]>
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> | |
59 | #include <linux/workqueue.h> | |
6c271ce2 | 60 | #include <linux/kthread.h> |
2b188cc1 | 61 | #include <linux/blkdev.h> |
edafccee | 62 | #include <linux/bvec.h> |
2b188cc1 JA |
63 | #include <linux/net.h> |
64 | #include <net/sock.h> | |
65 | #include <net/af_unix.h> | |
6b06314c | 66 | #include <net/scm.h> |
2b188cc1 JA |
67 | #include <linux/anon_inodes.h> |
68 | #include <linux/sched/mm.h> | |
69 | #include <linux/uaccess.h> | |
70 | #include <linux/nospec.h> | |
edafccee JA |
71 | #include <linux/sizes.h> |
72 | #include <linux/hugetlb.h> | |
2b188cc1 JA |
73 | |
74 | #include <uapi/linux/io_uring.h> | |
75 | ||
76 | #include "internal.h" | |
77 | ||
5277deaa | 78 | #define IORING_MAX_ENTRIES 32768 |
33a107f0 | 79 | #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES) |
6b06314c | 80 | #define IORING_MAX_FIXED_FILES 1024 |
2b188cc1 JA |
81 | |
82 | struct io_uring { | |
83 | u32 head ____cacheline_aligned_in_smp; | |
84 | u32 tail ____cacheline_aligned_in_smp; | |
85 | }; | |
86 | ||
1e84b97b | 87 | /* |
75b28aff HV |
88 | * This data is shared with the application through the mmap at offsets |
89 | * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING. | |
1e84b97b SB |
90 | * |
91 | * The offsets to the member fields are published through struct | |
92 | * io_sqring_offsets when calling io_uring_setup. | |
93 | */ | |
75b28aff | 94 | struct io_rings { |
1e84b97b SB |
95 | /* |
96 | * Head and tail offsets into the ring; the offsets need to be | |
97 | * masked to get valid indices. | |
98 | * | |
75b28aff HV |
99 | * The kernel controls head of the sq ring and the tail of the cq ring, |
100 | * and the application controls tail of the sq ring and the head of the | |
101 | * cq ring. | |
1e84b97b | 102 | */ |
75b28aff | 103 | struct io_uring sq, cq; |
1e84b97b | 104 | /* |
75b28aff | 105 | * Bitmasks to apply to head and tail offsets (constant, equals |
1e84b97b SB |
106 | * ring_entries - 1) |
107 | */ | |
75b28aff HV |
108 | u32 sq_ring_mask, cq_ring_mask; |
109 | /* Ring sizes (constant, power of 2) */ | |
110 | u32 sq_ring_entries, cq_ring_entries; | |
1e84b97b SB |
111 | /* |
112 | * Number of invalid entries dropped by the kernel due to | |
113 | * invalid index stored in array | |
114 | * | |
115 | * Written by the kernel, shouldn't be modified by the | |
116 | * application (i.e. get number of "new events" by comparing to | |
117 | * cached value). | |
118 | * | |
119 | * After a new SQ head value was read by the application this | |
120 | * counter includes all submissions that were dropped reaching | |
121 | * the new SQ head (and possibly more). | |
122 | */ | |
75b28aff | 123 | u32 sq_dropped; |
1e84b97b SB |
124 | /* |
125 | * Runtime flags | |
126 | * | |
127 | * Written by the kernel, shouldn't be modified by the | |
128 | * application. | |
129 | * | |
130 | * The application needs a full memory barrier before checking | |
131 | * for IORING_SQ_NEED_WAKEUP after updating the sq tail. | |
132 | */ | |
75b28aff | 133 | u32 sq_flags; |
1e84b97b SB |
134 | /* |
135 | * Number of completion events lost because the queue was full; | |
136 | * this should be avoided by the application by making sure | |
137 | * there are not more requests pending thatn there is space in | |
138 | * the completion queue. | |
139 | * | |
140 | * Written by the kernel, shouldn't be modified by the | |
141 | * application (i.e. get number of "new events" by comparing to | |
142 | * cached value). | |
143 | * | |
144 | * As completion events come in out of order this counter is not | |
145 | * ordered with any other data. | |
146 | */ | |
75b28aff | 147 | u32 cq_overflow; |
1e84b97b SB |
148 | /* |
149 | * Ring buffer of completion events. | |
150 | * | |
151 | * The kernel writes completion events fresh every time they are | |
152 | * produced, so the application is allowed to modify pending | |
153 | * entries. | |
154 | */ | |
75b28aff | 155 | struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp; |
2b188cc1 JA |
156 | }; |
157 | ||
edafccee JA |
158 | struct io_mapped_ubuf { |
159 | u64 ubuf; | |
160 | size_t len; | |
161 | struct bio_vec *bvec; | |
162 | unsigned int nr_bvecs; | |
163 | }; | |
164 | ||
31b51510 JA |
165 | struct async_list { |
166 | spinlock_t lock; | |
167 | atomic_t cnt; | |
168 | struct list_head list; | |
169 | ||
170 | struct file *file; | |
6d5d5ac5 | 171 | off_t io_start; |
9310a7ba | 172 | size_t io_len; |
31b51510 JA |
173 | }; |
174 | ||
2b188cc1 JA |
175 | struct io_ring_ctx { |
176 | struct { | |
177 | struct percpu_ref refs; | |
178 | } ____cacheline_aligned_in_smp; | |
179 | ||
180 | struct { | |
181 | unsigned int flags; | |
182 | bool compat; | |
183 | bool account_mem; | |
184 | ||
75b28aff HV |
185 | /* |
186 | * Ring buffer of indices into array of io_uring_sqe, which is | |
187 | * mmapped by the application using the IORING_OFF_SQES offset. | |
188 | * | |
189 | * This indirection could e.g. be used to assign fixed | |
190 | * io_uring_sqe entries to operations and only submit them to | |
191 | * the queue when needed. | |
192 | * | |
193 | * The kernel modifies neither the indices array nor the entries | |
194 | * array. | |
195 | */ | |
196 | u32 *sq_array; | |
2b188cc1 JA |
197 | unsigned cached_sq_head; |
198 | unsigned sq_entries; | |
199 | unsigned sq_mask; | |
6c271ce2 | 200 | unsigned sq_thread_idle; |
498ccd9e | 201 | unsigned cached_sq_dropped; |
2b188cc1 | 202 | struct io_uring_sqe *sq_sqes; |
de0617e4 JA |
203 | |
204 | struct list_head defer_list; | |
5262f567 | 205 | struct list_head timeout_list; |
2b188cc1 JA |
206 | } ____cacheline_aligned_in_smp; |
207 | ||
208 | /* IO offload */ | |
54a91f3b | 209 | struct workqueue_struct *sqo_wq[2]; |
6c271ce2 | 210 | struct task_struct *sqo_thread; /* if using sq thread polling */ |
2b188cc1 | 211 | struct mm_struct *sqo_mm; |
6c271ce2 | 212 | wait_queue_head_t sqo_wait; |
a4c0b3de | 213 | struct completion sqo_thread_started; |
2b188cc1 JA |
214 | |
215 | struct { | |
2b188cc1 | 216 | unsigned cached_cq_tail; |
498ccd9e | 217 | atomic_t cached_cq_overflow; |
2b188cc1 JA |
218 | unsigned cq_entries; |
219 | unsigned cq_mask; | |
220 | struct wait_queue_head cq_wait; | |
221 | struct fasync_struct *cq_fasync; | |
9b402849 | 222 | struct eventfd_ctx *cq_ev_fd; |
5262f567 | 223 | atomic_t cq_timeouts; |
2b188cc1 JA |
224 | } ____cacheline_aligned_in_smp; |
225 | ||
75b28aff HV |
226 | struct io_rings *rings; |
227 | ||
6b06314c JA |
228 | /* |
229 | * If used, fixed file set. Writers must ensure that ->refs is dead, | |
230 | * readers must ensure that ->refs is alive as long as the file* is | |
231 | * used. Only updated through io_uring_register(2). | |
232 | */ | |
233 | struct file **user_files; | |
234 | unsigned nr_user_files; | |
235 | ||
edafccee JA |
236 | /* if used, fixed mapped user buffers */ |
237 | unsigned nr_user_bufs; | |
238 | struct io_mapped_ubuf *user_bufs; | |
239 | ||
2b188cc1 JA |
240 | struct user_struct *user; |
241 | ||
242 | struct completion ctx_done; | |
243 | ||
244 | struct { | |
245 | struct mutex uring_lock; | |
246 | wait_queue_head_t wait; | |
247 | } ____cacheline_aligned_in_smp; | |
248 | ||
249 | struct { | |
250 | spinlock_t completion_lock; | |
def596e9 JA |
251 | bool poll_multi_file; |
252 | /* | |
253 | * ->poll_list is protected by the ctx->uring_lock for | |
254 | * io_uring instances that don't use IORING_SETUP_SQPOLL. | |
255 | * For SQPOLL, only the single threaded io_sq_thread() will | |
256 | * manipulate the list, hence no extra locking is needed there. | |
257 | */ | |
258 | struct list_head poll_list; | |
221c5eb2 | 259 | struct list_head cancel_list; |
2b188cc1 JA |
260 | } ____cacheline_aligned_in_smp; |
261 | ||
31b51510 JA |
262 | struct async_list pending_async[2]; |
263 | ||
2b188cc1 JA |
264 | #if defined(CONFIG_UNIX) |
265 | struct socket *ring_sock; | |
266 | #endif | |
267 | }; | |
268 | ||
269 | struct sqe_submit { | |
270 | const struct io_uring_sqe *sqe; | |
271 | unsigned short index; | |
8776f3fa | 272 | u32 sequence; |
2b188cc1 | 273 | bool has_user; |
ba5290cc | 274 | bool in_async; |
6c271ce2 | 275 | bool needs_fixed_file; |
2b188cc1 JA |
276 | }; |
277 | ||
09bb8394 JA |
278 | /* |
279 | * First field must be the file pointer in all the | |
280 | * iocb unions! See also 'struct kiocb' in <linux/fs.h> | |
281 | */ | |
221c5eb2 JA |
282 | struct io_poll_iocb { |
283 | struct file *file; | |
284 | struct wait_queue_head *head; | |
285 | __poll_t events; | |
8c838788 | 286 | bool done; |
221c5eb2 JA |
287 | bool canceled; |
288 | struct wait_queue_entry wait; | |
289 | }; | |
290 | ||
5262f567 JA |
291 | struct io_timeout { |
292 | struct file *file; | |
293 | struct hrtimer timer; | |
294 | }; | |
295 | ||
09bb8394 JA |
296 | /* |
297 | * NOTE! Each of the iocb union members has the file pointer | |
298 | * as the first entry in their struct definition. So you can | |
299 | * access the file pointer through any of the sub-structs, | |
300 | * or directly as just 'ki_filp' in this struct. | |
301 | */ | |
2b188cc1 | 302 | struct io_kiocb { |
221c5eb2 | 303 | union { |
09bb8394 | 304 | struct file *file; |
221c5eb2 JA |
305 | struct kiocb rw; |
306 | struct io_poll_iocb poll; | |
5262f567 | 307 | struct io_timeout timeout; |
221c5eb2 | 308 | }; |
2b188cc1 JA |
309 | |
310 | struct sqe_submit submit; | |
311 | ||
312 | struct io_ring_ctx *ctx; | |
313 | struct list_head list; | |
9e645e11 | 314 | struct list_head link_list; |
2b188cc1 | 315 | unsigned int flags; |
c16361c1 | 316 | refcount_t refs; |
8449eeda | 317 | #define REQ_F_NOWAIT 1 /* must not punt to workers */ |
def596e9 | 318 | #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */ |
6b06314c | 319 | #define REQ_F_FIXED_FILE 4 /* ctx owns file */ |
31b51510 | 320 | #define REQ_F_SEQ_PREV 8 /* sequential with previous */ |
e2033e33 SB |
321 | #define REQ_F_IO_DRAIN 16 /* drain existing IO first */ |
322 | #define REQ_F_IO_DRAINED 32 /* drain done */ | |
9e645e11 | 323 | #define REQ_F_LINK 64 /* linked sqes */ |
f7b76ac9 ZL |
324 | #define REQ_F_LINK_DONE 128 /* linked sqes done */ |
325 | #define REQ_F_FAIL_LINK 256 /* fail rest of links */ | |
4fe2c963 | 326 | #define REQ_F_SHADOW_DRAIN 512 /* link-drain shadow req */ |
5262f567 | 327 | #define REQ_F_TIMEOUT 1024 /* timeout request */ |
491381ce JA |
328 | #define REQ_F_ISREG 2048 /* regular file */ |
329 | #define REQ_F_MUST_PUNT 4096 /* must be punted even for NONBLOCK */ | |
2b188cc1 | 330 | u64 user_data; |
9e645e11 | 331 | u32 result; |
de0617e4 | 332 | u32 sequence; |
2b188cc1 JA |
333 | |
334 | struct work_struct work; | |
335 | }; | |
336 | ||
337 | #define IO_PLUG_THRESHOLD 2 | |
def596e9 | 338 | #define IO_IOPOLL_BATCH 8 |
2b188cc1 | 339 | |
9a56a232 JA |
340 | struct io_submit_state { |
341 | struct blk_plug plug; | |
342 | ||
2579f913 JA |
343 | /* |
344 | * io_kiocb alloc cache | |
345 | */ | |
346 | void *reqs[IO_IOPOLL_BATCH]; | |
347 | unsigned int free_reqs; | |
348 | unsigned int cur_req; | |
349 | ||
9a56a232 JA |
350 | /* |
351 | * File reference cache | |
352 | */ | |
353 | struct file *file; | |
354 | unsigned int fd; | |
355 | unsigned int has_refs; | |
356 | unsigned int used_refs; | |
357 | unsigned int ios_left; | |
358 | }; | |
359 | ||
de0617e4 | 360 | static void io_sq_wq_submit_work(struct work_struct *work); |
5262f567 JA |
361 | static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data, |
362 | long res); | |
4fe2c963 | 363 | static void __io_free_req(struct io_kiocb *req); |
de0617e4 | 364 | |
2b188cc1 JA |
365 | static struct kmem_cache *req_cachep; |
366 | ||
367 | static const struct file_operations io_uring_fops; | |
368 | ||
369 | struct sock *io_uring_get_socket(struct file *file) | |
370 | { | |
371 | #if defined(CONFIG_UNIX) | |
372 | if (file->f_op == &io_uring_fops) { | |
373 | struct io_ring_ctx *ctx = file->private_data; | |
374 | ||
375 | return ctx->ring_sock->sk; | |
376 | } | |
377 | #endif | |
378 | return NULL; | |
379 | } | |
380 | EXPORT_SYMBOL(io_uring_get_socket); | |
381 | ||
382 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) | |
383 | { | |
384 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
385 | ||
386 | complete(&ctx->ctx_done); | |
387 | } | |
388 | ||
389 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) | |
390 | { | |
391 | struct io_ring_ctx *ctx; | |
31b51510 | 392 | int i; |
2b188cc1 JA |
393 | |
394 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
395 | if (!ctx) | |
396 | return NULL; | |
397 | ||
21482896 RG |
398 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
399 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) { | |
2b188cc1 JA |
400 | kfree(ctx); |
401 | return NULL; | |
402 | } | |
403 | ||
404 | ctx->flags = p->flags; | |
405 | init_waitqueue_head(&ctx->cq_wait); | |
406 | init_completion(&ctx->ctx_done); | |
a4c0b3de | 407 | init_completion(&ctx->sqo_thread_started); |
2b188cc1 JA |
408 | mutex_init(&ctx->uring_lock); |
409 | init_waitqueue_head(&ctx->wait); | |
31b51510 JA |
410 | for (i = 0; i < ARRAY_SIZE(ctx->pending_async); i++) { |
411 | spin_lock_init(&ctx->pending_async[i].lock); | |
412 | INIT_LIST_HEAD(&ctx->pending_async[i].list); | |
413 | atomic_set(&ctx->pending_async[i].cnt, 0); | |
414 | } | |
2b188cc1 | 415 | spin_lock_init(&ctx->completion_lock); |
def596e9 | 416 | INIT_LIST_HEAD(&ctx->poll_list); |
221c5eb2 | 417 | INIT_LIST_HEAD(&ctx->cancel_list); |
de0617e4 | 418 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 419 | INIT_LIST_HEAD(&ctx->timeout_list); |
2b188cc1 JA |
420 | return ctx; |
421 | } | |
422 | ||
7adf4eaf JA |
423 | static inline bool __io_sequence_defer(struct io_ring_ctx *ctx, |
424 | struct io_kiocb *req) | |
425 | { | |
498ccd9e JA |
426 | return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped |
427 | + atomic_read(&ctx->cached_cq_overflow); | |
7adf4eaf JA |
428 | } |
429 | ||
de0617e4 JA |
430 | static inline bool io_sequence_defer(struct io_ring_ctx *ctx, |
431 | struct io_kiocb *req) | |
432 | { | |
7adf4eaf | 433 | if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) != REQ_F_IO_DRAIN) |
de0617e4 JA |
434 | return false; |
435 | ||
7adf4eaf | 436 | return __io_sequence_defer(ctx, req); |
de0617e4 JA |
437 | } |
438 | ||
7adf4eaf | 439 | static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx) |
de0617e4 JA |
440 | { |
441 | struct io_kiocb *req; | |
442 | ||
7adf4eaf JA |
443 | req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list); |
444 | if (req && !io_sequence_defer(ctx, req)) { | |
de0617e4 JA |
445 | list_del_init(&req->list); |
446 | return req; | |
447 | } | |
448 | ||
449 | return NULL; | |
450 | } | |
451 | ||
5262f567 JA |
452 | static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx) |
453 | { | |
7adf4eaf JA |
454 | struct io_kiocb *req; |
455 | ||
456 | req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list); | |
457 | if (req && !__io_sequence_defer(ctx, req)) { | |
458 | list_del_init(&req->list); | |
459 | return req; | |
460 | } | |
461 | ||
462 | return NULL; | |
5262f567 JA |
463 | } |
464 | ||
de0617e4 | 465 | static void __io_commit_cqring(struct io_ring_ctx *ctx) |
2b188cc1 | 466 | { |
75b28aff | 467 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 468 | |
75b28aff | 469 | if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) { |
2b188cc1 | 470 | /* order cqe stores with ring update */ |
75b28aff | 471 | smp_store_release(&rings->cq.tail, ctx->cached_cq_tail); |
2b188cc1 | 472 | |
2b188cc1 JA |
473 | if (wq_has_sleeper(&ctx->cq_wait)) { |
474 | wake_up_interruptible(&ctx->cq_wait); | |
475 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
476 | } | |
477 | } | |
478 | } | |
479 | ||
18d9be1a JA |
480 | static inline void io_queue_async_work(struct io_ring_ctx *ctx, |
481 | struct io_kiocb *req) | |
482 | { | |
6cc47d1d | 483 | int rw = 0; |
54a91f3b | 484 | |
6cc47d1d JA |
485 | if (req->submit.sqe) { |
486 | switch (req->submit.sqe->opcode) { | |
487 | case IORING_OP_WRITEV: | |
488 | case IORING_OP_WRITE_FIXED: | |
489 | rw = !(req->rw.ki_flags & IOCB_DIRECT); | |
490 | break; | |
491 | } | |
54a91f3b JA |
492 | } |
493 | ||
494 | queue_work(ctx->sqo_wq[rw], &req->work); | |
18d9be1a JA |
495 | } |
496 | ||
5262f567 JA |
497 | static void io_kill_timeout(struct io_kiocb *req) |
498 | { | |
499 | int ret; | |
500 | ||
501 | ret = hrtimer_try_to_cancel(&req->timeout.timer); | |
502 | if (ret != -1) { | |
503 | atomic_inc(&req->ctx->cq_timeouts); | |
504 | list_del(&req->list); | |
505 | io_cqring_fill_event(req->ctx, req->user_data, 0); | |
506 | __io_free_req(req); | |
507 | } | |
508 | } | |
509 | ||
510 | static void io_kill_timeouts(struct io_ring_ctx *ctx) | |
511 | { | |
512 | struct io_kiocb *req, *tmp; | |
513 | ||
514 | spin_lock_irq(&ctx->completion_lock); | |
515 | list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list) | |
516 | io_kill_timeout(req); | |
517 | spin_unlock_irq(&ctx->completion_lock); | |
518 | } | |
519 | ||
de0617e4 JA |
520 | static void io_commit_cqring(struct io_ring_ctx *ctx) |
521 | { | |
522 | struct io_kiocb *req; | |
523 | ||
5262f567 JA |
524 | while ((req = io_get_timeout_req(ctx)) != NULL) |
525 | io_kill_timeout(req); | |
526 | ||
de0617e4 JA |
527 | __io_commit_cqring(ctx); |
528 | ||
529 | while ((req = io_get_deferred_req(ctx)) != NULL) { | |
4fe2c963 JL |
530 | if (req->flags & REQ_F_SHADOW_DRAIN) { |
531 | /* Just for drain, free it. */ | |
532 | __io_free_req(req); | |
533 | continue; | |
534 | } | |
de0617e4 | 535 | req->flags |= REQ_F_IO_DRAINED; |
18d9be1a | 536 | io_queue_async_work(ctx, req); |
de0617e4 JA |
537 | } |
538 | } | |
539 | ||
2b188cc1 JA |
540 | static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx) |
541 | { | |
75b28aff | 542 | struct io_rings *rings = ctx->rings; |
2b188cc1 JA |
543 | unsigned tail; |
544 | ||
545 | tail = ctx->cached_cq_tail; | |
115e12e5 SB |
546 | /* |
547 | * writes to the cq entry need to come after reading head; the | |
548 | * control dependency is enough as we're using WRITE_ONCE to | |
549 | * fill the cq entry | |
550 | */ | |
75b28aff | 551 | if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries) |
2b188cc1 JA |
552 | return NULL; |
553 | ||
554 | ctx->cached_cq_tail++; | |
75b28aff | 555 | return &rings->cqes[tail & ctx->cq_mask]; |
2b188cc1 JA |
556 | } |
557 | ||
558 | static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data, | |
c71ffb67 | 559 | long res) |
2b188cc1 JA |
560 | { |
561 | struct io_uring_cqe *cqe; | |
562 | ||
563 | /* | |
564 | * If we can't get a cq entry, userspace overflowed the | |
565 | * submission (by quite a lot). Increment the overflow count in | |
566 | * the ring. | |
567 | */ | |
568 | cqe = io_get_cqring(ctx); | |
569 | if (cqe) { | |
570 | WRITE_ONCE(cqe->user_data, ki_user_data); | |
571 | WRITE_ONCE(cqe->res, res); | |
c71ffb67 | 572 | WRITE_ONCE(cqe->flags, 0); |
2b188cc1 | 573 | } else { |
