91f3b JA |
603 | } |
604 | ||
94ae5e77 | 605 | *link = io_prep_linked_timeout(req); |
561fb04a JA |
606 | return do_hashed; |
607 | } | |
608 | ||
a197f664 | 609 | static inline void io_queue_async_work(struct io_kiocb *req) |
561fb04a | 610 | { |
a197f664 | 611 | struct io_ring_ctx *ctx = req->ctx; |
94ae5e77 JA |
612 | struct io_kiocb *link; |
613 | bool do_hashed; | |
614 | ||
615 | do_hashed = io_prep_async_work(req, &link); | |
561fb04a JA |
616 | |
617 | trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work, | |
618 | req->flags); | |
619 | if (!do_hashed) { | |
620 | io_wq_enqueue(ctx->io_wq, &req->work); | |
621 | } else { | |
622 | io_wq_enqueue_hashed(ctx->io_wq, &req->work, | |
623 | file_inode(req->file)); | |
624 | } | |
94ae5e77 JA |
625 | |
626 | if (link) | |
627 | io_queue_linked_timeout(link); | |
18d9be1a JA |
628 | } |
629 | ||
5262f567 JA |
630 | static void io_kill_timeout(struct io_kiocb *req) |
631 | { | |
632 | int ret; | |
633 | ||
2d28390a | 634 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); |
5262f567 JA |
635 | if (ret != -1) { |
636 | atomic_inc(&req->ctx->cq_timeouts); | |
842f9612 | 637 | list_del_init(&req->list); |
78e19bbe | 638 | io_cqring_fill_event(req, 0); |
ec9c02ad | 639 | io_put_req(req); |
5262f567 JA |
640 | } |
641 | } | |
642 | ||
643 | static void io_kill_timeouts(struct io_ring_ctx *ctx) | |
644 | { | |
645 | struct io_kiocb *req, *tmp; | |
646 | ||
647 | spin_lock_irq(&ctx->completion_lock); | |
648 | list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list) | |
649 | io_kill_timeout(req); | |
650 | spin_unlock_irq(&ctx->completion_lock); | |
651 | } | |
652 | ||
de0617e4 JA |
653 | static void io_commit_cqring(struct io_ring_ctx *ctx) |
654 | { | |
655 | struct io_kiocb *req; | |
656 | ||
5262f567 JA |
657 | while ((req = io_get_timeout_req(ctx)) != NULL) |
658 | io_kill_timeout(req); | |
659 | ||
de0617e4 JA |
660 | __io_commit_cqring(ctx); |
661 | ||
662 | while ((req = io_get_deferred_req(ctx)) != NULL) { | |
663 | req->flags |= REQ_F_IO_DRAINED; | |
a197f664 | 664 | io_queue_async_work(req); |
de0617e4 JA |
665 | } |
666 | } | |
667 | ||
2b188cc1 JA |
668 | static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx) |
669 | { | |
75b28aff | 670 | struct io_rings *rings = ctx->rings; |
2b188cc1 JA |
671 | unsigned tail; |
672 | ||
673 | tail = ctx->cached_cq_tail; | |
115e12e5 SB |
674 | /* |
675 | * writes to the cq entry need to come after reading head; the | |
676 | * control dependency is enough as we're using WRITE_ONCE to | |
677 | * fill the cq entry | |
678 | */ | |
75b28aff | 679 | if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries) |
2b188cc1 JA |
680 | return NULL; |
681 | ||
682 | ctx->cached_cq_tail++; | |
75b28aff | 683 | return &rings->cqes[tail & ctx->cq_mask]; |
2b188cc1 JA |
684 | } |
685 | ||
1d7bb1d5 JA |
686 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
687 | { | |
688 | if (waitqueue_active(&ctx->wait)) | |
689 | wake_up(&ctx->wait); | |
690 | if (waitqueue_active(&ctx->sqo_wait)) | |
691 | wake_up(&ctx->sqo_wait); | |
692 | if (ctx->cq_ev_fd) | |
693 | eventfd_signal(ctx->cq_ev_fd, 1); | |
694 | } | |
695 | ||
c4a2ed72 JA |
696 | /* Returns true if there are no backlogged entries after the flush */ |
697 | static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force) | |
1d7bb1d5 JA |
698 | { |
699 | struct io_rings *rings = ctx->rings; | |
700 | struct io_uring_cqe *cqe; | |
701 | struct io_kiocb *req; | |
702 | unsigned long flags; | |
703 | LIST_HEAD(list); | |
704 | ||
705 | if (!force) { | |
706 | if (list_empty_careful(&ctx->cq_overflow_list)) | |
c4a2ed72 | 707 | return true; |
1d7bb1d5 JA |
708 | if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) == |
709 | rings->cq_ring_entries)) | |
c4a2ed72 | 710 | return false; |
1d7bb1d5 JA |
711 | } |
712 | ||
713 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
714 | ||
715 | /* if force is set, the ring is going away. always drop after that */ | |
716 | if (force) | |
717 | ctx->cq_overflow_flushed = true; | |
718 | ||
c4a2ed72 | 719 | cqe = NULL; |
1d7bb1d5 JA |
720 | while (!list_empty(&ctx->cq_overflow_list)) { |
721 | cqe = io_get_cqring(ctx); | |
722 | if (!cqe && !force) | |
723 | break; | |
724 | ||
725 | req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb, | |
726 | list); | |
727 | list_move(&req->list, &list); | |
728 | if (cqe) { | |
729 | WRITE_ONCE(cqe->user_data, req->user_data); | |
730 | WRITE_ONCE(cqe->res, req->result); | |
731 | WRITE_ONCE(cqe->flags, 0); | |
732 | } else { | |
733 | WRITE_ONCE(ctx->rings->cq_overflow, | |
734 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
735 | } | |
736 | } | |
737 | ||
738 | io_commit_cqring(ctx); | |
739 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
740 | io_cqring_ev_posted(ctx); | |
741 | ||
742 | while (!list_empty(&list)) { | |
743 | req = list_first_entry(&list, struct io_kiocb, list); | |
744 | list_del(&req->list); | |
ec9c02ad | 745 | io_put_req(req); |
1d7bb1d5 | 746 | } |
c4a2ed72 JA |
747 | |
748 | return cqe != NULL; | |
1d7bb1d5 JA |
749 | } |
750 | ||
78e19bbe | 751 | static void io_cqring_fill_event(struct io_kiocb *req, long res) |
2b188cc1 | 752 | { |
78e19bbe | 753 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
754 | struct io_uring_cqe *cqe; |
755 | ||
78e19bbe | 756 | trace_io_uring_complete(ctx, req->user_data, res); |
51c3ff62 | 757 | |
2b188cc1 JA |
758 | /* |
759 | * If we can't get a cq entry, userspace overflowed the | |
760 | * submission (by quite a lot). Increment the overflow count in | |
761 | * the ring. | |
762 | */ | |
763 | cqe = io_get_cqring(ctx); | |
1d7bb1d5 | 764 | if (likely(cqe)) { |
78e19bbe | 765 | WRITE_ONCE(cqe->user_data, req->user_data); |
2b188cc1 | 766 | WRITE_ONCE(cqe->res, res); |
c71ffb67 | 767 | WRITE_ONCE(cqe->flags, 0); |
1d7bb1d5 | 768 | } else if (ctx->cq_overflow_flushed) { |
498ccd9e JA |
769 | WRITE_ONCE(ctx->rings->cq_overflow, |
770 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
1d7bb1d5 JA |
771 | } else { |
772 | refcount_inc(&req->refs); | |
773 | req->result = res; | |
774 | list_add_tail(&req->list, &ctx->cq_overflow_list); | |
2b188cc1 JA |
775 | } |
776 | } | |
777 | ||
78e19bbe | 778 | static void io_cqring_add_event(struct io_kiocb *req, long res) |
2b188cc1 | 779 | { |
78e19bbe | 780 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
781 | unsigned long flags; |
782 | ||
783 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
78e19bbe | 784 | io_cqring_fill_event(req, res); |
2b188cc1 JA |
785 | io_commit_cqring(ctx); |
786 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
787 | ||
8c838788 | 788 | io_cqring_ev_posted(ctx); |
2b188cc1 JA |
789 | } |
790 | ||
0ddf92e8 JA |
791 | static inline bool io_is_fallback_req(struct io_kiocb *req) |
792 | { | |
793 | return req == (struct io_kiocb *) | |
794 | ((unsigned long) req->ctx->fallback_req & ~1UL); | |
795 | } | |
796 | ||
797 | static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx) | |
798 | { | |
799 | struct io_kiocb *req; | |
800 | ||
801 | req = ctx->fallback_req; | |
802 | if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req)) | |
803 | return req; | |
804 | ||
805 | return NULL; | |
806 | } | |
807 | ||
2579f913 JA |
808 | static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx, |
809 | struct io_submit_state *state) | |
2b188cc1 | 810 | { |
fd6fab2c | 811 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
2b188cc1 JA |
812 | struct io_kiocb *req; |
813 | ||
814 | if (!percpu_ref_tryget(&ctx->refs)) | |
815 | return NULL; | |
816 | ||
2579f913 | 817 | if (!state) { |
fd6fab2c | 818 | req = kmem_cache_alloc(req_cachep, gfp); |
2579f913 | 819 | if (unlikely(!req)) |
0ddf92e8 | 820 | goto fallback; |
2579f913 JA |
821 | } else if (!state->free_reqs) { |
822 | size_t sz; | |
823 | int ret; | |
824 | ||
825 | sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs)); | |
fd6fab2c JA |
826 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs); |
827 | ||
828 | /* | |
829 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
830 | * retry single alloc to be on the safe side. | |
831 | */ | |
832 | if (unlikely(ret <= 0)) { | |
833 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
834 | if (!state->reqs[0]) | |
0ddf92e8 | 835 | goto fallback; |
fd6fab2c JA |
836 | ret = 1; |
837 | } | |
2579f913 JA |
838 | state->free_reqs = ret - 1; |
839 | state->cur_req = 1; | |
840 | req = state->reqs[0]; | |
841 | } else { | |
842 | req = state->reqs[state->cur_req]; | |
843 | state->free_reqs--; | |
844 | state->cur_req++; | |
2b188cc1 JA |
845 | } |
846 | ||
0ddf92e8 | 847 | got_it: |
1a6b74fc | 848 | req->io = NULL; |
cf6fd4bd | 849 | req->ring_file = NULL; |
60c112b0 | 850 | req->file = NULL; |
2579f913 JA |
851 | req->ctx = ctx; |
852 | req->flags = 0; | |
e65ef56d JA |
853 | /* one is dropped after submission, the other at completion */ |
854 | refcount_set(&req->refs, 2); | |
9e645e11 | 855 | req->result = 0; |
561fb04a | 856 | INIT_IO_WORK(&req->work, io_wq_submit_work); |
2579f913 | 857 | return req; |
0ddf92e8 JA |
858 | fallback: |
859 | req = io_get_fallback_req(ctx); | |
860 | if (req) | |
861 | goto got_it; | |
6805b32e | 862 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
863 | return NULL; |
864 | } | |
865 | ||
def596e9 JA |
866 | static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr) |
867 | { | |
868 | if (*nr) { | |
869 | kmem_cache_free_bulk(req_cachep, *nr, reqs); | |
6805b32e | 870 | percpu_ref_put_many(&ctx->refs, *nr); |
def596e9 JA |
871 | *nr = 0; |
872 | } | |
873 | } | |
874 | ||
9e645e11 | 875 | static void __io_free_req(struct io_kiocb *req) |
2b188cc1 | 876 | { |
fcb323cc JA |
877 | struct io_ring_ctx *ctx = req->ctx; |
878 | ||
1a6b74fc JA |
879 | if (req->io) |
880 | kfree(req->io); | |
09bb8394 JA |
881 | if (req->file && !(req->flags & REQ_F_FIXED_FILE)) |
882 | fput(req->file); | |
fcb323cc JA |
883 | if (req->flags & REQ_F_INFLIGHT) { |
884 | unsigned long flags; | |
885 | ||
886 | spin_lock_irqsave(&ctx->inflight_lock, flags); | |
887 | list_del(&req->inflight_entry); | |
888 | if (waitqueue_active(&ctx->inflight_wait)) | |
889 | wake_up(&ctx->inflight_wait); | |
890 | spin_unlock_irqrestore(&ctx->inflight_lock, flags); | |
891 | } | |
892 | percpu_ref_put(&ctx->refs); | |
0ddf92e8 JA |
893 | if (likely(!io_is_fallback_req(req))) |
894 | kmem_cache_free(req_cachep, req); | |
895 | else | |
896 | clear_bit_unlock(0, (unsigned long *) ctx->fallback_req); | |
e65ef56d JA |
897 | } |
898 | ||
a197f664 | 899 | static bool io_link_cancel_timeout(struct io_kiocb *req) |
2665abfd | 900 | { |
a197f664 | 901 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd JA |
902 | int ret; |
903 | ||
2d28390a | 904 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); |
2665abfd | 905 | if (ret != -1) { |
78e19bbe | 906 | io_cqring_fill_event(req, -ECANCELED); |
2665abfd JA |
907 | io_commit_cqring(ctx); |
908 | req->flags &= ~REQ_F_LINK; | |
ec9c02ad | 909 | io_put_req(req); |
2665abfd JA |
910 | return true; |
911 | } | |
912 | ||
913 | return false; | |
e65ef56d JA |
914 | } |
915 | ||
ba816ad6 | 916 | static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr) |
9e645e11 | 917 | { |
2665abfd | 918 | struct io_ring_ctx *ctx = req->ctx; |
9e645e11 | 919 | struct io_kiocb *nxt; |
2665abfd | 920 | bool wake_ev = false; |
9e645e11 | 921 | |
4d7dd462 JA |
922 | /* Already got next link */ |
923 | if (req->flags & REQ_F_LINK_NEXT) | |
924 | return; | |
925 | ||
9e645e11 JA |
926 | /* |
927 | * The list should never be empty when we are called here. But could | |
928 | * potentially happen if the chain is messed up, check to be on the | |
929 | * safe side. | |
930 | */ | |
931 | nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list); | |
2665abfd | 932 | while (nxt) { |
76a46e06 | 933 | list_del_init(&nxt->list); |
94ae5e77 JA |
934 | |
935 | if ((req->flags & REQ_F_LINK_TIMEOUT) && | |
936 | (nxt->flags & REQ_F_TIMEOUT)) { | |
937 | wake_ev |= io_link_cancel_timeout(nxt); | |
938 | nxt = list_first_entry_or_null(&req->link_list, | |
939 | struct io_kiocb, list); | |
940 | req->flags &= ~REQ_F_LINK_TIMEOUT; | |
941 | continue; | |
942 | } | |
9e645e11 JA |
943 | if (!list_empty(&req->link_list)) { |
944 | INIT_LIST_HEAD(&nxt->link_list); | |
945 | list_splice(&req->link_list, &nxt->link_list); | |
946 | nxt->flags |= REQ_F_LINK; | |
947 | } | |
948 | ||
b18fdf71 | 949 | *nxtptr = nxt; |
94ae5e77 | 950 | break; |
9e645e11 | 951 | } |
2665abfd | 952 | |
4d7dd462 | 953 | req->flags |= REQ_F_LINK_NEXT; |
2665abfd JA |
954 | if (wake_ev) |
955 | io_cqring_ev_posted(ctx); | |
9e645e11 JA |
956 | } |
957 | ||
958 | /* | |
959 | * Called if REQ_F_LINK is set, and we fail the head request | |
960 | */ | |
961 | static void io_fail_links(struct io_kiocb *req) | |
962 | { | |
2665abfd | 963 | struct io_ring_ctx *ctx = req->ctx; |
9e645e11 | 964 | struct io_kiocb *link; |
2665abfd JA |
965 | unsigned long flags; |
966 | ||
967 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
9e645e11 JA |
968 | |
969 | while (!list_empty(&req->link_list)) { | |
970 | link = list_first_entry(&req->link_list, struct io_kiocb, list); | |
2665abfd | 971 | list_del_init(&link->list); |
9e645e11 | 972 | |
c826bd7a | 973 | trace_io_uring_fail_link(req, link); |
2665abfd JA |
974 | |
975 | if ((req->flags & REQ_F_LINK_TIMEOUT) && | |
cf6fd4bd | 976 | link->sqe->opcode == IORING_OP_LINK_TIMEOUT) { |
a197f664 | 977 | io_link_cancel_timeout(link); |
2665abfd | 978 | } else { |
78e19bbe | 979 | io_cqring_fill_event(link, -ECANCELED); |
978db57e | 980 | __io_double_put_req(link); |
2665abfd | 981 | } |
5d960724 | 982 | req->flags &= ~REQ_F_LINK_TIMEOUT; |
9e645e11 | 983 | } |
2665abfd JA |
984 | |
985 | io_commit_cqring(ctx); | |
986 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
987 | io_cqring_ev_posted(ctx); | |
9e645e11 JA |
988 | } |
989 | ||
4d7dd462 | 990 | static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt) |
9e645e11 | 991 | { |
4d7dd462 | 992 | if (likely(!(req->flags & REQ_F_LINK))) |
2665abfd | 993 | return; |
2665abfd | 994 | |
9e645e11 JA |
995 | /* |
996 | * If LINK is set, we have dependent requests in this chain. If we | |
997 | * didn't fail this request, queue the first one up, moving any other | |
998 | * dependencies to the next request. In case of failure, fail the rest | |
999 | * of the chain. | |
1000 | */ | |
2665abfd JA |
1001 | if (req->flags & REQ_F_FAIL_LINK) { |
1002 | io_fail_links(req); | |
7c9e7f0f JA |
1003 | } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) == |
1004 | REQ_F_LINK_TIMEOUT) { | |
2665abfd JA |
1005 | struct io_ring_ctx *ctx = req->ctx; |
1006 | unsigned long flags; | |
1007 | ||
1008 | /* | |
1009 | * If this is a timeout link, we could be racing with the | |
1010 | * timeout timer. Grab the completion lock for this case to | |
7c9e7f0f | 1011 | * protect against that. |
2665abfd JA |
1012 | */ |
1013 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
1014 | io_req_link_next(req, nxt); | |
1015 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1016 | } else { | |
1017 | io_req_link_next(req, nxt); | |
9e645e11 | 1018 | } |
4d7dd462 | 1019 | } |
9e645e11 | 1020 | |
c69f8dbe JL |
1021 | static void io_free_req(struct io_kiocb *req) |
1022 | { | |
944e58bf PB |
1023 | struct io_kiocb *nxt = NULL; |
1024 | ||
1025 | io_req_find_next(req, &nxt); | |
70cf9f32 | 1026 | __io_free_req(req); |
944e58bf PB |
1027 | |
1028 | if (nxt) | |
1029 | io_queue_async_work(nxt); | |
c69f8dbe JL |
1030 | } |
1031 | ||
ba816ad6 JA |
1032 | /* |
1033 | * Drop reference to request, return next in chain (if there is one) if this | |
1034 | * was the last reference to this request. | |
1035 | */ | |
f9bd67f6 | 1036 | __attribute__((nonnull)) |
ec9c02ad | 1037 | static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr) |
e65ef56d | 1038 | { |
f9bd67f6 | 1039 | io_req_find_next(req, nxtptr); |
4d7dd462 | 1040 | |
e65ef56d | 1041 | if (refcount_dec_and_test(&req->refs)) |
4d7dd462 | 1042 | __io_free_req(req); |
2b188cc1 JA |
1043 | } |
1044 | ||
e65ef56d JA |
1045 | static void io_put_req(struct io_kiocb *req) |
1046 | { | |
1047 | if (refcount_dec_and_test(&req->refs)) | |
1048 | io_free_req(req); | |
2b188cc1 JA |
1049 | } |
1050 | ||
978db57e JA |
1051 | /* |
1052 | * Must only be used if we don't need to care about links, usually from | |
1053 | * within the completion handling itself. | |
1054 | */ | |
1055 | static void __io_double_put_req(struct io_kiocb *req) | |
78e19bbe JA |
1056 | { |
1057 | /* drop both submit and complete references */ | |
1058 | if (refcount_sub_and_test(2, &req->refs)) | |
1059 | __io_free_req(req); | |
1060 | } | |
1061 | ||
978db57e JA |
1062 | static void io_double_put_req(struct io_kiocb *req) |
1063 | { | |
1064 | /* drop both submit and complete references */ | |
1065 | if (refcount_sub_and_test(2, &req->refs)) | |
1066 | io_free_req(req); | |
1067 | } | |
1068 | ||
1d7bb1d5 | 1069 | static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush) |
a3a0e43f | 1070 | { |
84f97dc2 JA |
1071 | struct io_rings *rings = ctx->rings; |
1072 | ||
1d7bb1d5 JA |
1073 | /* |
1074 | * noflush == true is from the waitqueue handler, just ensure we wake | |
1075 | * up the task, and the next invocation will flush the entries. We | |
1076 | * cannot safely to it from here. | |
1077 | */ | |
1078 | if (noflush && !list_empty(&ctx->cq_overflow_list)) | |
1079 | return -1U; | |
1080 | ||
1081 | io_cqring_overflow_flush(ctx, false); | |
1082 | ||
a3a0e43f JA |
1083 | /* See comment at the top of this file */ |
1084 | smp_rmb(); | |
75b28aff | 1085 | return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head); |
a3a0e43f JA |
1086 | } |
1087 | ||
fb5ccc98 PB |
1088 | static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) |
1089 | { | |
1090 | struct io_rings *rings = ctx->rings; | |
1091 | ||
1092 | /* make sure SQ entry isn't read before tail */ | |
1093 | return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; | |
1094 | } | |
1095 | ||
def596e9 JA |
1096 | /* |
1097 | * Find and free completed poll iocbs | |
1098 | */ | |
1099 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1100 | struct list_head *done) | |
1101 | { | |
1102 | void *reqs[IO_IOPOLL_BATCH]; | |
1103 | struct io_kiocb *req; | |
09bb8394 | 1104 | int to_free; |
def596e9 | 1105 | |
09bb8394 | 1106 | to_free = 0; |
def596e9 JA |
1107 | while (!list_empty(done)) { |
1108 | req = list_first_entry(done, struct io_kiocb, list); | |
1109 | list_del(&req->list); | |
1110 | ||
78e19bbe | 1111 | io_cqring_fill_event(req, req->result); |
def596e9 JA |
1112 | (*nr_events)++; |
1113 | ||
09bb8394 JA |
1114 | if (refcount_dec_and_test(&req->refs)) { |
1115 | /* If we're not using fixed files, we have to pair the | |
1116 | * completion part with the file put. Use regular | |
1117 | * completions for those, only batch free for fixed | |
9e645e11 | 1118 | * file and non-linked commands. |
09bb8394 | 1119 | */ |
1a6b74fc JA |
1120 | if (((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) == |
1121 | REQ_F_FIXED_FILE) && !io_is_fallback_req(req) && | |
1122 | !req->io) { | |
09bb8394 JA |
1123 | reqs[to_free++] = req; |
1124 | if (to_free == ARRAY_SIZE(reqs)) | |
1125 | io_free_req_many(ctx, reqs, &to_free); | |
6b06314c | 1126 | } else { |
09bb8394 | 1127 | io_free_req(req); |
6b06314c | 1128 | } |
9a56a232 | 1129 | } |
def596e9 | 1130 | } |
def596e9 | 1131 | |
09bb8394 | 1132 | io_commit_cqring(ctx); |
def596e9 JA |
1133 | io_free_req_many(ctx, reqs, &to_free); |
1134 | } | |
1135 | ||
1136 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1137 | long min) | |
1138 | { | |
1139 | struct io_kiocb *req, *tmp; | |
1140 | LIST_HEAD(done); | |
1141 | bool spin; | |
1142 | int ret; | |
1143 | ||
1144 | /* | |
1145 | * Only spin for completions if we don't have multiple devices hanging | |
1146 | * off our complete list, and we're under the requested amount. | |
1147 | */ | |
1148 | spin = !ctx->poll_multi_file && *nr_events < min; | |
1149 | ||
1150 | ret = 0; | |
1151 | list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) { | |
1152 | struct kiocb *kiocb = &req->rw; | |
1153 | ||
1154 | /* | |
1155 | * Move completed entries to our local list. If we find a | |
1156 | * request that requires polling, break out and complete | |
1157 | * the done list first, if we have entries there. | |
1158 | */ | |
1159 | if (req->flags & REQ_F_IOPOLL_COMPLETED) { | |
1160 | list_move_tail(&req->list, &done); | |
1161 | continue; | |
1162 | } | |
1163 | if (!list_empty(&done)) | |
1164 | break; | |
1165 | ||
1166 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
1167 | if (ret < 0) | |
1168 | break; | |
1169 | ||
1170 | if (ret && spin) | |
1171 | spin = false; | |
1172 | ret = 0; | |
1173 | } | |
