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 |
6b06314c | 79 | #define IORING_MAX_FIXED_FILES 1024 |
2b188cc1 JA |
80 | |
81 | struct io_uring { | |
82 | u32 head ____cacheline_aligned_in_smp; | |
83 | u32 tail ____cacheline_aligned_in_smp; | |
84 | }; | |
85 | ||
1e84b97b | 86 | /* |
75b28aff HV |
87 | * This data is shared with the application through the mmap at offsets |
88 | * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING. | |
1e84b97b SB |
89 | * |
90 | * The offsets to the member fields are published through struct | |
91 | * io_sqring_offsets when calling io_uring_setup. | |
92 | */ | |
75b28aff | 93 | struct io_rings { |
1e84b97b SB |
94 | /* |
95 | * Head and tail offsets into the ring; the offsets need to be | |
96 | * masked to get valid indices. | |
97 | * | |
75b28aff HV |
98 | * The kernel controls head of the sq ring and the tail of the cq ring, |
99 | * and the application controls tail of the sq ring and the head of the | |
100 | * cq ring. | |
1e84b97b | 101 | */ |
75b28aff | 102 | struct io_uring sq, cq; |
1e84b97b | 103 | /* |
75b28aff | 104 | * Bitmasks to apply to head and tail offsets (constant, equals |
1e84b97b SB |
105 | * ring_entries - 1) |
106 | */ | |
75b28aff HV |
107 | u32 sq_ring_mask, cq_ring_mask; |
108 | /* Ring sizes (constant, power of 2) */ | |
109 | u32 sq_ring_entries, cq_ring_entries; | |
1e84b97b SB |
110 | /* |
111 | * Number of invalid entries dropped by the kernel due to | |
112 | * invalid index stored in array | |
113 | * | |
114 | * Written by the kernel, shouldn't be modified by the | |
115 | * application (i.e. get number of "new events" by comparing to | |
116 | * cached value). | |
117 | * | |
118 | * After a new SQ head value was read by the application this | |
119 | * counter includes all submissions that were dropped reaching | |
120 | * the new SQ head (and possibly more). | |
121 | */ | |
75b28aff | 122 | u32 sq_dropped; |
1e84b97b SB |
123 | /* |
124 | * Runtime flags | |
125 | * | |
126 | * Written by the kernel, shouldn't be modified by the | |
127 | * application. | |
128 | * | |
129 | * The application needs a full memory barrier before checking | |
130 | * for IORING_SQ_NEED_WAKEUP after updating the sq tail. | |
131 | */ | |
75b28aff | 132 | u32 sq_flags; |
1e84b97b SB |
133 | /* |
134 | * Number of completion events lost because the queue was full; | |
135 | * this should be avoided by the application by making sure | |
136 | * there are not more requests pending thatn there is space in | |
137 | * the completion queue. | |
138 | * | |
139 | * Written by the kernel, shouldn't be modified by the | |
140 | * application (i.e. get number of "new events" by comparing to | |
141 | * cached value). | |
142 | * | |
143 | * As completion events come in out of order this counter is not | |
144 | * ordered with any other data. | |
145 | */ | |
75b28aff | 146 | u32 cq_overflow; |
1e84b97b SB |
147 | /* |
148 | * Ring buffer of completion events. | |
149 | * | |
150 | * The kernel writes completion events fresh every time they are | |
151 | * produced, so the application is allowed to modify pending | |
152 | * entries. | |
153 | */ | |