498ccd9e JA |
574 | WRITE_ONCE(ctx->rings->cq_overflow, |
575 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
2b188cc1 JA |
576 | } |
577 | } | |
578 | ||
8c838788 JA |
579 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
580 | { | |
581 | if (waitqueue_active(&ctx->wait)) | |
582 | wake_up(&ctx->wait); | |
583 | if (waitqueue_active(&ctx->sqo_wait)) | |
584 | wake_up(&ctx->sqo_wait); | |
9b402849 JA |
585 | if (ctx->cq_ev_fd) |
586 | eventfd_signal(ctx->cq_ev_fd, 1); | |
8c838788 JA |
587 | } |
588 | ||
589 | static void io_cqring_add_event(struct io_ring_ctx *ctx, u64 user_data, | |
c71ffb67 | 590 | long res) |
2b188cc1 JA |
591 | { |
592 | unsigned long flags; | |
593 | ||
594 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
c71ffb67 | 595 | io_cqring_fill_event(ctx, user_data, res); |
2b188cc1 JA |
596 | io_commit_cqring(ctx); |
597 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
598 | ||
8c838788 | 599 | io_cqring_ev_posted(ctx); |
2b188cc1 JA |
600 | } |
601 | ||
2579f913 JA |
602 | static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx, |
603 | struct io_submit_state *state) | |
2b188cc1 | 604 | { |
fd6fab2c | 605 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
2b188cc1 JA |
606 | struct io_kiocb *req; |
607 | ||
608 | if (!percpu_ref_tryget(&ctx->refs)) | |
609 | return NULL; | |
610 | ||
2579f913 | 611 | if (!state) { |
fd6fab2c | 612 | req = kmem_cache_alloc(req_cachep, gfp); |
2579f913 JA |
613 | if (unlikely(!req)) |
614 | goto out; | |
615 | } else if (!state->free_reqs) { | |
616 | size_t sz; | |
617 | int ret; | |
618 | ||
619 | sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs)); | |
fd6fab2c JA |
620 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs); |
621 | ||
622 | /* | |
623 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
624 | * retry single alloc to be on the safe side. | |
625 | */ | |
626 | if (unlikely(ret <= 0)) { | |
627 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
628 | if (!state->reqs[0]) | |
629 | goto out; | |
630 | ret = 1; | |
631 | } | |
2579f913 JA |
632 | state->free_reqs = ret - 1; |
633 | state->cur_req = 1; | |
634 | req = state->reqs[0]; | |
635 | } else { | |
636 | req = state->reqs[state->cur_req]; | |
637 | state->free_reqs--; | |
638 | state->cur_req++; | |
2b188cc1 JA |
639 | } |
640 | ||
60c112b0 | 641 | req->file = NULL; |
2579f913 JA |
642 | req->ctx = ctx; |
643 | req->flags = 0; | |
e65ef56d JA |
644 | /* one is dropped after submission, the other at completion */ |
645 | refcount_set(&req->refs, 2); | |
9e645e11 | 646 | req->result = 0; |
2579f913 JA |
647 | return req; |
648 | out: | |
6805b32e | 649 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
650 | return NULL; |
651 | } | |
652 | ||
def596e9 JA |
653 | static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr) |
654 | { | |
655 | if (*nr) { | |
656 | kmem_cache_free_bulk(req_cachep, *nr, reqs); | |
6805b32e | 657 | percpu_ref_put_many(&ctx->refs, *nr); |
def596e9 JA |
658 | *nr = 0; |
659 | } | |
660 | } | |
661 | ||
9e645e11 | 662 | static void __io_free_req(struct io_kiocb *req) |
2b188cc1 | 663 | { |
09bb8394 JA |
664 | if (req->file && !(req->flags & REQ_F_FIXED_FILE)) |
665 | fput(req->file); | |
6805b32e | 666 | percpu_ref_put(&req->ctx->refs); |
e65ef56d JA |
667 | kmem_cache_free(req_cachep, req); |
668 | } | |
669 | ||
ba816ad6 | 670 | static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr) |
9e645e11 JA |
671 | { |
672 | struct io_kiocb *nxt; | |
673 | ||
674 | /* | |
675 | * The list should never be empty when we are called here. But could | |
676 | * potentially happen if the chain is messed up, check to be on the | |
677 | * safe side. | |
678 | */ | |
679 | nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list); | |
680 | if (nxt) { | |
681 | list_del(&nxt->list); | |
682 | if (!list_empty(&req->link_list)) { | |
683 | INIT_LIST_HEAD(&nxt->link_list); | |
684 | list_splice(&req->link_list, &nxt->link_list); | |
685 | nxt->flags |= REQ_F_LINK; | |
686 | } | |
687 | ||
f7b76ac9 | 688 | nxt->flags |= REQ_F_LINK_DONE; |
ba816ad6 JA |
689 | /* |
690 | * If we're in async work, we can continue processing the chain | |
691 | * in this context instead of having to queue up new async work. | |
692 | */ | |
693 | if (nxtptr && current_work()) { | |
694 | *nxtptr = nxt; | |
695 | } else { | |
696 | INIT_WORK(&nxt->work, io_sq_wq_submit_work); | |
697 | io_queue_async_work(req->ctx, nxt); | |
698 | } | |
9e645e11 JA |
699 | } |
700 | } | |
701 | ||
702 | /* | |
703 | * Called if REQ_F_LINK is set, and we fail the head request | |
704 | */ | |
705 | static void io_fail_links(struct io_kiocb *req) | |
706 | { | |
707 | struct io_kiocb *link; | |
708 | ||
709 | while (!list_empty(&req->link_list)) { | |
710 | link = list_first_entry(&req->link_list, struct io_kiocb, list); | |
711 | list_del(&link->list); | |
712 | ||
713 | io_cqring_add_event(req->ctx, link->user_data, -ECANCELED); | |
714 | __io_free_req(link); | |
715 | } | |
716 | } | |
717 | ||
ba816ad6 | 718 | static void io_free_req(struct io_kiocb *req, struct io_kiocb **nxt) |
9e645e11 JA |
719 | { |
720 | /* | |
721 | * If LINK is set, we have dependent requests in this chain. If we | |
722 | * didn't fail this request, queue the first one up, moving any other | |
723 | * dependencies to the next request. In case of failure, fail the rest | |
724 | * of the chain. | |
725 | */ | |
726 | if (req->flags & REQ_F_LINK) { | |
727 | if (req->flags & REQ_F_FAIL_LINK) | |
728 | io_fail_links(req); | |
729 | else | |
ba816ad6 | 730 | io_req_link_next(req, nxt); |
9e645e11 JA |
731 | } |
732 | ||
733 | __io_free_req(req); | |
734 | } | |
735 | ||
ba816ad6 JA |
736 | /* |
737 | * Drop reference to request, return next in chain (if there is one) if this | |
738 | * was the last reference to this request. | |
739 | */ | |
740 | static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req) | |
e65ef56d | 741 | { |
ba816ad6 JA |
742 | struct io_kiocb *nxt = NULL; |
743 | ||
e65ef56d | 744 | if (refcount_dec_and_test(&req->refs)) |
ba816ad6 JA |
745 | io_free_req(req, &nxt); |
746 | ||
747 | return nxt; | |
748 | } | |
749 | ||
750 | static void io_put_req(struct io_kiocb *req, struct io_kiocb **nxtptr) | |
751 | { | |
752 | struct io_kiocb *nxt; | |
753 | ||
754 | nxt = io_put_req_find_next(req); | |
755 | if (nxt) { | |
756 | if (nxtptr) { | |
757 | *nxtptr = nxt; | |
758 | } else { | |
759 | INIT_WORK(&nxt->work, io_sq_wq_submit_work); | |
760 | io_queue_async_work(nxt->ctx, nxt); | |
761 | } | |
762 | } | |
2b188cc1 JA |
763 | } |
764 | ||
75b28aff | 765 | static unsigned io_cqring_events(struct io_rings *rings) |
a3a0e43f JA |
766 | { |
767 | /* See comment at the top of this file */ | |
768 | smp_rmb(); | |
75b28aff | 769 | return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head); |
a3a0e43f JA |
770 | } |
771 | ||
fb5ccc98 PB |
772 | static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) |
773 | { | |
774 | struct io_rings *rings = ctx->rings; | |
775 | ||
776 | /* make sure SQ entry isn't read before tail */ | |
777 | return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; | |
778 | } | |
779 | ||
def596e9 JA |
780 | /* |
781 | * Find and free completed poll iocbs | |
782 | */ | |
783 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
784 | struct list_head *done) | |
785 | { | |
786 | void *reqs[IO_IOPOLL_BATCH]; | |
787 | struct io_kiocb *req; | |
09bb8394 | 788 | int to_free; |
def596e9 | 789 | |
09bb8394 | 790 | to_free = 0; |
def596e9 JA |
791 | while (!list_empty(done)) { |
792 | req = list_first_entry(done, struct io_kiocb, list); | |
793 | list_del(&req->list); | |
794 | ||
9e645e11 | 795 | io_cqring_fill_event(ctx, req->user_data, req->result); |
def596e9 JA |
796 | (*nr_events)++; |
797 | ||
09bb8394 JA |
798 | if (refcount_dec_and_test(&req->refs)) { |
799 | /* If we're not using fixed files, we have to pair the | |
800 | * completion part with the file put. Use regular | |
801 | * completions for those, only batch free for fixed | |
9e645e11 | 802 | * file and non-linked commands. |
09bb8394 | 803 | */ |
9e645e11 JA |
804 | if ((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) == |
805 | REQ_F_FIXED_FILE) { | |
09bb8394 JA |
806 | reqs[to_free++] = req; |
807 | if (to_free == ARRAY_SIZE(reqs)) | |
808 | io_free_req_many(ctx, reqs, &to_free); | |
6b06314c | 809 | } else { |
ba816ad6 | 810 | io_free_req(req, NULL); |
6b06314c | 811 | } |
9a56a232 | 812 | } |
def596e9 | 813 | } |
def596e9 | 814 | |
09bb8394 | 815 | io_commit_cqring(ctx); |
def596e9 JA |
816 | io_free_req_many(ctx, reqs, &to_free); |
817 | } | |
818 | ||
819 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
820 | long min) | |
821 | { | |
822 | struct io_kiocb *req, *tmp; | |
823 | LIST_HEAD(done); | |
824 | bool spin; | |
825 | int ret; | |
826 | ||
827 | /* | |
828 | * Only spin for completions if we don't have multiple devices hanging | |
829 | * off our complete list, and we're under the requested amount. | |
830 | */ | |
831 | spin = !ctx->poll_multi_file && *nr_events < min; | |
832 | ||
833 | ret = 0; | |
834 | list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) { | |
835 | struct kiocb *kiocb = &req->rw; | |
836 | ||
837 | /* | |
838 | * Move completed entries to our local list. If we find a | |
839 | * request that requires polling, break out and complete | |
840 | * the done list first, if we have entries there. | |
841 | */ | |
842 | if (req->flags & REQ_F_IOPOLL_COMPLETED) { | |
843 | list_move_tail(&req->list, &done); | |
844 | continue; | |
845 | } | |
846 | if (!list_empty(&done)) | |
847 | break; | |
848 | ||
849 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
850 | if (ret < 0) | |
851 | break; | |
852 | ||
853 | if (ret && spin) | |
854 | spin = false; | |
855 | ret = 0; | |
856 | } | |
857 | ||
858 | if (!list_empty(&done)) | |
859 | io_iopoll_complete(ctx, nr_events, &done); | |
860 | ||
861 | return ret; | |
862 | } | |
863 | ||
864 | /* | |
865 | * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a | |
866 | * non-spinning poll check - we'll still enter the driver poll loop, but only | |
867 | * as a non-spinning completion check. | |
868 | */ | |
869 | static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
870 | long min) | |
871 | { | |
08f5439f | 872 | while (!list_empty(&ctx->poll_list) && !need_resched()) { |
def596e9 JA |
873 | int ret; |
874 | ||
875 | ret = io_do_iopoll(ctx, nr_events, min); | |
876 | if (ret < 0) | |
877 | return ret; | |
878 | if (!min || *nr_events >= min) | |
879 | return 0; | |
880 | } | |
881 | ||
882 | return 1; | |
883 | } | |
884 | ||
885 | /* | |
886 | * We can't just wait for polled events to come to us, we have to actively | |
887 | * find and complete them. | |
888 | */ | |
889 | static void io_iopoll_reap_events(struct io_ring_ctx *ctx) | |
890 | { | |
891 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
892 | return; | |
893 | ||
894 | mutex_lock(&ctx->uring_lock); | |
895 | while (!list_empty(&ctx->poll_list)) { | |
896 | unsigned int nr_events = 0; | |
897 | ||
898 | io_iopoll_getevents(ctx, &nr_events, 1); | |
08f5439f JA |
899 | |
900 | /* | |
901 | * Ensure we allow local-to-the-cpu processing to take place, | |
902 | * in this case we need to ensure that we reap all events. | |
903 | */ | |
904 | cond_resched(); | |
def596e9 JA |
905 | } |
906 | mutex_unlock(&ctx->uring_lock); | |
907 | } | |
908 | ||
2b2ed975 JA |
909 | static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, |
910 | long min) | |
def596e9 | 911 | { |
2b2ed975 | 912 | int iters = 0, ret = 0; |
500f9fba | 913 | |
def596e9 JA |
914 | do { |
915 | int tmin = 0; | |
916 | ||
a3a0e43f JA |
917 | /* |
918 | * Don't enter poll loop if we already have events pending. | |
919 | * If we do, we can potentially be spinning for commands that | |
920 | * already triggered a CQE (eg in error). | |
921 | */ | |
75b28aff | 922 | if (io_cqring_events(ctx->rings)) |
a3a0e43f JA |
923 | break; |
924 | ||
500f9fba JA |
925 | /* |
926 | * If a submit got punted to a workqueue, we can have the | |
927 | * application entering polling for a command before it gets | |
928 | * issued. That app will hold the uring_lock for the duration | |
929 | * of the poll right here, so we need to take a breather every | |
930 | * now and then to ensure that the issue has a chance to add | |
931 | * the poll to the issued list. Otherwise we can spin here | |
932 | * forever, while the workqueue is stuck trying to acquire the | |
933 | * very same mutex. | |
934 | */ | |
935 | if (!(++iters & 7)) { | |
936 | mutex_unlock(&ctx->uring_lock); | |
937 | mutex_lock(&ctx->uring_lock); | |
938 | } | |
939 | ||
def596e9 JA |
940 | if (*nr_events < min) |
941 | tmin = min - *nr_events; | |
942 | ||
943 | ret = io_iopoll_getevents(ctx, nr_events, tmin); | |
944 | if (ret <= 0) | |
945 | break; | |
946 | ret = 0; | |
947 | } while (min && !*nr_events && !need_resched()); | |
948 | ||
2b2ed975 JA |
949 | return ret; |
950 | } | |
951 | ||
952 | static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, | |
953 | long min) | |
954 | { | |
955 | int ret; | |
956 | ||
957 | /* | |
958 | * We disallow the app entering submit/complete with polling, but we | |
959 | * still need to lock the ring to prevent racing with polled issue | |
960 | * that got punted to a workqueue. | |
961 | */ | |
962 | mutex_lock(&ctx->uring_lock); | |
963 | ret = __io_iopoll_check(ctx, nr_events, min); | |
500f9fba | 964 | mutex_unlock(&ctx->uring_lock); |
def596e9 JA |
965 | return ret; |
966 | } | |
967 | ||
491381ce | 968 | static void kiocb_end_write(struct io_kiocb *req) |
2b188cc1 | 969 | { |
491381ce JA |
970 | /* |
971 | * Tell lockdep we inherited freeze protection from submission | |
972 | * thread. | |
973 | */ | |
974 | if (req->flags & REQ_F_ISREG) { | |
975 | struct inode *inode = file_inode(req->file); | |
2b188cc1 | 976 | |
491381ce | 977 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); |
2b188cc1 | 978 | } |
491381ce | 979 | file_end_write(req->file); |
2b188cc1 JA |
980 | } |
981 | ||
ba816ad6 | 982 | static void io_complete_rw_common(struct kiocb *kiocb, long res) |
2b188cc1 JA |
983 | { |
984 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
985 | ||
491381ce JA |
986 | if (kiocb->ki_flags & IOCB_WRITE) |
987 | kiocb_end_write(req); | |
2b188cc1 | 988 | |
9e645e11 JA |
989 | if ((req->flags & REQ_F_LINK) && res != req->result) |
990 | req->flags |= REQ_F_FAIL_LINK; | |
c71ffb67 | 991 | io_cqring_add_event(req->ctx, req->user_data, res); |
ba816ad6 JA |
992 | } |
993 | ||
994 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
995 | { | |
996 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
997 | ||
998 | io_complete_rw_common(kiocb, res); | |
999 | io_put_req(req, NULL); | |
1000 | } | |
1001 | ||
1002 | static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res) | |
1003 | { | |
1004 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
1005 | ||
1006 | io_complete_rw_common(kiocb, res); | |
1007 | return io_put_req_find_next(req); | |
2b188cc1 JA |
1008 | } |
1009 | ||
def596e9 JA |
1010 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
1011 | { | |
1012 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
1013 | ||
491381ce JA |
1014 | if (kiocb->ki_flags & IOCB_WRITE) |
1015 | kiocb_end_write(req); | |
def596e9 | 1016 | |
9e645e11 JA |
1017 | if ((req->flags & REQ_F_LINK) && res != req->result) |
1018 | req->flags |= REQ_F_FAIL_LINK; | |
1019 | req->result = res; | |
def596e9 JA |
1020 | if (res != -EAGAIN) |
1021 | req->flags |= REQ_F_IOPOLL_COMPLETED; | |
1022 | } | |
1023 | ||
1024 | /* | |
1025 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1026 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
1027 | * find it from a io_iopoll_getevents() thread before the issuer is done | |
1028 | * accessing the kiocb cookie. | |
1029 | */ | |
1030 | static void io_iopoll_req_issued(struct io_kiocb *req) | |
1031 | { | |
1032 | struct io_ring_ctx *ctx = req->ctx; | |
1033 | ||
1034 | /* | |
1035 | * Track whether we have multiple files in our lists. This will impact | |
1036 | * how we do polling eventually, not spinning if we're on potentially | |
1037 | * different devices. | |
1038 | */ | |
1039 | if (list_empty(&ctx->poll_list)) { | |
1040 | ctx->poll_multi_file = false; | |
1041 | } else if (!ctx->poll_multi_file) { | |
1042 | struct io_kiocb *list_req; | |
1043 | ||
1044 | list_req = list_first_entry(&ctx->poll_list, struct io_kiocb, | |
1045 | list); | |
1046 | if (list_req->rw.ki_filp != req->rw.ki_filp) | |
1047 | ctx->poll_multi_file = true; | |
1048 | } | |
1049 | ||
1050 | /* | |
1051 | * For fast devices, IO may have already completed. If it has, add | |
1052 | * it to the front so we find it first. | |
1053 | */ | |
1054 | if (req->flags & REQ_F_IOPOLL_COMPLETED) | |
1055 | list_add(&req->list, &ctx->poll_list); | |
1056 | else | |
1057 | list_add_tail(&req->list, &ctx->poll_list); | |
1058 | } | |
1059 | ||
3d6770fb | 1060 | static void io_file_put(struct io_submit_state *state) |
9a56a232 | 1061 | { |
3d6770fb | 1062 | if (state->file) { |
9a56a232 JA |
1063 | int diff = state->has_refs - state->used_refs; |
1064 | ||
1065 | if (diff) | |
1066 | fput_many(state->file, diff); | |
1067 | state->file = NULL; | |
1068 | } | |
1069 | } | |
1070 | ||
1071 | /* | |
1072 | * Get as many references to a file as we have IOs left in this submission, | |
1073 | * assuming most submissions are for one file, or at least that each file | |
1074 | * has more than one submission. | |
1075 | */ | |
1076 | static struct file *io_file_get(struct io_submit_state *state, int fd) | |
1077 | { | |
1078 | if (!state) | |
1079 | return fget(fd); | |
1080 | ||
1081 | if (state->file) { | |
1082 | if (state->fd == fd) { | |
1083 | state->used_refs++; | |
1084 | state->ios_left--; | |
1085 | return state->file; | |
1086 | } | |
3d6770fb | 1087 | io_file_put(state); |
9a56a232 JA |
1088 | } |
1089 | state->file = fget_many(fd, state->ios_left); | |
1090 | if (!state->file) | |
1091 | return NULL; | |
1092 | ||
1093 | state->fd = fd; | |
1094 | state->has_refs = state->ios_left; | |
1095 | state->used_refs = 1; | |
1096 | state->ios_left--; | |
1097 | return state->file; | |
1098 | } | |
1099 | ||
2b188cc1 JA |
1100 | /* |
1101 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1102 | * any file. For now, just ensure that anything potentially problematic is done | |
1103 | * inline. | |
1104 | */ | |
1105 | static bool io_file_supports_async(struct file *file) | |
1106 | { | |
1107 | umode_t mode = file_inode(file)->i_mode; | |
1108 | ||
1109 | if (S_ISBLK(mode) || S_ISCHR(mode)) | |
1110 | return true; | |
1111 | if (S_ISREG(mode) && file->f_op != &io_uring_fops) | |
1112 | return true; | |
1113 | ||
1114 | return false; | |
1115 | } | |
1116 | ||
6c271ce2 | 1117 | static int io_prep_rw(struct io_kiocb *req, const struct sqe_submit *s, |