1174 | ||
1175 | if (!list_empty(&done)) | |
1176 | io_iopoll_complete(ctx, nr_events, &done); | |
1177 | ||
1178 | return ret; | |
1179 | } | |
1180 | ||
1181 | /* | |
1182 | * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a | |
1183 | * non-spinning poll check - we'll still enter the driver poll loop, but only | |
1184 | * as a non-spinning completion check. | |
1185 | */ | |
1186 | static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1187 | long min) | |
1188 | { | |
08f5439f | 1189 | while (!list_empty(&ctx->poll_list) && !need_resched()) { |
def596e9 JA |
1190 | int ret; |
1191 | ||
1192 | ret = io_do_iopoll(ctx, nr_events, min); | |
1193 | if (ret < 0) | |
1194 | return ret; | |
1195 | if (!min || *nr_events >= min) | |
1196 | return 0; | |
1197 | } | |
1198 | ||
1199 | return 1; | |
1200 | } | |
1201 | ||
1202 | /* | |
1203 | * We can't just wait for polled events to come to us, we have to actively | |
1204 | * find and complete them. | |
1205 | */ | |
1206 | static void io_iopoll_reap_events(struct io_ring_ctx *ctx) | |
1207 | { | |
1208 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
1209 | return; | |
1210 | ||
1211 | mutex_lock(&ctx->uring_lock); | |
1212 | while (!list_empty(&ctx->poll_list)) { | |
1213 | unsigned int nr_events = 0; | |
1214 | ||
1215 | io_iopoll_getevents(ctx, &nr_events, 1); | |
08f5439f JA |
1216 | |
1217 | /* | |
1218 | * Ensure we allow local-to-the-cpu processing to take place, | |
1219 | * in this case we need to ensure that we reap all events. | |
1220 | */ | |
1221 | cond_resched(); | |
def596e9 JA |
1222 | } |
1223 | mutex_unlock(&ctx->uring_lock); | |
1224 | } | |
1225 | ||
2b2ed975 JA |
1226 | static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, |
1227 | long min) | |
def596e9 | 1228 | { |
2b2ed975 | 1229 | int iters = 0, ret = 0; |
500f9fba | 1230 | |
def596e9 JA |
1231 | do { |
1232 | int tmin = 0; | |
1233 | ||
a3a0e43f JA |
1234 | /* |
1235 | * Don't enter poll loop if we already have events pending. | |
1236 | * If we do, we can potentially be spinning for commands that | |
1237 | * already triggered a CQE (eg in error). | |
1238 | */ | |
1d7bb1d5 | 1239 | if (io_cqring_events(ctx, false)) |
a3a0e43f JA |
1240 | break; |
1241 | ||
500f9fba JA |
1242 | /* |
1243 | * If a submit got punted to a workqueue, we can have the | |
1244 | * application entering polling for a command before it gets | |
1245 | * issued. That app will hold the uring_lock for the duration | |
1246 | * of the poll right here, so we need to take a breather every | |
1247 | * now and then to ensure that the issue has a chance to add | |
1248 | * the poll to the issued list. Otherwise we can spin here | |
1249 | * forever, while the workqueue is stuck trying to acquire the | |
1250 | * very same mutex. | |
1251 | */ | |
1252 | if (!(++iters & 7)) { | |
1253 | mutex_unlock(&ctx->uring_lock); | |
1254 | mutex_lock(&ctx->uring_lock); | |
1255 | } | |
1256 | ||
def596e9 JA |
1257 | if (*nr_events < min) |
1258 | tmin = min - *nr_events; | |
1259 | ||
1260 | ret = io_iopoll_getevents(ctx, nr_events, tmin); | |
1261 | if (ret <= 0) | |
1262 | break; | |
1263 | ret = 0; | |
1264 | } while (min && !*nr_events && !need_resched()); | |
1265 | ||
2b2ed975 JA |
1266 | return ret; |
1267 | } | |
1268 | ||
1269 | static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, | |
1270 | long min) | |
1271 | { | |
1272 | int ret; | |
1273 | ||
1274 | /* | |
1275 | * We disallow the app entering submit/complete with polling, but we | |
1276 | * still need to lock the ring to prevent racing with polled issue | |
1277 | * that got punted to a workqueue. | |
1278 | */ | |
1279 | mutex_lock(&ctx->uring_lock); | |
1280 | ret = __io_iopoll_check(ctx, nr_events, min); | |
500f9fba | 1281 | mutex_unlock(&ctx->uring_lock); |
def596e9 JA |
1282 | return ret; |
1283 | } | |
1284 | ||
491381ce | 1285 | static void kiocb_end_write(struct io_kiocb *req) |
2b188cc1 | 1286 | { |
491381ce JA |
1287 | /* |
1288 | * Tell lockdep we inherited freeze protection from submission | |
1289 | * thread. | |
1290 | */ | |
1291 | if (req->flags & REQ_F_ISREG) { | |
1292 | struct inode *inode = file_inode(req->file); | |
2b188cc1 | 1293 | |
491381ce | 1294 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); |
2b188cc1 | 1295 | } |
491381ce | 1296 | file_end_write(req->file); |
2b188cc1 JA |
1297 | } |
1298 | ||
ba816ad6 | 1299 | static void io_complete_rw_common(struct kiocb *kiocb, long res) |
2b188cc1 JA |
1300 | { |
1301 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
1302 | ||
491381ce JA |
1303 | if (kiocb->ki_flags & IOCB_WRITE) |
1304 | kiocb_end_write(req); | |
2b188cc1 | 1305 | |
9e645e11 JA |
1306 | if ((req->flags & REQ_F_LINK) && res != req->result) |
1307 | req->flags |= REQ_F_FAIL_LINK; | |
78e19bbe | 1308 | io_cqring_add_event(req, res); |
ba816ad6 JA |
1309 | } |
1310 | ||
1311 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
1312 | { | |
1313 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
1314 | ||
1315 | io_complete_rw_common(kiocb, res); | |
e65ef56d | 1316 | io_put_req(req); |
2b188cc1 JA |
1317 | } |
1318 | ||
ba816ad6 JA |
1319 | static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res) |
1320 | { | |
1321 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
ec9c02ad | 1322 | struct io_kiocb *nxt = NULL; |
ba816ad6 JA |
1323 | |
1324 | io_complete_rw_common(kiocb, res); | |
ec9c02ad JL |
1325 | io_put_req_find_next(req, &nxt); |
1326 | ||
1327 | return nxt; | |
2b188cc1 JA |
1328 | } |
1329 | ||
def596e9 JA |
1330 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
1331 | { | |
1332 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
1333 | ||
491381ce JA |
1334 | if (kiocb->ki_flags & IOCB_WRITE) |
1335 | kiocb_end_write(req); | |
def596e9 | 1336 | |
9e645e11 JA |
1337 | if ((req->flags & REQ_F_LINK) && res != req->result) |
1338 | req->flags |= REQ_F_FAIL_LINK; | |
1339 | req->result = res; | |
def596e9 JA |
1340 | if (res != -EAGAIN) |
1341 | req->flags |= REQ_F_IOPOLL_COMPLETED; | |
1342 | } | |
1343 | ||
1344 | /* | |
1345 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1346 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
1347 | * find it from a io_iopoll_getevents() thread before the issuer is done | |
1348 | * accessing the kiocb cookie. | |
1349 | */ | |
1350 | static void io_iopoll_req_issued(struct io_kiocb *req) | |
1351 | { | |
1352 | struct io_ring_ctx *ctx = req->ctx; | |
1353 | ||
1354 | /* | |
1355 | * Track whether we have multiple files in our lists. This will impact | |
1356 | * how we do polling eventually, not spinning if we're on potentially | |
1357 | * different devices. | |
1358 | */ | |
1359 | if (list_empty(&ctx->poll_list)) { | |
1360 | ctx->poll_multi_file = false; | |
1361 | } else if (!ctx->poll_multi_file) { | |
1362 | struct io_kiocb *list_req; | |
1363 | ||
1364 | list_req = list_first_entry(&ctx->poll_list, struct io_kiocb, | |
1365 | list); | |
1366 | if (list_req->rw.ki_filp != req->rw.ki_filp) | |
1367 | ctx->poll_multi_file = true; | |
1368 | } | |
1369 | ||
1370 | /* | |
1371 | * For fast devices, IO may have already completed. If it has, add | |
1372 | * it to the front so we find it first. | |
1373 | */ | |
1374 | if (req->flags & REQ_F_IOPOLL_COMPLETED) | |
1375 | list_add(&req->list, &ctx->poll_list); | |
1376 | else | |
1377 | list_add_tail(&req->list, &ctx->poll_list); | |
1378 | } | |
1379 | ||
3d6770fb | 1380 | static void io_file_put(struct io_submit_state *state) |
9a56a232 | 1381 | { |
3d6770fb | 1382 | if (state->file) { |
9a56a232 JA |
1383 | int diff = state->has_refs - state->used_refs; |
1384 | ||
1385 | if (diff) | |
1386 | fput_many(state->file, diff); | |
1387 | state->file = NULL; | |
1388 | } | |
1389 | } | |
1390 | ||
1391 | /* | |
1392 | * Get as many references to a file as we have IOs left in this submission, | |
1393 | * assuming most submissions are for one file, or at least that each file | |
1394 | * has more than one submission. | |
1395 | */ | |
1396 | static struct file *io_file_get(struct io_submit_state *state, int fd) | |
1397 | { | |
1398 | if (!state) | |
1399 | return fget(fd); | |
1400 | ||
1401 | if (state->file) { | |
1402 | if (state->fd == fd) { | |
1403 | state->used_refs++; | |
1404 | state->ios_left--; | |
1405 | return state->file; | |
1406 | } | |
3d6770fb | 1407 | io_file_put(state); |
9a56a232 JA |
1408 | } |
1409 | state->file = fget_many(fd, state->ios_left); | |
1410 | if (!state->file) | |
1411 | return NULL; | |
1412 | ||
1413 | state->fd = fd; | |
1414 | state->has_refs = state->ios_left; | |
1415 | state->used_refs = 1; | |
1416 | state->ios_left--; | |
1417 | return state->file; | |
1418 | } | |
1419 | ||
2b188cc1 JA |
1420 | /* |
1421 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1422 | * any file. For now, just ensure that anything potentially problematic is done | |
1423 | * inline. | |
1424 | */ | |
1425 | static bool io_file_supports_async(struct file *file) | |
1426 | { | |
1427 | umode_t mode = file_inode(file)->i_mode; | |
1428 | ||
1429 | if (S_ISBLK(mode) || S_ISCHR(mode)) | |
1430 | return true; | |
1431 | if (S_ISREG(mode) && file->f_op != &io_uring_fops) | |
1432 | return true; | |
1433 | ||
1434 | return false; | |
1435 | } | |
1436 | ||
267bc904 | 1437 | static int io_prep_rw(struct io_kiocb *req, bool force_nonblock) |
2b188cc1 | 1438 | { |
cf6fd4bd | 1439 | const struct io_uring_sqe *sqe = req->sqe; |
def596e9 | 1440 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 | 1441 | struct kiocb *kiocb = &req->rw; |
09bb8394 JA |
1442 | unsigned ioprio; |
1443 | int ret; | |
2b188cc1 | 1444 | |
09bb8394 JA |
1445 | if (!req->file) |
1446 | return -EBADF; | |
2b188cc1 | 1447 | |
491381ce JA |
1448 | if (S_ISREG(file_inode(req->file)->i_mode)) |
1449 | req->flags |= REQ_F_ISREG; | |
1450 | ||
2b188cc1 JA |
1451 | kiocb->ki_pos = READ_ONCE(sqe->off); |
1452 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); | |
1453 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); | |
1454 | ||
1455 | ioprio = READ_ONCE(sqe->ioprio); | |
1456 | if (ioprio) { | |
1457 | ret = ioprio_check_cap(ioprio); | |
1458 | if (ret) | |
09bb8394 | 1459 | return ret; |
2b188cc1 JA |
1460 | |
1461 | kiocb->ki_ioprio = ioprio; | |
1462 | } else | |
1463 | kiocb->ki_ioprio = get_current_ioprio(); | |
1464 | ||
1465 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
1466 | if (unlikely(ret)) | |
09bb8394 | 1467 | return ret; |
8449eeda SB |
1468 | |
1469 | /* don't allow async punt if RWF_NOWAIT was requested */ | |
491381ce JA |
1470 | if ((kiocb->ki_flags & IOCB_NOWAIT) || |
1471 | (req->file->f_flags & O_NONBLOCK)) | |
8449eeda SB |
1472 | req->flags |= REQ_F_NOWAIT; |
1473 | ||
1474 | if (force_nonblock) | |
2b188cc1 | 1475 | kiocb->ki_flags |= IOCB_NOWAIT; |
8449eeda | 1476 | |
def596e9 | 1477 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
1478 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
1479 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 1480 | return -EOPNOTSUPP; |
2b188cc1 | 1481 | |
def596e9 JA |
1482 | kiocb->ki_flags |= IOCB_HIPRI; |
1483 | kiocb->ki_complete = io_complete_rw_iopoll; | |
6873e0bd | 1484 | req->result = 0; |
def596e9 | 1485 | } else { |
09bb8394 JA |
1486 | if (kiocb->ki_flags & IOCB_HIPRI) |
1487 | return -EINVAL; | |
def596e9 JA |
1488 | kiocb->ki_complete = io_complete_rw; |
1489 | } | |
2b188cc1 | 1490 | return 0; |
2b188cc1 JA |
1491 | } |
1492 | ||
1493 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
1494 | { | |
1495 | switch (ret) { | |
1496 | case -EIOCBQUEUED: | |
1497 | break; | |
1498 | case -ERESTARTSYS: | |
1499 | case -ERESTARTNOINTR: | |
1500 | case -ERESTARTNOHAND: | |
1501 | case -ERESTART_RESTARTBLOCK: | |
1502 | /* | |
1503 | * We can't just restart the syscall, since previously | |
1504 | * submitted sqes may already be in progress. Just fail this | |
1505 | * IO with EINTR. | |
1506 | */ | |
1507 | ret = -EINTR; | |
1508 | /* fall through */ | |
1509 | default: | |
1510 | kiocb->ki_complete(kiocb, ret, 0); | |
1511 | } | |
1512 | } | |
1513 | ||
ba816ad6 JA |
1514 | static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt, |
1515 | bool in_async) | |
1516 | { | |
f9bd67f6 | 1517 | if (in_async && ret >= 0 && kiocb->ki_complete == io_complete_rw) |
ba816ad6 JA |
1518 | *nxt = __io_complete_rw(kiocb, ret); |
1519 | else | |
1520 | io_rw_done(kiocb, ret); | |
1521 | } | |
1522 | ||
7d009165 PB |
1523 | static ssize_t io_import_fixed(struct io_ring_ctx *ctx, int rw, |
1524 | const struct io_uring_sqe *sqe, | |
1525 | struct iov_iter *iter) | |
edafccee JA |
1526 | { |
1527 | size_t len = READ_ONCE(sqe->len); | |
1528 | struct io_mapped_ubuf *imu; | |
1529 | unsigned index, buf_index; | |
1530 | size_t offset; | |
1531 | u64 buf_addr; | |
1532 | ||
1533 | /* attempt to use fixed buffers without having provided iovecs */ | |
1534 | if (unlikely(!ctx->user_bufs)) | |
1535 | return -EFAULT; | |
1536 | ||
1537 | buf_index = READ_ONCE(sqe->buf_index); | |
1538 | if (unlikely(buf_index >= ctx->nr_user_bufs)) | |
1539 | return -EFAULT; | |
1540 | ||
1541 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); | |
1542 | imu = &ctx->user_bufs[index]; | |
1543 | buf_addr = READ_ONCE(sqe->addr); | |
1544 | ||
1545 | /* overflow */ | |
1546 | if (buf_addr + len < buf_addr) | |
1547 | return -EFAULT; | |
1548 | /* not inside the mapped region */ | |
1549 | if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len) | |
1550 | return -EFAULT; | |
1551 | ||
1552 | /* | |
1553 | * May not be a start of buffer, set size appropriately | |
1554 | * and advance us to the beginning. | |
1555 | */ | |
1556 | offset = buf_addr - imu->ubuf; | |
1557 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
bd11b3a3 JA |
1558 | |
1559 | if (offset) { | |
1560 | /* | |
1561 | * Don't use iov_iter_advance() here, as it's really slow for | |
1562 | * using the latter parts of a big fixed buffer - it iterates | |
1563 | * over each segment manually. We can cheat a bit here, because | |
1564 | * we know that: | |
1565 | * | |
1566 | * 1) it's a BVEC iter, we set it up | |
1567 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
1568 | * first and last bvec | |
1569 | * | |
1570 | * So just find our index, and adjust the iterator afterwards. | |
1571 | * If the offset is within the first bvec (or the whole first | |
1572 | * bvec, just use iov_iter_advance(). This makes it easier | |
1573 | * since we can just skip the first segment, which may not | |
1574 | * be PAGE_SIZE aligned. | |
1575 | */ | |
1576 | const struct bio_vec *bvec = imu->bvec; | |
1577 | ||
1578 | if (offset <= bvec->bv_len) { | |
1579 | iov_iter_advance(iter, offset); | |
1580 | } else { | |
1581 | unsigned long seg_skip; | |
1582 | ||
1583 | /* skip first vec */ | |
1584 | offset -= bvec->bv_len; | |
1585 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
1586 | ||
1587 | iter->bvec = bvec + seg_skip; | |
1588 | iter->nr_segs -= seg_skip; | |
99c79f66 | 1589 | iter->count -= bvec->bv_len + offset; |
bd11b3a3 | 1590 | iter->iov_offset = offset & ~PAGE_MASK; |
bd11b3a3 JA |
1591 | } |
1592 | } | |
1593 | ||
5e559561 | 1594 | return len; |
edafccee JA |
1595 | } |
1596 | ||
cf6fd4bd PB |
1597 | static ssize_t io_import_iovec(int rw, struct io_kiocb *req, |
1598 | struct iovec **iovec, struct iov_iter *iter) | |
2b188cc1 | 1599 | { |
cf6fd4bd | 1600 | const struct io_uring_sqe *sqe = req->sqe; |
2b188cc1 JA |
1601 | void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
1602 | size_t sqe_len = READ_ONCE(sqe->len); | |
edafccee JA |
1603 | u8 opcode; |
1604 | ||
1605 | /* | |
1606 | * We're reading ->opcode for the second time, but the first read | |
1607 | * doesn't care whether it's _FIXED or not, so it doesn't matter | |
1608 | * whether ->opcode changes concurrently. The first read does care | |
1609 | * about whether it is a READ or a WRITE, so we don't trust this read | |
1610 | * for that purpose and instead let the caller pass in the read/write | |
1611 | * flag. | |
1612 | */ | |
1613 | opcode = READ_ONCE(sqe->opcode); | |
7d009165 | 1614 | if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) { |
edafccee | 1615 | *iovec = NULL; |
7d009165 | 1616 | return io_import_fixed(req->ctx, rw, sqe, iter); |
edafccee | 1617 | } |
2b188cc1 | 1618 | |
f67676d1 JA |
1619 | if (req->io) { |
1620 | struct io_async_rw *iorw = &req->io->rw; | |
1621 | ||
1622 | *iovec = iorw->iov; | |
1623 | iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size); | |
1624 | if (iorw->iov == iorw->fast_iov) | |
1625 | *iovec = NULL; | |
1626 | return iorw->size; | |
1627 | } | |
1628 | ||
cf6fd4bd | 1629 | if (!req->has_user) |
2b188cc1 JA |
1630 | return -EFAULT; |
1631 | ||
1632 | #ifdef CONFIG_COMPAT | |
cf6fd4bd | 1633 | if (req->ctx->compat) |
2b188cc1 JA |
1634 | return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV, |
1635 | iovec, iter); | |
1636 | #endif | |
1637 | ||
1638 | return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter); | |
1639 | } | |
1640 | ||
31b51510 | 1641 | /* |
32960613 JA |
1642 | * For files that don't have ->read_iter() and ->write_iter(), handle them |
1643 | * by looping over ->read() or ->write() manually. | |
31b51510 | 1644 | */ |
32960613 JA |
1645 | static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb, |
1646 | struct iov_iter *iter) | |
1647 | { | |
1648 | ssize_t ret = 0; | |
1649 | ||
1650 | /* | |
1651 | * Don't support polled IO through this interface, and we can't | |
1652 | * support non-blocking either. For the latter, this just causes | |
1653 | * the kiocb to be handled from an async context. | |
1654 | */ | |
1655 | if (kiocb->ki_flags & IOCB_HIPRI) | |
1656 | return -EOPNOTSUPP; | |
1657 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
1658 | return -EAGAIN; | |
1659 | ||
1660 | while (iov_iter_count(iter)) { | |
311ae9e1 | 1661 | struct iovec iovec; |
32960613 JA |
1662 | ssize_t nr; |
1663 | ||
311ae9e1 PB |
1664 | if (!iov_iter_is_bvec(iter)) { |
1665 | iovec = iov_iter_iovec(iter); | |
1666 | } else { | |
1667 | /* fixed buffers import bvec */ | |
1668 | iovec.iov_base = kmap(iter->bvec->bv_page) | |
1669 | + iter->iov_offset; | |
1670 | iovec.iov_len = min(iter->count, | |
1671 | iter->bvec->bv_len - iter->iov_offset); | |
1672 | } | |
1673 | ||
32960613 JA |
1674 | if (rw == READ) { |
1675 | nr = file->f_op->read(file, iovec.iov_base, | |
1676 | iovec.iov_len, &kiocb->ki_pos); | |
1677 | } else { | |
1678 | nr = file->f_op->write(file, iovec.iov_base, | |
1679 | iovec.iov_len, &kiocb->ki_pos); | |
1680 | } | |
1681 | ||
311ae9e1 PB |
1682 | if (iov_iter_is_bvec(iter)) |
1683 | kunmap(iter->bvec->bv_page); | |
1684 | ||
32960613 JA |
1685 | if (nr < 0) { |
1686 | if (!ret) | |
1687 | ret = nr; | |
1688 | break; | |
1689 | } | |
1690 | ret += nr; | |
1691 | if (nr != iovec.iov_len) | |
1692 | break; | |
1693 | iov_iter_advance(iter, nr); | |
1694 | } | |
1695 | ||
1696 | return ret; | |
1697 | } | |
1698 | ||
f67676d1 JA |
1699 | static void io_req_map_io(struct io_kiocb *req, ssize_t io_size, |
1700 | struct iovec *iovec, struct iovec *fast_iov, | |
1701 | struct iov_iter *iter) | |
1702 | { | |
1703 | req->io->rw.nr_segs = iter->nr_segs; | |
1704 | req->io->rw.size = io_size; | |
1705 | req->io->rw.iov = iovec; | |
1706 | if (!req->io->rw.iov) { | |
1707 | req->io->rw.iov = req->io->rw.fast_iov; | |
1708 | memcpy(req->io->rw.iov, fast_iov, | |
1709 | sizeof(struct iovec) * iter->nr_segs); | |
1710 | } | |
1711 | } | |
1712 | ||
1713 | static int io_setup_async_io(struct io_kiocb *req, ssize_t io_size, | |
1714 | struct iovec *iovec, struct iovec *fast_iov, | |
1715 | struct iov_iter *iter) | |
1716 | { | |
1717 | req->io = kmalloc(sizeof(*req->io), GFP_KERNEL); | |
1718 | if (req->io) { | |
1719 | io_req_map_io(req, io_size, iovec, fast_iov, iter); | |
1720 | memcpy(&req->io->sqe, req->sqe, sizeof(req->io->sqe)); | |
1721 | req->sqe = &req->io->sqe; | |
1722 | return 0; | |
1723 | } | |
1724 | ||
1725 | return -ENOMEM; | |
1726 | } | |
1727 | ||
1728 | static int io_read_prep(struct io_kiocb *req, struct iovec **iovec, | |
1729 | struct iov_iter *iter, bool force_nonblock) | |
1730 | { | |
1731 | ssize_t ret; | |
1732 | ||
1733 | ret = io_prep_rw(req, force_nonblock); | |
1734 | if (ret) | |
1735 | return ret; | |
1736 | ||
1737 | if (unlikely(!(req->file->f_mode & FMODE_READ))) | |
1738 | return -EBADF; | |
1739 | ||
1740 | return io_import_iovec(READ, req, iovec, iter); | |
1741 | } | |
1742 | ||
267bc904 | 1743 | static int io_read(struct io_kiocb *req, struct io_kiocb **nxt, |
8358e3a8 | 1744 | bool force_nonblock) |
2b188cc1 JA |
1745 | { |
1746 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1747 | struct kiocb *kiocb = &req->rw; | |
1748 | struct iov_iter iter; | |
1749 | struct file *file; | |