75b28aff | 154 | struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp; |
2b188cc1 JA |
155 | }; |
156 | ||
edafccee JA |
157 | struct io_mapped_ubuf { |
158 | u64 ubuf; | |
159 | size_t len; | |
160 | struct bio_vec *bvec; | |
161 | unsigned int nr_bvecs; | |
162 | }; | |
163 | ||
31b51510 JA |
164 | struct async_list { |
165 | spinlock_t lock; | |
166 | atomic_t cnt; | |
167 | struct list_head list; | |
168 | ||
169 | struct file *file; | |
6d5d5ac5 | 170 | off_t io_start; |
9310a7ba | 171 | size_t io_len; |
31b51510 JA |
172 | }; |
173 | ||
2b188cc1 JA |
174 | struct io_ring_ctx { |
175 | struct { | |
176 | struct percpu_ref refs; | |
177 | } ____cacheline_aligned_in_smp; | |
178 | ||
179 | struct { | |
180 | unsigned int flags; | |
181 | bool compat; | |
182 | bool account_mem; | |
183 | ||
75b28aff HV |
184 | /* |
185 | * Ring buffer of indices into array of io_uring_sqe, which is | |
186 | * mmapped by the application using the IORING_OFF_SQES offset. | |
187 | * | |
188 | * This indirection could e.g. be used to assign fixed | |
189 | * io_uring_sqe entries to operations and only submit them to | |
190 | * the queue when needed. | |
191 | * | |
192 | * The kernel modifies neither the indices array nor the entries | |
193 | * array. | |
194 | */ | |
195 | u32 *sq_array; | |
2b188cc1 JA |
196 | unsigned cached_sq_head; |
197 | unsigned sq_entries; | |
198 | unsigned sq_mask; | |
6c271ce2 | 199 | unsigned sq_thread_idle; |
498ccd9e | 200 | unsigned cached_sq_dropped; |
2b188cc1 | 201 | struct io_uring_sqe *sq_sqes; |
de0617e4 JA |
202 | |
203 | struct list_head defer_list; | |
5262f567 | 204 | struct list_head timeout_list; |
2b188cc1 JA |
205 | } ____cacheline_aligned_in_smp; |
206 | ||
207 | /* IO offload */ | |
54a91f3b | 208 | struct workqueue_struct *sqo_wq[2]; |
6c271ce2 | 209 | struct task_struct *sqo_thread; /* if using sq thread polling */ |
2b188cc1 | 210 | struct mm_struct *sqo_mm; |
6c271ce2 | 211 | wait_queue_head_t sqo_wait; |
a4c0b3de | 212 | struct completion sqo_thread_started; |
2b188cc1 JA |
213 | |
214 | struct { | |
2b188cc1 | 215 | unsigned cached_cq_tail; |
498ccd9e | 216 | atomic_t cached_cq_overflow; |
2b188cc1 JA |
217 | unsigned cq_entries; |
218 | unsigned cq_mask; | |
219 | struct wait_queue_head cq_wait; | |
220 | struct fasync_struct *cq_fasync; | |
9b402849 | 221 | struct eventfd_ctx *cq_ev_fd; |
5262f567 | 222 | atomic_t cq_timeouts; |
2b188cc1 JA |
223 | } ____cacheline_aligned_in_smp; |
224 | ||
75b28aff HV |
225 | struct io_rings *rings; |
226 | ||
6b06314c JA |
227 | /* |
228 | * If used, fixed file set. Writers must ensure that ->refs is dead, | |
229 | * readers must ensure that ->refs is alive as long as the file* is | |
230 | * used. Only updated through io_uring_register(2). | |
231 | */ | |
232 | struct file **user_files; | |
233 | unsigned nr_user_files; | |
234 | ||
edafccee JA |
235 | /* if used, fixed mapped user buffers */ |
236 | unsigned nr_user_bufs; | |
237 | struct io_mapped_ubuf *user_bufs; | |
238 | ||
2b188cc1 JA |
239 | struct user_struct *user; |
240 | ||
241 | struct completion ctx_done; | |
242 | ||
243 | struct { | |
244 | struct mutex uring_lock; | |
245 | wait_queue_head_t wait; | |
246 | } ____cacheline_aligned_in_smp; | |