8358e3a8 | 1118 | bool force_nonblock) |
2b188cc1 | 1119 | { |
6c271ce2 | 1120 | const struct io_uring_sqe *sqe = s->sqe; |
def596e9 | 1121 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 | 1122 | struct kiocb *kiocb = &req->rw; |
09bb8394 JA |
1123 | unsigned ioprio; |
1124 | int ret; | |
2b188cc1 | 1125 | |
09bb8394 JA |
1126 | if (!req->file) |
1127 | return -EBADF; | |
2b188cc1 | 1128 | |
491381ce JA |
1129 | if (S_ISREG(file_inode(req->file)->i_mode)) |
1130 | req->flags |= REQ_F_ISREG; | |
1131 | ||
1132 | /* | |
1133 | * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so | |
1134 | * we know to async punt it even if it was opened O_NONBLOCK | |
1135 | */ | |
1136 | if (force_nonblock && !io_file_supports_async(req->file)) { | |
1137 | req->flags |= REQ_F_MUST_PUNT; | |
1138 | return -EAGAIN; | |
1139 | } | |
6b06314c | 1140 | |
2b188cc1 JA |
1141 | kiocb->ki_pos = READ_ONCE(sqe->off); |
1142 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); | |
1143 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); | |
1144 | ||
1145 | ioprio = READ_ONCE(sqe->ioprio); | |
1146 | if (ioprio) { | |
1147 | ret = ioprio_check_cap(ioprio); | |
1148 | if (ret) | |
09bb8394 | 1149 | return ret; |
2b188cc1 JA |
1150 | |
1151 | kiocb->ki_ioprio = ioprio; | |
1152 | } else | |
1153 | kiocb->ki_ioprio = get_current_ioprio(); | |
1154 | ||
1155 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
1156 | if (unlikely(ret)) | |
09bb8394 | 1157 | return ret; |
8449eeda SB |
1158 | |
1159 | /* don't allow async punt if RWF_NOWAIT was requested */ | |
491381ce JA |
1160 | if ((kiocb->ki_flags & IOCB_NOWAIT) || |
1161 | (req->file->f_flags & O_NONBLOCK)) | |
8449eeda SB |
1162 | req->flags |= REQ_F_NOWAIT; |
1163 | ||
1164 | if (force_nonblock) | |
2b188cc1 | 1165 | kiocb->ki_flags |= IOCB_NOWAIT; |
8449eeda | 1166 | |
def596e9 | 1167 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
1168 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
1169 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 1170 | return -EOPNOTSUPP; |
2b188cc1 | 1171 | |
def596e9 JA |
1172 | kiocb->ki_flags |= IOCB_HIPRI; |
1173 | kiocb->ki_complete = io_complete_rw_iopoll; | |
1174 | } else { | |
09bb8394 JA |
1175 | if (kiocb->ki_flags & IOCB_HIPRI) |
1176 | return -EINVAL; | |
def596e9 JA |
1177 | kiocb->ki_complete = io_complete_rw; |
1178 | } | |
2b188cc1 | 1179 | return 0; |
2b188cc1 JA |
1180 | } |
1181 | ||
1182 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
1183 | { | |
1184 | switch (ret) { | |
1185 | case -EIOCBQUEUED: | |
1186 | break; | |
1187 | case -ERESTARTSYS: | |
1188 | case -ERESTARTNOINTR: | |
1189 | case -ERESTARTNOHAND: | |
1190 | case -ERESTART_RESTARTBLOCK: | |
1191 | /* | |
1192 | * We can't just restart the syscall, since previously | |
1193 | * submitted sqes may already be in progress. Just fail this | |
1194 | * IO with EINTR. | |
1195 | */ | |
1196 | ret = -EINTR; | |
1197 | /* fall through */ | |
1198 | default: | |
1199 | kiocb->ki_complete(kiocb, ret, 0); | |
1200 | } | |
1201 | } | |
1202 | ||
ba816ad6 JA |
1203 | static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt, |
1204 | bool in_async) | |
1205 | { | |
1206 | if (in_async && ret >= 0 && nxt && kiocb->ki_complete == io_complete_rw) | |
1207 | *nxt = __io_complete_rw(kiocb, ret); | |
1208 | else | |
1209 | io_rw_done(kiocb, ret); | |
1210 | } | |
1211 | ||
edafccee JA |
1212 | static int io_import_fixed(struct io_ring_ctx *ctx, int rw, |
1213 | const struct io_uring_sqe *sqe, | |
1214 | struct iov_iter *iter) | |
1215 | { | |
1216 | size_t len = READ_ONCE(sqe->len); | |
1217 | struct io_mapped_ubuf *imu; | |
1218 | unsigned index, buf_index; | |
1219 | size_t offset; | |
1220 | u64 buf_addr; | |
1221 | ||
1222 | /* attempt to use fixed buffers without having provided iovecs */ | |
1223 | if (unlikely(!ctx->user_bufs)) | |
1224 | return -EFAULT; | |
1225 | ||
1226 | buf_index = READ_ONCE(sqe->buf_index); | |
1227 | if (unlikely(buf_index >= ctx->nr_user_bufs)) | |
1228 | return -EFAULT; | |
1229 | ||
1230 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); | |
1231 | imu = &ctx->user_bufs[index]; | |
1232 | buf_addr = READ_ONCE(sqe->addr); | |
1233 | ||
1234 | /* overflow */ | |
1235 | if (buf_addr + len < buf_addr) | |
1236 | return -EFAULT; | |
1237 | /* not inside the mapped region */ | |
1238 | if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len) | |
1239 | return -EFAULT; | |
1240 | ||
1241 | /* | |
1242 | * May not be a start of buffer, set size appropriately | |
1243 | * and advance us to the beginning. | |
1244 | */ | |
1245 | offset = buf_addr - imu->ubuf; | |
1246 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
bd11b3a3 JA |
1247 | |
1248 | if (offset) { | |
1249 | /* | |
1250 | * Don't use iov_iter_advance() here, as it's really slow for | |
1251 | * using the latter parts of a big fixed buffer - it iterates | |
1252 | * over each segment manually. We can cheat a bit here, because | |
1253 | * we know that: | |
1254 | * | |
1255 | * 1) it's a BVEC iter, we set it up | |
1256 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
1257 | * first and last bvec | |
1258 | * | |
1259 | * So just find our index, and adjust the iterator afterwards. | |
1260 | * If the offset is within the first bvec (or the whole first | |
1261 | * bvec, just use iov_iter_advance(). This makes it easier | |
1262 | * since we can just skip the first segment, which may not | |
1263 | * be PAGE_SIZE aligned. | |
1264 | */ | |
1265 | const struct bio_vec *bvec = imu->bvec; | |
1266 | ||
1267 | if (offset <= bvec->bv_len) { | |
1268 | iov_iter_advance(iter, offset); | |
1269 | } else { | |
1270 | unsigned long seg_skip; | |
1271 | ||
1272 | /* skip first vec */ | |
1273 | offset -= bvec->bv_len; | |
1274 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
1275 | ||
1276 | iter->bvec = bvec + seg_skip; | |
1277 | iter->nr_segs -= seg_skip; | |
99c79f66 | 1278 | iter->count -= bvec->bv_len + offset; |
bd11b3a3 | 1279 | iter->iov_offset = offset & ~PAGE_MASK; |
bd11b3a3 JA |
1280 | } |
1281 | } | |
1282 | ||
edafccee JA |
1283 | return 0; |
1284 | } | |
1285 | ||
87e5e6da JA |
1286 | static ssize_t io_import_iovec(struct io_ring_ctx *ctx, int rw, |
1287 | const struct sqe_submit *s, struct iovec **iovec, | |
1288 | struct iov_iter *iter) | |
2b188cc1 JA |
1289 | { |
1290 | const struct io_uring_sqe *sqe = s->sqe; | |
1291 | void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
1292 | size_t sqe_len = READ_ONCE(sqe->len); | |
edafccee JA |
1293 | u8 opcode; |
1294 | ||
1295 | /* | |
1296 | * We're reading ->opcode for the second time, but the first read | |
1297 | * doesn't care whether it's _FIXED or not, so it doesn't matter | |
1298 | * whether ->opcode changes concurrently. The first read does care | |
1299 | * about whether it is a READ or a WRITE, so we don't trust this read | |
1300 | * for that purpose and instead let the caller pass in the read/write | |
1301 | * flag. | |
1302 | */ | |
1303 | opcode = READ_ONCE(sqe->opcode); | |
1304 | if (opcode == IORING_OP_READ_FIXED || | |
1305 | opcode == IORING_OP_WRITE_FIXED) { | |
87e5e6da | 1306 | ssize_t ret = io_import_fixed(ctx, rw, sqe, iter); |
edafccee JA |
1307 | *iovec = NULL; |
1308 | return ret; | |
1309 | } | |
2b188cc1 JA |
1310 | |
1311 | if (!s->has_user) | |
1312 | return -EFAULT; | |
1313 | ||
1314 | #ifdef CONFIG_COMPAT | |
1315 | if (ctx->compat) | |
1316 | return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV, | |
1317 | iovec, iter); | |
1318 | #endif | |
1319 | ||
1320 | return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter); | |
1321 | } | |
1322 | ||
6d5d5ac5 JA |
1323 | static inline bool io_should_merge(struct async_list *al, struct kiocb *kiocb) |
1324 | { | |
1325 | if (al->file == kiocb->ki_filp) { | |
1326 | off_t start, end; | |
1327 | ||
1328 | /* | |
1329 | * Allow merging if we're anywhere in the range of the same | |
1330 | * page. Generally this happens for sub-page reads or writes, | |
1331 | * and it's beneficial to allow the first worker to bring the | |
1332 | * page in and the piggy backed work can then work on the | |
1333 | * cached page. | |
1334 | */ | |
1335 | start = al->io_start & PAGE_MASK; | |
1336 | end = (al->io_start + al->io_len + PAGE_SIZE - 1) & PAGE_MASK; | |
1337 | if (kiocb->ki_pos >= start && kiocb->ki_pos <= end) | |
1338 | return true; | |
1339 | } | |
1340 | ||
1341 | al->file = NULL; | |
1342 | return false; | |
1343 | } | |
1344 | ||
31b51510 JA |
1345 | /* |
1346 | * Make a note of the last file/offset/direction we punted to async | |
1347 | * context. We'll use this information to see if we can piggy back a | |
1348 | * sequential request onto the previous one, if it's still hasn't been | |
1349 | * completed by the async worker. | |
1350 | */ | |
1351 | static void io_async_list_note(int rw, struct io_kiocb *req, size_t len) | |
1352 | { | |
1353 | struct async_list *async_list = &req->ctx->pending_async[rw]; | |
1354 | struct kiocb *kiocb = &req->rw; | |
1355 | struct file *filp = kiocb->ki_filp; | |
31b51510 | 1356 | |
6d5d5ac5 | 1357 | if (io_should_merge(async_list, kiocb)) { |
9310a7ba | 1358 | unsigned long max_bytes; |
31b51510 JA |
1359 | |
1360 | /* Use 8x RA size as a decent limiter for both reads/writes */ | |
9310a7ba ZL |
1361 | max_bytes = filp->f_ra.ra_pages << (PAGE_SHIFT + 3); |
1362 | if (!max_bytes) | |
1363 | max_bytes = VM_READAHEAD_PAGES << (PAGE_SHIFT + 3); | |
1364 | ||
1365 | /* If max len are exceeded, reset the state */ | |
1366 | if (async_list->io_len + len <= max_bytes) { | |
31b51510 | 1367 | req->flags |= REQ_F_SEQ_PREV; |
9310a7ba | 1368 | async_list->io_len += len; |
31b51510 | 1369 | } else { |
6d5d5ac5 | 1370 | async_list->file = NULL; |
31b51510 JA |
1371 | } |
1372 | } | |
1373 | ||
1374 | /* New file? Reset state. */ | |
1375 | if (async_list->file != filp) { | |
6d5d5ac5 JA |
1376 | async_list->io_start = kiocb->ki_pos; |
1377 | async_list->io_len = len; | |
31b51510 JA |
1378 | async_list->file = filp; |
1379 | } | |
31b51510 JA |
1380 | } |
1381 | ||
32960613 JA |
1382 | /* |
1383 | * For files that don't have ->read_iter() and ->write_iter(), handle them | |
1384 | * by looping over ->read() or ->write() manually. | |
1385 | */ | |
1386 | static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb, | |
1387 | struct iov_iter *iter) | |
1388 | { | |
1389 | ssize_t ret = 0; | |
1390 | ||
1391 | /* | |
1392 | * Don't support polled IO through this interface, and we can't | |
1393 | * support non-blocking either. For the latter, this just causes | |
1394 | * the kiocb to be handled from an async context. | |
1395 | */ | |
1396 | if (kiocb->ki_flags & IOCB_HIPRI) | |
1397 | return -EOPNOTSUPP; | |
1398 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
1399 | return -EAGAIN; | |
1400 | ||
1401 | while (iov_iter_count(iter)) { | |
1402 | struct iovec iovec = iov_iter_iovec(iter); | |
1403 | ssize_t nr; | |
1404 | ||
1405 | if (rw == READ) { | |
1406 | nr = file->f_op->read(file, iovec.iov_base, | |
1407 | iovec.iov_len, &kiocb->ki_pos); | |
1408 | } else { | |
1409 | nr = file->f_op->write(file, iovec.iov_base, | |
1410 | iovec.iov_len, &kiocb->ki_pos); | |
1411 | } | |
1412 | ||
1413 | if (nr < 0) { | |
1414 | if (!ret) | |
1415 | ret = nr; | |
1416 | break; | |
1417 | } | |
1418 | ret += nr; | |
1419 | if (nr != iovec.iov_len) | |
1420 | break; | |
1421 | iov_iter_advance(iter, nr); | |
1422 | } | |
1423 | ||
1424 | return ret; | |
1425 | } | |
1426 | ||
e0c5c576 | 1427 | static int io_read(struct io_kiocb *req, const struct sqe_submit *s, |
ba816ad6 | 1428 | struct io_kiocb **nxt, bool force_nonblock) |
2b188cc1 JA |
1429 | { |
1430 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1431 | struct kiocb *kiocb = &req->rw; | |
1432 | struct iov_iter iter; | |
1433 | struct file *file; | |
31b51510 | 1434 | size_t iov_count; |
9d93a3f5 | 1435 | ssize_t read_size, ret; |
2b188cc1 | 1436 | |
8358e3a8 | 1437 | ret = io_prep_rw(req, s, force_nonblock); |
2b188cc1 JA |
1438 | if (ret) |
1439 | return ret; | |
1440 | file = kiocb->ki_filp; | |
1441 | ||
2b188cc1 | 1442 | if (unlikely(!(file->f_mode & FMODE_READ))) |
09bb8394 | 1443 | return -EBADF; |
2b188cc1 JA |
1444 | |
1445 | ret = io_import_iovec(req->ctx, READ, s, &iovec, &iter); | |
87e5e6da | 1446 | if (ret < 0) |
09bb8394 | 1447 | return ret; |
2b188cc1 | 1448 | |
9d93a3f5 | 1449 | read_size = ret; |
9e645e11 JA |
1450 | if (req->flags & REQ_F_LINK) |
1451 | req->result = read_size; | |
1452 | ||
31b51510 JA |
1453 | iov_count = iov_iter_count(&iter); |
1454 | ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 JA |
1455 | if (!ret) { |
1456 | ssize_t ret2; | |
1457 | ||
32960613 JA |
1458 | if (file->f_op->read_iter) |
1459 | ret2 = call_read_iter(file, kiocb, &iter); | |
1460 | else | |
1461 | ret2 = loop_rw_iter(READ, file, kiocb, &iter); | |
1462 | ||
9d93a3f5 JA |
1463 | /* |
1464 | * In case of a short read, punt to async. This can happen | |
1465 | * if we have data partially cached. Alternatively we can | |
1466 | * return the short read, in which case the application will | |
1467 | * need to issue another SQE and wait for it. That SQE will | |
1468 | * need async punt anyway, so it's more efficient to do it | |
1469 | * here. | |
1470 | */ | |
491381ce JA |
1471 | if (force_nonblock && !(req->flags & REQ_F_NOWAIT) && |
1472 | (req->flags & REQ_F_ISREG) && | |
1473 | ret2 > 0 && ret2 < read_size) | |
9d93a3f5 JA |
1474 | ret2 = -EAGAIN; |
1475 | /* Catch -EAGAIN return for forced non-blocking submission */ | |
31b51510 | 1476 | if (!force_nonblock || ret2 != -EAGAIN) { |
ba5290cc | 1477 | kiocb_done(kiocb, ret2, nxt, s->in_async); |
31b51510 | 1478 | } else { |
ba5290cc | 1479 | if (!s->in_async) |
31b51510 | 1480 | io_async_list_note(READ, req, iov_count); |
2b188cc1 | 1481 | ret = -EAGAIN; |
31b51510 | 1482 | } |
2b188cc1 JA |
1483 | } |
1484 | kfree(iovec); | |
2b188cc1 JA |
1485 | return ret; |
1486 | } | |
1487 | ||
e0c5c576 | 1488 | static int io_write(struct io_kiocb *req, const struct sqe_submit *s, |
ba816ad6 | 1489 | struct io_kiocb **nxt, bool force_nonblock) |
2b188cc1 JA |
1490 | { |
1491 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1492 | struct kiocb *kiocb = &req->rw; | |
1493 | struct iov_iter iter; | |
1494 | struct file *file; | |
31b51510 | 1495 | size_t iov_count; |
87e5e6da | 1496 | ssize_t ret; |
2b188cc1 | 1497 | |
8358e3a8 | 1498 | ret = io_prep_rw(req, s, force_nonblock); |
2b188cc1 JA |
1499 | if (ret) |
1500 | return ret; | |
2b188cc1 | 1501 | |
2b188cc1 JA |
1502 | file = kiocb->ki_filp; |
1503 | if (unlikely(!(file->f_mode & FMODE_WRITE))) | |
09bb8394 | 1504 | return -EBADF; |
2b188cc1 JA |
1505 | |
1506 | ret = io_import_iovec(req->ctx, WRITE, s, &iovec, &iter); | |
87e5e6da | 1507 | if (ret < 0) |
09bb8394 | 1508 | return ret; |
2b188cc1 | 1509 | |
9e645e11 JA |
1510 | if (req->flags & REQ_F_LINK) |
1511 | req->result = ret; | |
1512 | ||
31b51510 JA |
1513 | iov_count = iov_iter_count(&iter); |
1514 | ||
1515 | ret = -EAGAIN; | |
1516 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT)) { | |
ba5290cc | 1517 | if (!s->in_async) |
31b51510 JA |
1518 | io_async_list_note(WRITE, req, iov_count); |
1519 | goto out_free; | |
1520 | } | |
1521 | ||
1522 | ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 | 1523 | if (!ret) { |
9bf7933f RP |
1524 | ssize_t ret2; |
1525 | ||
2b188cc1 JA |
1526 | /* |
1527 | * Open-code file_start_write here to grab freeze protection, | |
1528 | * which will be released by another thread in | |
1529 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
1530 | * released so that it doesn't complain about the held lock when | |
1531 | * we return to userspace. | |
1532 | */ | |
491381ce | 1533 | if (req->flags & REQ_F_ISREG) { |
2b188cc1 JA |
1534 | __sb_start_write(file_inode(file)->i_sb, |
1535 | SB_FREEZE_WRITE, true); | |
1536 | __sb_writers_release(file_inode(file)->i_sb, | |
1537 | SB_FREEZE_WRITE); | |
1538 | } | |
1539 | kiocb->ki_flags |= IOCB_WRITE; | |
9bf7933f | 1540 | |
32960613 JA |
1541 | if (file->f_op->write_iter) |
1542 | ret2 = call_write_iter(file, kiocb, &iter); | |
1543 | else | |
1544 | ret2 = loop_rw_iter(WRITE, file, kiocb, &iter); | |
9bf7933f | 1545 | if (!force_nonblock || ret2 != -EAGAIN) { |
ba5290cc | 1546 | kiocb_done(kiocb, ret2, nxt, s->in_async); |
9bf7933f | 1547 | } else { |
ba5290cc | 1548 | if (!s->in_async) |
9bf7933f RP |
1549 | io_async_list_note(WRITE, req, iov_count); |
1550 | ret = -EAGAIN; | |
1551 | } | |
2b188cc1 | 1552 | } |
31b51510 | 1553 | out_free: |
2b188cc1 | 1554 | kfree(iovec); |
2b188cc1 JA |
1555 | return ret; |
1556 | } | |
1557 | ||
1558 | /* | |
1559 | * IORING_OP_NOP just posts a completion event, nothing else. | |
1560 | */ | |
1561 | static int io_nop(struct io_kiocb *req, u64 user_data) | |
1562 | { | |
1563 | struct io_ring_ctx *ctx = req->ctx; | |
1564 | long err = 0; | |
1565 | ||
def596e9 JA |
1566 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
1567 | return -EINVAL; | |
1568 | ||
c71ffb67 | 1569 | io_cqring_add_event(ctx, user_data, err); |
ba816ad6 | 1570 | io_put_req(req, NULL); |
2b188cc1 JA |
1571 | return 0; |
1572 | } | |
1573 | ||
c992fe29 CH |
1574 | static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1575 | { | |
6b06314c | 1576 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 1577 | |
09bb8394 JA |
1578 | if (!req->file) |
1579 | return -EBADF; | |
c992fe29 | 1580 | |
6b06314c | 1581 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 1582 | return -EINVAL; |
edafccee | 1583 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
1584 | return -EINVAL; |
1585 | ||
c992fe29 CH |
1586 | return 0; |
1587 | } | |
1588 | ||
1589 | static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
ba816ad6 | 1590 | struct io_kiocb **nxt, bool force_nonblock) |
c992fe29 CH |
1591 | { |
1592 | loff_t sqe_off = READ_ONCE(sqe->off); | |
1593 | loff_t sqe_len = READ_ONCE(sqe->len); | |
1594 | loff_t end = sqe_off + sqe_len; | |
1595 | unsigned fsync_flags; | |
1596 | int ret; | |
1597 | ||
1598 | fsync_flags = READ_ONCE(sqe->fsync_flags); | |
1599 | if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC)) | |
1600 | return -EINVAL; | |
1601 | ||
1602 | ret = io_prep_fsync(req, sqe); | |
1603 | if (ret) | |
1604 | return ret; | |
1605 | ||
1606 | /* fsync always requires a blocking context */ | |
1607 | if (force_nonblock) | |
1608 | return -EAGAIN; | |
1609 | ||
1610 | ret = vfs_fsync_range(req->rw.ki_filp, sqe_off, | |
1611 | end > 0 ? end : LLONG_MAX, | |
1612 | fsync_flags & IORING_FSYNC_DATASYNC); | |
1613 | ||
9e645e11 JA |
1614 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1615 | req->flags |= REQ_F_FAIL_LINK; | |
c71ffb67 | 1616 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
ba816ad6 | 1617 | io_put_req(req, nxt); |
c992fe29 CH |
1618 | return 0; |
1619 | } | |
1620 | ||
5d17b4a4 JA |
1621 | static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1622 | { | |