31b51510 | 1750 | size_t iov_count; |
f67676d1 | 1751 | ssize_t io_size, ret; |
2b188cc1 | 1752 | |
f67676d1 JA |
1753 | if (!req->io) { |
1754 | ret = io_read_prep(req, &iovec, &iter, force_nonblock); | |
1755 | if (ret < 0) | |
1756 | return ret; | |
1757 | } else { | |
1758 | ret = io_import_iovec(READ, req, &iovec, &iter); | |
1759 | if (ret < 0) | |
1760 | return ret; | |
1761 | } | |
2b188cc1 | 1762 | |
f67676d1 JA |
1763 | file = req->file; |
1764 | io_size = ret; | |
9e645e11 | 1765 | if (req->flags & REQ_F_LINK) |
f67676d1 JA |
1766 | req->result = io_size; |
1767 | ||
1768 | /* | |
1769 | * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so | |
1770 | * we know to async punt it even if it was opened O_NONBLOCK | |
1771 | */ | |
1772 | if (force_nonblock && !io_file_supports_async(file)) { | |
1773 | req->flags |= REQ_F_MUST_PUNT; | |
1774 | goto copy_iov; | |
1775 | } | |
9e645e11 | 1776 | |
31b51510 JA |
1777 | iov_count = iov_iter_count(&iter); |
1778 | ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 JA |
1779 | if (!ret) { |
1780 | ssize_t ret2; | |
1781 | ||
32960613 JA |
1782 | if (file->f_op->read_iter) |
1783 | ret2 = call_read_iter(file, kiocb, &iter); | |
1784 | else | |
1785 | ret2 = loop_rw_iter(READ, file, kiocb, &iter); | |
1786 | ||
9d93a3f5 JA |
1787 | /* |
1788 | * In case of a short read, punt to async. This can happen | |
1789 | * if we have data partially cached. Alternatively we can | |
1790 | * return the short read, in which case the application will | |
1791 | * need to issue another SQE and wait for it. That SQE will | |
1792 | * need async punt anyway, so it's more efficient to do it | |
1793 | * here. | |
1794 | */ | |
491381ce JA |
1795 | if (force_nonblock && !(req->flags & REQ_F_NOWAIT) && |
1796 | (req->flags & REQ_F_ISREG) && | |
f67676d1 | 1797 | ret2 > 0 && ret2 < io_size) |
9d93a3f5 JA |
1798 | ret2 = -EAGAIN; |
1799 | /* Catch -EAGAIN return for forced non-blocking submission */ | |
f67676d1 | 1800 | if (!force_nonblock || ret2 != -EAGAIN) { |
cf6fd4bd | 1801 | kiocb_done(kiocb, ret2, nxt, req->in_async); |
f67676d1 JA |
1802 | } else { |
1803 | copy_iov: | |
1804 | ret = io_setup_async_io(req, io_size, iovec, | |
1805 | inline_vecs, &iter); | |
1806 | if (ret) | |
1807 | goto out_free; | |
1808 | return -EAGAIN; | |
1809 | } | |
2b188cc1 | 1810 | } |
f67676d1 | 1811 | out_free: |
2b188cc1 | 1812 | kfree(iovec); |
2b188cc1 JA |
1813 | return ret; |
1814 | } | |
1815 | ||
f67676d1 JA |
1816 | static int io_write_prep(struct io_kiocb *req, struct iovec **iovec, |
1817 | struct iov_iter *iter, bool force_nonblock) | |
1818 | { | |
1819 | ssize_t ret; | |
1820 | ||
1821 | ret = io_prep_rw(req, force_nonblock); | |
1822 | if (ret) | |
1823 | return ret; | |
1824 | ||
1825 | if (unlikely(!(req->file->f_mode & FMODE_WRITE))) | |
1826 | return -EBADF; | |
1827 | ||
1828 | return io_import_iovec(WRITE, req, iovec, iter); | |
1829 | } | |
1830 | ||
267bc904 | 1831 | static int io_write(struct io_kiocb *req, struct io_kiocb **nxt, |
8358e3a8 | 1832 | bool force_nonblock) |
2b188cc1 JA |
1833 | { |
1834 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1835 | struct kiocb *kiocb = &req->rw; | |
1836 | struct iov_iter iter; | |
1837 | struct file *file; | |
31b51510 | 1838 | size_t iov_count; |
f67676d1 | 1839 | ssize_t ret, io_size; |
2b188cc1 | 1840 | |
f67676d1 JA |
1841 | if (!req->io) { |
1842 | ret = io_write_prep(req, &iovec, &iter, force_nonblock); | |
1843 | if (ret < 0) | |
1844 | return ret; | |
1845 | } else { | |
1846 | ret = io_import_iovec(WRITE, req, &iovec, &iter); | |
1847 | if (ret < 0) | |
1848 | return ret; | |
1849 | } | |
2b188cc1 | 1850 | |
2b188cc1 | 1851 | file = kiocb->ki_filp; |
f67676d1 | 1852 | io_size = ret; |
9e645e11 | 1853 | if (req->flags & REQ_F_LINK) |
f67676d1 | 1854 | req->result = io_size; |
9e645e11 | 1855 | |
f67676d1 JA |
1856 | /* |
1857 | * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so | |
1858 | * we know to async punt it even if it was opened O_NONBLOCK | |
1859 | */ | |
1860 | if (force_nonblock && !io_file_supports_async(req->file)) { | |
1861 | req->flags |= REQ_F_MUST_PUNT; | |
1862 | goto copy_iov; | |
1863 | } | |
31b51510 | 1864 | |
561fb04a | 1865 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT)) |
f67676d1 | 1866 | goto copy_iov; |
31b51510 | 1867 | |
f67676d1 | 1868 | iov_count = iov_iter_count(&iter); |
31b51510 | 1869 | ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count); |
2b188cc1 | 1870 | if (!ret) { |
9bf7933f RP |
1871 | ssize_t ret2; |
1872 | ||
2b188cc1 JA |
1873 | /* |
1874 | * Open-code file_start_write here to grab freeze protection, | |
1875 | * which will be released by another thread in | |
1876 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
1877 | * released so that it doesn't complain about the held lock when | |
1878 | * we return to userspace. | |
1879 | */ | |
491381ce | 1880 | if (req->flags & REQ_F_ISREG) { |
2b188cc1 JA |
1881 | __sb_start_write(file_inode(file)->i_sb, |
1882 | SB_FREEZE_WRITE, true); | |
1883 | __sb_writers_release(file_inode(file)->i_sb, | |
1884 | SB_FREEZE_WRITE); | |
1885 | } | |
1886 | kiocb->ki_flags |= IOCB_WRITE; | |
9bf7933f | 1887 | |
32960613 JA |
1888 | if (file->f_op->write_iter) |
1889 | ret2 = call_write_iter(file, kiocb, &iter); | |
1890 | else | |
1891 | ret2 = loop_rw_iter(WRITE, file, kiocb, &iter); | |
f67676d1 | 1892 | if (!force_nonblock || ret2 != -EAGAIN) { |
cf6fd4bd | 1893 | kiocb_done(kiocb, ret2, nxt, req->in_async); |
f67676d1 JA |
1894 | } else { |
1895 | copy_iov: | |
1896 | ret = io_setup_async_io(req, io_size, iovec, | |
1897 | inline_vecs, &iter); | |
1898 | if (ret) | |
1899 | goto out_free; | |
1900 | return -EAGAIN; | |
1901 | } | |
2b188cc1 | 1902 | } |
31b51510 | 1903 | out_free: |
2b188cc1 | 1904 | kfree(iovec); |
2b188cc1 JA |
1905 | return ret; |
1906 | } | |
1907 | ||
1908 | /* | |
1909 | * IORING_OP_NOP just posts a completion event, nothing else. | |
1910 | */ | |
78e19bbe | 1911 | static int io_nop(struct io_kiocb *req) |
2b188cc1 JA |
1912 | { |
1913 | struct io_ring_ctx *ctx = req->ctx; | |
2b188cc1 | 1914 | |
def596e9 JA |
1915 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
1916 | return -EINVAL; | |
1917 | ||
78e19bbe | 1918 | io_cqring_add_event(req, 0); |
e65ef56d | 1919 | io_put_req(req); |
2b188cc1 JA |
1920 | return 0; |
1921 | } | |
1922 | ||
c992fe29 CH |
1923 | static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1924 | { | |
6b06314c | 1925 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 1926 | |
09bb8394 JA |
1927 | if (!req->file) |
1928 | return -EBADF; | |
c992fe29 | 1929 | |
6b06314c | 1930 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 1931 | return -EINVAL; |
edafccee | 1932 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
1933 | return -EINVAL; |
1934 | ||
c992fe29 CH |
1935 | return 0; |
1936 | } | |
1937 | ||
1938 | static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
ba816ad6 | 1939 | struct io_kiocb **nxt, bool force_nonblock) |
c992fe29 CH |
1940 | { |
1941 | loff_t sqe_off = READ_ONCE(sqe->off); | |
1942 | loff_t sqe_len = READ_ONCE(sqe->len); | |
1943 | loff_t end = sqe_off + sqe_len; | |
1944 | unsigned fsync_flags; | |
1945 | int ret; | |
1946 | ||
1947 | fsync_flags = READ_ONCE(sqe->fsync_flags); | |
1948 | if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC)) | |
1949 | return -EINVAL; | |
1950 | ||
1951 | ret = io_prep_fsync(req, sqe); | |
1952 | if (ret) | |
1953 | return ret; | |
1954 | ||
1955 | /* fsync always requires a blocking context */ | |
1956 | if (force_nonblock) | |
1957 | return -EAGAIN; | |
1958 | ||
1959 | ret = vfs_fsync_range(req->rw.ki_filp, sqe_off, | |
1960 | end > 0 ? end : LLONG_MAX, | |
1961 | fsync_flags & IORING_FSYNC_DATASYNC); | |
1962 | ||
9e645e11 JA |
1963 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1964 | req->flags |= REQ_F_FAIL_LINK; | |
78e19bbe | 1965 | io_cqring_add_event(req, ret); |
ec9c02ad | 1966 | io_put_req_find_next(req, nxt); |
c992fe29 CH |
1967 | return 0; |
1968 | } | |
1969 | ||
5d17b4a4 JA |
1970 | static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1971 | { | |
1972 | struct io_ring_ctx *ctx = req->ctx; | |
1973 | int ret = 0; | |
1974 | ||
1975 | if (!req->file) | |
1976 | return -EBADF; | |
5d17b4a4 JA |
1977 | |
1978 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
1979 | return -EINVAL; | |
1980 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) | |
1981 | return -EINVAL; | |
1982 | ||
5d17b4a4 JA |
1983 | return ret; |
1984 | } | |
1985 | ||
1986 | static int io_sync_file_range(struct io_kiocb *req, | |
1987 | const struct io_uring_sqe *sqe, | |
ba816ad6 | 1988 | struct io_kiocb **nxt, |
5d17b4a4 JA |
1989 | bool force_nonblock) |
1990 | { | |
1991 | loff_t sqe_off; | |
1992 | loff_t sqe_len; | |
1993 | unsigned flags; | |
1994 | int ret; | |
1995 | ||
1996 | ret = io_prep_sfr(req, sqe); | |
1997 | if (ret) | |
1998 | return ret; | |
1999 | ||
2000 | /* sync_file_range always requires a blocking context */ | |
2001 | if (force_nonblock) | |
2002 | return -EAGAIN; | |
2003 | ||
2004 | sqe_off = READ_ONCE(sqe->off); | |
2005 | sqe_len = READ_ONCE(sqe->len); | |
2006 | flags = READ_ONCE(sqe->sync_range_flags); | |
2007 | ||
2008 | ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags); | |
2009 | ||
9e645e11 JA |
2010 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
2011 | req->flags |= REQ_F_FAIL_LINK; | |
78e19bbe | 2012 | io_cqring_add_event(req, ret); |
ec9c02ad | 2013 | io_put_req_find_next(req, nxt); |
5d17b4a4 JA |
2014 | return 0; |
2015 | } | |
2016 | ||
03b1230c JA |
2017 | static int io_sendmsg_prep(struct io_kiocb *req, struct io_async_ctx *io) |
2018 | { | |
0fa03c62 | 2019 | #if defined(CONFIG_NET) |
03b1230c JA |
2020 | const struct io_uring_sqe *sqe = req->sqe; |
2021 | struct user_msghdr __user *msg; | |
2022 | unsigned flags; | |
2023 | ||
2024 | flags = READ_ONCE(sqe->msg_flags); | |
2025 | msg = (struct user_msghdr __user *)(unsigned long) READ_ONCE(sqe->addr); | |
2026 | return sendmsg_copy_msghdr(&io->msg.msg, msg, flags, &io->msg.iov); | |
2027 | #else | |
2028 | return 0; | |
2029 | #endif | |
2030 | } | |
2031 | ||
2032 | static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
2033 | struct io_kiocb **nxt, bool force_nonblock) | |
aa1fa28f | 2034 | { |
03b1230c | 2035 | #if defined(CONFIG_NET) |
0fa03c62 JA |
2036 | struct socket *sock; |
2037 | int ret; | |
2038 | ||
2039 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
2040 | return -EINVAL; | |
2041 | ||
2042 | sock = sock_from_file(req->file, &ret); | |
2043 | if (sock) { | |
03b1230c JA |
2044 | struct io_async_ctx io, *copy; |
2045 | struct sockaddr_storage addr; | |
2046 | struct msghdr *kmsg; | |
0fa03c62 JA |
2047 | unsigned flags; |
2048 | ||
2049 | flags = READ_ONCE(sqe->msg_flags); | |
2050 | if (flags & MSG_DONTWAIT) | |
2051 | req->flags |= REQ_F_NOWAIT; | |
2052 | else if (force_nonblock) | |
2053 | flags |= MSG_DONTWAIT; | |
2054 | ||
03b1230c JA |
2055 | if (req->io) { |
2056 | kmsg = &req->io->msg.msg; | |
2057 | kmsg->msg_name = &addr; | |
2058 | } else { | |
2059 | kmsg = &io.msg.msg; | |
2060 | kmsg->msg_name = &addr; | |
2061 | io.msg.iov = io.msg.fast_iov; | |
2062 | ret = io_sendmsg_prep(req, &io); | |
2063 | if (ret) | |
2064 | goto out; | |
2065 | } | |
0fa03c62 | 2066 | |
03b1230c JA |
2067 | ret = __sys_sendmsg_sock(sock, kmsg, flags); |
2068 | if (force_nonblock && ret == -EAGAIN) { | |
2069 | copy = kmalloc(sizeof(*copy), GFP_KERNEL); | |
2070 | if (!copy) { | |
2071 | ret = -ENOMEM; | |
2072 | goto out; | |
2073 | } | |
2074 | memcpy(©->msg, &io.msg, sizeof(copy->msg)); | |
2075 | req->io = copy; | |
2076 | memcpy(&req->io->sqe, req->sqe, sizeof(*req->sqe)); | |
2077 | req->sqe = &req->io->sqe; | |
0fa03c62 | 2078 | return ret; |
03b1230c | 2079 | } |
441cdbd5 JA |
2080 | if (ret == -ERESTARTSYS) |
2081 | ret = -EINTR; | |
0fa03c62 JA |
2082 | } |
2083 | ||
03b1230c | 2084 | out: |
78e19bbe | 2085 | io_cqring_add_event(req, ret); |
f1f40853 JA |
2086 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
2087 | req->flags |= REQ_F_FAIL_LINK; | |
ec9c02ad | 2088 | io_put_req_find_next(req, nxt); |
5d17b4a4 | 2089 | return 0; |
03b1230c JA |
2090 | #else |
2091 | return -EOPNOTSUPP; | |
aa1fa28f | 2092 | #endif |
03b1230c | 2093 | } |
aa1fa28f | 2094 | |
03b1230c | 2095 | static int io_recvmsg_prep(struct io_kiocb *req, struct io_async_ctx *io) |
aa1fa28f JA |
2096 | { |
2097 | #if defined(CONFIG_NET) | |
03b1230c JA |
2098 | const struct io_uring_sqe *sqe = req->sqe; |
2099 | struct user_msghdr __user *msg; | |
2100 | unsigned flags; | |
2101 | ||
2102 | flags = READ_ONCE(sqe->msg_flags); | |
2103 | msg = (struct user_msghdr __user *)(unsigned long) READ_ONCE(sqe->addr); | |
2104 | return recvmsg_copy_msghdr(&io->msg.msg, msg, flags, &io->msg.uaddr, | |
2105 | &io->msg.iov); | |
aa1fa28f | 2106 | #else |
03b1230c | 2107 | return 0; |
aa1fa28f JA |
2108 | #endif |
2109 | } | |
2110 | ||
2111 | static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
ba816ad6 | 2112 | struct io_kiocb **nxt, bool force_nonblock) |
aa1fa28f JA |
2113 | { |
2114 | #if defined(CONFIG_NET) | |
03b1230c JA |
2115 | struct socket *sock; |
2116 | int ret; | |
2117 | ||
2118 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
2119 | return -EINVAL; | |
2120 | ||
2121 | sock = sock_from_file(req->file, &ret); | |
2122 | if (sock) { | |
2123 | struct user_msghdr __user *msg; | |
2124 | struct io_async_ctx io, *copy; | |
2125 | struct sockaddr_storage addr; | |
2126 | struct msghdr *kmsg; | |
2127 | unsigned flags; | |
2128 | ||
2129 | flags = READ_ONCE(sqe->msg_flags); | |
2130 | if (flags & MSG_DONTWAIT) | |
2131 | req->flags |= REQ_F_NOWAIT; | |
2132 | else if (force_nonblock) | |
2133 | flags |= MSG_DONTWAIT; | |
2134 | ||
2135 | msg = (struct user_msghdr __user *) (unsigned long) | |
2136 | READ_ONCE(sqe->addr); | |
2137 | if (req->io) { | |
2138 | kmsg = &req->io->msg.msg; | |
2139 | kmsg->msg_name = &addr; | |
2140 | } else { | |
2141 | kmsg = &io.msg.msg; | |
2142 | kmsg->msg_name = &addr; | |
2143 | io.msg.iov = io.msg.fast_iov; | |
2144 | ret = io_recvmsg_prep(req, &io); | |
2145 | if (ret) | |
2146 | goto out; | |
2147 | } | |
2148 | ||
2149 | ret = __sys_recvmsg_sock(sock, kmsg, msg, io.msg.uaddr, flags); | |
2150 | if (force_nonblock && ret == -EAGAIN) { | |
2151 | copy = kmalloc(sizeof(*copy), GFP_KERNEL); | |
2152 | if (!copy) { | |
2153 | ret = -ENOMEM; | |
2154 | goto out; | |
2155 | } | |
2156 | memcpy(copy, &io, sizeof(*copy)); | |
2157 | req->io = copy; | |
2158 | memcpy(&req->io->sqe, req->sqe, sizeof(*req->sqe)); | |
2159 | req->sqe = &req->io->sqe; | |
2160 | return ret; | |
2161 | } | |
2162 | if (ret == -ERESTARTSYS) | |
2163 | ret = -EINTR; | |
2164 | } | |
2165 | ||
2166 | out: | |
2167 | io_cqring_add_event(req, ret); | |
2168 | if (ret < 0 && (req->flags & REQ_F_LINK)) | |
2169 | req->flags |= REQ_F_FAIL_LINK; | |
2170 | io_put_req_find_next(req, nxt); | |
2171 | return 0; | |
0fa03c62 JA |
2172 | #else |
2173 | return -EOPNOTSUPP; | |
2174 | #endif | |
2175 | } | |
5d17b4a4 | 2176 | |
17f2fe35 JA |
2177 | static int io_accept(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2178 | struct io_kiocb **nxt, bool force_nonblock) | |
2179 | { | |
2180 | #if defined(CONFIG_NET) | |
2181 | struct sockaddr __user *addr; | |
2182 | int __user *addr_len; | |
2183 | unsigned file_flags; | |
2184 | int flags, ret; | |
2185 | ||
2186 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) | |
2187 | return -EINVAL; | |
8042d6ce | 2188 | if (sqe->ioprio || sqe->len || sqe->buf_index) |
17f2fe35 JA |
2189 | return -EINVAL; |
2190 | ||
2191 | addr = (struct sockaddr __user *) (unsigned long) READ_ONCE(sqe->addr); | |
2192 | addr_len = (int __user *) (unsigned long) READ_ONCE(sqe->addr2); | |
2193 | flags = READ_ONCE(sqe->accept_flags); | |
2194 | file_flags = force_nonblock ? O_NONBLOCK : 0; | |
2195 | ||
2196 | ret = __sys_accept4_file(req->file, file_flags, addr, addr_len, flags); | |
2197 | if (ret == -EAGAIN && force_nonblock) { | |
2198 | req->work.flags |= IO_WQ_WORK_NEEDS_FILES; | |
2199 | return -EAGAIN; | |
2200 | } | |
8e3cca12 JA |
2201 | if (ret == -ERESTARTSYS) |
2202 | ret = -EINTR; | |
17f2fe35 JA |
2203 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
2204 | req->flags |= REQ_F_FAIL_LINK; | |
78e19bbe | 2205 | io_cqring_add_event(req, ret); |
ec9c02ad | 2206 | io_put_req_find_next(req, nxt); |
17f2fe35 | 2207 | return 0; |
0fa03c62 JA |
2208 | #else |
2209 | return -EOPNOTSUPP; | |
2210 | #endif | |
2211 | } | |
5d17b4a4 | 2212 | |
f499a021 JA |
2213 | static int io_connect_prep(struct io_kiocb *req, struct io_async_ctx *io) |
2214 | { | |
2215 | #if defined(CONFIG_NET) | |
2216 | const struct io_uring_sqe *sqe = req->sqe; | |
2217 | struct sockaddr __user *addr; | |
2218 | int addr_len; | |
2219 | ||
2220 | addr = (struct sockaddr __user *) (unsigned long) READ_ONCE(sqe->addr); | |
2221 | addr_len = READ_ONCE(sqe->addr2); | |
2222 | return move_addr_to_kernel(addr, addr_len, &io->connect.address); | |
2223 | #else | |
2224 | return 0; | |
2225 | #endif | |
2226 | } | |
2227 | ||
f8e85cf2 JA |
2228 | static int io_connect(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2229 | struct io_kiocb **nxt, bool force_nonblock) | |
2230 | { | |
2231 | #if defined(CONFIG_NET) | |
f499a021 | 2232 | struct io_async_ctx __io, *io; |
f8e85cf2 JA |
2233 | unsigned file_flags; |
2234 | int addr_len, ret; | |
2235 | ||
2236 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) | |
2237 | return -EINVAL; | |
2238 | if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags) | |
2239 | return -EINVAL; | |
2240 | ||
f8e85cf2 JA |
2241 | addr_len = READ_ONCE(sqe->addr2); |
2242 | file_flags = force_nonblock ? O_NONBLOCK : 0; | |
2243 | ||
f499a021 JA |
2244 | if (req->io) { |
2245 | io = req->io; | |
2246 | } else { | |
2247 | ret = io_connect_prep(req, &__io); | |
2248 | if (ret) | |
2249 | goto out; | |
2250 | io = &__io; | |
2251 | } | |
2252 | ||
2253 | ret = __sys_connect_file(req->file, &io->connect.address, addr_len, | |
2254 | file_flags); | |
87f80d62 | 2255 | if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) { |
f499a021 JA |
2256 | io = kmalloc(sizeof(*io), GFP_KERNEL); |
2257 | if (!io) { | |
2258 | ret = -ENOMEM; | |
2259 | goto out; | |
2260 | } | |
2261 | memcpy(&io->connect, &__io.connect, sizeof(io->connect)); | |
2262 | req->io = io; | |
2263 | memcpy(&io->sqe, req->sqe, sizeof(*req->sqe)); | |
2264 | req->sqe = &io->sqe; | |
f8e85cf2 | 2265 | return -EAGAIN; |
f499a021 | 2266 | } |
f8e85cf2 JA |
2267 | if (ret == -ERESTARTSYS) |
2268 | ret = -EINTR; | |
f499a021 | 2269 | out: |
f8e85cf2 JA |
2270 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
2271 | req->flags |= REQ_F_FAIL_LINK; | |
2272 | io_cqring_add_event(req, ret); | |
2273 | io_put_req_find_next(req, nxt); | |
2274 | return 0; | |
2275 | #else | |
2276 | return -EOPNOTSUPP; | |
2277 | #endif | |
2278 | } | |
2279 | ||
221c5eb2 JA |
2280 | static void io_poll_remove_one(struct io_kiocb *req) |
2281 | { | |
2282 | struct io_poll_iocb *poll = &req->poll; | |
2283 | ||
2284 | spin_lock(&poll->head->lock); | |
2285 | WRITE_ONCE(poll->canceled, true); | |
e944475e JA |
2286 | if (!list_empty(&poll->wait->entry)) { |
2287 | list_del_init(&poll->wait->entry); | |
a197f664 | 2288 | io_queue_async_work(req); |
221c5eb2 JA |
2289 | } |
2290 | spin_unlock(&poll->head->lock); | |
78076bb6 | 2291 | hash_del(&req->hash_node); |
221c5eb2 JA |
2292 | } |
2293 | ||
2294 | static void io_poll_remove_all(struct io_ring_ctx *ctx) | |
2295 | { | |
78076bb6 | 2296 | struct hlist_node *tmp; |
221c5eb2 | 2297 | struct io_kiocb *req; |
78076bb6 | 2298 | int i; |
221c5eb2 JA |
2299 | |
2300 | spin_lock_irq(&ctx->completion_lock); | |
78076bb6 JA |
2301 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { |
2302 | struct hlist_head *list; | |
2303 | ||
2304 | list = &ctx->cancel_hash[i]; | |
2305 | hlist_for_each_entry_safe(req, tmp, list, hash_node) | |
2306 | io_poll_remove_one(req); | |
221c5eb2 JA |
2307 | } |
2308 | spin_unlock_irq(&ctx->completion_lock); | |
2309 | } | |
2310 | ||
47f46768 JA |
2311 | static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr) |
2312 | { | |
78076bb6 | 2313 | struct hlist_head *list; |
47f46768 JA |
2314 | struct io_kiocb *req; |
2315 | ||
78076bb6 JA |
2316 | list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)]; |
2317 | hlist_for_each_entry(req, list, hash_node) { | |
2318 | if (sqe_addr == req->user_data) { | |
eac406c6 JA |
2319 | io_poll_remove_one(req); |
2320 | return 0; | |
2321 | } | |
47f46768 JA |
2322 | } |
2323 | ||
2324 | return -ENOENT; | |
2325 | } | |
2326 | ||
221c5eb2 JA |
2327 | /* |
2328 | * Find a running poll command that matches one specified in sqe->addr, | |