247 | ||
248 | struct { | |
249 | spinlock_t completion_lock; | |
def596e9 JA |
250 | bool poll_multi_file; |
251 | /* | |
252 | * ->poll_list is protected by the ctx->uring_lock for | |
253 | * io_uring instances that don't use IORING_SETUP_SQPOLL. | |
254 | * For SQPOLL, only the single threaded io_sq_thread() will | |
255 | * manipulate the list, hence no extra locking is needed there. | |
256 | */ | |
257 | struct list_head poll_list; | |
221c5eb2 | 258 | struct list_head cancel_list; |
2b188cc1 JA |
259 | } ____cacheline_aligned_in_smp; |
260 | ||
31b51510 JA |
261 | struct async_list pending_async[2]; |
262 | ||
2b188cc1 JA |
263 | #if defined(CONFIG_UNIX) |
264 | struct socket *ring_sock; | |
265 | #endif | |
266 | }; | |
267 | ||
268 | struct sqe_submit { | |
269 | const struct io_uring_sqe *sqe; | |
270 | unsigned short index; | |
8776f3fa | 271 | u32 sequence; |
2b188cc1 | 272 | bool has_user; |
def596e9 | 273 | bool needs_lock; |
6c271ce2 | 274 | bool needs_fixed_file; |
2b188cc1 JA |
275 | }; |
276 | ||
09bb8394 JA |
277 | /* |
278 | * First field must be the file pointer in all the | |
279 | * iocb unions! See also 'struct kiocb' in <linux/fs.h> | |
280 | */ | |
221c5eb2 JA |
281 | struct io_poll_iocb { |
282 | struct file *file; | |
283 | struct wait_queue_head *head; | |
284 | __poll_t events; | |
8c838788 | 285 | bool done; |
221c5eb2 JA |
286 | bool canceled; |
287 | struct wait_queue_entry wait; | |
288 | }; | |
289 | ||
5262f567 JA |
290 | struct io_timeout { |
291 | struct file *file; | |
292 | struct hrtimer timer; | |
293 | }; | |
294 | ||
09bb8394 JA |
295 | /* |
296 | * NOTE! Each of the iocb union members has the file pointer | |
297 | * as the first entry in their struct definition. So you can | |
298 | * access the file pointer through any of the sub-structs, | |
299 | * or directly as just 'ki_filp' in this struct. | |
300 | */ | |
2b188cc1 | 301 | struct io_kiocb { |
221c5eb2 | 302 | union { |
09bb8394 | 303 | struct file *file; |
221c5eb2 JA |
304 | struct kiocb rw; |
305 | struct io_poll_iocb poll; | |
5262f567 | 306 | struct io_timeout timeout; |
221c5eb2 | 307 | }; |
2b188cc1 JA |
308 | |
309 | struct sqe_submit submit; | |
310 | ||
311 | struct io_ring_ctx *ctx; | |
312 | struct list_head list; | |
9e645e11 | 313 | struct list_head link_list; |
2b188cc1 | 314 | unsigned int flags; |
c16361c1 | 315 | refcount_t refs; |
8449eeda | 316 | #define REQ_F_NOWAIT 1 /* must not punt to workers */ |
def596e9 | 317 | #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */ |
6b06314c | 318 | #define REQ_F_FIXED_FILE 4 /* ctx owns file */ |
31b51510 | 319 | #define REQ_F_SEQ_PREV 8 /* sequential with previous */ |
e2033e33 SB |
320 | #define REQ_F_IO_DRAIN 16 /* drain existing IO first */ |
321 | #define REQ_F_IO_DRAINED 32 /* drain done */ | |
9e645e11 | 322 | #define REQ_F_LINK 64 /* linked sqes */ |
f7b76ac9 ZL |
323 | #define REQ_F_LINK_DONE 128 /* linked sqes done */ |
324 | #define REQ_F_FAIL_LINK 256 /* fail rest of links */ | |
4fe2c963 | 325 | #define REQ_F_SHADOW_DRAIN 512 /* link-drain shadow req */ |
5262f567 | 326 | #define REQ_F_TIMEOUT 1024 /* timeout request */ |
491381ce JA |