1623 | struct io_ring_ctx *ctx = req->ctx; | |
1624 | int ret = 0; | |
1625 | ||
1626 | if (!req->file) | |
1627 | return -EBADF; | |
5d17b4a4 JA |
1628 | |
1629 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
1630 | return -EINVAL; | |
1631 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) | |
1632 | return -EINVAL; | |
1633 | ||
5d17b4a4 JA |
1634 | return ret; |
1635 | } | |
1636 | ||
1637 | static int io_sync_file_range(struct io_kiocb *req, | |
1638 | const struct io_uring_sqe *sqe, | |
ba816ad6 | 1639 | struct io_kiocb **nxt, |
5d17b4a4 JA |
1640 | bool force_nonblock) |
1641 | { | |
1642 | loff_t sqe_off; | |
1643 | loff_t sqe_len; | |
1644 | unsigned flags; | |
1645 | int ret; | |
1646 | ||
1647 | ret = io_prep_sfr(req, sqe); | |
1648 | if (ret) | |
1649 | return ret; | |
1650 | ||
1651 | /* sync_file_range always requires a blocking context */ | |
1652 | if (force_nonblock) | |
1653 | return -EAGAIN; | |
1654 | ||
1655 | sqe_off = READ_ONCE(sqe->off); | |
1656 | sqe_len = READ_ONCE(sqe->len); | |
1657 | flags = READ_ONCE(sqe->sync_range_flags); | |
1658 | ||
1659 | ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags); | |
1660 | ||
9e645e11 JA |
1661 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1662 | req->flags |= REQ_F_FAIL_LINK; | |
c71ffb67 | 1663 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
ba816ad6 | 1664 | io_put_req(req, nxt); |
5d17b4a4 JA |
1665 | return 0; |
1666 | } | |
1667 | ||
0fa03c62 | 1668 | #if defined(CONFIG_NET) |
aa1fa28f | 1669 | static int io_send_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
ba816ad6 | 1670 | struct io_kiocb **nxt, bool force_nonblock, |
aa1fa28f JA |
1671 | long (*fn)(struct socket *, struct user_msghdr __user *, |
1672 | unsigned int)) | |
1673 | { | |
0fa03c62 JA |
1674 | struct socket *sock; |
1675 | int ret; | |
1676 | ||
1677 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1678 | return -EINVAL; | |
1679 | ||
1680 | sock = sock_from_file(req->file, &ret); | |
1681 | if (sock) { | |
1682 | struct user_msghdr __user *msg; | |
1683 | unsigned flags; | |
1684 | ||
1685 | flags = READ_ONCE(sqe->msg_flags); | |
1686 | if (flags & MSG_DONTWAIT) | |
1687 | req->flags |= REQ_F_NOWAIT; | |
1688 | else if (force_nonblock) | |
1689 | flags |= MSG_DONTWAIT; | |
1690 | ||
1691 | msg = (struct user_msghdr __user *) (unsigned long) | |
1692 | READ_ONCE(sqe->addr); | |
1693 | ||
aa1fa28f | 1694 | ret = fn(sock, msg, flags); |
0fa03c62 JA |
1695 | if (force_nonblock && ret == -EAGAIN) |
1696 | return ret; | |
1697 | } | |
1698 | ||
c71ffb67 | 1699 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
ba816ad6 | 1700 | io_put_req(req, nxt); |
5d17b4a4 JA |
1701 | return 0; |
1702 | } | |
aa1fa28f JA |
1703 | #endif |
1704 | ||
1705 | static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
ba816ad6 | 1706 | struct io_kiocb **nxt, bool force_nonblock) |
aa1fa28f JA |
1707 | { |
1708 | #if defined(CONFIG_NET) | |
ba816ad6 JA |
1709 | return io_send_recvmsg(req, sqe, nxt, force_nonblock, |
1710 | __sys_sendmsg_sock); | |
aa1fa28f JA |
1711 | #else |
1712 | return -EOPNOTSUPP; | |
1713 | #endif | |
1714 | } | |
1715 | ||
1716 | static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
ba816ad6 | 1717 | struct io_kiocb **nxt, bool force_nonblock) |
aa1fa28f JA |
1718 | { |
1719 | #if defined(CONFIG_NET) | |
ba816ad6 JA |
1720 | return io_send_recvmsg(req, sqe, nxt, force_nonblock, |
1721 | __sys_recvmsg_sock); | |
0fa03c62 JA |
1722 | #else |
1723 | return -EOPNOTSUPP; | |
1724 | #endif | |
1725 | } | |
5d17b4a4 | 1726 | |
221c5eb2 JA |
1727 | static void io_poll_remove_one(struct io_kiocb *req) |
1728 | { | |
1729 | struct io_poll_iocb *poll = &req->poll; | |
1730 | ||
1731 | spin_lock(&poll->head->lock); | |
1732 | WRITE_ONCE(poll->canceled, true); | |
1733 | if (!list_empty(&poll->wait.entry)) { | |
1734 | list_del_init(&poll->wait.entry); | |
18d9be1a | 1735 | io_queue_async_work(req->ctx, req); |
221c5eb2 JA |
1736 | } |
1737 | spin_unlock(&poll->head->lock); | |
1738 | ||
1739 | list_del_init(&req->list); | |
1740 | } | |
1741 | ||
1742 | static void io_poll_remove_all(struct io_ring_ctx *ctx) | |
1743 | { | |
1744 | struct io_kiocb *req; | |
1745 | ||
1746 | spin_lock_irq(&ctx->completion_lock); | |
1747 | while (!list_empty(&ctx->cancel_list)) { | |
1748 | req = list_first_entry(&ctx->cancel_list, struct io_kiocb,list); | |
1749 | io_poll_remove_one(req); | |
1750 | } | |
1751 | spin_unlock_irq(&ctx->completion_lock); | |
1752 | } | |
1753 | ||
1754 | /* | |
1755 | * Find a running poll command that matches one specified in sqe->addr, | |
1756 | * and remove it if found. | |
1757 | */ | |
1758 | static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
1759 | { | |
1760 | struct io_ring_ctx *ctx = req->ctx; | |
1761 | struct io_kiocb *poll_req, *next; | |
1762 | int ret = -ENOENT; | |
1763 | ||
1764 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1765 | return -EINVAL; | |
1766 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index || | |
1767 | sqe->poll_events) | |
1768 | return -EINVAL; | |
1769 | ||
1770 | spin_lock_irq(&ctx->completion_lock); | |
1771 | list_for_each_entry_safe(poll_req, next, &ctx->cancel_list, list) { | |
1772 | if (READ_ONCE(sqe->addr) == poll_req->user_data) { | |
1773 | io_poll_remove_one(poll_req); | |
1774 | ret = 0; | |
1775 | break; | |
1776 | } | |
1777 | } | |
1778 | spin_unlock_irq(&ctx->completion_lock); | |
1779 | ||
c71ffb67 | 1780 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
ba816ad6 | 1781 | io_put_req(req, NULL); |
221c5eb2 JA |
1782 | return 0; |
1783 | } | |
1784 | ||
8c838788 JA |
1785 | static void io_poll_complete(struct io_ring_ctx *ctx, struct io_kiocb *req, |
1786 | __poll_t mask) | |
221c5eb2 | 1787 | { |
8c838788 | 1788 | req->poll.done = true; |
c71ffb67 | 1789 | io_cqring_fill_event(ctx, req->user_data, mangle_poll(mask)); |
8c838788 | 1790 | io_commit_cqring(ctx); |
221c5eb2 JA |
1791 | } |
1792 | ||
1793 | static void io_poll_complete_work(struct work_struct *work) | |
1794 | { | |
1795 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
1796 | struct io_poll_iocb *poll = &req->poll; | |
1797 | struct poll_table_struct pt = { ._key = poll->events }; | |
1798 | struct io_ring_ctx *ctx = req->ctx; | |
1799 | __poll_t mask = 0; | |
1800 | ||
1801 | if (!READ_ONCE(poll->canceled)) | |
1802 | mask = vfs_poll(poll->file, &pt) & poll->events; | |
1803 | ||
1804 | /* | |
1805 | * Note that ->ki_cancel callers also delete iocb from active_reqs after | |
1806 | * calling ->ki_cancel. We need the ctx_lock roundtrip here to | |
1807 | * synchronize with them. In the cancellation case the list_del_init | |
1808 | * itself is not actually needed, but harmless so we keep it in to | |
1809 | * avoid further branches in the fast path. | |
1810 | */ | |
1811 | spin_lock_irq(&ctx->completion_lock); | |
1812 | if (!mask && !READ_ONCE(poll->canceled)) { | |
1813 | add_wait_queue(poll->head, &poll->wait); | |
1814 | spin_unlock_irq(&ctx->completion_lock); | |
1815 | return; | |
1816 | } | |
1817 | list_del_init(&req->list); | |
8c838788 | 1818 | io_poll_complete(ctx, req, mask); |
221c5eb2 JA |
1819 | spin_unlock_irq(&ctx->completion_lock); |
1820 | ||
8c838788 | 1821 | io_cqring_ev_posted(ctx); |
ba816ad6 | 1822 | io_put_req(req, NULL); |
221c5eb2 JA |
1823 | } |
1824 | ||
1825 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
1826 | void *key) | |
1827 | { | |
1828 | struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb, | |
1829 | wait); | |
1830 | struct io_kiocb *req = container_of(poll, struct io_kiocb, poll); | |
1831 | struct io_ring_ctx *ctx = req->ctx; | |
1832 | __poll_t mask = key_to_poll(key); | |
8c838788 | 1833 | unsigned long flags; |
221c5eb2 JA |
1834 | |
1835 | /* for instances that support it check for an event match first: */ | |
8c838788 JA |
1836 | if (mask && !(mask & poll->events)) |
1837 | return 0; | |
221c5eb2 | 1838 | |
8c838788 | 1839 | list_del_init(&poll->wait.entry); |
221c5eb2 | 1840 | |
8c838788 JA |
1841 | if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) { |
1842 | list_del(&req->list); | |
1843 | io_poll_complete(ctx, req, mask); | |
1844 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
221c5eb2 | 1845 | |
8c838788 | 1846 | io_cqring_ev_posted(ctx); |
ba816ad6 | 1847 | io_put_req(req, NULL); |
8c838788 | 1848 | } else { |
18d9be1a | 1849 | io_queue_async_work(ctx, req); |
221c5eb2 JA |
1850 | } |
1851 | ||
221c5eb2 JA |
1852 | return 1; |
1853 | } | |
1854 | ||
1855 | struct io_poll_table { | |
1856 | struct poll_table_struct pt; | |
1857 | struct io_kiocb *req; | |
1858 | int error; | |
1859 | }; | |
1860 | ||
1861 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
1862 | struct poll_table_struct *p) | |
1863 | { | |
1864 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
1865 | ||
1866 | if (unlikely(pt->req->poll.head)) { | |
1867 | pt->error = -EINVAL; | |
1868 | return; | |
1869 | } | |
1870 | ||
1871 | pt->error = 0; | |
1872 | pt->req->poll.head = head; | |
1873 | add_wait_queue(head, &pt->req->poll.wait); | |
1874 | } | |
1875 | ||
1876 | static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
1877 | { | |
1878 | struct io_poll_iocb *poll = &req->poll; | |
1879 | struct io_ring_ctx *ctx = req->ctx; | |
1880 | struct io_poll_table ipt; | |
8c838788 | 1881 | bool cancel = false; |
221c5eb2 JA |
1882 | __poll_t mask; |
1883 | u16 events; | |
221c5eb2 JA |
1884 | |
1885 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1886 | return -EINVAL; | |
1887 | if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
1888 | return -EINVAL; | |
09bb8394 JA |
1889 | if (!poll->file) |
1890 | return -EBADF; | |
221c5eb2 | 1891 | |
6cc47d1d | 1892 | req->submit.sqe = NULL; |
221c5eb2 JA |
1893 | INIT_WORK(&req->work, io_poll_complete_work); |
1894 | events = READ_ONCE(sqe->poll_events); | |
1895 | poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP; | |
1896 | ||
221c5eb2 | 1897 | poll->head = NULL; |
8c838788 | 1898 | poll->done = false; |
221c5eb2 JA |
1899 | poll->canceled = false; |
1900 | ||
1901 | ipt.pt._qproc = io_poll_queue_proc; | |
1902 | ipt.pt._key = poll->events; | |
1903 | ipt.req = req; | |
1904 | ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */ | |
1905 | ||
1906 | /* initialized the list so that we can do list_empty checks */ | |
1907 | INIT_LIST_HEAD(&poll->wait.entry); | |
1908 | init_waitqueue_func_entry(&poll->wait, io_poll_wake); | |
1909 | ||
36703247 JA |
1910 | INIT_LIST_HEAD(&req->list); |
1911 | ||
221c5eb2 | 1912 | mask = vfs_poll(poll->file, &ipt.pt) & poll->events; |
221c5eb2 JA |
1913 | |
1914 | spin_lock_irq(&ctx->completion_lock); | |
8c838788 JA |
1915 | if (likely(poll->head)) { |
1916 | spin_lock(&poll->head->lock); | |
1917 | if (unlikely(list_empty(&poll->wait.entry))) { | |
1918 | if (ipt.error) | |
1919 | cancel = true; | |
1920 | ipt.error = 0; | |
1921 | mask = 0; | |
1922 | } | |
1923 | if (mask || ipt.error) | |
1924 | list_del_init(&poll->wait.entry); | |
1925 | else if (cancel) | |
1926 | WRITE_ONCE(poll->canceled, true); | |
1927 | else if (!poll->done) /* actually waiting for an event */ | |
1928 | list_add_tail(&req->list, &ctx->cancel_list); | |
1929 | spin_unlock(&poll->head->lock); | |
1930 | } | |
1931 | if (mask) { /* no async, we'd stolen it */ | |
221c5eb2 | 1932 | ipt.error = 0; |
8c838788 | 1933 | io_poll_complete(ctx, req, mask); |
221c5eb2 | 1934 | } |
221c5eb2 JA |
1935 | spin_unlock_irq(&ctx->completion_lock); |
1936 | ||
8c838788 JA |
1937 | if (mask) { |
1938 | io_cqring_ev_posted(ctx); | |
ba816ad6 | 1939 | io_put_req(req, NULL); |
221c5eb2 | 1940 | } |
8c838788 | 1941 | return ipt.error; |
221c5eb2 JA |
1942 | } |
1943 | ||
5262f567 JA |
1944 | static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer) |
1945 | { | |
1946 | struct io_ring_ctx *ctx; | |
11365043 | 1947 | struct io_kiocb *req; |
5262f567 | 1948 | unsigned long flags; |
11365043 | 1949 | bool comp; |
5262f567 JA |
1950 | |
1951 | req = container_of(timer, struct io_kiocb, timeout.timer); | |
1952 | ctx = req->ctx; | |
1953 | atomic_inc(&ctx->cq_timeouts); | |
1954 | ||
1955 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
ef03681a | 1956 | /* |
11365043 JA |
1957 | * We could be racing with timeout deletion. If the list is empty, |
1958 | * then timeout lookup already found it and will be handling it. | |
ef03681a | 1959 | */ |
11365043 JA |
1960 | comp = !list_empty(&req->list); |
1961 | if (comp) { | |
1962 | struct io_kiocb *prev; | |
5262f567 | 1963 | |
11365043 JA |
1964 | /* |
1965 | * Adjust the reqs sequence before the current one because it | |
1966 | * will consume a slot in the cq_ring and the the cq_tail | |
1967 | * pointer will be increased, otherwise other timeout reqs may | |
1968 | * return in advance without waiting for enough wait_nr. | |
1969 | */ | |
1970 | prev = req; | |
1971 | list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list) | |
1972 | prev->sequence++; | |
1973 | ||
1974 | list_del_init(&req->list); | |
1975 | io_cqring_fill_event(ctx, req->user_data, -ETIME); | |
1976 | io_commit_cqring(ctx); | |
1977 | } | |
5262f567 JA |
1978 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
1979 | ||
11365043 JA |
1980 | if (comp) { |
1981 | io_cqring_ev_posted(ctx); | |
1982 | io_put_req(req, NULL); | |
1983 | } | |
1984 | return HRTIMER_NORESTART; | |
1985 | } | |
1986 | ||
1987 | /* | |
1988 | * Remove or update an existing timeout command | |
1989 | */ | |
1990 | static int io_timeout_remove(struct io_kiocb *req, | |
1991 | const struct io_uring_sqe *sqe) | |
1992 | { | |
1993 | struct io_ring_ctx *ctx = req->ctx; | |
1994 | struct io_kiocb *treq; | |
1995 | int ret = -ENOENT; | |
1996 | __u64 user_data; | |
1997 | unsigned flags; | |
1998 | ||
1999 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
2000 | return -EINVAL; | |
2001 | if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len) | |
2002 | return -EINVAL; | |
2003 | flags = READ_ONCE(sqe->timeout_flags); | |
2004 | if (flags) | |
2005 | return -EINVAL; | |
2006 | ||
2007 | user_data = READ_ONCE(sqe->addr); | |
2008 | spin_lock_irq(&ctx->completion_lock); | |
2009 | list_for_each_entry(treq, &ctx->timeout_list, list) { | |
2010 | if (user_data == treq->user_data) { | |
2011 | list_del_init(&treq->list); | |
2012 | ret = 0; | |
2013 | break; | |
2014 | } | |
2015 | } | |
2016 | ||
2017 | /* didn't find timeout */ | |
2018 | if (ret) { | |
2019 | fill_ev: | |
2020 | io_cqring_fill_event(ctx, req->user_data, ret); | |
2021 | io_commit_cqring(ctx); | |
2022 | spin_unlock_irq(&ctx->completion_lock); | |
2023 | io_cqring_ev_posted(ctx); | |
2024 | io_put_req(req, NULL); | |
2025 | return 0; | |
2026 | } | |
2027 | ||
2028 | ret = hrtimer_try_to_cancel(&treq->timeout.timer); | |
2029 | if (ret == -1) { | |
2030 | ret = -EBUSY; | |
2031 | goto fill_ev; | |
2032 | } | |
2033 | ||
2034 | io_cqring_fill_event(ctx, req->user_data, 0); | |
2035 | io_cqring_fill_event(ctx, treq->user_data, -ECANCELED); | |
2036 | io_commit_cqring(ctx); | |
2037 | spin_unlock_irq(&ctx->completion_lock); | |
5262f567 JA |
2038 | io_cqring_ev_posted(ctx); |
2039 | ||
11365043 | 2040 | io_put_req(treq, NULL); |
ba816ad6 | 2041 | io_put_req(req, NULL); |
11365043 | 2042 | return 0; |
5262f567 JA |
2043 | } |
2044 | ||
2045 | static int io_timeout(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
2046 | { | |
5da0fb1a | 2047 | unsigned count; |
5262f567 JA |
2048 | struct io_ring_ctx *ctx = req->ctx; |
2049 | struct list_head *entry; | |
a41525ab | 2050 | enum hrtimer_mode mode; |
bdf20073 | 2051 | struct timespec64 ts; |
a1f58ba4 | 2052 | unsigned span = 0; |
a41525ab | 2053 | unsigned flags; |
5262f567 JA |
2054 | |
2055 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
2056 | return -EINVAL; | |
a41525ab JA |
2057 | if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len != 1) |
2058 | return -EINVAL; | |
2059 | flags = READ_ONCE(sqe->timeout_flags); | |
2060 | if (flags & ~IORING_TIMEOUT_ABS) | |
5262f567 | 2061 | return -EINVAL; |
bdf20073 AB |
2062 | |
2063 | if (get_timespec64(&ts, u64_to_user_ptr(sqe->addr))) | |
5262f567 JA |
2064 | return -EFAULT; |
2065 | ||
11365043 JA |
2066 | if (flags & IORING_TIMEOUT_ABS) |
2067 | mode = HRTIMER_MODE_ABS; | |
2068 | else | |
2069 | mode = HRTIMER_MODE_REL; | |
2070 | ||
2071 | hrtimer_init(&req->timeout.timer, CLOCK_MONOTONIC, mode); | |
2072 | ||
5262f567 JA |
2073 | /* |
2074 | * sqe->off holds how many events that need to occur for this | |
2075 | * timeout event to be satisfied. | |
2076 | */ | |
2077 | count = READ_ONCE(sqe->off); | |
2078 | if (!count) | |
2079 | count = 1; | |
2080 | ||
2081 | req->sequence = ctx->cached_sq_head + count - 1; | |
5da0fb1a | 2082 | /* reuse it to store the count */ |
2083 | req->submit.sequence = count; | |
5262f567 JA |
2084 | req->flags |= REQ_F_TIMEOUT; |
2085 | ||
2086 | /* | |
2087 | * Insertion sort, ensuring the first entry in the list is always | |
2088 | * the one we need first. | |
2089 | */ | |
5262f567 JA |
2090 | spin_lock_irq(&ctx->completion_lock); |
2091 | list_for_each_prev(entry, &ctx->timeout_list) { | |
2092 | struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list); | |
5da0fb1a | 2093 | unsigned nxt_sq_head; |
2094 | long long tmp, tmp_nxt; | |
5262f567 | 2095 | |
5da0fb1a | 2096 | /* |
2097 | * Since cached_sq_head + count - 1 can overflow, use type long | |
2098 | * long to store it. | |
2099 | */ | |
2100 | tmp = (long long)ctx->cached_sq_head + count - 1; | |
2101 | nxt_sq_head = nxt->sequence - nxt->submit.sequence + 1; | |
2102 | tmp_nxt = (long long)nxt_sq_head + nxt->submit.sequence - 1; | |
2103 | ||
2104 | /* | |
2105 | * cached_sq_head may overflow, and it will never overflow twice | |
2106 | * once there is some timeout req still be valid. | |
2107 | */ | |
2108 | if (ctx->cached_sq_head < nxt_sq_head) | |
8b07a65a | 2109 | tmp += UINT_MAX; |
5da0fb1a | 2110 | |
a1f58ba4 | 2111 | if (tmp > tmp_nxt) |
5262f567 | 2112 | break; |
a1f58ba4 | 2113 | |
2114 | /* | |
2115 | * Sequence of reqs after the insert one and itself should | |
2116 | * be adjusted because each timeout req consumes a slot. | |
2117 | */ | |
2118 | span++; | |
2119 | nxt->sequence++; | |
5262f567 | 2120 | } |
a1f58ba4 | 2121 | req->sequence -= span; |
5262f567 JA |
2122 | list_add(&req->list, entry); |
2123 | spin_unlock_irq(&ctx->completion_lock); | |
5262f567 | 2124 | req->timeout.timer.function = io_timeout_fn; |
a41525ab | 2125 | hrtimer_start(&req->timeout.timer, timespec64_to_ktime(ts), mode); |
5262f567 JA |
2126 | return 0; |
2127 | } | |
2128 | ||
de0617e4 JA |
2129 | static int io_req_defer(struct io_ring_ctx *ctx, struct io_kiocb *req, |
2130 | const struct io_uring_sqe *sqe) | |
2131 | { | |
2132 | struct io_uring_sqe *sqe_copy; | |
2133 | ||
2134 | if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) | |
2135 | return 0; | |
2136 | ||