2329 | * and remove it if found. | |
2330 | */ | |
2331 | static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
2332 | { | |
2333 | struct io_ring_ctx *ctx = req->ctx; | |
47f46768 | 2334 | int ret; |
221c5eb2 JA |
2335 | |
2336 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
2337 | return -EINVAL; | |
2338 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index || | |
2339 | sqe->poll_events) | |
2340 | return -EINVAL; | |
2341 | ||
2342 | spin_lock_irq(&ctx->completion_lock); | |
47f46768 | 2343 | ret = io_poll_cancel(ctx, READ_ONCE(sqe->addr)); |
221c5eb2 JA |
2344 | spin_unlock_irq(&ctx->completion_lock); |
2345 | ||
78e19bbe | 2346 | io_cqring_add_event(req, ret); |
f1f40853 JA |
2347 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
2348 | req->flags |= REQ_F_FAIL_LINK; | |
e65ef56d | 2349 | io_put_req(req); |
221c5eb2 JA |
2350 | return 0; |
2351 | } | |
2352 | ||
b0dd8a41 | 2353 | static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error) |
221c5eb2 | 2354 | { |
a197f664 JL |
2355 | struct io_ring_ctx *ctx = req->ctx; |
2356 | ||
8c838788 | 2357 | req->poll.done = true; |
e944475e | 2358 | kfree(req->poll.wait); |
b0dd8a41 JA |
2359 | if (error) |
2360 | io_cqring_fill_event(req, error); | |
2361 | else | |
2362 | io_cqring_fill_event(req, mangle_poll(mask)); | |
8c838788 | 2363 | io_commit_cqring(ctx); |
221c5eb2 JA |
2364 | } |
2365 | ||
561fb04a | 2366 | static void io_poll_complete_work(struct io_wq_work **workptr) |
221c5eb2 | 2367 | { |
561fb04a | 2368 | struct io_wq_work *work = *workptr; |
221c5eb2 JA |
2369 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
2370 | struct io_poll_iocb *poll = &req->poll; | |
2371 | struct poll_table_struct pt = { ._key = poll->events }; | |
2372 | struct io_ring_ctx *ctx = req->ctx; | |
89723d0b | 2373 | struct io_kiocb *nxt = NULL; |
221c5eb2 | 2374 | __poll_t mask = 0; |
b0dd8a41 | 2375 | int ret = 0; |
221c5eb2 | 2376 | |
b0dd8a41 | 2377 | if (work->flags & IO_WQ_WORK_CANCEL) { |
561fb04a | 2378 | WRITE_ONCE(poll->canceled, true); |
b0dd8a41 JA |
2379 | ret = -ECANCELED; |
2380 | } else if (READ_ONCE(poll->canceled)) { | |
2381 | ret = -ECANCELED; | |
2382 | } | |
561fb04a | 2383 | |
b0dd8a41 | 2384 | if (ret != -ECANCELED) |
221c5eb2 JA |
2385 | mask = vfs_poll(poll->file, &pt) & poll->events; |
2386 | ||
2387 | /* | |
2388 | * Note that ->ki_cancel callers also delete iocb from active_reqs after | |
2389 | * calling ->ki_cancel. We need the ctx_lock roundtrip here to | |
2390 | * synchronize with them. In the cancellation case the list_del_init | |
2391 | * itself is not actually needed, but harmless so we keep it in to | |
2392 | * avoid further branches in the fast path. | |
2393 | */ | |
2394 | spin_lock_irq(&ctx->completion_lock); | |
b0dd8a41 | 2395 | if (!mask && ret != -ECANCELED) { |
e944475e | 2396 | add_wait_queue(poll->head, poll->wait); |
221c5eb2 JA |
2397 | spin_unlock_irq(&ctx->completion_lock); |
2398 | return; | |
2399 | } | |
78076bb6 | 2400 | hash_del(&req->hash_node); |
b0dd8a41 | 2401 | io_poll_complete(req, mask, ret); |
221c5eb2 JA |
2402 | spin_unlock_irq(&ctx->completion_lock); |
2403 | ||
8c838788 | 2404 | io_cqring_ev_posted(ctx); |
89723d0b | 2405 | |
fba38c27 JA |
2406 | if (ret < 0 && req->flags & REQ_F_LINK) |
2407 | req->flags |= REQ_F_FAIL_LINK; | |
ec9c02ad | 2408 | io_put_req_find_next(req, &nxt); |
89723d0b JA |
2409 | if (nxt) |
2410 | *workptr = &nxt->work; | |
221c5eb2 JA |
2411 | } |
2412 | ||
2413 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
2414 | void *key) | |
2415 | { | |
e944475e | 2416 | struct io_poll_iocb *poll = wait->private; |
221c5eb2 JA |
2417 | struct io_kiocb *req = container_of(poll, struct io_kiocb, poll); |
2418 | struct io_ring_ctx *ctx = req->ctx; | |
2419 | __poll_t mask = key_to_poll(key); | |
8c838788 | 2420 | unsigned long flags; |
221c5eb2 JA |
2421 | |
2422 | /* for instances that support it check for an event match first: */ | |
8c838788 JA |
2423 | if (mask && !(mask & poll->events)) |
2424 | return 0; | |
221c5eb2 | 2425 | |
e944475e | 2426 | list_del_init(&poll->wait->entry); |
221c5eb2 | 2427 | |
7c9e7f0f JA |
2428 | /* |
2429 | * Run completion inline if we can. We're using trylock here because | |
2430 | * we are violating the completion_lock -> poll wq lock ordering. | |
2431 | * If we have a link timeout we're going to need the completion_lock | |
2432 | * for finalizing the request, mark us as having grabbed that already. | |
2433 | */ | |
8c838788 | 2434 | if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) { |
78076bb6 | 2435 | hash_del(&req->hash_node); |
b0dd8a41 | 2436 | io_poll_complete(req, mask, 0); |
7c9e7f0f JA |
2437 | req->flags |= REQ_F_COMP_LOCKED; |
2438 | io_put_req(req); | |
8c838788 | 2439 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
221c5eb2 | 2440 | |
8c838788 | 2441 | io_cqring_ev_posted(ctx); |
8c838788 | 2442 | } else { |
a197f664 | 2443 | io_queue_async_work(req); |
221c5eb2 JA |
2444 | } |
2445 | ||
221c5eb2 JA |
2446 | return 1; |
2447 | } | |
2448 | ||
2449 | struct io_poll_table { | |
2450 | struct poll_table_struct pt; | |
2451 | struct io_kiocb *req; | |
2452 | int error; | |
2453 | }; | |
2454 | ||
2455 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
2456 | struct poll_table_struct *p) | |
2457 | { | |
2458 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
2459 | ||
2460 | if (unlikely(pt->req->poll.head)) { | |
2461 | pt->error = -EINVAL; | |
2462 | return; | |
2463 | } | |
2464 | ||
2465 | pt->error = 0; | |
2466 | pt->req->poll.head = head; | |
e944475e | 2467 | add_wait_queue(head, pt->req->poll.wait); |
221c5eb2 JA |
2468 | } |
2469 | ||
eac406c6 JA |
2470 | static void io_poll_req_insert(struct io_kiocb *req) |
2471 | { | |
2472 | struct io_ring_ctx *ctx = req->ctx; | |
78076bb6 JA |
2473 | struct hlist_head *list; |
2474 | ||
2475 | list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)]; | |
2476 | hlist_add_head(&req->hash_node, list); | |
eac406c6 JA |
2477 | } |
2478 | ||
89723d0b JA |
2479 | static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2480 | struct io_kiocb **nxt) | |
221c5eb2 JA |
2481 | { |
2482 | struct io_poll_iocb *poll = &req->poll; | |
2483 | struct io_ring_ctx *ctx = req->ctx; | |
2484 | struct io_poll_table ipt; | |
8c838788 | 2485 | bool cancel = false; |
221c5eb2 JA |
2486 | __poll_t mask; |
2487 | u16 events; | |
221c5eb2 JA |
2488 | |
2489 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
2490 | return -EINVAL; | |
2491 | if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
2492 | return -EINVAL; | |
09bb8394 JA |
2493 | if (!poll->file) |
2494 | return -EBADF; | |
221c5eb2 | 2495 | |
e944475e JA |
2496 | poll->wait = kmalloc(sizeof(*poll->wait), GFP_KERNEL); |
2497 | if (!poll->wait) | |
2498 | return -ENOMEM; | |
2499 | ||
1a6b74fc | 2500 | req->io = NULL; |
561fb04a | 2501 | INIT_IO_WORK(&req->work, io_poll_complete_work); |
221c5eb2 JA |
2502 | events = READ_ONCE(sqe->poll_events); |
2503 | poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP; | |
78076bb6 | 2504 | INIT_HLIST_NODE(&req->hash_node); |
221c5eb2 | 2505 | |
221c5eb2 | 2506 | poll->head = NULL; |
8c838788 | 2507 | poll->done = false; |
221c5eb2 JA |
2508 | poll->canceled = false; |
2509 | ||
2510 | ipt.pt._qproc = io_poll_queue_proc; | |
2511 | ipt.pt._key = poll->events; | |
2512 | ipt.req = req; | |
2513 | ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */ | |
2514 | ||
2515 | /* initialized the list so that we can do list_empty checks */ | |
e944475e JA |
2516 | INIT_LIST_HEAD(&poll->wait->entry); |
2517 | init_waitqueue_func_entry(poll->wait, io_poll_wake); | |
2518 | poll->wait->private = poll; | |
221c5eb2 | 2519 | |
36703247 JA |
2520 | INIT_LIST_HEAD(&req->list); |
2521 | ||
221c5eb2 | 2522 | mask = vfs_poll(poll->file, &ipt.pt) & poll->events; |
221c5eb2 JA |
2523 | |
2524 | spin_lock_irq(&ctx->completion_lock); | |
8c838788 JA |
2525 | if (likely(poll->head)) { |
2526 | spin_lock(&poll->head->lock); | |
e944475e | 2527 | if (unlikely(list_empty(&poll->wait->entry))) { |
8c838788 JA |
2528 | if (ipt.error) |
2529 | cancel = true; | |
2530 | ipt.error = 0; | |
2531 | mask = 0; | |
2532 | } | |
2533 | if (mask || ipt.error) | |
e944475e | 2534 | list_del_init(&poll->wait->entry); |
8c838788 JA |
2535 | else if (cancel) |
2536 | WRITE_ONCE(poll->canceled, true); | |
2537 | else if (!poll->done) /* actually waiting for an event */ | |
eac406c6 | 2538 | io_poll_req_insert(req); |
8c838788 JA |
2539 | spin_unlock(&poll->head->lock); |
2540 | } | |
2541 | if (mask) { /* no async, we'd stolen it */ | |
221c5eb2 | 2542 | ipt.error = 0; |
b0dd8a41 | 2543 | io_poll_complete(req, mask, 0); |
221c5eb2 | 2544 | } |
221c5eb2 JA |
2545 | spin_unlock_irq(&ctx->completion_lock); |
2546 | ||
8c838788 JA |
2547 | if (mask) { |
2548 | io_cqring_ev_posted(ctx); | |
ec9c02ad | 2549 | io_put_req_find_next(req, nxt); |
221c5eb2 | 2550 | } |
8c838788 | 2551 | return ipt.error; |
221c5eb2 JA |
2552 | } |
2553 | ||
5262f567 JA |
2554 | static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer) |
2555 | { | |
ad8a48ac JA |
2556 | struct io_timeout_data *data = container_of(timer, |
2557 | struct io_timeout_data, timer); | |
2558 | struct io_kiocb *req = data->req; | |
2559 | struct io_ring_ctx *ctx = req->ctx; | |
5262f567 JA |
2560 | unsigned long flags; |
2561 | ||
5262f567 JA |
2562 | atomic_inc(&ctx->cq_timeouts); |
2563 | ||
2564 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
ef03681a | 2565 | /* |
11365043 JA |
2566 | * We could be racing with timeout deletion. If the list is empty, |
2567 | * then timeout lookup already found it and will be handling it. | |
ef03681a | 2568 | */ |
842f9612 | 2569 | if (!list_empty(&req->list)) { |
11365043 | 2570 | struct io_kiocb *prev; |
5262f567 | 2571 | |
11365043 JA |
2572 | /* |
2573 | * Adjust the reqs sequence before the current one because it | |
2574 | * will consume a slot in the cq_ring and the the cq_tail | |
2575 | * pointer will be increased, otherwise other timeout reqs may | |
2576 | * return in advance without waiting for enough wait_nr. | |
2577 | */ | |
2578 | prev = req; | |
2579 | list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list) | |
2580 | prev->sequence++; | |
11365043 | 2581 | list_del_init(&req->list); |
11365043 | 2582 | } |
5262f567 | 2583 | |
78e19bbe | 2584 | io_cqring_fill_event(req, -ETIME); |
5262f567 JA |
2585 | io_commit_cqring(ctx); |
2586 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
2587 | ||
2588 | io_cqring_ev_posted(ctx); | |
f1f40853 JA |
2589 | if (req->flags & REQ_F_LINK) |
2590 | req->flags |= REQ_F_FAIL_LINK; | |
5262f567 JA |
2591 | io_put_req(req); |
2592 | return HRTIMER_NORESTART; | |
2593 | } | |
2594 | ||
47f46768 JA |
2595 | static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data) |
2596 | { | |
2597 | struct io_kiocb *req; | |
2598 | int ret = -ENOENT; | |
2599 | ||
2600 | list_for_each_entry(req, &ctx->timeout_list, list) { | |
2601 | if (user_data == req->user_data) { | |
2602 | list_del_init(&req->list); | |
2603 | ret = 0; | |
2604 | break; | |
2605 | } | |
2606 | } | |
2607 | ||
2608 | if (ret == -ENOENT) | |
2609 | return ret; | |
2610 | ||
2d28390a | 2611 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); |
47f46768 JA |
2612 | if (ret == -1) |
2613 | return -EALREADY; | |
2614 | ||
fba38c27 JA |
2615 | if (req->flags & REQ_F_LINK) |
2616 | req->flags |= REQ_F_FAIL_LINK; | |
47f46768 JA |
2617 | io_cqring_fill_event(req, -ECANCELED); |
2618 | io_put_req(req); | |
2619 | return 0; | |
2620 | } | |
2621 | ||
11365043 JA |
2622 | /* |
2623 | * Remove or update an existing timeout command | |
2624 | */ | |
2625 | static int io_timeout_remove(struct io_kiocb *req, | |
2626 | const struct io_uring_sqe *sqe) | |
2627 | { | |
2628 | struct io_ring_ctx *ctx = req->ctx; | |
11365043 | 2629 | unsigned flags; |
47f46768 | 2630 | int ret; |
11365043 JA |
2631 | |
2632 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
2633 | return -EINVAL; | |
2634 | if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len) | |
2635 | return -EINVAL; | |
2636 | flags = READ_ONCE(sqe->timeout_flags); | |
2637 | if (flags) | |
2638 | return -EINVAL; | |
2639 | ||
11365043 | 2640 | spin_lock_irq(&ctx->completion_lock); |
47f46768 | 2641 | ret = io_timeout_cancel(ctx, READ_ONCE(sqe->addr)); |
11365043 | 2642 | |
47f46768 | 2643 | io_cqring_fill_event(req, ret); |
11365043 JA |
2644 | io_commit_cqring(ctx); |
2645 | spin_unlock_irq(&ctx->completion_lock); | |
5262f567 | 2646 | io_cqring_ev_posted(ctx); |
47f46768 JA |
2647 | if (ret < 0 && req->flags & REQ_F_LINK) |
2648 | req->flags |= REQ_F_FAIL_LINK; | |
ec9c02ad | 2649 | io_put_req(req); |
11365043 | 2650 | return 0; |
5262f567 JA |
2651 | } |
2652 | ||
2d28390a JA |
2653 | static int io_timeout_prep(struct io_kiocb *req, struct io_async_ctx *io, |
2654 | bool is_timeout_link) | |
5262f567 | 2655 | { |
cf6fd4bd | 2656 | const struct io_uring_sqe *sqe = req->sqe; |
ad8a48ac | 2657 | struct io_timeout_data *data; |
a41525ab | 2658 | unsigned flags; |
5262f567 | 2659 | |
ad8a48ac | 2660 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5262f567 | 2661 | return -EINVAL; |
ad8a48ac | 2662 | if (sqe->ioprio || sqe->buf_index || sqe->len != 1) |
a41525ab | 2663 | return -EINVAL; |
2d28390a JA |
2664 | if (sqe->off && is_timeout_link) |
2665 | return -EINVAL; | |
a41525ab JA |
2666 | flags = READ_ONCE(sqe->timeout_flags); |
2667 | if (flags & ~IORING_TIMEOUT_ABS) | |
5262f567 | 2668 | return -EINVAL; |
bdf20073 | 2669 | |
2d28390a | 2670 | data = &io->timeout; |
ad8a48ac | 2671 | data->req = req; |
ad8a48ac JA |
2672 | req->flags |= REQ_F_TIMEOUT; |
2673 | ||
2674 | if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr))) | |
5262f567 JA |
2675 | return -EFAULT; |
2676 | ||
11365043 | 2677 | if (flags & IORING_TIMEOUT_ABS) |
ad8a48ac | 2678 | data->mode = HRTIMER_MODE_ABS; |
11365043 | 2679 | else |
ad8a48ac | 2680 | data->mode = HRTIMER_MODE_REL; |
11365043 | 2681 | |
ad8a48ac | 2682 | hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode); |
2d28390a | 2683 | req->io = io; |
ad8a48ac JA |
2684 | return 0; |
2685 | } | |
2686 | ||
2687 | static int io_timeout(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
2688 | { | |
2689 | unsigned count; | |
2690 | struct io_ring_ctx *ctx = req->ctx; | |
2691 | struct io_timeout_data *data; | |
2d28390a | 2692 | struct io_async_ctx *io; |
ad8a48ac JA |
2693 | struct list_head *entry; |
2694 | unsigned span = 0; | |
ad8a48ac | 2695 | |
2d28390a JA |
2696 | io = req->io; |
2697 | if (!io) { | |
2698 | int ret; | |
2699 | ||
2700 | io = kmalloc(sizeof(*io), GFP_KERNEL); | |
2701 | if (!io) | |
2702 | return -ENOMEM; | |
2703 | ret = io_timeout_prep(req, io, false); | |
2704 | if (ret) { | |
2705 | kfree(io); | |
2706 | return ret; | |
2707 | } | |
2708 | } | |
2709 | data = &req->io->timeout; | |
93bd25bb | 2710 | |
5262f567 JA |
2711 | /* |
2712 | * sqe->off holds how many events that need to occur for this | |
93bd25bb JA |
2713 | * timeout event to be satisfied. If it isn't set, then this is |
2714 | * a pure timeout request, sequence isn't used. | |
5262f567 JA |
2715 | */ |
2716 | count = READ_ONCE(sqe->off); | |
93bd25bb JA |
2717 | if (!count) { |
2718 | req->flags |= REQ_F_TIMEOUT_NOSEQ; | |
2719 | spin_lock_irq(&ctx->completion_lock); | |
2720 | entry = ctx->timeout_list.prev; | |
2721 | goto add; | |
2722 | } | |
5262f567 JA |
2723 | |
2724 | req->sequence = ctx->cached_sq_head + count - 1; | |
2d28390a | 2725 | data->seq_offset = count; |
5262f567 JA |
2726 | |
2727 | /* | |
2728 | * Insertion sort, ensuring the first entry in the list is always | |
2729 | * the one we need first. | |
2730 | */ | |
5262f567 JA |
2731 | spin_lock_irq(&ctx->completion_lock); |
2732 | list_for_each_prev(entry, &ctx->timeout_list) { | |
2733 | struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list); | |
5da0fb1a | 2734 | unsigned nxt_sq_head; |
2735 | long long tmp, tmp_nxt; | |
2d28390a | 2736 | u32 nxt_offset = nxt->io->timeout.seq_offset; |
5262f567 | 2737 | |
93bd25bb JA |
2738 | if (nxt->flags & REQ_F_TIMEOUT_NOSEQ) |
2739 | continue; | |
2740 | ||
5da0fb1a | 2741 | /* |
2742 | * Since cached_sq_head + count - 1 can overflow, use type long | |
2743 | * long to store it. | |
2744 | */ | |
2745 | tmp = (long long)ctx->cached_sq_head + count - 1; | |
cc42e0ac PB |
2746 | nxt_sq_head = nxt->sequence - nxt_offset + 1; |
2747 | tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1; | |
5da0fb1a | 2748 | |
2749 | /* | |
2750 | * cached_sq_head may overflow, and it will never overflow twice | |
2751 | * once there is some timeout req still be valid. | |
2752 | */ | |
2753 | if (ctx->cached_sq_head < nxt_sq_head) | |
8b07a65a | 2754 | tmp += UINT_MAX; |
5da0fb1a | 2755 | |
a1f58ba4 | 2756 | if (tmp > tmp_nxt) |
5262f567 | 2757 | break; |
a1f58ba4 | 2758 | |
2759 | /* | |
2760 | * Sequence of reqs after the insert one and itself should | |
2761 | * be adjusted because each timeout req consumes a slot. | |
2762 | */ | |
2763 | span++; | |
2764 | nxt->sequence++; | |
5262f567 | 2765 | } |
a1f58ba4 | 2766 | req->sequence -= span; |
93bd25bb | 2767 | add: |
5262f567 | 2768 | list_add(&req->list, entry); |
ad8a48ac JA |
2769 | data->timer.function = io_timeout_fn; |
2770 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode); | |
5262f567 | 2771 | spin_unlock_irq(&ctx->completion_lock); |
5262f567 JA |
2772 | return 0; |
2773 | } | |
5262f567 | 2774 | |
62755e35 JA |
2775 | static bool io_cancel_cb(struct io_wq_work *work, void *data) |
2776 | { | |
2777 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
2778 | ||
2779 | return req->user_data == (unsigned long) data; | |
2780 | } | |
2781 | ||
e977d6d3 | 2782 | static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr) |
62755e35 | 2783 | { |
62755e35 | 2784 | enum io_wq_cancel cancel_ret; |
62755e35 JA |
2785 | int ret = 0; |
2786 | ||
62755e35 JA |
2787 | cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr); |
2788 | switch (cancel_ret) { | |
2789 | case IO_WQ_CANCEL_OK: | |
2790 | ret = 0; | |
2791 | break; | |
2792 | case IO_WQ_CANCEL_RUNNING: | |
2793 | ret = -EALREADY; | |
2794 | break; | |
2795 | case IO_WQ_CANCEL_NOTFOUND: | |
2796 | ret = -ENOENT; | |
2797 | break; | |
2798 | } | |
2799 | ||
e977d6d3 JA |
2800 | return ret; |
2801 | } | |
2802 | ||
47f46768 JA |
2803 | static void io_async_find_and_cancel(struct io_ring_ctx *ctx, |
2804 | struct io_kiocb *req, __u64 sqe_addr, | |
b0dd8a41 | 2805 | struct io_kiocb **nxt, int success_ret) |
47f46768 JA |
2806 | { |
2807 | unsigned long flags; | |
2808 | int ret; | |
2809 | ||
2810 | ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr); | |
2811 | if (ret != -ENOENT) { | |
2812 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
2813 | goto done; | |
2814 | } | |
2815 | ||
2816 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
2817 | ret = io_timeout_cancel(ctx, sqe_addr); | |
2818 | if (ret != -ENOENT) | |
2819 | goto done; | |
2820 | ret = io_poll_cancel(ctx, sqe_addr); | |
2821 | done: | |
b0dd8a41 JA |
2822 | if (!ret) |
2823 | ret = success_ret; | |
47f46768 JA |
2824 | io_cqring_fill_event(req, ret); |
2825 | io_commit_cqring(ctx); | |
2826 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
2827 | io_cqring_ev_posted(ctx); | |
2828 | ||
2829 | if (ret < 0 && (req->flags & REQ_F_LINK)) | |
2830 | req->flags |= REQ_F_FAIL_LINK; | |
2831 | io_put_req_find_next(req, nxt); | |
2832 | } | |
2833 | ||
e977d6d3 JA |
2834 | static int io_async_cancel(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2835 | struct io_kiocb **nxt) | |
2836 | { | |
2837 | struct io_ring_ctx *ctx = req->ctx; | |
e977d6d3 JA |
2838 | |
2839 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
2840 | return -EINVAL; | |
2841 | if (sqe->flags || sqe->ioprio || sqe->off || sqe->len || | |
2842 | sqe->cancel_flags) | |
2843 | return -EINVAL; | |
2844 | ||
b0dd8a41 | 2845 | io_async_find_and_cancel(ctx, req, READ_ONCE(sqe->addr), nxt, 0); |
5262f567 JA |
2846 | return 0; |
2847 | } | |
2848 | ||
f67676d1 JA |
2849 | static int io_req_defer_prep(struct io_kiocb *req, struct io_async_ctx *io) |
2850 | { | |
2851 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
2852 | struct iov_iter iter; | |
2853 | ssize_t ret; | |
2854 | ||
2855 | memcpy(&io->sqe, req->sqe, sizeof(io->sqe)); | |
2856 | req->sqe = &io->sqe; | |
2857 | ||
2858 | switch (io->sqe.opcode) { | |
2859 | case IORING_OP_READV: | |
2860 | case IORING_OP_READ_FIXED: | |
2861 | ret = io_read_prep(req, &iovec, &iter, true); | |
2862 | break; | |
2863 | case IORING_OP_WRITEV: | |
2864 | case IORING_OP_WRITE_FIXED: | |
2865 | ret = io_write_prep(req, &iovec, &iter, true); | |
2866 | break; | |
03b1230c JA |