327 | #define REQ_F_ISREG 2048 /* regular file */ |
328 | #define REQ_F_MUST_PUNT 4096 /* must be punted even for NONBLOCK */ | |
2b188cc1 | 329 | u64 user_data; |
9e645e11 | 330 | u32 result; |
de0617e4 | 331 | u32 sequence; |
2b188cc1 JA |
332 | |
333 | struct work_struct work; | |
334 | }; | |
335 | ||
336 | #define IO_PLUG_THRESHOLD 2 | |
def596e9 | 337 | #define IO_IOPOLL_BATCH 8 |
2b188cc1 | 338 | |
9a56a232 JA |
339 | struct io_submit_state { |
340 | struct blk_plug plug; | |
341 | ||
2579f913 JA |
342 | /* |
343 | * io_kiocb alloc cache | |
344 | */ | |
345 | void *reqs[IO_IOPOLL_BATCH]; | |
346 | unsigned int free_reqs; | |
347 | unsigned int cur_req; | |
348 | ||
9a56a232 JA |
349 | /* |
350 | * File reference cache | |
351 | */ | |
352 | struct file *file; | |
353 | unsigned int fd; | |
354 | unsigned int has_refs; | |
355 | unsigned int used_refs; | |
356 | unsigned int ios_left; | |
357 | }; | |
358 | ||
de0617e4 | 359 | static void io_sq_wq_submit_work(struct work_struct *work); |
5262f567 JA |
360 | static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data, |
361 | long res); | |
4fe2c963 | 362 | static void __io_free_req(struct io_kiocb *req); |
de0617e4 | 363 | |
2b188cc1 JA |
364 | static struct kmem_cache *req_cachep; |
365 | ||
366 | static const struct file_operations io_uring_fops; | |
367 | ||
368 | struct sock *io_uring_get_socket(struct file *file) | |
369 | { | |
370 | #if defined(CONFIG_UNIX) | |
371 | if (file->f_op == &io_uring_fops) { | |
372 | struct io_ring_ctx *ctx = file->private_data; | |
373 | ||
374 | return ctx->ring_sock->sk; | |
375 | } | |
376 | #endif | |
377 | return NULL; | |
378 | } | |
379 | EXPORT_SYMBOL(io_uring_get_socket); | |
380 | ||
381 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) | |
382 | { | |
383 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
384 | ||
385 | complete(&ctx->ctx_done); | |
386 | } | |
387 | ||
388 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) | |
389 | { | |
390 | struct io_ring_ctx *ctx; | |
31b51510 | 391 | int i; |
2b188cc1 JA |
392 | |
393 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
394 | if (!ctx) | |
395 | return NULL; | |
396 | ||
21482896 RG |
397 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
398 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) { | |
2b188cc1 JA |
399 | kfree(ctx); |
400 | return NULL; | |
401 | } | |
402 | ||
403 | ctx->flags = p->flags; | |
404 | init_waitqueue_head(&ctx->cq_wait); | |
405 | init_completion(&ctx->ctx_done); | |
a4c0b3de | 406 | init_completion(&ctx->sqo_thread_started); |
2b188cc1 JA |
407 | mutex_init(&ctx->uring_lock); |
408 | init_waitqueue_head(&ctx->wait); | |
31b51510 JA |
409 | for (i = 0; i < ARRAY_SIZE(ctx->pending_async); i++) { |
410 | spin_lock_init(&ctx->pending_async[i].lock); | |
411 | INIT_LIST_HEAD(&ctx->pending_async[i].list); | |
412 | atomic_set(&ctx->pending_async[i].cnt, 0); | |
413 | } | |
2b188cc1 | 414 | spin_lock_init(&ctx->completion_lock); |
def596e9 | 415 | INIT_LIST_HEAD(&ctx->poll_list); |
221c5eb2 | 416 | INIT_LIST_HEAD(&ctx->cancel_list); |
de0617e4 | 417 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 418 | INIT_LIST_HEAD(&ctx->timeout_list); |
2b188cc1 JA |