2137 | sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL); | |
2138 | if (!sqe_copy) | |
2139 | return -EAGAIN; | |
2140 | ||
2141 | spin_lock_irq(&ctx->completion_lock); | |
2142 | if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) { | |
2143 | spin_unlock_irq(&ctx->completion_lock); | |
2144 | kfree(sqe_copy); | |
2145 | return 0; | |
2146 | } | |
2147 | ||
2148 | memcpy(sqe_copy, sqe, sizeof(*sqe_copy)); | |
2149 | req->submit.sqe = sqe_copy; | |
2150 | ||
2151 | INIT_WORK(&req->work, io_sq_wq_submit_work); | |
2152 | list_add_tail(&req->list, &ctx->defer_list); | |
2153 | spin_unlock_irq(&ctx->completion_lock); | |
2154 | return -EIOCBQUEUED; | |
2155 | } | |
2156 | ||
2b188cc1 | 2157 | static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
ba816ad6 JA |
2158 | const struct sqe_submit *s, struct io_kiocb **nxt, |
2159 | bool force_nonblock) | |
2b188cc1 | 2160 | { |
e0c5c576 | 2161 | int ret, opcode; |
2b188cc1 | 2162 | |
9e645e11 JA |
2163 | req->user_data = READ_ONCE(s->sqe->user_data); |
2164 | ||
2b188cc1 JA |
2165 | if (unlikely(s->index >= ctx->sq_entries)) |
2166 | return -EINVAL; | |
2b188cc1 JA |
2167 | |
2168 | opcode = READ_ONCE(s->sqe->opcode); | |
2169 | switch (opcode) { | |
2170 | case IORING_OP_NOP: | |
2171 | ret = io_nop(req, req->user_data); | |
2172 | break; | |
2173 | case IORING_OP_READV: | |
edafccee JA |
2174 | if (unlikely(s->sqe->buf_index)) |
2175 | return -EINVAL; | |
ba816ad6 | 2176 | ret = io_read(req, s, nxt, force_nonblock); |
2b188cc1 JA |
2177 | break; |
2178 | case IORING_OP_WRITEV: | |
edafccee JA |
2179 | if (unlikely(s->sqe->buf_index)) |
2180 | return -EINVAL; | |
ba816ad6 | 2181 | ret = io_write(req, s, nxt, force_nonblock); |
edafccee JA |
2182 | break; |
2183 | case IORING_OP_READ_FIXED: | |
ba816ad6 | 2184 | ret = io_read(req, s, nxt, force_nonblock); |
edafccee JA |
2185 | break; |
2186 | case IORING_OP_WRITE_FIXED: | |
ba816ad6 | 2187 | ret = io_write(req, s, nxt, force_nonblock); |
2b188cc1 | 2188 | break; |
c992fe29 | 2189 | case IORING_OP_FSYNC: |
ba816ad6 | 2190 | ret = io_fsync(req, s->sqe, nxt, force_nonblock); |
c992fe29 | 2191 | break; |
221c5eb2 JA |
2192 | case IORING_OP_POLL_ADD: |
2193 | ret = io_poll_add(req, s->sqe); | |
2194 | break; | |
2195 | case IORING_OP_POLL_REMOVE: | |
2196 | ret = io_poll_remove(req, s->sqe); | |
2197 | break; | |
5d17b4a4 | 2198 | case IORING_OP_SYNC_FILE_RANGE: |
ba816ad6 | 2199 | ret = io_sync_file_range(req, s->sqe, nxt, force_nonblock); |
5d17b4a4 | 2200 | break; |
0fa03c62 | 2201 | case IORING_OP_SENDMSG: |
ba816ad6 | 2202 | ret = io_sendmsg(req, s->sqe, nxt, force_nonblock); |
0fa03c62 | 2203 | break; |
aa1fa28f | 2204 | case IORING_OP_RECVMSG: |
ba816ad6 | 2205 | ret = io_recvmsg(req, s->sqe, nxt, force_nonblock); |
aa1fa28f | 2206 | break; |
5262f567 JA |
2207 | case IORING_OP_TIMEOUT: |
2208 | ret = io_timeout(req, s->sqe); | |
2209 | break; | |
11365043 JA |
2210 | case IORING_OP_TIMEOUT_REMOVE: |
2211 | ret = io_timeout_remove(req, s->sqe); | |
2212 | break; | |
2b188cc1 JA |
2213 | default: |
2214 | ret = -EINVAL; | |
2215 | break; | |
2216 | } | |
2217 | ||
def596e9 JA |
2218 | if (ret) |
2219 | return ret; | |
2220 | ||
2221 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
9e645e11 | 2222 | if (req->result == -EAGAIN) |
def596e9 JA |
2223 | return -EAGAIN; |
2224 | ||
2225 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
ba5290cc | 2226 | if (s->in_async) |
def596e9 JA |
2227 | mutex_lock(&ctx->uring_lock); |
2228 | io_iopoll_req_issued(req); | |
ba5290cc | 2229 | if (s->in_async) |
def596e9 JA |
2230 | mutex_unlock(&ctx->uring_lock); |
2231 | } | |
2232 | ||
2233 | return 0; | |
2b188cc1 JA |
2234 | } |
2235 | ||
31b51510 JA |
2236 | static struct async_list *io_async_list_from_sqe(struct io_ring_ctx *ctx, |
2237 | const struct io_uring_sqe *sqe) | |
2238 | { | |
2239 | switch (sqe->opcode) { | |
2240 | case IORING_OP_READV: | |
2241 | case IORING_OP_READ_FIXED: | |
2242 | return &ctx->pending_async[READ]; | |
2243 | case IORING_OP_WRITEV: | |
2244 | case IORING_OP_WRITE_FIXED: | |
2245 | return &ctx->pending_async[WRITE]; | |
2246 | default: | |
2247 | return NULL; | |
2248 | } | |
2249 | } | |
2250 | ||
edafccee JA |
2251 | static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe) |
2252 | { | |
2253 | u8 opcode = READ_ONCE(sqe->opcode); | |
2254 | ||
2255 | return !(opcode == IORING_OP_READ_FIXED || | |
2256 | opcode == IORING_OP_WRITE_FIXED); | |
2257 | } | |
2258 | ||
2b188cc1 JA |
2259 | static void io_sq_wq_submit_work(struct work_struct *work) |
2260 | { | |
2261 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
2b188cc1 | 2262 | struct io_ring_ctx *ctx = req->ctx; |
31b51510 JA |
2263 | struct mm_struct *cur_mm = NULL; |
2264 | struct async_list *async_list; | |
2265 | LIST_HEAD(req_list); | |
edafccee | 2266 | mm_segment_t old_fs; |
2b188cc1 JA |
2267 | int ret; |
2268 | ||
31b51510 JA |
2269 | async_list = io_async_list_from_sqe(ctx, req->submit.sqe); |
2270 | restart: | |
2271 | do { | |
2272 | struct sqe_submit *s = &req->submit; | |
2273 | const struct io_uring_sqe *sqe = s->sqe; | |
d0ee8791 | 2274 | unsigned int flags = req->flags; |
ba816ad6 | 2275 | struct io_kiocb *nxt = NULL; |
2b188cc1 | 2276 | |
8449eeda | 2277 | /* Ensure we clear previously set non-block flag */ |
31b51510 JA |
2278 | req->rw.ki_flags &= ~IOCB_NOWAIT; |
2279 | ||
2280 | ret = 0; | |
2281 | if (io_sqe_needs_user(sqe) && !cur_mm) { | |
2282 | if (!mmget_not_zero(ctx->sqo_mm)) { | |
2283 | ret = -EFAULT; | |
2284 | } else { | |
2285 | cur_mm = ctx->sqo_mm; | |
2286 | use_mm(cur_mm); | |
2287 | old_fs = get_fs(); | |
2288 | set_fs(USER_DS); | |
2289 | } | |
2290 | } | |
2291 | ||
2292 | if (!ret) { | |
2293 | s->has_user = cur_mm != NULL; | |
ba5290cc | 2294 | s->in_async = true; |
31b51510 | 2295 | do { |
ba816ad6 | 2296 | ret = __io_submit_sqe(ctx, req, s, &nxt, false); |
31b51510 JA |
2297 | /* |
2298 | * We can get EAGAIN for polled IO even though | |
2299 | * we're forcing a sync submission from here, | |
2300 | * since we can't wait for request slots on the | |
2301 | * block side. | |
2302 | */ | |
2303 | if (ret != -EAGAIN) | |
2304 | break; | |
2305 | cond_resched(); | |
2306 | } while (1); | |
2307 | } | |
817869d2 JA |
2308 | |
2309 | /* drop submission reference */ | |
ba816ad6 | 2310 | io_put_req(req, NULL); |
817869d2 | 2311 | |
31b51510 | 2312 | if (ret) { |
c71ffb67 | 2313 | io_cqring_add_event(ctx, sqe->user_data, ret); |
ba816ad6 | 2314 | io_put_req(req, NULL); |
31b51510 JA |
2315 | } |
2316 | ||
2317 | /* async context always use a copy of the sqe */ | |
2318 | kfree(sqe); | |
2319 | ||
ba816ad6 JA |
2320 | /* if a dependent link is ready, do that as the next one */ |
2321 | if (!ret && nxt) { | |
2322 | req = nxt; | |
2323 | continue; | |
2324 | } | |
2325 | ||
f7b76ac9 | 2326 | /* req from defer and link list needn't decrease async cnt */ |
d0ee8791 | 2327 | if (flags & (REQ_F_IO_DRAINED | REQ_F_LINK_DONE)) |
f7b76ac9 ZL |
2328 | goto out; |
2329 | ||
31b51510 JA |
2330 | if (!async_list) |
2331 | break; | |
2332 | if (!list_empty(&req_list)) { | |
2333 | req = list_first_entry(&req_list, struct io_kiocb, | |
2334 | list); | |
2335 | list_del(&req->list); | |
2336 | continue; | |
2337 | } | |
2338 | if (list_empty(&async_list->list)) | |
2339 | break; | |
2340 | ||
2341 | req = NULL; | |
2342 | spin_lock(&async_list->lock); | |
2343 | if (list_empty(&async_list->list)) { | |
2344 | spin_unlock(&async_list->lock); | |
2345 | break; | |
2346 | } | |
2347 | list_splice_init(&async_list->list, &req_list); | |
2348 | spin_unlock(&async_list->lock); | |
2349 | ||
2350 | req = list_first_entry(&req_list, struct io_kiocb, list); | |
2351 | list_del(&req->list); | |
2352 | } while (req); | |
edafccee JA |
2353 | |
2354 | /* | |
31b51510 JA |
2355 | * Rare case of racing with a submitter. If we find the count has |
2356 | * dropped to zero AND we have pending work items, then restart | |
2357 | * the processing. This is a tiny race window. | |
edafccee | 2358 | */ |
31b51510 JA |
2359 | if (async_list) { |
2360 | ret = atomic_dec_return(&async_list->cnt); | |
2361 | while (!ret && !list_empty(&async_list->list)) { | |
2362 | spin_lock(&async_list->lock); | |
2363 | atomic_inc(&async_list->cnt); | |
2364 | list_splice_init(&async_list->list, &req_list); | |
2365 | spin_unlock(&async_list->lock); | |
2366 | ||
2367 | if (!list_empty(&req_list)) { | |
2368 | req = list_first_entry(&req_list, | |
2369 | struct io_kiocb, list); | |
2370 | list_del(&req->list); | |
2371 | goto restart; | |
2372 | } | |
2373 | ret = atomic_dec_return(&async_list->cnt); | |
edafccee | 2374 | } |
edafccee | 2375 | } |
2b188cc1 | 2376 | |
f7b76ac9 | 2377 | out: |
31b51510 | 2378 | if (cur_mm) { |
edafccee | 2379 | set_fs(old_fs); |
31b51510 JA |
2380 | unuse_mm(cur_mm); |
2381 | mmput(cur_mm); | |
2b188cc1 | 2382 | } |
31b51510 | 2383 | } |
2b188cc1 | 2384 | |
31b51510 JA |
2385 | /* |
2386 | * See if we can piggy back onto previously submitted work, that is still | |
2387 | * running. We currently only allow this if the new request is sequential | |
2388 | * to the previous one we punted. | |
2389 | */ | |
2390 | static bool io_add_to_prev_work(struct async_list *list, struct io_kiocb *req) | |
2391 | { | |
6d5d5ac5 | 2392 | bool ret; |
31b51510 JA |
2393 | |
2394 | if (!list) | |
2395 | return false; | |
2396 | if (!(req->flags & REQ_F_SEQ_PREV)) | |
2397 | return false; | |
2398 | if (!atomic_read(&list->cnt)) | |
2399 | return false; | |
2400 | ||
2401 | ret = true; | |
2402 | spin_lock(&list->lock); | |
2403 | list_add_tail(&req->list, &list->list); | |
c0e48f9d ZL |
2404 | /* |
2405 | * Ensure we see a simultaneous modification from io_sq_wq_submit_work() | |
2406 | */ | |
2407 | smp_mb(); | |
31b51510 JA |
2408 | if (!atomic_read(&list->cnt)) { |
2409 | list_del_init(&req->list); | |
2410 | ret = false; | |
2411 | } | |
2412 | spin_unlock(&list->lock); | |
2413 | return ret; | |
2b188cc1 JA |
2414 | } |
2415 | ||
09bb8394 JA |
2416 | static bool io_op_needs_file(const struct io_uring_sqe *sqe) |
2417 | { | |
2418 | int op = READ_ONCE(sqe->opcode); | |
2419 | ||
2420 | switch (op) { | |
2421 | case IORING_OP_NOP: | |
2422 | case IORING_OP_POLL_REMOVE: | |
2423 | return false; | |
2424 | default: | |
2425 | return true; | |
2426 | } | |
2427 | } | |
2428 | ||
2429 | static int io_req_set_file(struct io_ring_ctx *ctx, const struct sqe_submit *s, | |
2430 | struct io_submit_state *state, struct io_kiocb *req) | |
2431 | { | |
2432 | unsigned flags; | |
2433 | int fd; | |
2434 | ||
2435 | flags = READ_ONCE(s->sqe->flags); | |
2436 | fd = READ_ONCE(s->sqe->fd); | |
2437 | ||
4fe2c963 | 2438 | if (flags & IOSQE_IO_DRAIN) |
de0617e4 | 2439 | req->flags |= REQ_F_IO_DRAIN; |
4fe2c963 JL |
2440 | /* |
2441 | * All io need record the previous position, if LINK vs DARIN, | |
2442 | * it can be used to mark the position of the first IO in the | |
2443 | * link list. | |
2444 | */ | |
2445 | req->sequence = s->sequence; | |
de0617e4 | 2446 | |
60c112b0 | 2447 | if (!io_op_needs_file(s->sqe)) |
09bb8394 | 2448 | return 0; |
09bb8394 JA |
2449 | |
2450 | if (flags & IOSQE_FIXED_FILE) { | |
2451 | if (unlikely(!ctx->user_files || | |
2452 | (unsigned) fd >= ctx->nr_user_files)) | |
2453 | return -EBADF; | |
08a45173 JA |
2454 | if (!ctx->user_files[fd]) |
2455 | return -EBADF; | |
09bb8394 JA |
2456 | req->file = ctx->user_files[fd]; |
2457 | req->flags |= REQ_F_FIXED_FILE; | |
2458 | } else { | |
2459 | if (s->needs_fixed_file) | |
2460 | return -EBADF; | |
2461 | req->file = io_file_get(state, fd); | |
2462 | if (unlikely(!req->file)) | |
2463 | return -EBADF; | |
2464 | } | |
2465 | ||
2466 | return 0; | |
2467 | } | |
2468 | ||
4fe2c963 | 2469 | static int __io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
bc808bce | 2470 | struct sqe_submit *s) |
2b188cc1 | 2471 | { |
e0c5c576 | 2472 | int ret; |
2b188cc1 | 2473 | |
ba816ad6 | 2474 | ret = __io_submit_sqe(ctx, req, s, NULL, true); |
491381ce JA |
2475 | |
2476 | /* | |
2477 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
2478 | * doesn't support non-blocking read/write attempts | |
2479 | */ | |
2480 | if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) || | |
2481 | (req->flags & REQ_F_MUST_PUNT))) { | |
2b188cc1 JA |
2482 | struct io_uring_sqe *sqe_copy; |
2483 | ||
954dab19 | 2484 | sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL); |
2b188cc1 | 2485 | if (sqe_copy) { |
31b51510 JA |
2486 | struct async_list *list; |
2487 | ||
2b188cc1 | 2488 | s->sqe = sqe_copy; |
2b188cc1 | 2489 | memcpy(&req->submit, s, sizeof(*s)); |
31b51510 JA |
2490 | list = io_async_list_from_sqe(ctx, s->sqe); |
2491 | if (!io_add_to_prev_work(list, req)) { | |
2492 | if (list) | |
2493 | atomic_inc(&list->cnt); | |
2494 | INIT_WORK(&req->work, io_sq_wq_submit_work); | |
18d9be1a | 2495 | io_queue_async_work(ctx, req); |
31b51510 | 2496 | } |
e65ef56d JA |
2497 | |
2498 | /* | |
2499 | * Queued up for async execution, worker will release | |
9e645e11 | 2500 | * submit reference when the iocb is actually submitted. |
e65ef56d JA |
2501 | */ |
2502 | return 0; | |
2b188cc1 JA |
2503 | } |
2504 | } | |
e65ef56d JA |
2505 | |
2506 | /* drop submission reference */ | |
ba816ad6 | 2507 | io_put_req(req, NULL); |
e65ef56d JA |
2508 | |
2509 | /* and drop final reference, if we failed */ | |
9e645e11 JA |
2510 | if (ret) { |
2511 | io_cqring_add_event(ctx, req->user_data, ret); | |
2512 | if (req->flags & REQ_F_LINK) | |
2513 | req->flags |= REQ_F_FAIL_LINK; | |
ba816ad6 | 2514 | io_put_req(req, NULL); |
9e645e11 | 2515 | } |
2b188cc1 JA |
2516 | |
2517 | return ret; | |
2518 | } | |
2519 | ||
4fe2c963 | 2520 | static int io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
bc808bce | 2521 | struct sqe_submit *s) |
4fe2c963 JL |
2522 | { |
2523 | int ret; | |
2524 | ||
2525 | ret = io_req_defer(ctx, req, s->sqe); | |
2526 | if (ret) { | |
2527 | if (ret != -EIOCBQUEUED) { | |
ba816ad6 | 2528 | io_free_req(req, NULL); |
4fe2c963 JL |
2529 | io_cqring_add_event(ctx, s->sqe->user_data, ret); |
2530 | } | |
2531 | return 0; | |
2532 | } | |
2533 | ||
bc808bce | 2534 | return __io_queue_sqe(ctx, req, s); |
4fe2c963 JL |
2535 | } |
2536 | ||
2537 | static int io_queue_link_head(struct io_ring_ctx *ctx, struct io_kiocb *req, | |
bc808bce | 2538 | struct sqe_submit *s, struct io_kiocb *shadow) |
4fe2c963 JL |
2539 | { |
2540 | int ret; | |
2541 | int need_submit = false; | |
2542 | ||
2543 | if (!shadow) | |
bc808bce | 2544 | return io_queue_sqe(ctx, req, s); |
4fe2c963 JL |
2545 | |
2546 | /* | |
2547 | * Mark the first IO in link list as DRAIN, let all the following | |
2548 | * IOs enter the defer list. all IO needs to be completed before link | |
2549 | * list. | |
2550 | */ | |
2551 | req->flags |= REQ_F_IO_DRAIN; | |
2552 | ret = io_req_defer(ctx, req, s->sqe); | |
2553 | if (ret) { | |
2554 | if (ret != -EIOCBQUEUED) { | |
ba816ad6 | 2555 | io_free_req(req, NULL); |
7b20238d | 2556 | __io_free_req(shadow); |
4fe2c963 JL |
2557 | io_cqring_add_event(ctx, s->sqe->user_data, ret); |
2558 | return 0; | |
2559 | } | |
2560 | } else { | |
2561 | /* | |
2562 | * If ret == 0 means that all IOs in front of link io are | |
2563 | * running done. let's queue link head. | |
2564 | */ | |
2565 | need_submit = true; | |
2566 | } | |
2567 | ||
2568 | /* Insert shadow req to defer_list, blocking next IOs */ | |
2569 | spin_lock_irq(&ctx->completion_lock); | |
2570 | list_add_tail(&shadow->list, &ctx->defer_list); | |
2571 | spin_unlock_irq(&ctx->completion_lock); | |
2572 | ||
2573 | if (need_submit) | |
bc808bce | 2574 | return __io_queue_sqe(ctx, req, s); |
4fe2c963 JL |
2575 | |
2576 | return 0; | |
2577 | } | |
2578 | ||
9e645e11 JA |
2579 | #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK) |
2580 | ||
2581 | static void io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s, | |
bc808bce | 2582 | struct io_submit_state *state, struct io_kiocb **link) |
9e645e11 JA |
2583 | { |
2584 | struct io_uring_sqe *sqe_copy; | |
2585 | struct io_kiocb *req; | |
2586 | int ret; | |
2587 | ||
2588 | /* enforce forwards compatibility on users */ | |
2589 | if (unlikely(s->sqe->flags & ~SQE_VALID_FLAGS)) { | |
2590 | ret = -EINVAL; | |
2591 | goto err; | |
2592 | } | |
2593 | ||
2594 | req = io_get_req(ctx, state); | |
2595 | if (unlikely(!req)) { | |
2596 | ret = -EAGAIN; | |
2597 | goto err; | |
2598 | } | |
2599 | ||
2600 | ret = io_req_set_file(ctx, s, state, req); | |
2601 | if (unlikely(ret)) { | |
2602 | err_req: | |
ba816ad6 | 2603 | io_free_req(req, NULL); |
9e645e11 JA |
2604 | err: |
2605 | io_cqring_add_event(ctx, s->sqe->user_data, ret); | |
2606 | return; | |
2607 | } | |
2608 | ||
84d55dc5 PB |
2609 | req->user_data = s->sqe->user_data; |
2610 | ||
9e645e11 JA |
2611 | /* |
2612 | * If we already have a head request, queue this one for async | |
2613 | * submittal once the head completes. If we don't have a head but | |
2614 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
2615 | * submitted sync once the chain is complete. If none of those | |
2616 | * conditions are true (normal request), then just queue it. | |
2617 | */ | |
2618 | if (*link) { | |
2619 | struct io_kiocb *prev = *link; | |
2620 | ||
2621 | sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL); | |
2622 | if (!sqe_copy) { | |
2623 | ret = -EAGAIN; | |
2624 | goto err_req; | |
2625 | } | |
2626 | ||
2627 | s->sqe = sqe_copy; | |
2628 | memcpy(&req->submit, s, sizeof(*s)); | |
2629 | list_add_tail(&req->list, &prev->link_list); | |
2630 | } else if (s->sqe->flags & IOSQE_IO_LINK) { | |
2631 | req->flags |= REQ_F_LINK; | |
2632 | ||
2633 | memcpy(&req->submit, s, sizeof(*s)); | |
2634 | INIT_LIST_HEAD(&req->link_list); | |
2635 | *link = req; | |
2636 | } else { | |
bc808bce | 2637 | io_queue_sqe(ctx, req, s); |
9e645e11 JA |
2638 | } |
2639 | } | |
2640 | ||
9a56a232 JA |
2641 | /* |
2642 | * Batched submission is done, ensure local IO is flushed out. | |
2643 | */ | |
2644 | static void io_submit_state_end(struct io_submit_state *state) | |
2645 | { | |
2646 | blk_finish_plug(&state->plug); | |
3d6770fb | 2647 | io_file_put(state); |
2579f913 JA |
2648 | if (state->free_reqs) |
2649 | kmem_cache_free_bulk(req_cachep, state->free_reqs, | |
2650 | &state->reqs[state->cur_req]); | |
9a56a232 JA |
2651 | } |
2652 | ||
2653 | /* | |
2654 | * Start submission side cache. | |
2655 | */ | |
2656 | static void io_submit_state_start(struct io_submit_state *state, | |
2657 | struct io_ring_ctx *ctx, unsigned max_ios) | |
2658 | { | |
2659 | blk_start_plug(&state->plug); | |
2579f913 | 2660 | state->free_reqs = 0; |
9a56a232 JA |
2661 | state->file = NULL; |
2662 | state->ios_left = max_ios; | |
2663 | } | |
2664 | ||
2b188cc1 JA |
2665 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
2666 | { | |
75b28aff | 2667 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 2668 | |
75b28aff | 2669 | if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) { |