2867 | case IORING_OP_SENDMSG: |
2868 | ret = io_sendmsg_prep(req, io); | |
2869 | break; | |
2870 | case IORING_OP_RECVMSG: | |
2871 | ret = io_recvmsg_prep(req, io); | |
2872 | break; | |
f499a021 JA |
2873 | case IORING_OP_CONNECT: |
2874 | ret = io_connect_prep(req, io); | |
2875 | break; | |
2d28390a JA |
2876 | case IORING_OP_TIMEOUT: |
2877 | return io_timeout_prep(req, io, false); | |
2878 | case IORING_OP_LINK_TIMEOUT: | |
2879 | return io_timeout_prep(req, io, true); | |
f67676d1 JA |
2880 | default: |
2881 | req->io = io; | |
2882 | return 0; | |
2883 | } | |
2884 | ||
2885 | if (ret < 0) | |
2886 | return ret; | |
2887 | ||
2888 | req->io = io; | |
2889 | io_req_map_io(req, ret, iovec, inline_vecs, &iter); | |
2890 | return 0; | |
2891 | } | |
2892 | ||
a197f664 | 2893 | static int io_req_defer(struct io_kiocb *req) |
de0617e4 | 2894 | { |
a197f664 | 2895 | struct io_ring_ctx *ctx = req->ctx; |
1a6b74fc | 2896 | struct io_async_ctx *io; |
f67676d1 | 2897 | int ret; |
de0617e4 | 2898 | |
9d858b21 BL |
2899 | /* Still need defer if there is pending req in defer list. */ |
2900 | if (!req_need_defer(req) && list_empty(&ctx->defer_list)) | |
de0617e4 JA |
2901 | return 0; |
2902 | ||
1a6b74fc JA |
2903 | io = kmalloc(sizeof(*io), GFP_KERNEL); |
2904 | if (!io) | |
de0617e4 JA |
2905 | return -EAGAIN; |
2906 | ||
2d28390a JA |
2907 | ret = io_req_defer_prep(req, io); |
2908 | if (ret < 0) { | |
2909 | kfree(io); | |
2910 | return ret; | |
2911 | } | |
2912 | ||
de0617e4 | 2913 | spin_lock_irq(&ctx->completion_lock); |
9d858b21 | 2914 | if (!req_need_defer(req) && list_empty(&ctx->defer_list)) { |
de0617e4 | 2915 | spin_unlock_irq(&ctx->completion_lock); |
de0617e4 JA |
2916 | return 0; |
2917 | } | |
2918 | ||
915967f6 | 2919 | trace_io_uring_defer(ctx, req, req->user_data); |
de0617e4 JA |
2920 | list_add_tail(&req->list, &ctx->defer_list); |
2921 | spin_unlock_irq(&ctx->completion_lock); | |
2922 | return -EIOCBQUEUED; | |
2923 | } | |
2924 | ||
f9bd67f6 | 2925 | __attribute__((nonnull)) |
d732447f PB |
2926 | static int io_issue_sqe(struct io_kiocb *req, struct io_kiocb **nxt, |
2927 | bool force_nonblock) | |
2b188cc1 | 2928 | { |
e0c5c576 | 2929 | int ret, opcode; |
a197f664 | 2930 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 | 2931 | |
cf6fd4bd | 2932 | opcode = READ_ONCE(req->sqe->opcode); |
2b188cc1 JA |
2933 | switch (opcode) { |
2934 | case IORING_OP_NOP: | |
78e19bbe | 2935 | ret = io_nop(req); |
2b188cc1 JA |
2936 | break; |
2937 | case IORING_OP_READV: | |
cf6fd4bd | 2938 | if (unlikely(req->sqe->buf_index)) |
edafccee | 2939 | return -EINVAL; |
267bc904 | 2940 | ret = io_read(req, nxt, force_nonblock); |
2b188cc1 JA |
2941 | break; |
2942 | case IORING_OP_WRITEV: | |
cf6fd4bd | 2943 | if (unlikely(req->sqe->buf_index)) |
edafccee | 2944 | return -EINVAL; |
267bc904 | 2945 | ret = io_write(req, nxt, force_nonblock); |
edafccee JA |
2946 | break; |
2947 | case IORING_OP_READ_FIXED: | |
267bc904 | 2948 | ret = io_read(req, nxt, force_nonblock); |
edafccee JA |
2949 | break; |
2950 | case IORING_OP_WRITE_FIXED: | |
267bc904 | 2951 | ret = io_write(req, nxt, force_nonblock); |
2b188cc1 | 2952 | break; |
c992fe29 | 2953 | case IORING_OP_FSYNC: |
cf6fd4bd | 2954 | ret = io_fsync(req, req->sqe, nxt, force_nonblock); |
c992fe29 | 2955 | break; |
221c5eb2 | 2956 | case IORING_OP_POLL_ADD: |
cf6fd4bd | 2957 | ret = io_poll_add(req, req->sqe, nxt); |
221c5eb2 JA |
2958 | break; |
2959 | case IORING_OP_POLL_REMOVE: | |
cf6fd4bd | 2960 | ret = io_poll_remove(req, req->sqe); |
221c5eb2 | 2961 | break; |
5d17b4a4 | 2962 | case IORING_OP_SYNC_FILE_RANGE: |
cf6fd4bd | 2963 | ret = io_sync_file_range(req, req->sqe, nxt, force_nonblock); |
5d17b4a4 | 2964 | break; |
0fa03c62 | 2965 | case IORING_OP_SENDMSG: |
cf6fd4bd | 2966 | ret = io_sendmsg(req, req->sqe, nxt, force_nonblock); |
0fa03c62 | 2967 | break; |
aa1fa28f | 2968 | case IORING_OP_RECVMSG: |
cf6fd4bd | 2969 | ret = io_recvmsg(req, req->sqe, nxt, force_nonblock); |
aa1fa28f | 2970 | break; |
5262f567 | 2971 | case IORING_OP_TIMEOUT: |
cf6fd4bd | 2972 | ret = io_timeout(req, req->sqe); |
5262f567 | 2973 | break; |
11365043 | 2974 | case IORING_OP_TIMEOUT_REMOVE: |
cf6fd4bd | 2975 | ret = io_timeout_remove(req, req->sqe); |
11365043 | 2976 | break; |
17f2fe35 | 2977 | case IORING_OP_ACCEPT: |
cf6fd4bd | 2978 | ret = io_accept(req, req->sqe, nxt, force_nonblock); |
17f2fe35 | 2979 | break; |
f8e85cf2 | 2980 | case IORING_OP_CONNECT: |
cf6fd4bd | 2981 | ret = io_connect(req, req->sqe, nxt, force_nonblock); |
f8e85cf2 | 2982 | break; |
62755e35 | 2983 | case IORING_OP_ASYNC_CANCEL: |
cf6fd4bd | 2984 | ret = io_async_cancel(req, req->sqe, nxt); |
62755e35 | 2985 | break; |
2b188cc1 JA |
2986 | default: |
2987 | ret = -EINVAL; | |
2988 | break; | |
2989 | } | |
2990 | ||
def596e9 JA |
2991 | if (ret) |
2992 | return ret; | |
2993 | ||
2994 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
9e645e11 | 2995 | if (req->result == -EAGAIN) |
def596e9 JA |
2996 | return -EAGAIN; |
2997 | ||
2998 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
cf6fd4bd | 2999 | if (req->in_async) |
def596e9 JA |
3000 | mutex_lock(&ctx->uring_lock); |
3001 | io_iopoll_req_issued(req); | |
cf6fd4bd | 3002 | if (req->in_async) |
def596e9 JA |
3003 | mutex_unlock(&ctx->uring_lock); |
3004 | } | |
3005 | ||
3006 | return 0; | |
2b188cc1 JA |
3007 | } |
3008 | ||
b76da70f JA |
3009 | static void io_link_work_cb(struct io_wq_work **workptr) |
3010 | { | |
3011 | struct io_wq_work *work = *workptr; | |
3012 | struct io_kiocb *link = work->data; | |
3013 | ||
3014 | io_queue_linked_timeout(link); | |
3015 | work->func = io_wq_submit_work; | |
3016 | } | |
3017 | ||
561fb04a | 3018 | static void io_wq_submit_work(struct io_wq_work **workptr) |
2b188cc1 | 3019 | { |
561fb04a | 3020 | struct io_wq_work *work = *workptr; |
2b188cc1 | 3021 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
561fb04a JA |
3022 | struct io_kiocb *nxt = NULL; |
3023 | int ret = 0; | |
2b188cc1 | 3024 | |
561fb04a JA |
3025 | /* Ensure we clear previously set non-block flag */ |
3026 | req->rw.ki_flags &= ~IOCB_NOWAIT; | |
2b188cc1 | 3027 | |
561fb04a JA |
3028 | if (work->flags & IO_WQ_WORK_CANCEL) |
3029 | ret = -ECANCELED; | |
31b51510 | 3030 | |
561fb04a | 3031 | if (!ret) { |
cf6fd4bd PB |
3032 | req->has_user = (work->flags & IO_WQ_WORK_HAS_MM) != 0; |
3033 | req->in_async = true; | |
561fb04a | 3034 | do { |
d732447f | 3035 | ret = io_issue_sqe(req, &nxt, false); |
561fb04a JA |
3036 | /* |
3037 | * We can get EAGAIN for polled IO even though we're | |
3038 | * forcing a sync submission from here, since we can't | |
3039 | * wait for request slots on the block side. | |
3040 | */ | |
3041 | if (ret != -EAGAIN) | |
3042 | break; | |
3043 | cond_resched(); | |
3044 | } while (1); | |
3045 | } | |
31b51510 | 3046 | |
561fb04a | 3047 | /* drop submission reference */ |
ec9c02ad | 3048 | io_put_req(req); |
817869d2 | 3049 | |
561fb04a | 3050 | if (ret) { |
f1f40853 JA |
3051 | if (req->flags & REQ_F_LINK) |
3052 | req->flags |= REQ_F_FAIL_LINK; | |
78e19bbe | 3053 | io_cqring_add_event(req, ret); |
817869d2 | 3054 | io_put_req(req); |
edafccee | 3055 | } |
2b188cc1 | 3056 | |
561fb04a JA |
3057 | /* if a dependent link is ready, pass it back */ |
3058 | if (!ret && nxt) { | |
94ae5e77 JA |
3059 | struct io_kiocb *link; |
3060 | ||
3061 | io_prep_async_work(nxt, &link); | |
561fb04a | 3062 | *workptr = &nxt->work; |
b76da70f JA |
3063 | if (link) { |
3064 | nxt->work.flags |= IO_WQ_WORK_CB; | |
3065 | nxt->work.func = io_link_work_cb; | |
3066 | nxt->work.data = link; | |
3067 | } | |
31b51510 | 3068 | } |
2b188cc1 JA |
3069 | } |
3070 | ||
09bb8394 JA |
3071 | static bool io_op_needs_file(const struct io_uring_sqe *sqe) |
3072 | { | |
3073 | int op = READ_ONCE(sqe->opcode); | |
3074 | ||
3075 | switch (op) { | |
3076 | case IORING_OP_NOP: | |
3077 | case IORING_OP_POLL_REMOVE: | |
5683e540 | 3078 | case IORING_OP_TIMEOUT: |
a320e9fa PB |
3079 | case IORING_OP_TIMEOUT_REMOVE: |
3080 | case IORING_OP_ASYNC_CANCEL: | |
3081 | case IORING_OP_LINK_TIMEOUT: | |
09bb8394 JA |
3082 | return false; |
3083 | default: | |
3084 | return true; | |
3085 | } | |
3086 | } | |
3087 | ||
65e19f54 JA |
3088 | static inline struct file *io_file_from_index(struct io_ring_ctx *ctx, |
3089 | int index) | |
3090 | { | |
3091 | struct fixed_file_table *table; | |
3092 | ||
3093 | table = &ctx->file_table[index >> IORING_FILE_TABLE_SHIFT]; | |
3094 | return table->files[index & IORING_FILE_TABLE_MASK]; | |
3095 | } | |
3096 | ||
a197f664 | 3097 | static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req) |
09bb8394 | 3098 | { |
a197f664 | 3099 | struct io_ring_ctx *ctx = req->ctx; |
09bb8394 JA |
3100 | unsigned flags; |
3101 | int fd; | |
3102 | ||
cf6fd4bd PB |
3103 | flags = READ_ONCE(req->sqe->flags); |
3104 | fd = READ_ONCE(req->sqe->fd); | |
09bb8394 | 3105 | |
4fe2c963 | 3106 | if (flags & IOSQE_IO_DRAIN) |
de0617e4 | 3107 | req->flags |= REQ_F_IO_DRAIN; |
de0617e4 | 3108 | |
cf6fd4bd | 3109 | if (!io_op_needs_file(req->sqe)) |
09bb8394 | 3110 | return 0; |
09bb8394 JA |
3111 | |
3112 | if (flags & IOSQE_FIXED_FILE) { | |
65e19f54 | 3113 | if (unlikely(!ctx->file_table || |
09bb8394 JA |
3114 | (unsigned) fd >= ctx->nr_user_files)) |
3115 | return -EBADF; | |
b7620121 | 3116 | fd = array_index_nospec(fd, ctx->nr_user_files); |
65e19f54 JA |
3117 | req->file = io_file_from_index(ctx, fd); |
3118 | if (!req->file) | |
08a45173 | 3119 | return -EBADF; |
09bb8394 JA |
3120 | req->flags |= REQ_F_FIXED_FILE; |
3121 | } else { | |
cf6fd4bd | 3122 | if (req->needs_fixed_file) |
09bb8394 | 3123 | return -EBADF; |
c826bd7a | 3124 | trace_io_uring_file_get(ctx, fd); |
09bb8394 JA |
3125 | req->file = io_file_get(state, fd); |
3126 | if (unlikely(!req->file)) | |
3127 | return -EBADF; | |
3128 | } | |
3129 | ||
3130 | return 0; | |
3131 | } | |
3132 | ||
a197f664 | 3133 | static int io_grab_files(struct io_kiocb *req) |
fcb323cc JA |
3134 | { |
3135 | int ret = -EBADF; | |
a197f664 | 3136 | struct io_ring_ctx *ctx = req->ctx; |
fcb323cc JA |
3137 | |
3138 | rcu_read_lock(); | |
3139 | spin_lock_irq(&ctx->inflight_lock); | |
3140 | /* | |
3141 | * We use the f_ops->flush() handler to ensure that we can flush | |
3142 | * out work accessing these files if the fd is closed. Check if | |
3143 | * the fd has changed since we started down this path, and disallow | |
3144 | * this operation if it has. | |
3145 | */ | |
cf6fd4bd | 3146 | if (fcheck(req->ring_fd) == req->ring_file) { |
fcb323cc JA |
3147 | list_add(&req->inflight_entry, &ctx->inflight_list); |
3148 | req->flags |= REQ_F_INFLIGHT; | |
3149 | req->work.files = current->files; | |
3150 | ret = 0; | |
3151 | } | |
3152 | spin_unlock_irq(&ctx->inflight_lock); | |
3153 | rcu_read_unlock(); | |
3154 | ||
3155 | return ret; | |
3156 | } | |
3157 | ||
2665abfd | 3158 | static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer) |
2b188cc1 | 3159 | { |
ad8a48ac JA |
3160 | struct io_timeout_data *data = container_of(timer, |
3161 | struct io_timeout_data, timer); | |
3162 | struct io_kiocb *req = data->req; | |
2665abfd JA |
3163 | struct io_ring_ctx *ctx = req->ctx; |
3164 | struct io_kiocb *prev = NULL; | |
3165 | unsigned long flags; | |
2665abfd JA |
3166 | |
3167 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
3168 | ||
3169 | /* | |
3170 | * We don't expect the list to be empty, that will only happen if we | |
3171 | * race with the completion of the linked work. | |
3172 | */ | |
3173 | if (!list_empty(&req->list)) { | |
3174 | prev = list_entry(req->list.prev, struct io_kiocb, link_list); | |
5d960724 | 3175 | if (refcount_inc_not_zero(&prev->refs)) { |
76a46e06 | 3176 | list_del_init(&req->list); |
5d960724 JA |
3177 | prev->flags &= ~REQ_F_LINK_TIMEOUT; |
3178 | } else | |
76a46e06 | 3179 | prev = NULL; |
2665abfd JA |
3180 | } |
3181 | ||
3182 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
3183 | ||
3184 | if (prev) { | |
fba38c27 JA |
3185 | if (prev->flags & REQ_F_LINK) |
3186 | prev->flags |= REQ_F_FAIL_LINK; | |
b0dd8a41 JA |
3187 | io_async_find_and_cancel(ctx, req, prev->user_data, NULL, |
3188 | -ETIME); | |
76a46e06 | 3189 | io_put_req(prev); |
47f46768 JA |
3190 | } else { |
3191 | io_cqring_add_event(req, -ETIME); | |
3192 | io_put_req(req); | |
2665abfd | 3193 | } |
2665abfd JA |
3194 | return HRTIMER_NORESTART; |
3195 | } | |
3196 | ||
ad8a48ac | 3197 | static void io_queue_linked_timeout(struct io_kiocb *req) |
2665abfd | 3198 | { |
76a46e06 | 3199 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd | 3200 | |
76a46e06 JA |
3201 | /* |
3202 | * If the list is now empty, then our linked request finished before | |
3203 | * we got a chance to setup the timer | |
3204 | */ | |
3205 | spin_lock_irq(&ctx->completion_lock); | |
3206 | if (!list_empty(&req->list)) { | |
2d28390a | 3207 | struct io_timeout_data *data = &req->io->timeout; |
94ae5e77 | 3208 | |
ad8a48ac JA |
3209 | data->timer.function = io_link_timeout_fn; |
3210 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), | |
3211 | data->mode); | |
2665abfd | 3212 | } |
76a46e06 | 3213 | spin_unlock_irq(&ctx->completion_lock); |
2665abfd | 3214 | |
2665abfd | 3215 | /* drop submission reference */ |
76a46e06 JA |
3216 | io_put_req(req); |
3217 | } | |
2665abfd | 3218 | |
ad8a48ac | 3219 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) |
2665abfd JA |
3220 | { |
3221 | struct io_kiocb *nxt; | |
3222 | ||
3223 | if (!(req->flags & REQ_F_LINK)) | |
3224 | return NULL; | |
3225 | ||
3226 | nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list); | |
cf6fd4bd | 3227 | if (!nxt || nxt->sqe->opcode != IORING_OP_LINK_TIMEOUT) |
76a46e06 | 3228 | return NULL; |
2665abfd | 3229 | |
76a46e06 | 3230 | req->flags |= REQ_F_LINK_TIMEOUT; |
76a46e06 | 3231 | return nxt; |
2665abfd JA |
3232 | } |
3233 | ||
0e0702da | 3234 | static void __io_queue_sqe(struct io_kiocb *req) |
2b188cc1 | 3235 | { |
f9bd67f6 PB |
3236 | struct io_kiocb *linked_timeout = io_prep_linked_timeout(req); |
3237 | struct io_kiocb *nxt = NULL; | |
e0c5c576 | 3238 | int ret; |
2b188cc1 | 3239 | |
f9bd67f6 PB |
3240 | ret = io_issue_sqe(req, &nxt, true); |
3241 | if (nxt) | |
3242 | io_queue_async_work(nxt); | |
491381ce JA |
3243 | |
3244 | /* | |
3245 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
3246 | * doesn't support non-blocking read/write attempts | |
3247 | */ | |
3248 | if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) || | |
3249 | (req->flags & REQ_F_MUST_PUNT))) { | |
bbad27b2 PB |
3250 | if (req->work.flags & IO_WQ_WORK_NEEDS_FILES) { |
3251 | ret = io_grab_files(req); | |
3252 | if (ret) | |
3253 | goto err; | |
2b188cc1 | 3254 | } |
bbad27b2 PB |
3255 | |
3256 | /* | |
3257 | * Queued up for async execution, worker will release | |
3258 | * submit reference when the iocb is actually submitted. | |
3259 | */ | |
3260 | io_queue_async_work(req); | |
3261 | return; | |
2b188cc1 | 3262 | } |
e65ef56d | 3263 | |
fcb323cc | 3264 | err: |
76a46e06 | 3265 | /* drop submission reference */ |
ec9c02ad | 3266 | io_put_req(req); |
e65ef56d | 3267 | |
f9bd67f6 | 3268 | if (linked_timeout) { |
76a46e06 | 3269 | if (!ret) |
f9bd67f6 | 3270 | io_queue_linked_timeout(linked_timeout); |
76a46e06 | 3271 | else |
f9bd67f6 | 3272 | io_put_req(linked_timeout); |
76a46e06 JA |
3273 | } |
3274 | ||
e65ef56d | 3275 | /* and drop final reference, if we failed */ |
9e645e11 | 3276 | if (ret) { |
78e19bbe | 3277 | io_cqring_add_event(req, ret); |
9e645e11 JA |
3278 | if (req->flags & REQ_F_LINK) |
3279 | req->flags |= REQ_F_FAIL_LINK; | |
e65ef56d | 3280 | io_put_req(req); |
9e645e11 | 3281 | } |
2b188cc1 JA |
3282 | } |
3283 | ||
0e0702da | 3284 | static void io_queue_sqe(struct io_kiocb *req) |
4fe2c963 JL |
3285 | { |
3286 | int ret; | |
3287 | ||
1b4a51b6 PB |
3288 | if (unlikely(req->ctx->drain_next)) { |
3289 | req->flags |= REQ_F_IO_DRAIN; | |
3290 | req->ctx->drain_next = false; | |
3291 | } | |
3292 | req->ctx->drain_next = (req->flags & REQ_F_DRAIN_LINK); | |
3293 | ||
a197f664 | 3294 | ret = io_req_defer(req); |
4fe2c963 JL |
3295 | if (ret) { |
3296 | if (ret != -EIOCBQUEUED) { | |
78e19bbe | 3297 | io_cqring_add_event(req, ret); |
d3b35796 PB |
3298 | if (req->flags & REQ_F_LINK) |
3299 | req->flags |= REQ_F_FAIL_LINK; | |
78e19bbe | 3300 | io_double_put_req(req); |
4fe2c963 | 3301 | } |
0e0702da JA |
3302 | } else |
3303 | __io_queue_sqe(req); | |
4fe2c963 JL |
3304 | } |
3305 | ||
1b4a51b6 | 3306 | static inline void io_queue_link_head(struct io_kiocb *req) |
4fe2c963 | 3307 | { |
94ae5e77 | 3308 | if (unlikely(req->flags & REQ_F_FAIL_LINK)) { |
1b4a51b6 PB |
3309 | io_cqring_add_event(req, -ECANCELED); |
3310 | io_double_put_req(req); | |
3311 | } else | |
0e0702da | 3312 | io_queue_sqe(req); |
4fe2c963 JL |
3313 | } |
3314 | ||
1b4a51b6 | 3315 | |
9e645e11 JA |
3316 | #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK) |
3317 | ||
2e6e1fde | 3318 | static bool io_submit_sqe(struct io_kiocb *req, struct io_submit_state *state, |
a197f664 | 3319 | struct io_kiocb **link) |
9e645e11 | 3320 | { |
a197f664 | 3321 | struct io_ring_ctx *ctx = req->ctx; |
9e645e11 JA |
3322 | int ret; |
3323 | ||
cf6fd4bd | 3324 | req->user_data = req->sqe->user_data; |
78e19bbe | 3325 | |
9e645e11 | 3326 | /* enforce forwards compatibility on users */ |
cf6fd4bd | 3327 | if (unlikely(req->sqe->flags & ~SQE_VALID_FLAGS)) { |
9e645e11 | 3328 | ret = -EINVAL; |
196be95c | 3329 | goto err_req; |
9e645e11 JA |
3330 | } |
3331 | ||
a197f664 | 3332 | ret = io_req_set_file(state, req); |
9e645e11 JA |
3333 | if (unlikely(ret)) { |
3334 | err_req: | |
78e19bbe JA |
3335 | io_cqring_add_event(req, ret); |
3336 | io_double_put_req(req); | |
2e6e1fde | 3337 | return false; |
9e645e11 JA |
3338 | } |
3339 | ||
9e645e11 JA |
3340 | /* |
3341 | * If we already have a head request, queue this one for async | |
3342 | * submittal once the head completes. If we don't have a head but | |
3343 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
3344 | * submitted sync once the chain is complete. If none of those | |
3345 | * conditions are true (normal request), then just queue it. | |
3346 | */ | |
3347 | if (*link) { | |
3348 | struct io_kiocb *prev = *link; | |
1a6b74fc | 3349 | struct io_async_ctx *io; |
9e645e11 | 3350 | |
cf6fd4bd | 3351 | if (req->sqe->flags & IOSQE_IO_DRAIN) |
1b4a51b6 PB |
3352 | (*link)->flags |= REQ_F_DRAIN_LINK | REQ_F_IO_DRAIN; |
3353 | ||
1a6b74fc JA |
3354 | io = kmalloc(sizeof(*io), GFP_KERNEL); |
3355 | if (!io) { | |
9e645e11 JA |
3356 | ret = -EAGAIN; |
3357 | goto err_req; | |
3358 | } | |
3359 | ||
f67676d1 | 3360 | ret = io_req_defer_prep(req, io); |
2d28390a JA |
3361 | if (ret) { |
3362 | kfree(io); | |
3363 | prev->flags |= REQ_F_FAIL_LINK; | |
f67676d1 | 3364 | goto err_req; |
2d28390a | 3365 | } |
c826bd7a | 3366 | trace_io_uring_link(ctx, req, prev); |
9e645e11 | 3367 | list_add_tail(&req->list, &prev->link_list); |
cf6fd4bd | 3368 | } else if (req->sqe->flags & IOSQE_IO_LINK) { |
9e645e11 JA |
3369 | req->flags |= REQ_F_LINK; |
3370 | ||
9e645e11 JA |
3371 | INIT_LIST_HEAD(&req->link_list); |
3372 | *link = req; | |
3373 | } else { | |
a197f664 | 3374 | io_queue_sqe(req); |
9e645e11 | 3375 | } |
2e6e1fde PB |
3376 | |
3377 | return true; | |
9e645e11 JA |
3378 | } |
3379 | ||
9a56a232 JA |
3380 | /* |
3381 | * Batched submission is done, ensure local IO is flushed out. | |
3382 | */ | |
3383 | static void io_submit_state_end(struct io_submit_state *state) | |
3384 | { | |
3385 | blk_finish_plug(&state->plug); | |
3d6770fb | 3386 | io_file_put(state); |
2579f913 JA |
3387 | if (state->free_reqs) |
3388 | kmem_cache_free_bulk(req_cachep, state->free_reqs, | |
3389 | &state->reqs[state->cur_req]); | |
9a56a232 JA |
3390 | } |
3391 | ||
3392 | /* | |
3393 | * Start submission side cache. | |
3394 | */ | |
3395 | static void io_submit_state_start(struct io_submit_state *state, | |
22efde59 | 3396 | unsigned int max_ios) |
9a56a232 JA |
3397 | { |
3398 | blk_start_plug(&state->plug); | |
2579f913 | 3399 | state->free_reqs = 0; |
9a56a232 JA |
3400 | state->file = NULL; |
3401 | state->ios_left = max_ios; | |
3402 | } | |
3403 | ||
2b188cc1 JA |
3404 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
3405 | { | |
75b28aff | 3406 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 3407 | |
75b28aff | 3408 | if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) { |
2b188cc1 JA |
3409 | /* |
3410 | * Ensure any loads from the SQEs are done at this point, | |
3411 | * since once we write the new head, the application could | |
3412 | * write new data to them. | |
3413 | */ | |
75b28aff | 3414 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); |
2b188cc1 JA |
3415 | } |
3416 | } | |
3417 | ||
2b188cc1 JA |
3418 | /* |