419 | return ctx; |
420 | } | |
421 | ||
7adf4eaf JA |
422 | static inline bool __io_sequence_defer(struct io_ring_ctx *ctx, |
423 | struct io_kiocb *req) | |
424 | { | |
498ccd9e JA |
425 | return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped |
426 | + atomic_read(&ctx->cached_cq_overflow); | |
7adf4eaf JA |
427 | } |
428 | ||
de0617e4 JA |
429 | static inline bool io_sequence_defer(struct io_ring_ctx *ctx, |
430 | struct io_kiocb *req) | |
431 | { | |
7adf4eaf | 432 | if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) != REQ_F_IO_DRAIN) |
de0617e4 JA |
433 | return false; |
434 | ||
7adf4eaf | 435 | return __io_sequence_defer(ctx, req); |
de0617e4 JA |
436 | } |
437 | ||
7adf4eaf | 438 | static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx) |
de0617e4 JA |
439 | { |
440 | struct io_kiocb *req; | |
441 | ||
7adf4eaf JA |
442 | req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list); |
443 | if (req && !io_sequence_defer(ctx, req)) { | |
de0617e4 JA |
444 | list_del_init(&req->list); |
445 | return req; | |
446 | } | |
447 | ||
448 | return NULL; | |
449 | } | |
450 | ||
5262f567 JA |
451 | static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx) |
452 | { | |
7adf4eaf JA |
453 | struct io_kiocb *req; |
454 | ||
455 | req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list); | |
456 | if (req && !__io_sequence_defer(ctx, req)) { | |
457 | list_del_init(&req->list); | |
458 | return req; | |
459 | } | |
460 | ||
461 | return NULL; | |
5262f567 JA |
462 | } |
463 | ||
de0617e4 | 464 | static void __io_commit_cqring(struct io_ring_ctx *ctx) |
2b188cc1 | 465 | { |
75b28aff | 466 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 467 | |
75b28aff | 468 | if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) { |
2b188cc1 | 469 | /* order cqe stores with ring update */ |
75b28aff | 470 | smp_store_release(&rings->cq.tail, ctx->cached_cq_tail); |
2b188cc1 | 471 | |
2b188cc1 JA |
472 | if (wq_has_sleeper(&ctx->cq_wait)) { |
473 | wake_up_interruptible(&ctx->cq_wait); | |
474 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
475 | } | |
476 | } | |
477 | } | |
478 | ||
18d9be1a JA |
479 | static inline void io_queue_async_work(struct io_ring_ctx *ctx, |
480 | struct io_kiocb *req) | |
481 | { | |
6cc47d1d | 482 | int rw = 0; |
54a91f3b | 483 | |
6cc47d1d JA |
484 | if (req->submit.sqe) { |
485 | switch (req->submit.sqe->opcode) { | |
486 | case IORING_OP_WRITEV: | |
487 | case IORING_OP_WRITE_FIXED: | |
488 | rw = !(req->rw.ki_flags & IOCB_DIRECT); | |
489 | break; | |
490 | } | |
54a91f3b JA |
491 | } |
492 | ||
493 | queue_work(ctx->sqo_wq[rw], &req->work); | |
18d9be1a JA |
494 | } |
495 | ||
5262f567 JA |
496 | static void io_kill_timeout(struct io_kiocb *req) |
497 | { | |
498 | int ret; | |
499 | ||
500 | ret = hrtimer_try_to_cancel(&req->timeout.timer); | |
501 | if (ret != -1) { | |
502 | atomic_inc(&req->ctx->cq_timeouts); | |
503 | list_del(&req->list); | |
504 | io_cqring_fill_event(req->ctx, req->user_data, 0); | |
505 | __io_free_req(req); | |
506 | } | |
507 | } | |
508 | ||
509 | static void io_kill_timeouts(struct io_ring_ctx *ctx) | |
510 | { | |
511 | struct io_kiocb *req, *tmp; | |