2b188cc1 JA |
2670 | /* |
2671 | * Ensure any loads from the SQEs are done at this point, | |
2672 | * since once we write the new head, the application could | |
2673 | * write new data to them. | |
2674 | */ | |
75b28aff | 2675 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); |
2b188cc1 JA |
2676 | } |
2677 | } | |
2678 | ||
2b188cc1 JA |
2679 | /* |
2680 | * Fetch an sqe, if one is available. Note that s->sqe will point to memory | |
2681 | * that is mapped by userspace. This means that care needs to be taken to | |
2682 | * ensure that reads are stable, as we cannot rely on userspace always | |
2683 | * being a good citizen. If members of the sqe are validated and then later | |
2684 | * used, it's important that those reads are done through READ_ONCE() to | |
2685 | * prevent a re-load down the line. | |
2686 | */ | |
2687 | static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s) | |
2688 | { | |
75b28aff HV |
2689 | struct io_rings *rings = ctx->rings; |
2690 | u32 *sq_array = ctx->sq_array; | |
2b188cc1 JA |
2691 | unsigned head; |
2692 | ||
2693 | /* | |
2694 | * The cached sq head (or cq tail) serves two purposes: | |
2695 | * | |
2696 | * 1) allows us to batch the cost of updating the user visible | |
2697 | * head updates. | |
2698 | * 2) allows the kernel side to track the head on its own, even | |
2699 | * though the application is the one updating it. | |
2700 | */ | |
2701 | head = ctx->cached_sq_head; | |
e523a29c | 2702 | /* make sure SQ entry isn't read before tail */ |
75b28aff | 2703 | if (head == smp_load_acquire(&rings->sq.tail)) |
2b188cc1 JA |
2704 | return false; |
2705 | ||
75b28aff | 2706 | head = READ_ONCE(sq_array[head & ctx->sq_mask]); |
2b188cc1 JA |
2707 | if (head < ctx->sq_entries) { |
2708 | s->index = head; | |
2709 | s->sqe = &ctx->sq_sqes[head]; | |
8776f3fa | 2710 | s->sequence = ctx->cached_sq_head; |
2b188cc1 JA |
2711 | ctx->cached_sq_head++; |
2712 | return true; | |
2713 | } | |
2714 | ||
2715 | /* drop invalid entries */ | |
2716 | ctx->cached_sq_head++; | |
498ccd9e JA |
2717 | ctx->cached_sq_dropped++; |
2718 | WRITE_ONCE(rings->sq_dropped, ctx->cached_sq_dropped); | |
2b188cc1 JA |
2719 | return false; |
2720 | } | |
2721 | ||
fb5ccc98 PB |
2722 | static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr, |
2723 | bool has_user, bool mm_fault) | |
6c271ce2 JA |
2724 | { |
2725 | struct io_submit_state state, *statep = NULL; | |
9e645e11 | 2726 | struct io_kiocb *link = NULL; |
4fe2c963 | 2727 | struct io_kiocb *shadow_req = NULL; |
9e645e11 JA |
2728 | bool prev_was_link = false; |
2729 | int i, submitted = 0; | |
6c271ce2 JA |
2730 | |
2731 | if (nr > IO_PLUG_THRESHOLD) { | |
2732 | io_submit_state_start(&state, ctx, nr); | |
2733 | statep = &state; | |
2734 | } | |
2735 | ||
2736 | for (i = 0; i < nr; i++) { | |
fb5ccc98 PB |
2737 | struct sqe_submit s; |
2738 | ||
2739 | if (!io_get_sqring(ctx, &s)) | |
2740 | break; | |
2741 | ||
9e645e11 JA |
2742 | /* |
2743 | * If previous wasn't linked and we have a linked command, | |
2744 | * that's the end of the chain. Submit the previous link. | |
2745 | */ | |
2746 | if (!prev_was_link && link) { | |
bc808bce | 2747 | io_queue_link_head(ctx, link, &link->submit, shadow_req); |
9e645e11 | 2748 | link = NULL; |
5f5ad9ce | 2749 | shadow_req = NULL; |
9e645e11 | 2750 | } |
fb5ccc98 | 2751 | prev_was_link = (s.sqe->flags & IOSQE_IO_LINK) != 0; |
9e645e11 | 2752 | |
fb5ccc98 | 2753 | if (link && (s.sqe->flags & IOSQE_IO_DRAIN)) { |
4fe2c963 JL |
2754 | if (!shadow_req) { |
2755 | shadow_req = io_get_req(ctx, NULL); | |
a1041c27 JL |
2756 | if (unlikely(!shadow_req)) |
2757 | goto out; | |
4fe2c963 JL |
2758 | shadow_req->flags |= (REQ_F_IO_DRAIN | REQ_F_SHADOW_DRAIN); |
2759 | refcount_dec(&shadow_req->refs); | |
2760 | } | |
fb5ccc98 | 2761 | shadow_req->sequence = s.sequence; |
4fe2c963 JL |
2762 | } |
2763 | ||
a1041c27 | 2764 | out: |
6c271ce2 | 2765 | if (unlikely(mm_fault)) { |
fb5ccc98 | 2766 | io_cqring_add_event(ctx, s.sqe->user_data, |
9e645e11 | 2767 | -EFAULT); |
6c271ce2 | 2768 | } else { |
fb5ccc98 | 2769 | s.has_user = has_user; |
ba5290cc | 2770 | s.in_async = true; |
fb5ccc98 PB |
2771 | s.needs_fixed_file = true; |
2772 | io_submit_sqe(ctx, &s, statep, &link); | |
6c271ce2 | 2773 | submitted++; |
6c271ce2 | 2774 | } |
6c271ce2 JA |
2775 | } |
2776 | ||
9e645e11 | 2777 | if (link) |
bc808bce | 2778 | io_queue_link_head(ctx, link, &link->submit, shadow_req); |
6c271ce2 JA |
2779 | if (statep) |
2780 | io_submit_state_end(&state); | |
2781 | ||
2782 | return submitted; | |
2783 | } | |
2784 | ||
2785 | static int io_sq_thread(void *data) | |
2786 | { | |
6c271ce2 JA |
2787 | struct io_ring_ctx *ctx = data; |
2788 | struct mm_struct *cur_mm = NULL; | |
2789 | mm_segment_t old_fs; | |
2790 | DEFINE_WAIT(wait); | |
2791 | unsigned inflight; | |
2792 | unsigned long timeout; | |
2793 | ||
a4c0b3de JL |
2794 | complete(&ctx->sqo_thread_started); |
2795 | ||
6c271ce2 JA |
2796 | old_fs = get_fs(); |
2797 | set_fs(USER_DS); | |
2798 | ||
2799 | timeout = inflight = 0; | |
2bbcd6d3 | 2800 | while (!kthread_should_park()) { |
fb5ccc98 PB |
2801 | bool mm_fault = false; |
2802 | unsigned int to_submit; | |
6c271ce2 JA |
2803 | |
2804 | if (inflight) { | |
2805 | unsigned nr_events = 0; | |
2806 | ||
2807 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
2b2ed975 JA |
2808 | /* |
2809 | * inflight is the count of the maximum possible | |
2810 | * entries we submitted, but it can be smaller | |
2811 | * if we dropped some of them. If we don't have | |
2812 | * poll entries available, then we know that we | |
2813 | * have nothing left to poll for. Reset the | |
2814 | * inflight count to zero in that case. | |
2815 | */ | |
2816 | mutex_lock(&ctx->uring_lock); | |
2817 | if (!list_empty(&ctx->poll_list)) | |
2818 | __io_iopoll_check(ctx, &nr_events, 0); | |
2819 | else | |
2820 | inflight = 0; | |
2821 | mutex_unlock(&ctx->uring_lock); | |
6c271ce2 JA |
2822 | } else { |
2823 | /* | |
2824 | * Normal IO, just pretend everything completed. | |
2825 | * We don't have to poll completions for that. | |
2826 | */ | |
2827 | nr_events = inflight; | |
2828 | } | |
2829 | ||
2830 | inflight -= nr_events; | |
2831 | if (!inflight) | |
2832 | timeout = jiffies + ctx->sq_thread_idle; | |
2833 | } | |
2834 | ||
fb5ccc98 PB |
2835 | to_submit = io_sqring_entries(ctx); |
2836 | if (!to_submit) { | |
6c271ce2 JA |
2837 | /* |
2838 | * We're polling. If we're within the defined idle | |
2839 | * period, then let us spin without work before going | |
2840 | * to sleep. | |
2841 | */ | |
2842 | if (inflight || !time_after(jiffies, timeout)) { | |
9831a90c | 2843 | cond_resched(); |
6c271ce2 JA |
2844 | continue; |
2845 | } | |
2846 | ||
2847 | /* | |
2848 | * Drop cur_mm before scheduling, we can't hold it for | |
2849 | * long periods (or over schedule()). Do this before | |
2850 | * adding ourselves to the waitqueue, as the unuse/drop | |
2851 | * may sleep. | |
2852 | */ | |
2853 | if (cur_mm) { | |
2854 | unuse_mm(cur_mm); | |
2855 | mmput(cur_mm); | |
2856 | cur_mm = NULL; | |
2857 | } | |
2858 | ||
2859 | prepare_to_wait(&ctx->sqo_wait, &wait, | |
2860 | TASK_INTERRUPTIBLE); | |
2861 | ||
2862 | /* Tell userspace we may need a wakeup call */ | |
75b28aff | 2863 | ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP; |
0d7bae69 SB |
2864 | /* make sure to read SQ tail after writing flags */ |
2865 | smp_mb(); | |
6c271ce2 | 2866 | |
fb5ccc98 PB |
2867 | to_submit = io_sqring_entries(ctx); |
2868 | if (!to_submit) { | |
2bbcd6d3 | 2869 | if (kthread_should_park()) { |
6c271ce2 JA |
2870 | finish_wait(&ctx->sqo_wait, &wait); |
2871 | break; | |
2872 | } | |
2873 | if (signal_pending(current)) | |
2874 | flush_signals(current); | |
2875 | schedule(); | |
2876 | finish_wait(&ctx->sqo_wait, &wait); | |
2877 | ||
75b28aff | 2878 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
2879 | continue; |
2880 | } | |
2881 | finish_wait(&ctx->sqo_wait, &wait); | |
2882 | ||
75b28aff | 2883 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
2884 | } |
2885 | ||
6c271ce2 | 2886 | /* Unless all new commands are FIXED regions, grab mm */ |
fb5ccc98 | 2887 | if (!cur_mm) { |
6c271ce2 JA |
2888 | mm_fault = !mmget_not_zero(ctx->sqo_mm); |
2889 | if (!mm_fault) { | |
2890 | use_mm(ctx->sqo_mm); | |
2891 | cur_mm = ctx->sqo_mm; | |
2892 | } | |
2893 | } | |
2894 | ||
fb5ccc98 PB |
2895 | to_submit = min(to_submit, ctx->sq_entries); |
2896 | inflight += io_submit_sqes(ctx, to_submit, cur_mm != NULL, | |
2897 | mm_fault); | |
6c271ce2 JA |
2898 | |
2899 | /* Commit SQ ring head once we've consumed all SQEs */ | |
2900 | io_commit_sqring(ctx); | |
2901 | } | |
2902 | ||
2903 | set_fs(old_fs); | |
2904 | if (cur_mm) { | |
2905 | unuse_mm(cur_mm); | |
2906 | mmput(cur_mm); | |
2907 | } | |
06058632 | 2908 | |
2bbcd6d3 | 2909 | kthread_parkme(); |
06058632 | 2910 | |
6c271ce2 JA |
2911 | return 0; |
2912 | } | |
2913 | ||
bc808bce | 2914 | static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit) |
2b188cc1 | 2915 | { |
9a56a232 | 2916 | struct io_submit_state state, *statep = NULL; |
9e645e11 | 2917 | struct io_kiocb *link = NULL; |
4fe2c963 | 2918 | struct io_kiocb *shadow_req = NULL; |
9e645e11 | 2919 | bool prev_was_link = false; |
5c8b0b54 | 2920 | int i, submit = 0; |
2b188cc1 | 2921 | |
9a56a232 JA |
2922 | if (to_submit > IO_PLUG_THRESHOLD) { |
2923 | io_submit_state_start(&state, ctx, to_submit); | |
2924 | statep = &state; | |
2925 | } | |
2b188cc1 JA |
2926 | |
2927 | for (i = 0; i < to_submit; i++) { | |
2928 | struct sqe_submit s; | |
2929 | ||
2930 | if (!io_get_sqring(ctx, &s)) | |
2931 | break; | |
2932 | ||
9e645e11 JA |
2933 | /* |
2934 | * If previous wasn't linked and we have a linked command, | |
2935 | * that's the end of the chain. Submit the previous link. | |
2936 | */ | |
2937 | if (!prev_was_link && link) { | |
bc808bce | 2938 | io_queue_link_head(ctx, link, &link->submit, shadow_req); |
9e645e11 | 2939 | link = NULL; |
5f5ad9ce | 2940 | shadow_req = NULL; |
9e645e11 JA |
2941 | } |
2942 | prev_was_link = (s.sqe->flags & IOSQE_IO_LINK) != 0; | |
2943 | ||
4fe2c963 JL |
2944 | if (link && (s.sqe->flags & IOSQE_IO_DRAIN)) { |
2945 | if (!shadow_req) { | |
2946 | shadow_req = io_get_req(ctx, NULL); | |
a1041c27 JL |
2947 | if (unlikely(!shadow_req)) |
2948 | goto out; | |
4fe2c963 JL |
2949 | shadow_req->flags |= (REQ_F_IO_DRAIN | REQ_F_SHADOW_DRAIN); |
2950 | refcount_dec(&shadow_req->refs); | |
2951 | } | |
2952 | shadow_req->sequence = s.sequence; | |
2953 | } | |
2954 | ||
a1041c27 | 2955 | out: |
2b188cc1 | 2956 | s.has_user = true; |
ba5290cc | 2957 | s.in_async = false; |
6c271ce2 | 2958 | s.needs_fixed_file = false; |
5c8b0b54 | 2959 | submit++; |
bc808bce | 2960 | io_submit_sqe(ctx, &s, statep, &link); |
2b188cc1 | 2961 | } |
2b188cc1 | 2962 | |
9e645e11 | 2963 | if (link) |
bc808bce | 2964 | io_queue_link_head(ctx, link, &link->submit, shadow_req); |
9a56a232 JA |
2965 | if (statep) |
2966 | io_submit_state_end(statep); | |
2b188cc1 | 2967 | |
935d1e45 PB |
2968 | io_commit_sqring(ctx); |
2969 | ||
5c8b0b54 | 2970 | return submit; |
2b188cc1 JA |
2971 | } |
2972 | ||
bda52162 JA |
2973 | struct io_wait_queue { |
2974 | struct wait_queue_entry wq; | |
2975 | struct io_ring_ctx *ctx; | |
2976 | unsigned to_wait; | |
2977 | unsigned nr_timeouts; | |
2978 | }; | |
2979 | ||
2980 | static inline bool io_should_wake(struct io_wait_queue *iowq) | |
2981 | { | |
2982 | struct io_ring_ctx *ctx = iowq->ctx; | |
2983 | ||
2984 | /* | |
2985 | * Wake up if we have enough events, or if a timeout occured since we | |
2986 | * started waiting. For timeouts, we always want to return to userspace, | |
2987 | * regardless of event count. | |
2988 | */ | |
2989 | return io_cqring_events(ctx->rings) >= iowq->to_wait || | |
2990 | atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; | |
2991 | } | |
2992 | ||
2993 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
2994 | int wake_flags, void *key) | |
2995 | { | |
2996 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
2997 | wq); | |
2998 | ||
2999 | if (!io_should_wake(iowq)) | |
3000 | return -1; | |
3001 | ||
3002 | return autoremove_wake_function(curr, mode, wake_flags, key); | |
3003 | } | |
3004 | ||
2b188cc1 JA |
3005 | /* |
3006 | * Wait until events become available, if we don't already have some. The | |
3007 | * application must reap them itself, as they reside on the shared cq ring. | |
3008 | */ | |
3009 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
3010 | const sigset_t __user *sig, size_t sigsz) | |
3011 | { | |
bda52162 JA |
3012 | struct io_wait_queue iowq = { |
3013 | .wq = { | |
3014 | .private = current, | |
3015 | .func = io_wake_function, | |
3016 | .entry = LIST_HEAD_INIT(iowq.wq.entry), | |
3017 | }, | |
3018 | .ctx = ctx, | |
3019 | .to_wait = min_events, | |
3020 | }; | |
75b28aff | 3021 | struct io_rings *rings = ctx->rings; |
2b188cc1 JA |
3022 | int ret; |
3023 | ||
75b28aff | 3024 | if (io_cqring_events(rings) >= min_events) |
2b188cc1 JA |
3025 | return 0; |
3026 | ||
3027 | if (sig) { | |
9e75ad5d AB |
3028 | #ifdef CONFIG_COMPAT |
3029 | if (in_compat_syscall()) | |
3030 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 3031 | sigsz); |
9e75ad5d AB |
3032 | else |
3033 | #endif | |
b772434b | 3034 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 3035 | |
2b188cc1 JA |
3036 | if (ret) |
3037 | return ret; | |
3038 | } | |
3039 | ||
bda52162 JA |
3040 | ret = 0; |
3041 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); | |
3042 | do { | |
3043 | prepare_to_wait_exclusive(&ctx->wait, &iowq.wq, | |
3044 | TASK_INTERRUPTIBLE); | |
3045 | if (io_should_wake(&iowq)) | |
3046 | break; | |
3047 | schedule(); | |
3048 | if (signal_pending(current)) { | |
3049 | ret = -ERESTARTSYS; | |
3050 | break; | |
3051 | } | |
3052 | } while (1); | |
3053 | finish_wait(&ctx->wait, &iowq.wq); | |
3054 | ||
b772434b | 3055 | restore_saved_sigmask_unless(ret == -ERESTARTSYS); |
97abc889 ON |
3056 | if (ret == -ERESTARTSYS) |
3057 | ret = -EINTR; | |
2b188cc1 | 3058 | |
75b28aff | 3059 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
3060 | } |
3061 | ||
6b06314c JA |
3062 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
3063 | { | |
3064 | #if defined(CONFIG_UNIX) | |
3065 | if (ctx->ring_sock) { | |
3066 | struct sock *sock = ctx->ring_sock->sk; | |
3067 | struct sk_buff *skb; | |
3068 | ||
3069 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
3070 | kfree_skb(skb); | |
3071 | } | |
3072 | #else | |
3073 | int i; | |
3074 | ||
3075 | for (i = 0; i < ctx->nr_user_files; i++) | |
08a45173 JA |
3076 | if (ctx->user_files[i]) |
3077 | fput(ctx->user_files[i]); | |
6b06314c JA |
3078 | #endif |
3079 | } | |
3080 | ||
3081 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) | |
3082 | { | |
3083 | if (!ctx->user_files) | |
3084 | return -ENXIO; | |
3085 | ||
3086 | __io_sqe_files_unregister(ctx); | |
3087 | kfree(ctx->user_files); | |
3088 | ctx->user_files = NULL; | |
3089 | ctx->nr_user_files = 0; | |
3090 | return 0; | |
3091 | } | |
3092 | ||
6c271ce2 JA |
3093 | static void io_sq_thread_stop(struct io_ring_ctx *ctx) |
3094 | { | |
3095 | if (ctx->sqo_thread) { | |
a4c0b3de | 3096 | wait_for_completion(&ctx->sqo_thread_started); |
2bbcd6d3 RP |
3097 | /* |
3098 | * The park is a bit of a work-around, without it we get | |
3099 | * warning spews on shutdown with SQPOLL set and affinity | |
3100 | * set to a single CPU. | |
3101 | */ | |
06058632 | 3102 | kthread_park(ctx->sqo_thread); |
6c271ce2 JA |
3103 | kthread_stop(ctx->sqo_thread); |
3104 | ctx->sqo_thread = NULL; | |
3105 | } | |
3106 | } | |
3107 | ||
6b06314c JA |
3108 | static void io_finish_async(struct io_ring_ctx *ctx) |
3109 | { | |
54a91f3b JA |
3110 | int i; |
3111 | ||
6c271ce2 JA |
3112 | io_sq_thread_stop(ctx); |
3113 | ||
54a91f3b JA |
3114 | for (i = 0; i < ARRAY_SIZE(ctx->sqo_wq); i++) { |
3115 | if (ctx->sqo_wq[i]) { | |
3116 | destroy_workqueue(ctx->sqo_wq[i]); | |
3117 | ctx->sqo_wq[i] = NULL; | |
3118 | } | |
6b06314c JA |
3119 | } |
3120 | } | |
3121 | ||
3122 | #if defined(CONFIG_UNIX) | |
3123 | static void io_destruct_skb(struct sk_buff *skb) | |
3124 | { | |
3125 | struct io_ring_ctx *ctx = skb->sk->sk_user_data; | |
8a997340 JA |
3126 | int i; |
3127 | ||
3128 | for (i = 0; i < ARRAY_SIZE(ctx->sqo_wq); i++) | |
3129 | if (ctx->sqo_wq[i]) | |
3130 | flush_workqueue(ctx->sqo_wq[i]); | |
6b06314c | 3131 | |
6b06314c JA |
3132 | unix_destruct_scm(skb); |
3133 | } | |
3134 | ||
3135 | /* | |
3136 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
3137 | * the io_uring can be safely unregistered on process exit, even if we have | |
3138 | * loops in the file referencing. | |
3139 | */ | |
3140 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
3141 | { | |
3142 | struct sock *sk = ctx->ring_sock->sk; | |
3143 | struct scm_fp_list *fpl; | |
3144 | struct sk_buff *skb; | |
08a45173 | 3145 | int i, nr_files; |
6b06314c JA |
3146 | |
3147 | if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) { | |
3148 | unsigned long inflight = ctx->user->unix_inflight + nr; | |
3149 | ||
3150 | if (inflight > task_rlimit(current, RLIMIT_NOFILE)) | |
3151 | return -EMFILE; | |
3152 | } | |
3153 | ||
3154 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); | |
3155 | if (!fpl) | |
3156 | return -ENOMEM; | |
3157 | ||
3158 | skb = alloc_skb(0, GFP_KERNEL); | |
3159 | if (!skb) { | |
3160 | kfree(fpl); | |
3161 | return -ENOMEM; | |
3162 | } | |
3163 | ||
3164 | skb->sk = sk; | |
6b06314c | 3165 | |
08a45173 | 3166 | nr_files = 0; |
6b06314c JA |
3167 | fpl->user = get_uid(ctx->user); |
3168 | for (i = 0; i < nr; i++) { | |
08a45173 JA |
3169 | if (!ctx->user_files[i + offset]) |
3170 | continue; | |
3171 | fpl->fp[nr_files] = get_file(ctx->user_files[i + offset]); | |
3172 | unix_inflight(fpl->user, fpl->fp[nr_files]); | |
3173 | nr_files++; | |
6b06314c JA |
3174 | } |
3175 | ||
08a45173 JA |
3176 | if (nr_files) { |
3177 | fpl->max = SCM_MAX_FD; | |
3178 | fpl->count = nr_files; | |
3179 | UNIXCB(skb).fp = fpl; | |
3180 | skb->destructor = io_destruct_skb; | |
3181 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); | |
3182 | skb_queue_head(&sk->sk_receive_queue, skb); | |
6b06314c | 3183 | |
08a45173 JA |
3184 | for (i = 0; i < nr_files; i++) |
3185 | fput(fpl->fp[i]); | |
3186 | } else { | |
3187 | kfree_skb(skb); | |
3188 | kfree(fpl); | |
3189 | } | |
6b06314c JA |
3190 | |
3191 | return 0; | |
3192 | } | |
3193 | ||
3194 | /* | |
3195 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
3196 | * causes regular reference counting to break down. We rely on the UNIX | |
3197 | * garbage collection to take care of this problem for us. | |
3198 | */ | |
3199 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
3200 | { | |
3201 | unsigned left, total; | |
3202 | int ret = 0; | |
3203 | ||
3204 | total = 0; | |