3419 | * Fetch an sqe, if one is available. Note that s->sqe will point to memory | |
3420 | * that is mapped by userspace. This means that care needs to be taken to | |
3421 | * ensure that reads are stable, as we cannot rely on userspace always | |
3422 | * being a good citizen. If members of the sqe are validated and then later | |
3423 | * used, it's important that those reads are done through READ_ONCE() to | |
3424 | * prevent a re-load down the line. | |
3425 | */ | |
cf6fd4bd | 3426 | static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req) |
2b188cc1 | 3427 | { |
75b28aff HV |
3428 | struct io_rings *rings = ctx->rings; |
3429 | u32 *sq_array = ctx->sq_array; | |
2b188cc1 JA |
3430 | unsigned head; |
3431 | ||
3432 | /* | |
3433 | * The cached sq head (or cq tail) serves two purposes: | |
3434 | * | |
3435 | * 1) allows us to batch the cost of updating the user visible | |
3436 | * head updates. | |
3437 | * 2) allows the kernel side to track the head on its own, even | |
3438 | * though the application is the one updating it. | |
3439 | */ | |
3440 | head = ctx->cached_sq_head; | |
e523a29c | 3441 | /* make sure SQ entry isn't read before tail */ |
9835d6fa | 3442 | if (unlikely(head == smp_load_acquire(&rings->sq.tail))) |
2b188cc1 JA |
3443 | return false; |
3444 | ||
75b28aff | 3445 | head = READ_ONCE(sq_array[head & ctx->sq_mask]); |
9835d6fa | 3446 | if (likely(head < ctx->sq_entries)) { |
cf6fd4bd PB |
3447 | /* |
3448 | * All io need record the previous position, if LINK vs DARIN, | |
3449 | * it can be used to mark the position of the first IO in the | |
3450 | * link list. | |
3451 | */ | |
3452 | req->sequence = ctx->cached_sq_head; | |
3453 | req->sqe = &ctx->sq_sqes[head]; | |
2b188cc1 JA |
3454 | ctx->cached_sq_head++; |
3455 | return true; | |
3456 | } | |
3457 | ||
3458 | /* drop invalid entries */ | |
3459 | ctx->cached_sq_head++; | |
498ccd9e JA |
3460 | ctx->cached_sq_dropped++; |
3461 | WRITE_ONCE(rings->sq_dropped, ctx->cached_sq_dropped); | |
2b188cc1 JA |
3462 | return false; |
3463 | } | |
3464 | ||
fb5ccc98 | 3465 | static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr, |
ae9428ca PB |
3466 | struct file *ring_file, int ring_fd, |
3467 | struct mm_struct **mm, bool async) | |
6c271ce2 JA |
3468 | { |
3469 | struct io_submit_state state, *statep = NULL; | |
9e645e11 | 3470 | struct io_kiocb *link = NULL; |
9e645e11 | 3471 | int i, submitted = 0; |
95a1b3ff | 3472 | bool mm_fault = false; |
6c271ce2 | 3473 | |
c4a2ed72 JA |
3474 | /* if we have a backlog and couldn't flush it all, return BUSY */ |
3475 | if (!list_empty(&ctx->cq_overflow_list) && | |
3476 | !io_cqring_overflow_flush(ctx, false)) | |
1d7bb1d5 | 3477 | return -EBUSY; |
6c271ce2 JA |
3478 | |
3479 | if (nr > IO_PLUG_THRESHOLD) { | |
22efde59 | 3480 | io_submit_state_start(&state, nr); |
6c271ce2 JA |
3481 | statep = &state; |
3482 | } | |
3483 | ||
3484 | for (i = 0; i < nr; i++) { | |
196be95c | 3485 | struct io_kiocb *req; |
50585b9a | 3486 | unsigned int sqe_flags; |
fb5ccc98 | 3487 | |
196be95c PB |
3488 | req = io_get_req(ctx, statep); |
3489 | if (unlikely(!req)) { | |
3490 | if (!submitted) | |
3491 | submitted = -EAGAIN; | |
fb5ccc98 | 3492 | break; |
196be95c | 3493 | } |
cf6fd4bd | 3494 | if (!io_get_sqring(ctx, req)) { |
196be95c PB |
3495 | __io_free_req(req); |
3496 | break; | |
3497 | } | |
fb5ccc98 | 3498 | |
cf6fd4bd | 3499 | if (io_sqe_needs_user(req->sqe) && !*mm) { |
95a1b3ff PB |
3500 | mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm); |
3501 | if (!mm_fault) { | |
3502 | use_mm(ctx->sqo_mm); | |
3503 | *mm = ctx->sqo_mm; | |
3504 | } | |
9e645e11 | 3505 | } |
9e645e11 | 3506 | |
2e6e1fde | 3507 | submitted++; |
cf6fd4bd | 3508 | sqe_flags = req->sqe->flags; |
50585b9a | 3509 | |
cf6fd4bd PB |
3510 | req->ring_file = ring_file; |
3511 | req->ring_fd = ring_fd; | |
3512 | req->has_user = *mm != NULL; | |
3513 | req->in_async = async; | |
3514 | req->needs_fixed_file = async; | |
3515 | trace_io_uring_submit_sqe(ctx, req->sqe->user_data, | |
50585b9a | 3516 | true, async); |
2e6e1fde PB |
3517 | if (!io_submit_sqe(req, statep, &link)) |
3518 | break; | |
e5eb6366 PB |
3519 | /* |
3520 | * If previous wasn't linked and we have a linked command, | |
3521 | * that's the end of the chain. Submit the previous link. | |
3522 | */ | |
50585b9a | 3523 | if (!(sqe_flags & IOSQE_IO_LINK) && link) { |
1b4a51b6 | 3524 | io_queue_link_head(link); |
e5eb6366 | 3525 | link = NULL; |
6c271ce2 | 3526 | } |
6c271ce2 JA |
3527 | } |
3528 | ||
9e645e11 | 3529 | if (link) |
1b4a51b6 | 3530 | io_queue_link_head(link); |
6c271ce2 JA |
3531 | if (statep) |
3532 | io_submit_state_end(&state); | |
3533 | ||
ae9428ca PB |
3534 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
3535 | io_commit_sqring(ctx); | |
3536 | ||
6c271ce2 JA |
3537 | return submitted; |
3538 | } | |
3539 | ||
3540 | static int io_sq_thread(void *data) | |
3541 | { | |
6c271ce2 JA |
3542 | struct io_ring_ctx *ctx = data; |
3543 | struct mm_struct *cur_mm = NULL; | |
181e448d | 3544 | const struct cred *old_cred; |
6c271ce2 JA |
3545 | mm_segment_t old_fs; |
3546 | DEFINE_WAIT(wait); | |
3547 | unsigned inflight; | |
3548 | unsigned long timeout; | |
c1edbf5f | 3549 | int ret; |
6c271ce2 | 3550 | |
206aefde | 3551 | complete(&ctx->completions[1]); |
a4c0b3de | 3552 | |
6c271ce2 JA |
3553 | old_fs = get_fs(); |
3554 | set_fs(USER_DS); | |
181e448d | 3555 | old_cred = override_creds(ctx->creds); |
6c271ce2 | 3556 | |
c1edbf5f | 3557 | ret = timeout = inflight = 0; |
2bbcd6d3 | 3558 | while (!kthread_should_park()) { |
fb5ccc98 | 3559 | unsigned int to_submit; |
6c271ce2 JA |
3560 | |
3561 | if (inflight) { | |
3562 | unsigned nr_events = 0; | |
3563 | ||
3564 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
2b2ed975 JA |
3565 | /* |
3566 | * inflight is the count of the maximum possible | |
3567 | * entries we submitted, but it can be smaller | |
3568 | * if we dropped some of them. If we don't have | |
3569 | * poll entries available, then we know that we | |
3570 | * have nothing left to poll for. Reset the | |
3571 | * inflight count to zero in that case. | |
3572 | */ | |
3573 | mutex_lock(&ctx->uring_lock); | |
3574 | if (!list_empty(&ctx->poll_list)) | |
3575 | __io_iopoll_check(ctx, &nr_events, 0); | |
3576 | else | |
3577 | inflight = 0; | |
3578 | mutex_unlock(&ctx->uring_lock); | |
6c271ce2 JA |
3579 | } else { |
3580 | /* | |
3581 | * Normal IO, just pretend everything completed. | |
3582 | * We don't have to poll completions for that. | |
3583 | */ | |
3584 | nr_events = inflight; | |
3585 | } | |
3586 | ||
3587 | inflight -= nr_events; | |
3588 | if (!inflight) | |
3589 | timeout = jiffies + ctx->sq_thread_idle; | |
3590 | } | |
3591 | ||
fb5ccc98 | 3592 | to_submit = io_sqring_entries(ctx); |
c1edbf5f JA |
3593 | |
3594 | /* | |
3595 | * If submit got -EBUSY, flag us as needing the application | |
3596 | * to enter the kernel to reap and flush events. | |
3597 | */ | |
3598 | if (!to_submit || ret == -EBUSY) { | |
6c271ce2 JA |
3599 | /* |
3600 | * We're polling. If we're within the defined idle | |
3601 | * period, then let us spin without work before going | |
c1edbf5f JA |
3602 | * to sleep. The exception is if we got EBUSY doing |
3603 | * more IO, we should wait for the application to | |
3604 | * reap events and wake us up. | |
6c271ce2 | 3605 | */ |
c1edbf5f JA |
3606 | if (inflight || |
3607 | (!time_after(jiffies, timeout) && ret != -EBUSY)) { | |
9831a90c | 3608 | cond_resched(); |
6c271ce2 JA |
3609 | continue; |
3610 | } | |
3611 | ||
3612 | /* | |
3613 | * Drop cur_mm before scheduling, we can't hold it for | |
3614 | * long periods (or over schedule()). Do this before | |
3615 | * adding ourselves to the waitqueue, as the unuse/drop | |
3616 | * may sleep. | |
3617 | */ | |
3618 | if (cur_mm) { | |
3619 | unuse_mm(cur_mm); | |
3620 | mmput(cur_mm); | |
3621 | cur_mm = NULL; | |
3622 | } | |
3623 | ||
3624 | prepare_to_wait(&ctx->sqo_wait, &wait, | |
3625 | TASK_INTERRUPTIBLE); | |
3626 | ||
3627 | /* Tell userspace we may need a wakeup call */ | |
75b28aff | 3628 | ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP; |
0d7bae69 SB |
3629 | /* make sure to read SQ tail after writing flags */ |
3630 | smp_mb(); | |
6c271ce2 | 3631 | |
fb5ccc98 | 3632 | to_submit = io_sqring_entries(ctx); |
c1edbf5f | 3633 | if (!to_submit || ret == -EBUSY) { |
2bbcd6d3 | 3634 | if (kthread_should_park()) { |
6c271ce2 JA |
3635 | finish_wait(&ctx->sqo_wait, &wait); |
3636 | break; | |
3637 | } | |
3638 | if (signal_pending(current)) | |
3639 | flush_signals(current); | |
3640 | schedule(); | |
3641 | finish_wait(&ctx->sqo_wait, &wait); | |
3642 | ||
75b28aff | 3643 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
3644 | continue; |
3645 | } | |
3646 | finish_wait(&ctx->sqo_wait, &wait); | |
3647 | ||
75b28aff | 3648 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
3649 | } |
3650 | ||
fb5ccc98 | 3651 | to_submit = min(to_submit, ctx->sq_entries); |
1d7bb1d5 JA |
3652 | ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true); |
3653 | if (ret > 0) | |
3654 | inflight += ret; | |
6c271ce2 JA |
3655 | } |
3656 | ||
3657 | set_fs(old_fs); | |
3658 | if (cur_mm) { | |
3659 | unuse_mm(cur_mm); | |
3660 | mmput(cur_mm); | |
3661 | } | |
181e448d | 3662 | revert_creds(old_cred); |
06058632 | 3663 | |
2bbcd6d3 | 3664 | kthread_parkme(); |
06058632 | 3665 | |
6c271ce2 JA |
3666 | return 0; |
3667 | } | |
3668 | ||
bda52162 JA |
3669 | struct io_wait_queue { |
3670 | struct wait_queue_entry wq; | |
3671 | struct io_ring_ctx *ctx; | |
3672 | unsigned to_wait; | |
3673 | unsigned nr_timeouts; | |
3674 | }; | |
3675 | ||
1d7bb1d5 | 3676 | static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush) |
bda52162 JA |
3677 | { |
3678 | struct io_ring_ctx *ctx = iowq->ctx; | |
3679 | ||
3680 | /* | |
3681 | * Wake up if we have enough events, or if a timeout occured since we | |
3682 | * started waiting. For timeouts, we always want to return to userspace, | |
3683 | * regardless of event count. | |
3684 | */ | |
1d7bb1d5 | 3685 | return io_cqring_events(ctx, noflush) >= iowq->to_wait || |
bda52162 JA |
3686 | atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
3687 | } | |
3688 | ||
3689 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
3690 | int wake_flags, void *key) | |
3691 | { | |
3692 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
3693 | wq); | |
3694 | ||
1d7bb1d5 JA |
3695 | /* use noflush == true, as we can't safely rely on locking context */ |
3696 | if (!io_should_wake(iowq, true)) | |
bda52162 JA |
3697 | return -1; |
3698 | ||
3699 | return autoremove_wake_function(curr, mode, wake_flags, key); | |
3700 | } | |
3701 | ||
2b188cc1 JA |
3702 | /* |
3703 | * Wait until events become available, if we don't already have some. The | |
3704 | * application must reap them itself, as they reside on the shared cq ring. | |
3705 | */ | |
3706 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
3707 | const sigset_t __user *sig, size_t sigsz) | |
3708 | { | |
bda52162 JA |
3709 | struct io_wait_queue iowq = { |
3710 | .wq = { | |
3711 | .private = current, | |
3712 | .func = io_wake_function, | |
3713 | .entry = LIST_HEAD_INIT(iowq.wq.entry), | |
3714 | }, | |
3715 | .ctx = ctx, | |
3716 | .to_wait = min_events, | |
3717 | }; | |
75b28aff | 3718 | struct io_rings *rings = ctx->rings; |
e9ffa5c2 | 3719 | int ret = 0; |
2b188cc1 | 3720 | |
1d7bb1d5 | 3721 | if (io_cqring_events(ctx, false) >= min_events) |
2b188cc1 JA |
3722 | return 0; |
3723 | ||
3724 | if (sig) { | |
9e75ad5d AB |
3725 | #ifdef CONFIG_COMPAT |
3726 | if (in_compat_syscall()) | |
3727 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 3728 | sigsz); |
9e75ad5d AB |
3729 | else |
3730 | #endif | |
b772434b | 3731 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 3732 | |
2b188cc1 JA |
3733 | if (ret) |
3734 | return ret; | |
3735 | } | |
3736 | ||
bda52162 | 3737 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
c826bd7a | 3738 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 JA |
3739 | do { |
3740 | prepare_to_wait_exclusive(&ctx->wait, &iowq.wq, | |
3741 | TASK_INTERRUPTIBLE); | |
1d7bb1d5 | 3742 | if (io_should_wake(&iowq, false)) |
bda52162 JA |
3743 | break; |
3744 | schedule(); | |
3745 | if (signal_pending(current)) { | |
e9ffa5c2 | 3746 | ret = -EINTR; |
bda52162 JA |
3747 | break; |
3748 | } | |
3749 | } while (1); | |
3750 | finish_wait(&ctx->wait, &iowq.wq); | |
3751 | ||
e9ffa5c2 | 3752 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 3753 | |
75b28aff | 3754 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
3755 | } |
3756 | ||
6b06314c JA |
3757 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
3758 | { | |
3759 | #if defined(CONFIG_UNIX) | |
3760 | if (ctx->ring_sock) { | |
3761 | struct sock *sock = ctx->ring_sock->sk; | |
3762 | struct sk_buff *skb; | |
3763 | ||
3764 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
3765 | kfree_skb(skb); | |
3766 | } | |
3767 | #else | |
3768 | int i; | |
3769 | ||
65e19f54 JA |
3770 | for (i = 0; i < ctx->nr_user_files; i++) { |
3771 | struct file *file; | |
3772 | ||
3773 | file = io_file_from_index(ctx, i); | |
3774 | if (file) | |
3775 | fput(file); | |
3776 | } | |
6b06314c JA |
3777 | #endif |
3778 | } | |
3779 | ||
3780 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) | |
3781 | { | |
65e19f54 JA |
3782 | unsigned nr_tables, i; |
3783 | ||
3784 | if (!ctx->file_table) | |
6b06314c JA |
3785 | return -ENXIO; |
3786 | ||
3787 | __io_sqe_files_unregister(ctx); | |
65e19f54 JA |
3788 | nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE); |
3789 | for (i = 0; i < nr_tables; i++) | |
3790 | kfree(ctx->file_table[i].files); | |
3791 | kfree(ctx->file_table); | |
3792 | ctx->file_table = NULL; | |
6b06314c JA |
3793 | ctx->nr_user_files = 0; |
3794 | return 0; | |
3795 | } | |
3796 | ||
6c271ce2 JA |
3797 | static void io_sq_thread_stop(struct io_ring_ctx *ctx) |
3798 | { | |
3799 | if (ctx->sqo_thread) { | |
206aefde | 3800 | wait_for_completion(&ctx->completions[1]); |
2bbcd6d3 RP |
3801 | /* |
3802 | * The park is a bit of a work-around, without it we get | |
3803 | * warning spews on shutdown with SQPOLL set and affinity | |
3804 | * set to a single CPU. | |
3805 | */ | |
06058632 | 3806 | kthread_park(ctx->sqo_thread); |
6c271ce2 JA |
3807 | kthread_stop(ctx->sqo_thread); |
3808 | ctx->sqo_thread = NULL; | |
3809 | } | |
3810 | } | |
3811 | ||
6b06314c JA |
3812 | static void io_finish_async(struct io_ring_ctx *ctx) |
3813 | { | |
6c271ce2 JA |
3814 | io_sq_thread_stop(ctx); |
3815 | ||
561fb04a JA |
3816 | if (ctx->io_wq) { |
3817 | io_wq_destroy(ctx->io_wq); | |
3818 | ctx->io_wq = NULL; | |
6b06314c JA |
3819 | } |
3820 | } | |
3821 | ||
3822 | #if defined(CONFIG_UNIX) | |
3823 | static void io_destruct_skb(struct sk_buff *skb) | |
3824 | { | |
3825 | struct io_ring_ctx *ctx = skb->sk->sk_user_data; | |
8a997340 | 3826 | |
561fb04a JA |
3827 | if (ctx->io_wq) |
3828 | io_wq_flush(ctx->io_wq); | |
6b06314c | 3829 | |
6b06314c JA |
3830 | unix_destruct_scm(skb); |
3831 | } | |
3832 | ||
3833 | /* | |
3834 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
3835 | * the io_uring can be safely unregistered on process exit, even if we have | |
3836 | * loops in the file referencing. | |
3837 | */ | |
3838 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
3839 | { | |
3840 | struct sock *sk = ctx->ring_sock->sk; | |
3841 | struct scm_fp_list *fpl; | |
3842 | struct sk_buff *skb; | |
08a45173 | 3843 | int i, nr_files; |
6b06314c JA |
3844 | |
3845 | if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) { | |
3846 | unsigned long inflight = ctx->user->unix_inflight + nr; | |
3847 | ||
3848 | if (inflight > task_rlimit(current, RLIMIT_NOFILE)) | |
3849 | return -EMFILE; | |
3850 | } | |
3851 | ||
3852 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); | |
3853 | if (!fpl) | |
3854 | return -ENOMEM; | |
3855 | ||
3856 | skb = alloc_skb(0, GFP_KERNEL); | |
3857 | if (!skb) { | |
3858 | kfree(fpl); | |
3859 | return -ENOMEM; | |
3860 | } | |
3861 | ||
3862 | skb->sk = sk; | |
6b06314c | 3863 | |
08a45173 | 3864 | nr_files = 0; |
6b06314c JA |
3865 | fpl->user = get_uid(ctx->user); |
3866 | for (i = 0; i < nr; i++) { | |
65e19f54 JA |
3867 | struct file *file = io_file_from_index(ctx, i + offset); |
3868 | ||
3869 | if (!file) | |
08a45173 | 3870 | continue; |
65e19f54 | 3871 | fpl->fp[nr_files] = get_file(file); |
08a45173 JA |
3872 | unix_inflight(fpl->user, fpl->fp[nr_files]); |
3873 | nr_files++; | |
6b06314c JA |
3874 | } |
3875 | ||
08a45173 JA |
3876 | if (nr_files) { |
3877 | fpl->max = SCM_MAX_FD; | |
3878 | fpl->count = nr_files; | |
3879 | UNIXCB(skb).fp = fpl; | |
3880 | skb->destructor = io_destruct_skb; | |
3881 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); | |
3882 | skb_queue_head(&sk->sk_receive_queue, skb); | |
6b06314c | 3883 | |
08a45173 JA |
3884 | for (i = 0; i < nr_files; i++) |
3885 | fput(fpl->fp[i]); | |
3886 | } else { | |
3887 | kfree_skb(skb); | |
3888 | kfree(fpl); | |
3889 | } | |
6b06314c JA |
3890 | |
3891 | return 0; | |
3892 | } | |
3893 | ||
3894 | /* | |
3895 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
3896 | * causes regular reference counting to break down. We rely on the UNIX | |
3897 | * garbage collection to take care of this problem for us. | |
3898 | */ | |
3899 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
3900 | { | |
3901 | unsigned left, total; | |
3902 | int ret = 0; | |
3903 | ||
3904 | total = 0; | |
3905 | left = ctx->nr_user_files; | |
3906 | while (left) { | |
3907 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
3908 | |
3909 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
3910 | if (ret) | |
3911 | break; | |
3912 | left -= this_files; | |
3913 | total += this_files; | |
3914 | } | |
3915 | ||
3916 | if (!ret) | |
3917 | return 0; | |
3918 | ||
3919 | while (total < ctx->nr_user_files) { | |
65e19f54 JA |
3920 | struct file *file = io_file_from_index(ctx, total); |
3921 | ||
3922 | if (file) | |
3923 | fput(file); | |
6b06314c JA |
3924 | total++; |
3925 | } | |
3926 | ||
3927 | return ret; | |
3928 | } | |
3929 | #else | |
3930 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
3931 | { | |
3932 | return 0; | |
3933 | } | |
3934 | #endif | |
3935 | ||
65e19f54 JA |
3936 | static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables, |
3937 | unsigned nr_files) | |
3938 | { | |
3939 | int i; | |
3940 | ||
3941 | for (i = 0; i < nr_tables; i++) { | |
3942 | struct fixed_file_table *table = &ctx->file_table[i]; | |
3943 | unsigned this_files; | |
3944 | ||
3945 | this_files = min(nr_files, IORING_MAX_FILES_TABLE); | |
3946 | table->files = kcalloc(this_files, sizeof(struct file *), | |
3947 | GFP_KERNEL); | |
3948 | if (!table->files) | |
3949 | break; | |
3950 | nr_files -= this_files; | |
3951 | } | |
3952 | ||
3953 | if (i == nr_tables) | |
3954 | return 0; | |
3955 | ||
3956 | for (i = 0; i < nr_tables; i++) { | |
3957 | struct fixed_file_table *table = &ctx->file_table[i]; | |
3958 | kfree(table->files); | |
3959 | } | |
3960 | return 1; | |
3961 | } | |
3962 | ||
6b06314c JA |
3963 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, |
3964 | unsigned nr_args) | |
3965 | { | |
3966 | __s32 __user *fds = (__s32 __user *) arg; | |
65e19f54 | 3967 | unsigned nr_tables; |
6b06314c JA |
3968 | int fd, ret = 0; |
3969 | unsigned i; | |
3970 | ||
65e19f54 | 3971 | if (ctx->file_table) |
6b06314c JA |
3972 | return -EBUSY; |
3973 | if (!nr_args) | |
3974 | return -EINVAL; | |
3975 | if (nr_args > IORING_MAX_FIXED_FILES) | |
3976 | return -EMFILE; | |
3977 | ||
65e19f54 JA |
3978 | nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE); |
3979 | ctx->file_table = kcalloc(nr_tables, sizeof(struct fixed_file_table), | |
3980 | GFP_KERNEL); | |
3981 | if (!ctx->file_table) | |
6b06314c JA |
3982 | return -ENOMEM; |
3983 | ||
65e19f54 JA |
3984 | if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) { |
3985 | kfree(ctx->file_table); | |
46568e9b | 3986 | ctx->file_table = NULL; |
65e19f54 JA |
3987 | return -ENOMEM; |
3988 | } | |
3989 | ||
08a45173 | 3990 | for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { |
65e19f54 JA |
3991 | struct fixed_file_table *table; |
3992 | unsigned index; | |
3993 | ||
6b06314c JA |
3994 | ret = -EFAULT; |
3995 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) | |
3996 | break; | |
08a45173 JA |
3997 | /* allow sparse sets */ |
3998 | if (fd == -1) { | |
3999 | ret = 0; | |
4000 | continue; | |
4001 | } | |
6b06314c | 4002 | |
65e19f54 JA |
4003 | table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT]; |
4004 | index = i & IORING_FILE_TABLE_MASK; | |
4005 | table->files[index] = fget(fd); | |
6b06314c JA |
4006 | |