512 | ||
513 | spin_lock_irq(&ctx->completion_lock); | |
514 | list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list) | |
515 | io_kill_timeout(req); | |
516 | spin_unlock_irq(&ctx->completion_lock); | |
517 | } | |
518 | ||
de0617e4 JA |
519 | static void io_commit_cqring(struct io_ring_ctx *ctx) |
520 | { | |
521 | struct io_kiocb *req; | |
522 | ||
5262f567 JA |
523 | while ((req = io_get_timeout_req(ctx)) != NULL) |
524 | io_kill_timeout(req); | |
525 | ||
de0617e4 JA |
526 | __io_commit_cqring(ctx); |
527 | ||
528 | while ((req = io_get_deferred_req(ctx)) != NULL) { | |
4fe2c963 JL |
529 | if (req->flags & REQ_F_SHADOW_DRAIN) { |
530 | /* Just for drain, free it. */ | |
531 | __io_free_req(req); | |
532 | continue; | |
533 | } | |
de0617e4 | 534 | req->flags |= REQ_F_IO_DRAINED; |
18d9be1a | 535 | io_queue_async_work(ctx, req); |
de0617e4 JA |
536 | } |
537 | } | |
538 | ||
2b188cc1 JA |
539 | static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx) |
540 | { | |
75b28aff | 541 | struct io_rings *rings = ctx->rings; |
2b188cc1 JA |
542 | unsigned tail; |
543 | ||
544 | tail = ctx->cached_cq_tail; | |
115e12e5 SB |
545 | /* |
546 | * writes to the cq entry need to come after reading head; the | |
547 | * control dependency is enough as we're using WRITE_ONCE to | |
548 | * fill the cq entry | |
549 | */ | |
75b28aff | 550 | if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries) |
2b188cc1 JA |
551 | return NULL; |
552 | ||
553 | ctx->cached_cq_tail++; | |
75b28aff | 554 | return &rings->cqes[tail & ctx->cq_mask]; |
2b188cc1 JA |
555 | } |
556 | ||
557 | static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data, | |
c71ffb67 | 558 | long res) |
2b188cc1 JA |
559 | { |
560 | struct io_uring_cqe *cqe; | |
561 | ||
562 | /* | |
563 | * If we can't get a cq entry, userspace overflowed the | |
564 | * submission (by quite a lot). Increment the overflow count in | |
565 | * the ring. | |
566 | */ | |
567 | cqe = io_get_cqring(ctx); | |
568 | if (cqe) { | |
569 | WRITE_ONCE(cqe->user_data, ki_user_data); | |
570 | WRITE_ONCE(cqe->res, res); | |
c71ffb67 | 571 | WRITE_ONCE(cqe->flags, 0); |
2b188cc1 | 572 | } else { |
498ccd9e JA |
573 | WRITE_ONCE(ctx->rings->cq_overflow, |
574 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
2b188cc1 JA |
575 | } |
576 | } | |
577 | ||
8c838788 JA |
578 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
579 | { | |
580 | if (waitqueue_active(&ctx->wait)) | |
581 | wake_up(&ctx->wait); | |
582 | if (waitqueue_active(&ctx->sqo_wait)) | |
583 | wake_up(&ctx->sqo_wait); | |
9b402849 JA |
584 | if (ctx->cq_ev_fd) |
585 | eventfd_signal(ctx->cq_ev_fd, 1); | |
8c838788 JA |
586 | } |
587 | ||
588 | static void io_cqring_add_event(struct io_ring_ctx *ctx, u64 user_data, | |
c71ffb67 | 589 | long res) |
2b188cc1 JA |
590 | { |
591 | unsigned long flags; | |
592 | ||
593 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
c71ffb67 | 594 | io_cqring_fill_event(ctx, user_data, res); |
2b188cc1 JA |
595 | io_commit_cqring(ctx); |
596 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
597 | ||
8c838788 | 598 | io_cqring_ev_posted(ctx); |
2b188cc1 JA |
599 | } |
600 | ||
2579f913 JA |