3205 | left = ctx->nr_user_files; | |
3206 | while (left) { | |
3207 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
3208 | |
3209 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
3210 | if (ret) | |
3211 | break; | |
3212 | left -= this_files; | |
3213 | total += this_files; | |
3214 | } | |
3215 | ||
3216 | if (!ret) | |
3217 | return 0; | |
3218 | ||
3219 | while (total < ctx->nr_user_files) { | |
08a45173 JA |
3220 | if (ctx->user_files[total]) |
3221 | fput(ctx->user_files[total]); | |
6b06314c JA |
3222 | total++; |
3223 | } | |
3224 | ||
3225 | return ret; | |
3226 | } | |
3227 | #else | |
3228 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
3229 | { | |
3230 | return 0; | |
3231 | } | |
3232 | #endif | |
3233 | ||
3234 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, | |
3235 | unsigned nr_args) | |
3236 | { | |
3237 | __s32 __user *fds = (__s32 __user *) arg; | |
3238 | int fd, ret = 0; | |
3239 | unsigned i; | |
3240 | ||
3241 | if (ctx->user_files) | |
3242 | return -EBUSY; | |
3243 | if (!nr_args) | |
3244 | return -EINVAL; | |
3245 | if (nr_args > IORING_MAX_FIXED_FILES) | |
3246 | return -EMFILE; | |
3247 | ||
3248 | ctx->user_files = kcalloc(nr_args, sizeof(struct file *), GFP_KERNEL); | |
3249 | if (!ctx->user_files) | |
3250 | return -ENOMEM; | |
3251 | ||
08a45173 | 3252 | for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { |
6b06314c JA |
3253 | ret = -EFAULT; |
3254 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) | |
3255 | break; | |
08a45173 JA |
3256 | /* allow sparse sets */ |
3257 | if (fd == -1) { | |
3258 | ret = 0; | |
3259 | continue; | |
3260 | } | |
6b06314c JA |
3261 | |
3262 | ctx->user_files[i] = fget(fd); | |
3263 | ||
3264 | ret = -EBADF; | |
3265 | if (!ctx->user_files[i]) | |
3266 | break; | |
3267 | /* | |
3268 | * Don't allow io_uring instances to be registered. If UNIX | |
3269 | * isn't enabled, then this causes a reference cycle and this | |
3270 | * instance can never get freed. If UNIX is enabled we'll | |
3271 | * handle it just fine, but there's still no point in allowing | |
3272 | * a ring fd as it doesn't support regular read/write anyway. | |
3273 | */ | |
3274 | if (ctx->user_files[i]->f_op == &io_uring_fops) { | |
3275 | fput(ctx->user_files[i]); | |
3276 | break; | |
3277 | } | |
6b06314c JA |
3278 | ret = 0; |
3279 | } | |
3280 | ||
3281 | if (ret) { | |
3282 | for (i = 0; i < ctx->nr_user_files; i++) | |
08a45173 JA |
3283 | if (ctx->user_files[i]) |
3284 | fput(ctx->user_files[i]); | |
6b06314c JA |
3285 | |
3286 | kfree(ctx->user_files); | |
25adf50f | 3287 | ctx->user_files = NULL; |
6b06314c JA |
3288 | ctx->nr_user_files = 0; |
3289 | return ret; | |
3290 | } | |
3291 | ||
3292 | ret = io_sqe_files_scm(ctx); | |
3293 | if (ret) | |
3294 | io_sqe_files_unregister(ctx); | |
3295 | ||
3296 | return ret; | |
3297 | } | |
3298 | ||
c3a31e60 JA |
3299 | static void io_sqe_file_unregister(struct io_ring_ctx *ctx, int index) |
3300 | { | |
3301 | #if defined(CONFIG_UNIX) | |
3302 | struct file *file = ctx->user_files[index]; | |
3303 | struct sock *sock = ctx->ring_sock->sk; | |
3304 | struct sk_buff_head list, *head = &sock->sk_receive_queue; | |
3305 | struct sk_buff *skb; | |
3306 | int i; | |
3307 | ||
3308 | __skb_queue_head_init(&list); | |
3309 | ||
3310 | /* | |
3311 | * Find the skb that holds this file in its SCM_RIGHTS. When found, | |
3312 | * remove this entry and rearrange the file array. | |
3313 | */ | |
3314 | skb = skb_dequeue(head); | |
3315 | while (skb) { | |
3316 | struct scm_fp_list *fp; | |
3317 | ||
3318 | fp = UNIXCB(skb).fp; | |
3319 | for (i = 0; i < fp->count; i++) { | |
3320 | int left; | |
3321 | ||
3322 | if (fp->fp[i] != file) | |
3323 | continue; | |
3324 | ||
3325 | unix_notinflight(fp->user, fp->fp[i]); | |
3326 | left = fp->count - 1 - i; | |
3327 | if (left) { | |
3328 | memmove(&fp->fp[i], &fp->fp[i + 1], | |
3329 | left * sizeof(struct file *)); | |
3330 | } | |
3331 | fp->count--; | |
3332 | if (!fp->count) { | |
3333 | kfree_skb(skb); | |
3334 | skb = NULL; | |
3335 | } else { | |
3336 | __skb_queue_tail(&list, skb); | |
3337 | } | |
3338 | fput(file); | |
3339 | file = NULL; | |
3340 | break; | |
3341 | } | |
3342 | ||
3343 | if (!file) | |
3344 | break; | |
3345 | ||
3346 | __skb_queue_tail(&list, skb); | |
3347 | ||
3348 | skb = skb_dequeue(head); | |
3349 | } | |
3350 | ||
3351 | if (skb_peek(&list)) { | |
3352 | spin_lock_irq(&head->lock); | |
3353 | while ((skb = __skb_dequeue(&list)) != NULL) | |
3354 | __skb_queue_tail(head, skb); | |
3355 | spin_unlock_irq(&head->lock); | |
3356 | } | |
3357 | #else | |
3358 | fput(ctx->user_files[index]); | |
3359 | #endif | |
3360 | } | |
3361 | ||
3362 | static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file, | |
3363 | int index) | |
3364 | { | |
3365 | #if defined(CONFIG_UNIX) | |
3366 | struct sock *sock = ctx->ring_sock->sk; | |
3367 | struct sk_buff_head *head = &sock->sk_receive_queue; | |
3368 | struct sk_buff *skb; | |
3369 | ||
3370 | /* | |
3371 | * See if we can merge this file into an existing skb SCM_RIGHTS | |
3372 | * file set. If there's no room, fall back to allocating a new skb | |
3373 | * and filling it in. | |
3374 | */ | |
3375 | spin_lock_irq(&head->lock); | |
3376 | skb = skb_peek(head); | |
3377 | if (skb) { | |
3378 | struct scm_fp_list *fpl = UNIXCB(skb).fp; | |
3379 | ||
3380 | if (fpl->count < SCM_MAX_FD) { | |
3381 | __skb_unlink(skb, head); | |
3382 | spin_unlock_irq(&head->lock); | |
3383 | fpl->fp[fpl->count] = get_file(file); | |
3384 | unix_inflight(fpl->user, fpl->fp[fpl->count]); | |
3385 | fpl->count++; | |
3386 | spin_lock_irq(&head->lock); | |
3387 | __skb_queue_head(head, skb); | |
3388 | } else { | |
3389 | skb = NULL; | |
3390 | } | |
3391 | } | |
3392 | spin_unlock_irq(&head->lock); | |
3393 | ||
3394 | if (skb) { | |
3395 | fput(file); | |
3396 | return 0; | |
3397 | } | |
3398 | ||
3399 | return __io_sqe_files_scm(ctx, 1, index); | |
3400 | #else | |
3401 | return 0; | |
3402 | #endif | |
3403 | } | |
3404 | ||
3405 | static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg, | |
3406 | unsigned nr_args) | |
3407 | { | |
3408 | struct io_uring_files_update up; | |
3409 | __s32 __user *fds; | |
3410 | int fd, i, err; | |
3411 | __u32 done; | |
3412 | ||
3413 | if (!ctx->user_files) | |
3414 | return -ENXIO; | |
3415 | if (!nr_args) | |
3416 | return -EINVAL; | |
3417 | if (copy_from_user(&up, arg, sizeof(up))) | |
3418 | return -EFAULT; | |
3419 | if (check_add_overflow(up.offset, nr_args, &done)) | |
3420 | return -EOVERFLOW; | |
3421 | if (done > ctx->nr_user_files) | |
3422 | return -EINVAL; | |
3423 | ||
3424 | done = 0; | |
3425 | fds = (__s32 __user *) up.fds; | |
3426 | while (nr_args) { | |
3427 | err = 0; | |
3428 | if (copy_from_user(&fd, &fds[done], sizeof(fd))) { | |
3429 | err = -EFAULT; | |
3430 | break; | |
3431 | } | |
3432 | i = array_index_nospec(up.offset, ctx->nr_user_files); | |
3433 | if (ctx->user_files[i]) { | |
3434 | io_sqe_file_unregister(ctx, i); | |
3435 | ctx->user_files[i] = NULL; | |
3436 | } | |
3437 | if (fd != -1) { | |
3438 | struct file *file; | |
3439 | ||
3440 | file = fget(fd); | |
3441 | if (!file) { | |
3442 | err = -EBADF; | |
3443 | break; | |
3444 | } | |
3445 | /* | |
3446 | * Don't allow io_uring instances to be registered. If | |
3447 | * UNIX isn't enabled, then this causes a reference | |
3448 | * cycle and this instance can never get freed. If UNIX | |
3449 | * is enabled we'll handle it just fine, but there's | |
3450 | * still no point in allowing a ring fd as it doesn't | |
3451 | * support regular read/write anyway. | |
3452 | */ | |
3453 | if (file->f_op == &io_uring_fops) { | |
3454 | fput(file); | |
3455 | err = -EBADF; | |
3456 | break; | |
3457 | } | |
3458 | ctx->user_files[i] = file; | |
3459 | err = io_sqe_file_register(ctx, file, i); | |
3460 | if (err) | |
3461 | break; | |
3462 | } | |
3463 | nr_args--; | |
3464 | done++; | |
3465 | up.offset++; | |
3466 | } | |
3467 | ||
3468 | return done ? done : err; | |
3469 | } | |
3470 | ||
6c271ce2 JA |
3471 | static int io_sq_offload_start(struct io_ring_ctx *ctx, |
3472 | struct io_uring_params *p) | |
2b188cc1 JA |
3473 | { |
3474 | int ret; | |
3475 | ||
6c271ce2 | 3476 | init_waitqueue_head(&ctx->sqo_wait); |
2b188cc1 JA |
3477 | mmgrab(current->mm); |
3478 | ctx->sqo_mm = current->mm; | |
3479 | ||
6c271ce2 | 3480 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3ec482d1 JA |
3481 | ret = -EPERM; |
3482 | if (!capable(CAP_SYS_ADMIN)) | |
3483 | goto err; | |
3484 | ||
917257da JA |
3485 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
3486 | if (!ctx->sq_thread_idle) | |
3487 | ctx->sq_thread_idle = HZ; | |
3488 | ||
6c271ce2 | 3489 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 3490 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 3491 | |
917257da | 3492 | ret = -EINVAL; |
44a9bd18 JA |
3493 | if (cpu >= nr_cpu_ids) |
3494 | goto err; | |
7889f44d | 3495 | if (!cpu_online(cpu)) |
917257da JA |
3496 | goto err; |
3497 | ||
6c271ce2 JA |
3498 | ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread, |
3499 | ctx, cpu, | |
3500 | "io_uring-sq"); | |
3501 | } else { | |
3502 | ctx->sqo_thread = kthread_create(io_sq_thread, ctx, | |
3503 | "io_uring-sq"); | |
3504 | } | |
3505 | if (IS_ERR(ctx->sqo_thread)) { | |
3506 | ret = PTR_ERR(ctx->sqo_thread); | |
3507 | ctx->sqo_thread = NULL; | |
3508 | goto err; | |
3509 | } | |
3510 | wake_up_process(ctx->sqo_thread); | |
3511 | } else if (p->flags & IORING_SETUP_SQ_AFF) { | |
3512 | /* Can't have SQ_AFF without SQPOLL */ | |
3513 | ret = -EINVAL; | |
3514 | goto err; | |
3515 | } | |
3516 | ||
2b188cc1 | 3517 | /* Do QD, or 2 * CPUS, whatever is smallest */ |
54a91f3b JA |
3518 | ctx->sqo_wq[0] = alloc_workqueue("io_ring-wq", |
3519 | WQ_UNBOUND | WQ_FREEZABLE, | |
2b188cc1 | 3520 | min(ctx->sq_entries - 1, 2 * num_online_cpus())); |
54a91f3b JA |
3521 | if (!ctx->sqo_wq[0]) { |
3522 | ret = -ENOMEM; | |
3523 | goto err; | |
3524 | } | |
3525 | ||
3526 | /* | |
3527 | * This is for buffered writes, where we want to limit the parallelism | |
3528 | * due to file locking in file systems. As "normal" buffered writes | |
3529 | * should parellelize on writeout quite nicely, limit us to having 2 | |
3530 | * pending. This avoids massive contention on the inode when doing | |
3531 | * buffered async writes. | |
3532 | */ | |
3533 | ctx->sqo_wq[1] = alloc_workqueue("io_ring-write-wq", | |
3534 | WQ_UNBOUND | WQ_FREEZABLE, 2); | |
3535 | if (!ctx->sqo_wq[1]) { | |
2b188cc1 JA |
3536 | ret = -ENOMEM; |
3537 | goto err; | |
3538 | } | |
3539 | ||
3540 | return 0; | |
3541 | err: | |
54a91f3b | 3542 | io_finish_async(ctx); |
2b188cc1 JA |
3543 | mmdrop(ctx->sqo_mm); |
3544 | ctx->sqo_mm = NULL; | |
3545 | return ret; | |
3546 | } | |
3547 | ||
3548 | static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages) | |
3549 | { | |
3550 | atomic_long_sub(nr_pages, &user->locked_vm); | |
3551 | } | |
3552 | ||
3553 | static int io_account_mem(struct user_struct *user, unsigned long nr_pages) | |
3554 | { | |
3555 | unsigned long page_limit, cur_pages, new_pages; | |
3556 | ||
3557 | /* Don't allow more pages than we can safely lock */ | |
3558 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
3559 | ||
3560 | do { | |
3561 | cur_pages = atomic_long_read(&user->locked_vm); | |
3562 | new_pages = cur_pages + nr_pages; | |
3563 | if (new_pages > page_limit) | |
3564 | return -ENOMEM; | |
3565 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
3566 | new_pages) != cur_pages); | |
3567 | ||
3568 | return 0; | |
3569 | } | |
3570 | ||
3571 | static void io_mem_free(void *ptr) | |
3572 | { | |
52e04ef4 MR |
3573 | struct page *page; |
3574 | ||
3575 | if (!ptr) | |
3576 | return; | |
2b188cc1 | 3577 | |
52e04ef4 | 3578 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
3579 | if (put_page_testzero(page)) |
3580 | free_compound_page(page); | |
3581 | } | |
3582 | ||
3583 | static void *io_mem_alloc(size_t size) | |
3584 | { | |
3585 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
3586 | __GFP_NORETRY; | |
3587 | ||
3588 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
3589 | } | |
3590 | ||
75b28aff HV |
3591 | static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries, |
3592 | size_t *sq_offset) | |
3593 | { | |
3594 | struct io_rings *rings; | |
3595 | size_t off, sq_array_size; | |
3596 | ||
3597 | off = struct_size(rings, cqes, cq_entries); | |
3598 | if (off == SIZE_MAX) | |
3599 | return SIZE_MAX; | |
3600 | ||
3601 | #ifdef CONFIG_SMP | |
3602 | off = ALIGN(off, SMP_CACHE_BYTES); | |
3603 | if (off == 0) | |
3604 | return SIZE_MAX; | |
3605 | #endif | |
3606 | ||
3607 | sq_array_size = array_size(sizeof(u32), sq_entries); | |
3608 | if (sq_array_size == SIZE_MAX) | |
3609 | return SIZE_MAX; | |
3610 | ||
3611 | if (check_add_overflow(off, sq_array_size, &off)) | |
3612 | return SIZE_MAX; | |
3613 | ||
3614 | if (sq_offset) | |
3615 | *sq_offset = off; | |
3616 | ||
3617 | return off; | |
3618 | } | |
3619 | ||
2b188cc1 JA |
3620 | static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries) |
3621 | { | |
75b28aff | 3622 | size_t pages; |
2b188cc1 | 3623 | |
75b28aff HV |
3624 | pages = (size_t)1 << get_order( |
3625 | rings_size(sq_entries, cq_entries, NULL)); | |
3626 | pages += (size_t)1 << get_order( | |
3627 | array_size(sizeof(struct io_uring_sqe), sq_entries)); | |
2b188cc1 | 3628 | |
75b28aff | 3629 | return pages; |
2b188cc1 JA |
3630 | } |
3631 | ||
edafccee JA |
3632 | static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx) |
3633 | { | |
3634 | int i, j; | |
3635 | ||
3636 | if (!ctx->user_bufs) | |
3637 | return -ENXIO; | |
3638 | ||
3639 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
3640 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
3641 | ||
3642 | for (j = 0; j < imu->nr_bvecs; j++) | |
27c4d3a3 | 3643 | put_user_page(imu->bvec[j].bv_page); |
edafccee JA |
3644 | |
3645 | if (ctx->account_mem) | |
3646 | io_unaccount_mem(ctx->user, imu->nr_bvecs); | |
d4ef6475 | 3647 | kvfree(imu->bvec); |
edafccee JA |
3648 | imu->nr_bvecs = 0; |
3649 | } | |
3650 | ||
3651 | kfree(ctx->user_bufs); | |
3652 | ctx->user_bufs = NULL; | |
3653 | ctx->nr_user_bufs = 0; | |
3654 | return 0; | |
3655 | } | |
3656 | ||
3657 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
3658 | void __user *arg, unsigned index) | |
3659 | { | |
3660 | struct iovec __user *src; | |
3661 | ||
3662 | #ifdef CONFIG_COMPAT | |
3663 | if (ctx->compat) { | |
3664 | struct compat_iovec __user *ciovs; | |
3665 | struct compat_iovec ciov; | |
3666 | ||
3667 | ciovs = (struct compat_iovec __user *) arg; | |
3668 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
3669 | return -EFAULT; | |
3670 | ||
3671 | dst->iov_base = (void __user *) (unsigned long) ciov.iov_base; | |
3672 | dst->iov_len = ciov.iov_len; | |
3673 | return 0; | |
3674 | } | |
3675 | #endif | |
3676 | src = (struct iovec __user *) arg; | |
3677 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
3678 | return -EFAULT; | |
3679 | return 0; | |
3680 | } | |
3681 | ||
3682 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, | |
3683 | unsigned nr_args) | |
3684 | { | |
3685 | struct vm_area_struct **vmas = NULL; | |
3686 | struct page **pages = NULL; | |
3687 | int i, j, got_pages = 0; | |
3688 | int ret = -EINVAL; | |
3689 | ||
3690 | if (ctx->user_bufs) | |
3691 | return -EBUSY; | |
3692 | if (!nr_args || nr_args > UIO_MAXIOV) | |
3693 | return -EINVAL; | |
3694 | ||
3695 | ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf), | |
3696 | GFP_KERNEL); | |
3697 | if (!ctx->user_bufs) | |
3698 | return -ENOMEM; | |
3699 | ||
3700 | for (i = 0; i < nr_args; i++) { | |
3701 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
3702 | unsigned long off, start, end, ubuf; | |
3703 | int pret, nr_pages; | |
3704 | struct iovec iov; | |
3705 | size_t size; | |
3706 | ||
3707 | ret = io_copy_iov(ctx, &iov, arg, i); | |
3708 | if (ret) | |
a278682d | 3709 | goto err; |
edafccee JA |
3710 | |
3711 | /* | |
3712 | * Don't impose further limits on the size and buffer | |
3713 | * constraints here, we'll -EINVAL later when IO is | |
3714 | * submitted if they are wrong. | |
3715 | */ | |
3716 | ret = -EFAULT; | |
3717 | if (!iov.iov_base || !iov.iov_len) | |
3718 | goto err; | |
3719 | ||
3720 | /* arbitrary limit, but we need something */ | |
3721 | if (iov.iov_len > SZ_1G) | |
3722 | goto err; | |
3723 | ||
3724 | ubuf = (unsigned long) iov.iov_base; | |
3725 | end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
3726 | start = ubuf >> PAGE_SHIFT; | |
3727 | nr_pages = end - start; | |
3728 | ||
3729 | if (ctx->account_mem) { | |
3730 | ret = io_account_mem(ctx->user, nr_pages); | |
3731 | if (ret) | |
3732 | goto err; | |
3733 | } | |
3734 | ||
3735 | ret = 0; | |
3736 | if (!pages || nr_pages > got_pages) { | |
3737 | kfree(vmas); | |
3738 | kfree(pages); | |
d4ef6475 | 3739 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), |
edafccee | 3740 | GFP_KERNEL); |
d4ef6475 | 3741 | vmas = kvmalloc_array(nr_pages, |
edafccee JA |
3742 | sizeof(struct vm_area_struct *), |
3743 | GFP_KERNEL); | |
3744 | if (!pages || !vmas) { | |
3745 | ret = -ENOMEM; | |
3746 | if (ctx->account_mem) | |
3747 | io_unaccount_mem(ctx->user, nr_pages); | |
3748 | goto err; | |
3749 | } | |
3750 | got_pages = nr_pages; | |
3751 | } | |
3752 | ||
d4ef6475 | 3753 | imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec), |
edafccee JA |
3754 | GFP_KERNEL); |
3755 | ret = -ENOMEM; | |
3756 | if (!imu->bvec) { | |
3757 | if (ctx->account_mem) | |
3758 | io_unaccount_mem(ctx->user, nr_pages); | |
3759 | goto err; | |
3760 | } | |
3761 | ||
3762 | ret = 0; | |
3763 | down_read(¤t->mm->mmap_sem); | |
932f4a63 IW |
3764 | pret = get_user_pages(ubuf, nr_pages, |
3765 | FOLL_WRITE | FOLL_LONGTERM, | |
3766 | pages, vmas); | |
edafccee JA |
3767 | if (pret == nr_pages) { |
3768 | /* don't support file backed memory */ | |
3769 | for (j = 0; j < nr_pages; j++) { | |
3770 | struct vm_area_struct *vma = vmas[j]; | |
3771 | ||
3772 | if (vma->vm_file && | |
3773 | !is_file_hugepages(vma->vm_file)) { | |
3774 | ret = -EOPNOTSUPP; | |
3775 | break; | |
3776 | } | |
3777 | } | |
3778 | } else { | |
3779 | ret = pret < 0 ? pret : -EFAULT; | |
3780 | } | |
3781 | up_read(¤t->mm->mmap_sem); | |
3782 | if (ret) { | |
3783 | /* | |
3784 | * if we did partial map, or found file backed vmas, | |
3785 | * release any pages we did get | |
3786 | */ | |
27c4d3a3 JH |
3787 | if (pret > 0) |
3788 | put_user_pages(pages, pret); | |
edafccee JA |
3789 | if (ctx->account_mem) |
3790 | io_unaccount_mem(ctx->user, nr_pages); | |
d4ef6475 | 3791 | kvfree(imu->bvec); |
edafccee JA |
3792 | goto err; |
3793 | } | |
3794 | ||
3795 | off = ubuf & ~PAGE_MASK; | |
3796 | size = iov.iov_len; | |
3797 | for (j = 0; j < nr_pages; j++) { | |
3798 | size_t vec_len; | |
3799 | ||