4007 | ret = -EBADF; | |
65e19f54 | 4008 | if (!table->files[index]) |
6b06314c JA |
4009 | break; |
4010 | /* | |
4011 | * Don't allow io_uring instances to be registered. If UNIX | |
4012 | * isn't enabled, then this causes a reference cycle and this | |
4013 | * instance can never get freed. If UNIX is enabled we'll | |
4014 | * handle it just fine, but there's still no point in allowing | |
4015 | * a ring fd as it doesn't support regular read/write anyway. | |
4016 | */ | |
65e19f54 JA |
4017 | if (table->files[index]->f_op == &io_uring_fops) { |
4018 | fput(table->files[index]); | |
6b06314c JA |
4019 | break; |
4020 | } | |
6b06314c JA |
4021 | ret = 0; |
4022 | } | |
4023 | ||
4024 | if (ret) { | |
65e19f54 JA |
4025 | for (i = 0; i < ctx->nr_user_files; i++) { |
4026 | struct file *file; | |
6b06314c | 4027 | |
65e19f54 JA |
4028 | file = io_file_from_index(ctx, i); |
4029 | if (file) | |
4030 | fput(file); | |
4031 | } | |
4032 | for (i = 0; i < nr_tables; i++) | |
4033 | kfree(ctx->file_table[i].files); | |
6b06314c | 4034 | |
65e19f54 JA |
4035 | kfree(ctx->file_table); |
4036 | ctx->file_table = NULL; | |
6b06314c JA |
4037 | ctx->nr_user_files = 0; |
4038 | return ret; | |
4039 | } | |
4040 | ||
4041 | ret = io_sqe_files_scm(ctx); | |
4042 | if (ret) | |
4043 | io_sqe_files_unregister(ctx); | |
4044 | ||
4045 | return ret; | |
4046 | } | |
4047 | ||
c3a31e60 JA |
4048 | static void io_sqe_file_unregister(struct io_ring_ctx *ctx, int index) |
4049 | { | |
4050 | #if defined(CONFIG_UNIX) | |
65e19f54 | 4051 | struct file *file = io_file_from_index(ctx, index); |
c3a31e60 JA |
4052 | struct sock *sock = ctx->ring_sock->sk; |
4053 | struct sk_buff_head list, *head = &sock->sk_receive_queue; | |
4054 | struct sk_buff *skb; | |
4055 | int i; | |
4056 | ||
4057 | __skb_queue_head_init(&list); | |
4058 | ||
4059 | /* | |
4060 | * Find the skb that holds this file in its SCM_RIGHTS. When found, | |
4061 | * remove this entry and rearrange the file array. | |
4062 | */ | |
4063 | skb = skb_dequeue(head); | |
4064 | while (skb) { | |
4065 | struct scm_fp_list *fp; | |
4066 | ||
4067 | fp = UNIXCB(skb).fp; | |
4068 | for (i = 0; i < fp->count; i++) { | |
4069 | int left; | |
4070 | ||
4071 | if (fp->fp[i] != file) | |
4072 | continue; | |
4073 | ||
4074 | unix_notinflight(fp->user, fp->fp[i]); | |
4075 | left = fp->count - 1 - i; | |
4076 | if (left) { | |
4077 | memmove(&fp->fp[i], &fp->fp[i + 1], | |
4078 | left * sizeof(struct file *)); | |
4079 | } | |
4080 | fp->count--; | |
4081 | if (!fp->count) { | |
4082 | kfree_skb(skb); | |
4083 | skb = NULL; | |
4084 | } else { | |
4085 | __skb_queue_tail(&list, skb); | |
4086 | } | |
4087 | fput(file); | |
4088 | file = NULL; | |
4089 | break; | |
4090 | } | |
4091 | ||
4092 | if (!file) | |
4093 | break; | |
4094 | ||
4095 | __skb_queue_tail(&list, skb); | |
4096 | ||
4097 | skb = skb_dequeue(head); | |
4098 | } | |
4099 | ||
4100 | if (skb_peek(&list)) { | |
4101 | spin_lock_irq(&head->lock); | |
4102 | while ((skb = __skb_dequeue(&list)) != NULL) | |
4103 | __skb_queue_tail(head, skb); | |
4104 | spin_unlock_irq(&head->lock); | |
4105 | } | |
4106 | #else | |
65e19f54 | 4107 | fput(io_file_from_index(ctx, index)); |
c3a31e60 JA |
4108 | #endif |
4109 | } | |
4110 | ||
4111 | static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file, | |
4112 | int index) | |
4113 | { | |
4114 | #if defined(CONFIG_UNIX) | |
4115 | struct sock *sock = ctx->ring_sock->sk; | |
4116 | struct sk_buff_head *head = &sock->sk_receive_queue; | |
4117 | struct sk_buff *skb; | |
4118 | ||
4119 | /* | |
4120 | * See if we can merge this file into an existing skb SCM_RIGHTS | |
4121 | * file set. If there's no room, fall back to allocating a new skb | |
4122 | * and filling it in. | |
4123 | */ | |
4124 | spin_lock_irq(&head->lock); | |
4125 | skb = skb_peek(head); | |
4126 | if (skb) { | |
4127 | struct scm_fp_list *fpl = UNIXCB(skb).fp; | |
4128 | ||
4129 | if (fpl->count < SCM_MAX_FD) { | |
4130 | __skb_unlink(skb, head); | |
4131 | spin_unlock_irq(&head->lock); | |
4132 | fpl->fp[fpl->count] = get_file(file); | |
4133 | unix_inflight(fpl->user, fpl->fp[fpl->count]); | |
4134 | fpl->count++; | |
4135 | spin_lock_irq(&head->lock); | |
4136 | __skb_queue_head(head, skb); | |
4137 | } else { | |
4138 | skb = NULL; | |
4139 | } | |
4140 | } | |
4141 | spin_unlock_irq(&head->lock); | |
4142 | ||
4143 | if (skb) { | |
4144 | fput(file); | |
4145 | return 0; | |
4146 | } | |
4147 | ||
4148 | return __io_sqe_files_scm(ctx, 1, index); | |
4149 | #else | |
4150 | return 0; | |
4151 | #endif | |
4152 | } | |
4153 | ||
4154 | static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg, | |
4155 | unsigned nr_args) | |
4156 | { | |
4157 | struct io_uring_files_update up; | |
4158 | __s32 __user *fds; | |
4159 | int fd, i, err; | |
4160 | __u32 done; | |
4161 | ||
65e19f54 | 4162 | if (!ctx->file_table) |
c3a31e60 JA |
4163 | return -ENXIO; |
4164 | if (!nr_args) | |
4165 | return -EINVAL; | |
4166 | if (copy_from_user(&up, arg, sizeof(up))) | |
4167 | return -EFAULT; | |
4168 | if (check_add_overflow(up.offset, nr_args, &done)) | |
4169 | return -EOVERFLOW; | |
4170 | if (done > ctx->nr_user_files) | |
4171 | return -EINVAL; | |
4172 | ||
4173 | done = 0; | |
4174 | fds = (__s32 __user *) up.fds; | |
4175 | while (nr_args) { | |
65e19f54 JA |
4176 | struct fixed_file_table *table; |
4177 | unsigned index; | |
4178 | ||
c3a31e60 JA |
4179 | err = 0; |
4180 | if (copy_from_user(&fd, &fds[done], sizeof(fd))) { | |
4181 | err = -EFAULT; | |
4182 | break; | |
4183 | } | |
4184 | i = array_index_nospec(up.offset, ctx->nr_user_files); | |
65e19f54 JA |
4185 | table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT]; |
4186 | index = i & IORING_FILE_TABLE_MASK; | |
4187 | if (table->files[index]) { | |
c3a31e60 | 4188 | io_sqe_file_unregister(ctx, i); |
65e19f54 | 4189 | table->files[index] = NULL; |
c3a31e60 JA |
4190 | } |
4191 | if (fd != -1) { | |
4192 | struct file *file; | |
4193 | ||
4194 | file = fget(fd); | |
4195 | if (!file) { | |
4196 | err = -EBADF; | |
4197 | break; | |
4198 | } | |
4199 | /* | |
4200 | * Don't allow io_uring instances to be registered. If | |
4201 | * UNIX isn't enabled, then this causes a reference | |
4202 | * cycle and this instance can never get freed. If UNIX | |
4203 | * is enabled we'll handle it just fine, but there's | |
4204 | * still no point in allowing a ring fd as it doesn't | |
4205 | * support regular read/write anyway. | |
4206 | */ | |
4207 | if (file->f_op == &io_uring_fops) { | |
4208 | fput(file); | |
4209 | err = -EBADF; | |
4210 | break; | |
4211 | } | |
65e19f54 | 4212 | table->files[index] = file; |
c3a31e60 JA |
4213 | err = io_sqe_file_register(ctx, file, i); |
4214 | if (err) | |
4215 | break; | |
4216 | } | |
4217 | nr_args--; | |
4218 | done++; | |
4219 | up.offset++; | |
4220 | } | |
4221 | ||
4222 | return done ? done : err; | |
4223 | } | |
4224 | ||
7d723065 JA |
4225 | static void io_put_work(struct io_wq_work *work) |
4226 | { | |
4227 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
4228 | ||
4229 | io_put_req(req); | |
4230 | } | |
4231 | ||
4232 | static void io_get_work(struct io_wq_work *work) | |
4233 | { | |
4234 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
4235 | ||
4236 | refcount_inc(&req->refs); | |
4237 | } | |
4238 | ||
6c271ce2 JA |
4239 | static int io_sq_offload_start(struct io_ring_ctx *ctx, |
4240 | struct io_uring_params *p) | |
2b188cc1 | 4241 | { |
576a347b | 4242 | struct io_wq_data data; |
561fb04a | 4243 | unsigned concurrency; |
2b188cc1 JA |
4244 | int ret; |
4245 | ||
6c271ce2 | 4246 | init_waitqueue_head(&ctx->sqo_wait); |
2b188cc1 JA |
4247 | mmgrab(current->mm); |
4248 | ctx->sqo_mm = current->mm; | |
4249 | ||
6c271ce2 | 4250 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3ec482d1 JA |
4251 | ret = -EPERM; |
4252 | if (!capable(CAP_SYS_ADMIN)) | |
4253 | goto err; | |
4254 | ||
917257da JA |
4255 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
4256 | if (!ctx->sq_thread_idle) | |
4257 | ctx->sq_thread_idle = HZ; | |
4258 | ||
6c271ce2 | 4259 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 4260 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 4261 | |
917257da | 4262 | ret = -EINVAL; |
44a9bd18 JA |
4263 | if (cpu >= nr_cpu_ids) |
4264 | goto err; | |
7889f44d | 4265 | if (!cpu_online(cpu)) |
917257da JA |
4266 | goto err; |
4267 | ||
6c271ce2 JA |
4268 | ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread, |
4269 | ctx, cpu, | |
4270 | "io_uring-sq"); | |
4271 | } else { | |
4272 | ctx->sqo_thread = kthread_create(io_sq_thread, ctx, | |
4273 | "io_uring-sq"); | |
4274 | } | |
4275 | if (IS_ERR(ctx->sqo_thread)) { | |
4276 | ret = PTR_ERR(ctx->sqo_thread); | |
4277 | ctx->sqo_thread = NULL; | |
4278 | goto err; | |
4279 | } | |
4280 | wake_up_process(ctx->sqo_thread); | |
4281 | } else if (p->flags & IORING_SETUP_SQ_AFF) { | |
4282 | /* Can't have SQ_AFF without SQPOLL */ | |
4283 | ret = -EINVAL; | |
4284 | goto err; | |
4285 | } | |
4286 | ||
576a347b JA |
4287 | data.mm = ctx->sqo_mm; |
4288 | data.user = ctx->user; | |
181e448d | 4289 | data.creds = ctx->creds; |
576a347b JA |
4290 | data.get_work = io_get_work; |
4291 | data.put_work = io_put_work; | |
4292 | ||
561fb04a JA |
4293 | /* Do QD, or 4 * CPUS, whatever is smallest */ |
4294 | concurrency = min(ctx->sq_entries, 4 * num_online_cpus()); | |
576a347b | 4295 | ctx->io_wq = io_wq_create(concurrency, &data); |
975c99a5 JA |
4296 | if (IS_ERR(ctx->io_wq)) { |
4297 | ret = PTR_ERR(ctx->io_wq); | |
4298 | ctx->io_wq = NULL; | |
2b188cc1 JA |
4299 | goto err; |
4300 | } | |
4301 | ||
4302 | return 0; | |
4303 | err: | |
54a91f3b | 4304 | io_finish_async(ctx); |
2b188cc1 JA |
4305 | mmdrop(ctx->sqo_mm); |
4306 | ctx->sqo_mm = NULL; | |
4307 | return ret; | |
4308 | } | |
4309 | ||
4310 | static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages) | |
4311 | { | |
4312 | atomic_long_sub(nr_pages, &user->locked_vm); | |
4313 | } | |
4314 | ||
4315 | static int io_account_mem(struct user_struct *user, unsigned long nr_pages) | |
4316 | { | |
4317 | unsigned long page_limit, cur_pages, new_pages; | |
4318 | ||
4319 | /* Don't allow more pages than we can safely lock */ | |
4320 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
4321 | ||
4322 | do { | |
4323 | cur_pages = atomic_long_read(&user->locked_vm); | |
4324 | new_pages = cur_pages + nr_pages; | |
4325 | if (new_pages > page_limit) | |
4326 | return -ENOMEM; | |
4327 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
4328 | new_pages) != cur_pages); | |
4329 | ||
4330 | return 0; | |
4331 | } | |
4332 | ||
4333 | static void io_mem_free(void *ptr) | |
4334 | { | |
52e04ef4 MR |
4335 | struct page *page; |
4336 | ||
4337 | if (!ptr) | |
4338 | return; | |
2b188cc1 | 4339 | |
52e04ef4 | 4340 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
4341 | if (put_page_testzero(page)) |
4342 | free_compound_page(page); | |
4343 | } | |
4344 | ||
4345 | static void *io_mem_alloc(size_t size) | |
4346 | { | |
4347 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
4348 | __GFP_NORETRY; | |
4349 | ||
4350 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
4351 | } | |
4352 | ||
75b28aff HV |
4353 | static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries, |
4354 | size_t *sq_offset) | |
4355 | { | |
4356 | struct io_rings *rings; | |
4357 | size_t off, sq_array_size; | |
4358 | ||
4359 | off = struct_size(rings, cqes, cq_entries); | |
4360 | if (off == SIZE_MAX) | |
4361 | return SIZE_MAX; | |
4362 | ||
4363 | #ifdef CONFIG_SMP | |
4364 | off = ALIGN(off, SMP_CACHE_BYTES); | |
4365 | if (off == 0) | |
4366 | return SIZE_MAX; | |
4367 | #endif | |
4368 | ||
4369 | sq_array_size = array_size(sizeof(u32), sq_entries); | |
4370 | if (sq_array_size == SIZE_MAX) | |
4371 | return SIZE_MAX; | |
4372 | ||
4373 | if (check_add_overflow(off, sq_array_size, &off)) | |
4374 | return SIZE_MAX; | |
4375 | ||
4376 | if (sq_offset) | |
4377 | *sq_offset = off; | |
4378 | ||
4379 | return off; | |
4380 | } | |
4381 | ||
2b188cc1 JA |
4382 | static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries) |
4383 | { | |
75b28aff | 4384 | size_t pages; |
2b188cc1 | 4385 | |
75b28aff HV |
4386 | pages = (size_t)1 << get_order( |
4387 | rings_size(sq_entries, cq_entries, NULL)); | |
4388 | pages += (size_t)1 << get_order( | |
4389 | array_size(sizeof(struct io_uring_sqe), sq_entries)); | |
2b188cc1 | 4390 | |
75b28aff | 4391 | return pages; |
2b188cc1 JA |
4392 | } |
4393 | ||
edafccee JA |
4394 | static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx) |
4395 | { | |
4396 | int i, j; | |
4397 | ||
4398 | if (!ctx->user_bufs) | |
4399 | return -ENXIO; | |
4400 | ||
4401 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
4402 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
4403 | ||
4404 | for (j = 0; j < imu->nr_bvecs; j++) | |
27c4d3a3 | 4405 | put_user_page(imu->bvec[j].bv_page); |
edafccee JA |
4406 | |
4407 | if (ctx->account_mem) | |
4408 | io_unaccount_mem(ctx->user, imu->nr_bvecs); | |
d4ef6475 | 4409 | kvfree(imu->bvec); |
edafccee JA |
4410 | imu->nr_bvecs = 0; |
4411 | } | |
4412 | ||
4413 | kfree(ctx->user_bufs); | |
4414 | ctx->user_bufs = NULL; | |
4415 | ctx->nr_user_bufs = 0; | |
4416 | return 0; | |
4417 | } | |
4418 | ||
4419 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
4420 | void __user *arg, unsigned index) | |
4421 | { | |
4422 | struct iovec __user *src; | |
4423 | ||
4424 | #ifdef CONFIG_COMPAT | |
4425 | if (ctx->compat) { | |
4426 | struct compat_iovec __user *ciovs; | |
4427 | struct compat_iovec ciov; | |
4428 | ||
4429 | ciovs = (struct compat_iovec __user *) arg; | |
4430 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
4431 | return -EFAULT; | |
4432 | ||
4433 | dst->iov_base = (void __user *) (unsigned long) ciov.iov_base; | |
4434 | dst->iov_len = ciov.iov_len; | |
4435 | return 0; | |
4436 | } | |
4437 | #endif | |
4438 | src = (struct iovec __user *) arg; | |
4439 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
4440 | return -EFAULT; | |
4441 | return 0; | |
4442 | } | |
4443 | ||
4444 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, | |
4445 | unsigned nr_args) | |
4446 | { | |
4447 | struct vm_area_struct **vmas = NULL; | |
4448 | struct page **pages = NULL; | |
4449 | int i, j, got_pages = 0; | |
4450 | int ret = -EINVAL; | |
4451 | ||
4452 | if (ctx->user_bufs) | |
4453 | return -EBUSY; | |
4454 | if (!nr_args || nr_args > UIO_MAXIOV) | |
4455 | return -EINVAL; | |
4456 | ||
4457 | ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf), | |
4458 | GFP_KERNEL); | |
4459 | if (!ctx->user_bufs) | |
4460 | return -ENOMEM; | |
4461 | ||
4462 | for (i = 0; i < nr_args; i++) { | |
4463 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
4464 | unsigned long off, start, end, ubuf; | |
4465 | int pret, nr_pages; | |
4466 | struct iovec iov; | |
4467 | size_t size; | |
4468 | ||
4469 | ret = io_copy_iov(ctx, &iov, arg, i); | |
4470 | if (ret) | |
a278682d | 4471 | goto err; |
edafccee JA |
4472 | |
4473 | /* | |
4474 | * Don't impose further limits on the size and buffer | |
4475 | * constraints here, we'll -EINVAL later when IO is | |
4476 | * submitted if they are wrong. | |
4477 | */ | |
4478 | ret = -EFAULT; | |
4479 | if (!iov.iov_base || !iov.iov_len) | |
4480 | goto err; | |
4481 | ||
4482 | /* arbitrary limit, but we need something */ | |
4483 | if (iov.iov_len > SZ_1G) | |
4484 | goto err; | |
4485 | ||
4486 | ubuf = (unsigned long) iov.iov_base; | |
4487 | end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
4488 | start = ubuf >> PAGE_SHIFT; | |
4489 | nr_pages = end - start; | |
4490 | ||
4491 | if (ctx->account_mem) { | |
4492 | ret = io_account_mem(ctx->user, nr_pages); | |
4493 | if (ret) | |
4494 | goto err; | |
4495 | } | |
4496 | ||
4497 | ret = 0; | |
4498 | if (!pages || nr_pages > got_pages) { | |
4499 | kfree(vmas); | |
4500 | kfree(pages); | |
d4ef6475 | 4501 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), |
edafccee | 4502 | GFP_KERNEL); |
d4ef6475 | 4503 | vmas = kvmalloc_array(nr_pages, |
edafccee JA |
4504 | sizeof(struct vm_area_struct *), |
4505 | GFP_KERNEL); | |
4506 | if (!pages || !vmas) { | |
4507 | ret = -ENOMEM; | |
4508 | if (ctx->account_mem) | |
4509 | io_unaccount_mem(ctx->user, nr_pages); | |
4510 | goto err; | |
4511 | } | |
4512 | got_pages = nr_pages; | |
4513 | } | |
4514 | ||
d4ef6475 | 4515 | imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec), |
edafccee JA |
4516 | GFP_KERNEL); |
4517 | ret = -ENOMEM; | |
4518 | if (!imu->bvec) { | |
4519 | if (ctx->account_mem) | |
4520 | io_unaccount_mem(ctx->user, nr_pages); | |
4521 | goto err; | |
4522 | } | |
4523 | ||
4524 | ret = 0; | |
4525 | down_read(¤t->mm->mmap_sem); | |
932f4a63 IW |
4526 | pret = get_user_pages(ubuf, nr_pages, |
4527 | FOLL_WRITE | FOLL_LONGTERM, | |
4528 | pages, vmas); | |
edafccee JA |
4529 | if (pret == nr_pages) { |
4530 | /* don't support file backed memory */ | |
4531 | for (j = 0; j < nr_pages; j++) { | |
4532 | struct vm_area_struct *vma = vmas[j]; | |
4533 | ||
4534 | if (vma->vm_file && | |
4535 | !is_file_hugepages(vma->vm_file)) { | |
4536 | ret = -EOPNOTSUPP; | |
4537 | break; | |
4538 | } | |
4539 | } | |
4540 | } else { | |
4541 | ret = pret < 0 ? pret : -EFAULT; | |
4542 | } | |
4543 | up_read(¤t->mm->mmap_sem); | |
4544 | if (ret) { | |
4545 | /* | |
4546 | * if we did partial map, or found file backed vmas, | |
4547 | * release any pages we did get | |
4548 | */ | |
27c4d3a3 JH |
4549 | if (pret > 0) |
4550 | put_user_pages(pages, pret); | |
edafccee JA |
4551 | if (ctx->account_mem) |
4552 | io_unaccount_mem(ctx->user, nr_pages); | |
d4ef6475 | 4553 | kvfree(imu->bvec); |
edafccee JA |
4554 | goto err; |
4555 | } | |
4556 | ||
4557 | off = ubuf & ~PAGE_MASK; | |
4558 | size = iov.iov_len; | |
4559 | for (j = 0; j < nr_pages; j++) { | |
4560 | size_t vec_len; | |
4561 | ||
4562 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
4563 | imu->bvec[j].bv_page = pages[j]; | |
4564 | imu->bvec[j].bv_len = vec_len; | |
4565 | imu->bvec[j].bv_offset = off; | |
4566 | off = 0; | |
4567 | size -= vec_len; | |
4568 | } | |
4569 | /* store original address for later verification */ | |
4570 | imu->ubuf = ubuf; | |
4571 | imu->len = iov.iov_len; | |
4572 | imu->nr_bvecs = nr_pages; | |
4573 | ||
4574 | ctx->nr_user_bufs++; | |
4575 | } | |
d4ef6475 MR |
4576 | kvfree(pages); |
4577 | kvfree(vmas); | |
edafccee JA |
4578 | return 0; |
4579 | err: | |
d4ef6475 MR |
4580 | kvfree(pages); |
4581 | kvfree(vmas); | |
edafccee JA |
4582 | io_sqe_buffer_unregister(ctx); |
4583 | return ret; | |
4584 | } | |
4585 | ||
9b402849 JA |
4586 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
4587 | { | |
4588 | __s32 __user *fds = arg; | |
4589 | int fd; | |
4590 | ||
4591 | if (ctx->cq_ev_fd) | |
4592 | return -EBUSY; | |
4593 | ||
4594 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
4595 | return -EFAULT; | |
4596 | ||
4597 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
4598 | if (IS_ERR(ctx->cq_ev_fd)) { | |
4599 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
4600 | ctx->cq_ev_fd = NULL; | |
4601 | return ret; | |
4602 | } | |
4603 | ||
4604 | return 0; | |
4605 | } | |
4606 | ||
4607 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
4608 | { | |
4609 | if (ctx->cq_ev_fd) { | |
4610 | eventfd_ctx_put(ctx->cq_ev_fd); | |
4611 | ctx->cq_ev_fd = NULL; | |
4612 | return 0; | |
4613 | } | |
4614 | ||
4615 | return -ENXIO; | |
4616 | } | |
4617 | ||
2b188cc1 JA |
4618 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
4619 | { | |
6b06314c | 4620 | io_finish_async(ctx); |
2b188cc1 JA |
4621 | if (ctx->sqo_mm) |
4622 | mmdrop(ctx->sqo_mm); | |
def596e9 JA |
4623 | |
4624 | io_iopoll_reap_events(ctx); | |
edafccee | 4625 | io_sqe_buffer_unregister(ctx); |
6b06314c | 4626 | io_sqe_files_unregister(ctx); |
9b402849 | 4627 | io_eventfd_unregister(ctx); |
def596e9 | 4628 | |
2b188cc1 | 4629 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
4630 | if (ctx->ring_sock) { |
4631 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 4632 | sock_release(ctx->ring_sock); |
355e8d26 | 4633 | } |
2b188cc1 JA |
4634 | #endif |
4635 | ||
75b28aff | 4636 | io_mem_free(ctx->rings); |
2b188cc1 | 4637 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
4638 | |
4639 | percpu_ref_exit(&ctx->refs); | |
4640 | if (ctx->account_mem) | |
4641 | io_unaccount_mem(ctx->user, | |
4642 | ring_pages(ctx->sq_entries, ctx->cq_entries)); | |
4643 | free_uid(ctx->user); | |
181e448d | 4644 | put_cred(ctx->creds); |
206aefde | 4645 | kfree(ctx->completions); |
78076bb6 | 4646 | kfree(ctx->cancel_hash); |
0ddf92e8 | 4647 | kmem_cache_free(req_cachep, ctx->fallback_req); |
2b188cc1 JA |
4648 | kfree(ctx); |
4649 | } | |
4650 | ||
4651 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
4652 | { | |
4653 | struct io_ring_ctx *ctx = file->private_data; | |
4654 | __poll_t mask = 0; | |
4655 | ||
4656 | poll_wait(file, &ctx->cq_wait, wait); | |
4f7067c3 SB |
4657 | /* |
4658 | * synchronizes with barrier from wq_has_sleeper call in | |
4659 | * io_commit_cqring | |
4660 | */ | |
2b188cc1 | 4661 | smp_rmb(); |
75b28aff HV |
4662 | if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head != |
4663 | ctx->rings->sq_ring_entries) | |
2b188cc1 | 4664 | mask |= EPOLLOUT | EPOLLWRNORM; |
daa5de54 | 4665 | if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail) |
2b188cc1 JA |
4666 | mask |= EPOLLIN | EPOLLRDNORM; |
4667 | ||
4668 | return mask; | |
4669 | } | |
4670 | ||
4671 | static int io_uring_fasync(int fd, struct file *file, int on) | |
4672 | { | |
4673 | struct io_ring_ctx *ctx = file->private_data; | |
4674 | ||
4675 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
4676 | } | |
4677 | ||
4678 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) | |
4679 | { | |
4680 | mutex_lock(&ctx->uring_lock); | |
4681 | percpu_ref_kill(&ctx->refs); | |
4682 | mutex_unlock(&ctx->uring_lock); | |
4683 | ||
5262f567 | 4684 | io_kill_timeouts(ctx); |
221c5eb2 | 4685 | io_poll_remove_all(ctx); |
561fb04a JA |
4686 | |
4687 | if (ctx->io_wq) | |
4688 | io_wq_cancel_all(ctx->io_wq); | |
4689 | ||
def596e9 | 4690 | io_iopoll_reap_events(ctx); |
15dff286 JA |
4691 | /* if we failed setting up the ctx, we might not have any rings */ |
4692 | if (ctx->rings) | |
4693 | io_cqring_overflow_flush(ctx, true); | |
206aefde | 4694 | wait_for_completion(&ctx->completions[0]); |
2b188cc1 JA |
4695 | io_ring_ctx_free(ctx); |
4696 | } | |
4697 | ||
4698 | static int io_uring_release(struct inode *inode, struct file *file) | |
4699 | { | |
4700 | struct io_ring_ctx *ctx = file->private_data; | |
4701 | ||
4702 | file->private_data = NULL; | |
4703 | io_ring_ctx_wait_and_kill(ctx); | |
4704 | return 0; | |
4705 | } | |
4706 | ||
fcb323cc JA |
4707 | static void io_uring_cancel_files(struct io_ring_ctx *ctx, |
4708 | struct files_struct *files) | |
4709 | { | |
4710 | struct io_kiocb *req; | |
4711 | DEFINE_WAIT(wait); | |
4712 | ||
4713 | while (!list_empty_careful(&ctx->inflight_list)) { | |
768134d4 | 4714 | struct io_kiocb *cancel_req = NULL; |
fcb323cc JA |
4715 | |
4716 | spin_lock_irq(&ctx->inflight_lock); | |
4717 | list_for_each_entry(req, &ctx->inflight_list, inflight_entry) { | |
768134d4 JA |
4718 | if (req->work.files != files) |
4719 | continue; | |
4720 | /* req is being completed, ignore */ | |
4721 | if (!refcount_inc_not_zero(&req->refs)) | |
4722 | continue; | |
4723 | cancel_req = req; | |
4724 | break; | |
fcb323cc | 4725 | } |
768134d4 | 4726 | if (cancel_req) |
fcb323cc | 4727 | prepare_to_wait(&ctx->inflight_wait, &wait, |
768134d4 | 4728 | TASK_UNINTERRUPTIBLE); |
fcb323cc JA |
4729 | spin_unlock_irq(&ctx->inflight_lock); |
4730 | ||
768134d4 JA |
4731 | /* We need to keep going until we don't find a matching req */ |
4732 | if (!cancel_req) | |
fcb323cc | 4733 | break; |
2f6d9b9d BL |
4734 | |
4735 | io_wq_cancel_work(ctx->io_wq, &cancel_req->work); | |
4736 | io_put_req(cancel_req); | |
fcb323cc JA |
4737 | schedule(); |
4738 | } | |
768134d4 | 4739 | finish_wait(&ctx->inflight_wait, &wait); |
fcb323cc JA |
4740 | } |
4741 | ||
4742 | static int io_uring_flush(struct file *file, void *data) | |
4743 | { | |
4744 | struct io_ring_ctx *ctx = file->private_data; | |
4745 | ||
4746 | io_uring_cancel_files(ctx, data); | |
1d7bb1d5 JA |
4747 | if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) { |
4748 | io_cqring_overflow_flush(ctx, true); | |
fcb323cc | 4749 | io_wq_cancel_all(ctx->io_wq); |
1d7bb1d5 | 4750 | } |
fcb323cc JA |
4751 | return 0; |
4752 | } | |
4753 | ||
6c5c240e RP |
4754 | static void *io_uring_validate_mmap_request(struct file *file, |
4755 | loff_t pgoff, size_t sz) | |
2b188cc1 | 4756 | { |
2b188cc1 | 4757 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 4758 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
4759 | struct page *page; |
4760 | void *ptr; | |
4761 | ||
4762 | switch (offset) { | |
4763 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
4764 | case IORING_OFF_CQ_RING: |
4765 | ptr = ctx->rings; | |
2b188cc1 JA |
4766 | break; |
4767 | case IORING_OFF_SQES: | |
4768 | ptr = ctx->sq_sqes; | |
4769 | break; | |
2b188cc1 | 4770 | default: |
6c5c240e | 4771 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
4772 | } |
4773 | ||
4774 | page = virt_to_head_page(ptr); | |
a50b854e | 4775 | if (sz > page_size(page)) |
6c5c240e RP |
4776 | return ERR_PTR(-EINVAL); |
4777 | ||
4778 | return ptr; | |
4779 | } | |
4780 | ||
4781 | #ifdef CONFIG_MMU | |
4782 | ||
4783 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
4784 | { | |
4785 | size_t sz = vma->vm_end - vma->vm_start; | |
4786 | unsigned long pfn; | |
4787 | void *ptr; | |
4788 | ||
4789 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
4790 | if (IS_ERR(ptr)) | |
4791 | return PTR_ERR(ptr); | |
2b188cc1 JA |
4792 | |
4793 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
4794 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
4795 | } | |
4796 | ||
6c5c240e RP |
4797 | #else /* !CONFIG_MMU */ |
4798 | ||
4799 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
4800 | { | |
4801 | return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL; | |
4802 | } | |
4803 | ||
4804 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
4805 | { | |
4806 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
4807 | } | |
4808 | ||
4809 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
4810 | unsigned long addr, unsigned long len, | |
4811 | unsigned long pgoff, unsigned long flags) | |
4812 | { | |
4813 | void *ptr; | |
4814 | ||
4815 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
4816 | if (IS_ERR(ptr)) | |
4817 | return PTR_ERR(ptr); | |
4818 | ||
4819 | return (unsigned long) ptr; | |
4820 | } | |
4821 | ||
4822 | #endif /* !CONFIG_MMU */ | |
4823 | ||
2b188cc1 JA |
4824 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
4825 | u32, min_complete, u32, flags, const sigset_t __user *, sig, | |
4826 | size_t, sigsz) | |
4827 | { | |
4828 | struct io_ring_ctx *ctx; | |
4829 | long ret = -EBADF; | |
4830 | int submitted = 0; | |
4831 | struct fd f; | |
4832 | ||
6c271ce2 | 4833 | if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP)) |
2b188cc1 JA |
4834 | return -EINVAL; |
4835 | ||
4836 | f = fdget(fd); | |
4837 | if (!f.file) | |
4838 | return -EBADF; | |
4839 | ||
4840 | ret = -EOPNOTSUPP; | |
4841 | if (f.file->f_op != &io_uring_fops) | |
4842 | goto out_fput; | |
4843 | ||
4844 | ret = -ENXIO; | |
4845 | ctx = f.file->private_data; | |
4846 | if (!percpu_ref_tryget(&ctx->refs)) | |
4847 | goto out_fput; | |
4848 | ||
6c271ce2 JA |
4849 | /* |
4850 | * For SQ polling, the thread will do all submissions and completions. | |
4851 | * Just return the requested submit count, and wake the thread if | |
4852 | * we were asked to. | |
4853 | */ | |
b2a9eada | 4854 | ret = 0; |
6c271ce2 | 4855 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
c1edbf5f JA |
4856 | if (!list_empty_careful(&ctx->cq_overflow_list)) |
4857 | io_cqring_overflow_flush(ctx, false); | |
6c271ce2 JA |
4858 | if (flags & IORING_ENTER_SQ_WAKEUP) |
4859 | wake_up(&ctx->sqo_wait); | |
4860 | submitted = to_submit; | |
b2a9eada | 4861 | } else if (to_submit) { |
ae9428ca | 4862 | struct mm_struct *cur_mm; |
2b188cc1 | 4863 | |
ae9428ca | 4864 | to_submit = min(to_submit, ctx->sq_entries); |
2b188cc1 | 4865 | mutex_lock(&ctx->uring_lock); |
ae9428ca PB |
4866 | /* already have mm, so io_submit_sqes() won't try to grab it */ |
4867 | cur_mm = ctx->sqo_mm; | |
4868 | submitted = io_submit_sqes(ctx, to_submit, f.file, fd, | |
4869 | &cur_mm, false); | |
2b188cc1 | 4870 | mutex_unlock(&ctx->uring_lock); |
2b188cc1 JA |
4871 | } |
4872 | if (flags & IORING_ENTER_GETEVENTS) { | |
def596e9 JA |
4873 | unsigned nr_events = 0; |
4874 | ||
2b188cc1 JA |
4875 | min_complete = min(min_complete, ctx->cq_entries); |
4876 | ||
def596e9 | 4877 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 | 4878 | ret = io_iopoll_check(ctx, &nr_events, min_complete); |
def596e9 JA |
4879 | } else { |
4880 | ret = io_cqring_wait(ctx, min_complete, sig, sigsz); | |
4881 | } | |
2b188cc1 JA |
4882 | } |
4883 | ||
6805b32e | 4884 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
4885 | out_fput: |
4886 | fdput(f); | |
4887 | return submitted ? submitted : ret; | |
4888 | } | |
4889 | ||
4890 | static const struct file_operations io_uring_fops = { | |
4891 | .release = io_uring_release, | |
fcb323cc | 4892 | .flush = io_uring_flush, |
2b188cc1 | 4893 | .mmap = io_uring_mmap, |
6c5c240e RP |
4894 | #ifndef CONFIG_MMU |
4895 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
4896 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
4897 | #endif | |
2b188cc1 JA |
4898 | .poll = io_uring_poll, |
4899 | .fasync = io_uring_fasync, | |
4900 | }; | |
4901 | ||
4902 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
4903 | struct io_uring_params *p) | |
4904 | { | |
75b28aff HV |
4905 | struct io_rings *rings; |
4906 | size_t size, sq_array_offset; | |
2b188cc1 | 4907 | |
75b28aff HV |
4908 | size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset); |
4909 | if (size == SIZE_MAX) | |
4910 | return -EOVERFLOW; | |
4911 | ||
4912 | rings = io_mem_alloc(size); | |
4913 | if (!rings) | |
2b188cc1 JA |
4914 | return -ENOMEM; |
4915 | ||
75b28aff HV |
4916 | ctx->rings = rings; |
4917 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
4918 | rings->sq_ring_mask = p->sq_entries - 1; | |
4919 | rings->cq_ring_mask = p->cq_entries - 1; | |
4920 | rings->sq_ring_entries = p->sq_entries; | |
4921 | rings->cq_ring_entries = p->cq_entries; | |
4922 | ctx->sq_mask = rings->sq_ring_mask; | |
4923 | ctx->cq_mask = rings->cq_ring_mask; | |
4924 | ctx->sq_entries = rings->sq_ring_entries; | |
4925 | ctx->cq_entries = rings->cq_ring_entries; | |
2b188cc1 JA |
4926 | |
4927 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 JA |
4928 | if (size == SIZE_MAX) { |
4929 | io_mem_free(ctx->rings); | |
4930 | ctx->rings = NULL; | |
2b188cc1 | 4931 | return -EOVERFLOW; |
eb065d30 | 4932 | } |
2b188cc1 JA |
4933 | |
4934 | ctx->sq_sqes = io_mem_alloc(size); | |
eb065d30 JA |
4935 | if (!ctx->sq_sqes) { |
4936 | io_mem_free(ctx->rings); | |
4937 | ctx->rings = NULL; | |
2b188cc1 | 4938 | return -ENOMEM; |
eb065d30 | 4939 | } |
2b188cc1 | 4940 | |
2b188cc1 JA |
4941 | return 0; |
4942 | } | |
4943 | ||
4944 | /* | |
4945 | * Allocate an anonymous fd, this is what constitutes the application | |
4946 | * visible backing of an io_uring instance. The application mmaps this | |
4947 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
4948 | * we have to tie this fd to a socket for file garbage collection purposes. | |
4949 | */ | |
4950 | static int io_uring_get_fd(struct io_ring_ctx *ctx) | |
4951 | { | |
4952 | struct file *file; | |
4953 | int ret; | |
4954 | ||
4955 | #if defined(CONFIG_UNIX) | |
4956 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, | |
4957 | &ctx->ring_sock); | |
4958 | if (ret) | |
4959 | return ret; | |
4960 | #endif | |
4961 | ||
4962 | ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
4963 | if (ret < 0) | |
4964 | goto err; | |
4965 | ||
4966 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, | |
4967 | O_RDWR | O_CLOEXEC); | |
4968 | if (IS_ERR(file)) { | |
4969 | put_unused_fd(ret); | |
4970 | ret = PTR_ERR(file); | |
4971 | goto err; | |
4972 | } | |
4973 | ||
4974 | #if defined(CONFIG_UNIX) | |
4975 | ctx->ring_sock->file = file; | |
6b06314c | 4976 | ctx->ring_sock->sk->sk_user_data = ctx; |
2b188cc1 JA |
4977 | #endif |
4978 | fd_install(ret, file); | |
4979 | return ret; | |
4980 | err: | |
4981 | #if defined(CONFIG_UNIX) | |
4982 | sock_release(ctx->ring_sock); | |
4983 | ctx->ring_sock = NULL; | |
4984 | #endif | |
4985 | return ret; | |
4986 | } | |
4987 | ||
4988 | static int io_uring_create(unsigned entries, struct io_uring_params *p) | |
4989 | { | |
4990 | struct user_struct *user = NULL; | |
4991 | struct io_ring_ctx *ctx; | |
4992 | bool account_mem; | |
4993 | int ret; | |
4994 | ||
4995 | if (!entries || entries > IORING_MAX_ENTRIES) | |
4996 | return -EINVAL; | |
4997 | ||
4998 | /* | |
4999 | * Use twice as many entries for the CQ ring. It's possible for the | |
5000 | * application to drive a higher depth than the size of the SQ ring, | |
5001 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
5002 | * some flexibility in overcommitting a bit. If the application has |
5003 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
5004 | * of CQ ring entries manually. | |
2b188cc1 JA |
5005 | */ |
5006 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
5007 | if (p->flags & IORING_SETUP_CQSIZE) { |
5008 | /* | |
5009 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
5010 | * to a power-of-two, if it isn't already. We do NOT impose | |
5011 | * any cq vs sq ring sizing. | |
5012 | */ | |
5013 | if (p->cq_entries < p->sq_entries || p->cq_entries > IORING_MAX_CQ_ENTRIES) | |
5014 | return -EINVAL; | |
5015 | p->cq_entries = roundup_pow_of_two(p->cq_entries); | |
5016 | } else { | |
5017 | p->cq_entries = 2 * p->sq_entries; | |
5018 | } | |
2b188cc1 JA |
5019 | |
5020 | user = get_uid(current_user()); | |
5021 | account_mem = !capable(CAP_IPC_LOCK); | |
5022 | ||
5023 | if (account_mem) { | |
5024 | ret = io_account_mem(user, | |
5025 | ring_pages(p->sq_entries, p->cq_entries)); | |
5026 | if (ret) { | |
5027 | free_uid(user); | |
5028 | return ret; | |
5029 | } | |
5030 | } | |
5031 | ||
5032 | ctx = io_ring_ctx_alloc(p); | |
5033 | if (!ctx) { | |
5034 | if (account_mem) | |
5035 | io_unaccount_mem(user, ring_pages(p->sq_entries, | |
5036 | p->cq_entries)); | |
5037 | free_uid(user); | |
5038 | return -ENOMEM; | |
5039 | } | |
5040 | ctx->compat = in_compat_syscall(); | |
5041 | ctx->account_mem = account_mem; | |
5042 | ctx->user = user; | |
0b8c0ec7 | 5043 | ctx->creds = get_current_cred(); |
2b188cc1 JA |
5044 | |
5045 | ret = io_allocate_scq_urings(ctx, p); | |
5046 | if (ret) | |
5047 | goto err; | |
5048 | ||
6c271ce2 | 5049 | ret = io_sq_offload_start(ctx, p); |
2b188cc1 JA |
5050 | if (ret) |
5051 | goto err; | |
5052 | ||
2b188cc1 | 5053 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
5054 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
5055 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
5056 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
5057 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
5058 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
5059 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
5060 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
5061 | |
5062 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
5063 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
5064 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
5065 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
5066 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
5067 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
5068 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
ac90f249 | 5069 | |
044c1ab3 JA |
5070 | /* |
5071 | * Install ring fd as the very last thing, so we don't risk someone | |
5072 | * having closed it before we finish setup | |
5073 | */ | |
5074 | ret = io_uring_get_fd(ctx); | |
5075 | if (ret < 0) | |
5076 | goto err; | |
5077 | ||
da8c9690 JA |
5078 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
5079 | IORING_FEAT_SUBMIT_STABLE; | |
c826bd7a | 5080 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
5081 | return ret; |
5082 | err: | |
5083 | io_ring_ctx_wait_and_kill(ctx); | |
5084 | return ret; | |
5085 | } | |
5086 | ||
5087 | /* | |
5088 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
5089 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
5090 | * params structure passed in. | |
5091 | */ | |
5092 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
5093 | { | |
5094 | struct io_uring_params p; | |
5095 | long ret; | |
5096 | int i; | |
5097 | ||
5098 | if (copy_from_user(&p, params, sizeof(p))) | |
5099 | return -EFAULT; | |
5100 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
5101 | if (p.resv[i]) | |
5102 | return -EINVAL; | |
5103 | } | |
5104 | ||
6c271ce2 | 5105 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
33a107f0 | 5106 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE)) |
2b188cc1 JA |
5107 | return -EINVAL; |
5108 | ||
5109 | ret = io_uring_create(entries, &p); | |
5110 | if (ret < 0) | |
5111 | return ret; | |
5112 | ||
5113 | if (copy_to_user(params, &p, sizeof(p))) | |
5114 | return -EFAULT; | |
5115 | ||
5116 | return ret; | |
5117 | } | |
5118 | ||
5119 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
5120 | struct io_uring_params __user *, params) | |
5121 | { | |
5122 | return io_uring_setup(entries, params); | |
5123 | } | |
5124 | ||
edafccee JA |
5125 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
5126 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
5127 | __releases(ctx->uring_lock) |
5128 | __acquires(ctx->uring_lock) | |
edafccee JA |
5129 | { |
5130 | int ret; | |
5131 | ||
35fa71a0 JA |
5132 | /* |
5133 | * We're inside the ring mutex, if the ref is already dying, then | |
5134 | * someone else killed the ctx or is already going through | |
5135 | * io_uring_register(). | |
5136 | */ | |
5137 | if (percpu_ref_is_dying(&ctx->refs)) | |
5138 | return -ENXIO; | |
5139 | ||
edafccee | 5140 | percpu_ref_kill(&ctx->refs); |
b19062a5 JA |
5141 | |
5142 | /* | |
5143 | * Drop uring mutex before waiting for references to exit. If another | |
5144 | * thread is currently inside io_uring_enter() it might need to grab | |
5145 | * the uring_lock to make progress. If we hold it here across the drain | |
5146 | * wait, then we can deadlock. It's safe to drop the mutex here, since | |
5147 | * no new references will come in after we've killed the percpu ref. | |
5148 | */ | |
5149 | mutex_unlock(&ctx->uring_lock); | |
206aefde | 5150 | wait_for_completion(&ctx->completions[0]); |
b19062a5 | 5151 | mutex_lock(&ctx->uring_lock); |
edafccee JA |
5152 | |
5153 | switch (opcode) { | |
5154 | case IORING_REGISTER_BUFFERS: | |
5155 | ret = io_sqe_buffer_register(ctx, arg, nr_args); | |
5156 | break; | |
5157 | case IORING_UNREGISTER_BUFFERS: | |
5158 | ret = -EINVAL; | |
5159 | if (arg || nr_args) | |
5160 | break; | |
5161 | ret = io_sqe_buffer_unregister(ctx); | |
5162 | break; | |
6b06314c JA |
5163 | case IORING_REGISTER_FILES: |
5164 | ret = io_sqe_files_register(ctx, arg, nr_args); | |
5165 | break; | |
5166 | case IORING_UNREGISTER_FILES: | |
5167 | ret = -EINVAL; | |
5168 | if (arg || nr_args) | |
5169 | break; | |
5170 | ret = io_sqe_files_unregister(ctx); | |
5171 | break; | |
c3a31e60 JA |
5172 | case IORING_REGISTER_FILES_UPDATE: |
5173 | ret = io_sqe_files_update(ctx, arg, nr_args); | |
5174 | break; | |
9b402849 JA |
5175 | case IORING_REGISTER_EVENTFD: |
5176 | ret = -EINVAL; | |
5177 | if (nr_args != 1) | |
5178 | break; | |
5179 | ret = io_eventfd_register(ctx, arg); | |
5180 | break; | |
5181 | case IORING_UNREGISTER_EVENTFD: | |
5182 | ret = -EINVAL; | |
5183 | if (arg || nr_args) | |
5184 | break; | |
5185 | ret = io_eventfd_unregister(ctx); | |
5186 | break; | |
edafccee JA |
5187 | default: |
5188 | ret = -EINVAL; | |
5189 | break; | |
5190 | } | |
5191 | ||
5192 | /* bring the ctx back to life */ | |
206aefde | 5193 | reinit_completion(&ctx->completions[0]); |
edafccee JA |
5194 | percpu_ref_reinit(&ctx->refs); |
5195 | return ret; | |
5196 | } | |
5197 | ||
5198 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
5199 | void __user *, arg, unsigned int, nr_args) | |
5200 | { | |
5201 | struct io_ring_ctx *ctx; | |
5202 | long ret = -EBADF; | |
5203 | struct fd f; | |
5204 | ||
5205 | f = fdget(fd); | |
5206 | if (!f.file) | |
5207 | return -EBADF; | |
5208 | ||
5209 | ret = -EOPNOTSUPP; | |
5210 | if (f.file->f_op != &io_uring_fops) | |
5211 | goto out_fput; | |
5212 | ||
5213 | ctx = f.file->private_data; | |
5214 | ||
5215 | mutex_lock(&ctx->uring_lock); | |
5216 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
5217 | mutex_unlock(&ctx->uring_lock); | |
c826bd7a DD |
5218 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, |
5219 | ctx->cq_ev_fd != NULL, ret); | |
edafccee JA |
5220 | out_fput: |
5221 | fdput(f); | |
5222 | return ret; | |
5223 | } | |
5224 | ||
2b188cc1 JA |
5225 | static int __init io_uring_init(void) |
5226 | { | |
5227 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC); | |
5228 | return 0; | |
5229 | }; | |
5230 | __initcall(io_uring_init); |