3800 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
3801 | imu->bvec[j].bv_page = pages[j]; | |
3802 | imu->bvec[j].bv_len = vec_len; | |
3803 | imu->bvec[j].bv_offset = off; | |
3804 | off = 0; | |
3805 | size -= vec_len; | |
3806 | } | |
3807 | /* store original address for later verification */ | |
3808 | imu->ubuf = ubuf; | |
3809 | imu->len = iov.iov_len; | |
3810 | imu->nr_bvecs = nr_pages; | |
3811 | ||
3812 | ctx->nr_user_bufs++; | |
3813 | } | |
d4ef6475 MR |
3814 | kvfree(pages); |
3815 | kvfree(vmas); | |
edafccee JA |
3816 | return 0; |
3817 | err: | |
d4ef6475 MR |
3818 | kvfree(pages); |
3819 | kvfree(vmas); | |
edafccee JA |
3820 | io_sqe_buffer_unregister(ctx); |
3821 | return ret; | |
3822 | } | |
3823 | ||
9b402849 JA |
3824 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
3825 | { | |
3826 | __s32 __user *fds = arg; | |
3827 | int fd; | |
3828 | ||
3829 | if (ctx->cq_ev_fd) | |
3830 | return -EBUSY; | |
3831 | ||
3832 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
3833 | return -EFAULT; | |
3834 | ||
3835 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
3836 | if (IS_ERR(ctx->cq_ev_fd)) { | |
3837 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
3838 | ctx->cq_ev_fd = NULL; | |
3839 | return ret; | |
3840 | } | |
3841 | ||
3842 | return 0; | |
3843 | } | |
3844 | ||
3845 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
3846 | { | |
3847 | if (ctx->cq_ev_fd) { | |
3848 | eventfd_ctx_put(ctx->cq_ev_fd); | |
3849 | ctx->cq_ev_fd = NULL; | |
3850 | return 0; | |
3851 | } | |
3852 | ||
3853 | return -ENXIO; | |
3854 | } | |
3855 | ||
2b188cc1 JA |
3856 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
3857 | { | |
6b06314c | 3858 | io_finish_async(ctx); |
2b188cc1 JA |
3859 | if (ctx->sqo_mm) |
3860 | mmdrop(ctx->sqo_mm); | |
def596e9 JA |
3861 | |
3862 | io_iopoll_reap_events(ctx); | |
edafccee | 3863 | io_sqe_buffer_unregister(ctx); |
6b06314c | 3864 | io_sqe_files_unregister(ctx); |
9b402849 | 3865 | io_eventfd_unregister(ctx); |
def596e9 | 3866 | |
2b188cc1 | 3867 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
3868 | if (ctx->ring_sock) { |
3869 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 3870 | sock_release(ctx->ring_sock); |
355e8d26 | 3871 | } |
2b188cc1 JA |
3872 | #endif |
3873 | ||
75b28aff | 3874 | io_mem_free(ctx->rings); |
2b188cc1 | 3875 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
3876 | |
3877 | percpu_ref_exit(&ctx->refs); | |
3878 | if (ctx->account_mem) | |
3879 | io_unaccount_mem(ctx->user, | |
3880 | ring_pages(ctx->sq_entries, ctx->cq_entries)); | |
3881 | free_uid(ctx->user); | |
3882 | kfree(ctx); | |
3883 | } | |
3884 | ||
3885 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
3886 | { | |
3887 | struct io_ring_ctx *ctx = file->private_data; | |
3888 | __poll_t mask = 0; | |
3889 | ||
3890 | poll_wait(file, &ctx->cq_wait, wait); | |
4f7067c3 SB |
3891 | /* |
3892 | * synchronizes with barrier from wq_has_sleeper call in | |
3893 | * io_commit_cqring | |
3894 | */ | |
2b188cc1 | 3895 | smp_rmb(); |
75b28aff HV |
3896 | if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head != |
3897 | ctx->rings->sq_ring_entries) | |
2b188cc1 | 3898 | mask |= EPOLLOUT | EPOLLWRNORM; |
daa5de54 | 3899 | if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail) |
2b188cc1 JA |
3900 | mask |= EPOLLIN | EPOLLRDNORM; |
3901 | ||
3902 | return mask; | |
3903 | } | |
3904 | ||
3905 | static int io_uring_fasync(int fd, struct file *file, int on) | |
3906 | { | |
3907 | struct io_ring_ctx *ctx = file->private_data; | |
3908 | ||
3909 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
3910 | } | |
3911 | ||
3912 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) | |
3913 | { | |
3914 | mutex_lock(&ctx->uring_lock); | |
3915 | percpu_ref_kill(&ctx->refs); | |
3916 | mutex_unlock(&ctx->uring_lock); | |
3917 | ||
5262f567 | 3918 | io_kill_timeouts(ctx); |
221c5eb2 | 3919 | io_poll_remove_all(ctx); |
def596e9 | 3920 | io_iopoll_reap_events(ctx); |
2b188cc1 JA |
3921 | wait_for_completion(&ctx->ctx_done); |
3922 | io_ring_ctx_free(ctx); | |
3923 | } | |
3924 | ||
3925 | static int io_uring_release(struct inode *inode, struct file *file) | |
3926 | { | |
3927 | struct io_ring_ctx *ctx = file->private_data; | |
3928 | ||
3929 | file->private_data = NULL; | |
3930 | io_ring_ctx_wait_and_kill(ctx); | |
3931 | return 0; | |
3932 | } | |
3933 | ||
3934 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
3935 | { | |
3936 | loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT; | |
3937 | unsigned long sz = vma->vm_end - vma->vm_start; | |
3938 | struct io_ring_ctx *ctx = file->private_data; | |
3939 | unsigned long pfn; | |
3940 | struct page *page; | |
3941 | void *ptr; | |
3942 | ||
3943 | switch (offset) { | |
3944 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
3945 | case IORING_OFF_CQ_RING: |
3946 | ptr = ctx->rings; | |
2b188cc1 JA |
3947 | break; |
3948 | case IORING_OFF_SQES: | |
3949 | ptr = ctx->sq_sqes; | |
3950 | break; | |
2b188cc1 JA |
3951 | default: |
3952 | return -EINVAL; | |
3953 | } | |
3954 | ||
3955 | page = virt_to_head_page(ptr); | |
a50b854e | 3956 | if (sz > page_size(page)) |
2b188cc1 JA |
3957 | return -EINVAL; |
3958 | ||
3959 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
3960 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
3961 | } | |
3962 | ||
3963 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, | |
3964 | u32, min_complete, u32, flags, const sigset_t __user *, sig, | |
3965 | size_t, sigsz) | |
3966 | { | |
3967 | struct io_ring_ctx *ctx; | |
3968 | long ret = -EBADF; | |
3969 | int submitted = 0; | |
3970 | struct fd f; | |
3971 | ||
6c271ce2 | 3972 | if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP)) |
2b188cc1 JA |
3973 | return -EINVAL; |
3974 | ||
3975 | f = fdget(fd); | |
3976 | if (!f.file) | |
3977 | return -EBADF; | |
3978 | ||
3979 | ret = -EOPNOTSUPP; | |
3980 | if (f.file->f_op != &io_uring_fops) | |
3981 | goto out_fput; | |
3982 | ||
3983 | ret = -ENXIO; | |
3984 | ctx = f.file->private_data; | |
3985 | if (!percpu_ref_tryget(&ctx->refs)) | |
3986 | goto out_fput; | |
3987 | ||
6c271ce2 JA |
3988 | /* |
3989 | * For SQ polling, the thread will do all submissions and completions. | |
3990 | * Just return the requested submit count, and wake the thread if | |
3991 | * we were asked to. | |
3992 | */ | |
b2a9eada | 3993 | ret = 0; |
6c271ce2 JA |
3994 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3995 | if (flags & IORING_ENTER_SQ_WAKEUP) | |
3996 | wake_up(&ctx->sqo_wait); | |
3997 | submitted = to_submit; | |
b2a9eada | 3998 | } else if (to_submit) { |
2b188cc1 JA |
3999 | to_submit = min(to_submit, ctx->sq_entries); |
4000 | ||
4001 | mutex_lock(&ctx->uring_lock); | |
bc808bce | 4002 | submitted = io_ring_submit(ctx, to_submit); |
2b188cc1 | 4003 | mutex_unlock(&ctx->uring_lock); |
2b188cc1 JA |
4004 | } |
4005 | if (flags & IORING_ENTER_GETEVENTS) { | |
def596e9 JA |
4006 | unsigned nr_events = 0; |
4007 | ||
2b188cc1 JA |
4008 | min_complete = min(min_complete, ctx->cq_entries); |
4009 | ||
def596e9 | 4010 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 | 4011 | ret = io_iopoll_check(ctx, &nr_events, min_complete); |
def596e9 JA |
4012 | } else { |
4013 | ret = io_cqring_wait(ctx, min_complete, sig, sigsz); | |
4014 | } | |
2b188cc1 JA |
4015 | } |
4016 | ||
6805b32e | 4017 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
4018 | out_fput: |
4019 | fdput(f); | |
4020 | return submitted ? submitted : ret; | |
4021 | } | |
4022 | ||
4023 | static const struct file_operations io_uring_fops = { | |
4024 | .release = io_uring_release, | |
4025 | .mmap = io_uring_mmap, | |
4026 | .poll = io_uring_poll, | |
4027 | .fasync = io_uring_fasync, | |
4028 | }; | |
4029 | ||
4030 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
4031 | struct io_uring_params *p) | |
4032 | { | |
75b28aff HV |
4033 | struct io_rings *rings; |
4034 | size_t size, sq_array_offset; | |
2b188cc1 | 4035 | |
75b28aff HV |
4036 | size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset); |
4037 | if (size == SIZE_MAX) | |
4038 | return -EOVERFLOW; | |
4039 | ||
4040 | rings = io_mem_alloc(size); | |
4041 | if (!rings) | |
2b188cc1 JA |
4042 | return -ENOMEM; |
4043 | ||
75b28aff HV |
4044 | ctx->rings = rings; |
4045 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
4046 | rings->sq_ring_mask = p->sq_entries - 1; | |
4047 | rings->cq_ring_mask = p->cq_entries - 1; | |
4048 | rings->sq_ring_entries = p->sq_entries; | |
4049 | rings->cq_ring_entries = p->cq_entries; | |
4050 | ctx->sq_mask = rings->sq_ring_mask; | |
4051 | ctx->cq_mask = rings->cq_ring_mask; | |
4052 | ctx->sq_entries = rings->sq_ring_entries; | |
4053 | ctx->cq_entries = rings->cq_ring_entries; | |
2b188cc1 JA |
4054 | |
4055 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
4056 | if (size == SIZE_MAX) | |
4057 | return -EOVERFLOW; | |
4058 | ||
4059 | ctx->sq_sqes = io_mem_alloc(size); | |
52e04ef4 | 4060 | if (!ctx->sq_sqes) |
2b188cc1 | 4061 | return -ENOMEM; |
2b188cc1 | 4062 | |
2b188cc1 JA |
4063 | return 0; |
4064 | } | |
4065 | ||
4066 | /* | |
4067 | * Allocate an anonymous fd, this is what constitutes the application | |
4068 | * visible backing of an io_uring instance. The application mmaps this | |
4069 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
4070 | * we have to tie this fd to a socket for file garbage collection purposes. | |
4071 | */ | |
4072 | static int io_uring_get_fd(struct io_ring_ctx *ctx) | |
4073 | { | |
4074 | struct file *file; | |
4075 | int ret; | |
4076 | ||
4077 | #if defined(CONFIG_UNIX) | |
4078 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, | |
4079 | &ctx->ring_sock); | |
4080 | if (ret) | |
4081 | return ret; | |
4082 | #endif | |
4083 | ||
4084 | ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
4085 | if (ret < 0) | |
4086 | goto err; | |
4087 | ||
4088 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, | |
4089 | O_RDWR | O_CLOEXEC); | |
4090 | if (IS_ERR(file)) { | |
4091 | put_unused_fd(ret); | |
4092 | ret = PTR_ERR(file); | |
4093 | goto err; | |
4094 | } | |
4095 | ||
4096 | #if defined(CONFIG_UNIX) | |
4097 | ctx->ring_sock->file = file; | |
6b06314c | 4098 | ctx->ring_sock->sk->sk_user_data = ctx; |
2b188cc1 JA |
4099 | #endif |
4100 | fd_install(ret, file); | |
4101 | return ret; | |
4102 | err: | |
4103 | #if defined(CONFIG_UNIX) | |
4104 | sock_release(ctx->ring_sock); | |
4105 | ctx->ring_sock = NULL; | |
4106 | #endif | |
4107 | return ret; | |
4108 | } | |
4109 | ||
4110 | static int io_uring_create(unsigned entries, struct io_uring_params *p) | |
4111 | { | |
4112 | struct user_struct *user = NULL; | |
4113 | struct io_ring_ctx *ctx; | |
4114 | bool account_mem; | |
4115 | int ret; | |
4116 | ||
4117 | if (!entries || entries > IORING_MAX_ENTRIES) | |
4118 | return -EINVAL; | |
4119 | ||
4120 | /* | |
4121 | * Use twice as many entries for the CQ ring. It's possible for the | |
4122 | * application to drive a higher depth than the size of the SQ ring, | |
4123 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
4124 | * some flexibility in overcommitting a bit. If the application has |
4125 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
4126 | * of CQ ring entries manually. | |
2b188cc1 JA |
4127 | */ |
4128 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
4129 | if (p->flags & IORING_SETUP_CQSIZE) { |
4130 | /* | |
4131 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
4132 | * to a power-of-two, if it isn't already. We do NOT impose | |
4133 | * any cq vs sq ring sizing. | |
4134 | */ | |
4135 | if (p->cq_entries < p->sq_entries || p->cq_entries > IORING_MAX_CQ_ENTRIES) | |
4136 | return -EINVAL; | |
4137 | p->cq_entries = roundup_pow_of_two(p->cq_entries); | |
4138 | } else { | |
4139 | p->cq_entries = 2 * p->sq_entries; | |
4140 | } | |
2b188cc1 JA |
4141 | |
4142 | user = get_uid(current_user()); | |
4143 | account_mem = !capable(CAP_IPC_LOCK); | |
4144 | ||
4145 | if (account_mem) { | |
4146 | ret = io_account_mem(user, | |
4147 | ring_pages(p->sq_entries, p->cq_entries)); | |
4148 | if (ret) { | |
4149 | free_uid(user); | |
4150 | return ret; | |
4151 | } | |
4152 | } | |
4153 | ||
4154 | ctx = io_ring_ctx_alloc(p); | |
4155 | if (!ctx) { | |
4156 | if (account_mem) | |
4157 | io_unaccount_mem(user, ring_pages(p->sq_entries, | |
4158 | p->cq_entries)); | |
4159 | free_uid(user); | |
4160 | return -ENOMEM; | |
4161 | } | |
4162 | ctx->compat = in_compat_syscall(); | |
4163 | ctx->account_mem = account_mem; | |
4164 | ctx->user = user; | |
4165 | ||
4166 | ret = io_allocate_scq_urings(ctx, p); | |
4167 | if (ret) | |
4168 | goto err; | |
4169 | ||
6c271ce2 | 4170 | ret = io_sq_offload_start(ctx, p); |
2b188cc1 JA |
4171 | if (ret) |
4172 | goto err; | |
4173 | ||
2b188cc1 | 4174 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
4175 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
4176 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
4177 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
4178 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
4179 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
4180 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
4181 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
4182 | |
4183 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
4184 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
4185 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
4186 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
4187 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
4188 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
4189 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
ac90f249 | 4190 | |
044c1ab3 JA |
4191 | /* |
4192 | * Install ring fd as the very last thing, so we don't risk someone | |
4193 | * having closed it before we finish setup | |
4194 | */ | |
4195 | ret = io_uring_get_fd(ctx); | |
4196 | if (ret < 0) | |
4197 | goto err; | |
4198 | ||
ac90f249 | 4199 | p->features = IORING_FEAT_SINGLE_MMAP; |
2b188cc1 JA |
4200 | return ret; |
4201 | err: | |
4202 | io_ring_ctx_wait_and_kill(ctx); | |
4203 | return ret; | |
4204 | } | |
4205 | ||
4206 | /* | |
4207 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
4208 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
4209 | * params structure passed in. | |
4210 | */ | |
4211 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
4212 | { | |
4213 | struct io_uring_params p; | |
4214 | long ret; | |
4215 | int i; | |
4216 | ||
4217 | if (copy_from_user(&p, params, sizeof(p))) | |
4218 | return -EFAULT; | |
4219 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
4220 | if (p.resv[i]) | |
4221 | return -EINVAL; | |
4222 | } | |
4223 | ||
6c271ce2 | 4224 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
33a107f0 | 4225 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE)) |
2b188cc1 JA |
4226 | return -EINVAL; |
4227 | ||
4228 | ret = io_uring_create(entries, &p); | |
4229 | if (ret < 0) | |
4230 | return ret; | |
4231 | ||
4232 | if (copy_to_user(params, &p, sizeof(p))) | |
4233 | return -EFAULT; | |
4234 | ||
4235 | return ret; | |
4236 | } | |
4237 | ||
4238 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
4239 | struct io_uring_params __user *, params) | |
4240 | { | |
4241 | return io_uring_setup(entries, params); | |
4242 | } | |
4243 | ||
edafccee JA |
4244 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
4245 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
4246 | __releases(ctx->uring_lock) |
4247 | __acquires(ctx->uring_lock) | |
edafccee JA |
4248 | { |
4249 | int ret; | |
4250 | ||
35fa71a0 JA |
4251 | /* |
4252 | * We're inside the ring mutex, if the ref is already dying, then | |
4253 | * someone else killed the ctx or is already going through | |
4254 | * io_uring_register(). | |
4255 | */ | |
4256 | if (percpu_ref_is_dying(&ctx->refs)) | |
4257 | return -ENXIO; | |
4258 | ||
edafccee | 4259 | percpu_ref_kill(&ctx->refs); |
b19062a5 JA |
4260 | |
4261 | /* | |
4262 | * Drop uring mutex before waiting for references to exit. If another | |
4263 | * thread is currently inside io_uring_enter() it might need to grab | |
4264 | * the uring_lock to make progress. If we hold it here across the drain | |
4265 | * wait, then we can deadlock. It's safe to drop the mutex here, since | |
4266 | * no new references will come in after we've killed the percpu ref. | |
4267 | */ | |
4268 | mutex_unlock(&ctx->uring_lock); | |
edafccee | 4269 | wait_for_completion(&ctx->ctx_done); |
b19062a5 | 4270 | mutex_lock(&ctx->uring_lock); |
edafccee JA |
4271 | |
4272 | switch (opcode) { | |
4273 | case IORING_REGISTER_BUFFERS: | |
4274 | ret = io_sqe_buffer_register(ctx, arg, nr_args); | |
4275 | break; | |
4276 | case IORING_UNREGISTER_BUFFERS: | |
4277 | ret = -EINVAL; | |
4278 | if (arg || nr_args) | |
4279 | break; | |
4280 | ret = io_sqe_buffer_unregister(ctx); | |
4281 | break; | |
6b06314c JA |
4282 | case IORING_REGISTER_FILES: |
4283 | ret = io_sqe_files_register(ctx, arg, nr_args); | |
4284 | break; | |
4285 | case IORING_UNREGISTER_FILES: | |
4286 | ret = -EINVAL; | |
4287 | if (arg || nr_args) | |
4288 | break; | |
4289 | ret = io_sqe_files_unregister(ctx); | |
4290 | break; | |
c3a31e60 JA |
4291 | case IORING_REGISTER_FILES_UPDATE: |
4292 | ret = io_sqe_files_update(ctx, arg, nr_args); | |
4293 | break; | |
9b402849 JA |
4294 | case IORING_REGISTER_EVENTFD: |
4295 | ret = -EINVAL; | |
4296 | if (nr_args != 1) | |
4297 | break; | |
4298 | ret = io_eventfd_register(ctx, arg); | |
4299 | break; | |
4300 | case IORING_UNREGISTER_EVENTFD: | |
4301 | ret = -EINVAL; | |
4302 | if (arg || nr_args) | |
4303 | break; | |
4304 | ret = io_eventfd_unregister(ctx); | |
4305 | break; | |
edafccee JA |
4306 | default: |
4307 | ret = -EINVAL; | |
4308 | break; | |
4309 | } | |
4310 | ||
4311 | /* bring the ctx back to life */ | |
4312 | reinit_completion(&ctx->ctx_done); | |
4313 | percpu_ref_reinit(&ctx->refs); | |
4314 | return ret; | |
4315 | } | |
4316 | ||
4317 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
4318 | void __user *, arg, unsigned int, nr_args) | |
4319 | { | |
4320 | struct io_ring_ctx *ctx; | |
4321 | long ret = -EBADF; | |
4322 | struct fd f; | |
4323 | ||
4324 | f = fdget(fd); | |
4325 | if (!f.file) | |
4326 | return -EBADF; | |
4327 | ||
4328 | ret = -EOPNOTSUPP; | |
4329 | if (f.file->f_op != &io_uring_fops) | |
4330 | goto out_fput; | |
4331 | ||
4332 | ctx = f.file->private_data; | |
4333 | ||
4334 | mutex_lock(&ctx->uring_lock); | |
4335 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
4336 | mutex_unlock(&ctx->uring_lock); | |
4337 | out_fput: | |
4338 | fdput(f); | |
4339 | return ret; | |
4340 | } | |
4341 | ||
2b188cc1 JA |
4342 | static int __init io_uring_init(void) |
4343 | { | |
4344 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC); | |
4345 | return 0; | |
4346 | }; | |
4347 | __initcall(io_uring_init); |