601 | static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx, |
602 | struct io_submit_state *state) | |
2b188cc1 | 603 | { |
fd6fab2c | 604 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
2b188cc1 JA |
605 | struct io_kiocb *req; |
606 | ||
607 | if (!percpu_ref_tryget(&ctx->refs)) | |
608 | return NULL; | |
609 | ||
2579f913 | 610 | if (!state) { |
fd6fab2c | 611 | req = kmem_cache_alloc(req_cachep, gfp); |
2579f913 JA |
612 | if (unlikely(!req)) |
613 | goto out; | |
614 | } else if (!state->free_reqs) { | |
615 | size_t sz; | |
616 | int ret; | |
617 | ||
618 | sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs)); | |
fd6fab2c JA |
619 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs); |
620 | ||
621 | /* | |
622 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
623 | * retry single alloc to be on the safe side. | |
624 | */ | |
625 | if (unlikely(ret <= 0)) { | |
626 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
627 | if (!state->reqs[0]) | |
628 | goto out; | |
629 | ret = 1; | |
630 | } | |
2579f913 JA |
631 | state->free_reqs = ret - 1; |
632 | state->cur_req = 1; | |
633 | req = state->reqs[0]; | |
634 | } else { | |
635 | req = state->reqs[state->cur_req]; | |
636 | state->free_reqs--; | |
637 | state->cur_req++; | |
2b188cc1 JA |
638 | } |
639 | ||
60c112b0 | 640 | req->file = NULL; |
2579f913 JA |
641 | req->ctx = ctx; |
642 | req->flags = 0; | |
e65ef56d JA |
643 | /* one is dropped after submission, the other at completion */ |
644 | refcount_set(&req->refs, 2); | |
9e645e11 | 645 | req->result = 0; |
2579f913 JA |
646 | return req; |
647 | out: | |
6805b32e | 648 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
649 | return NULL; |
650 | } | |
651 | ||
def596e9 JA |
652 | static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr) |
653 | { | |
654 | if (*nr) { | |
655 | kmem_cache_free_bulk(req_cachep, *nr, reqs); | |
6805b32e | 656 | percpu_ref_put_many(&ctx->refs, *nr); |
def596e9 JA |
657 | *nr = 0; |
658 | } | |
659 | } | |
660 | ||
9e645e11 | 661 | static void __io_free_req(struct io_kiocb *req) |
2b188cc1 | 662 | { |
09bb8394 JA |
663 | if (req->file && !(req->flags & REQ_F_FIXED_FILE)) |
664 | fput(req->file); | |
6805b32e | 665 | percpu_ref_put(&req->ctx->refs); |
e65ef56d JA |
666 | kmem_cache_free(req_cachep, req); |
667 | } | |
668 | ||
9e645e11 JA |
669 | static void io_req_link_next(struct io_kiocb *req) |
670 | { | |
671 | struct io_kiocb *nxt; | |
672 | ||
673 | /* | |
674 | * The list should never be empty when we are called here. But could | |
675 | * potentially happen if the chain is messed up, check to be on the | |
676 | * safe side. | |
677 | */ | |
678 | nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list); | |
679 | if (nxt) { | |
680 | list_del(&nxt->list); | |
681 | if (!list_empty(&req->link_list)) { | |
682 | INIT_LIST_HEAD(&nxt->link_list); | |
683 | list_splice(&req->link_list, &nxt->link_list); | |
684 | nxt->flags |= REQ_F_LINK; | |
685 | } | |
686 | ||
f7b76ac9 | 687 | nxt->flags |= REQ_F_LINK_DONE; |
9e645e11 | 688 | INIT_WORK(&nxt->work, io_sq_wq_submit_work); |
18d9be1a | 689 | io_queue_async_work(req->ctx, nxt); |
9e645e11 JA |
690 | } |
691 | } | |
692 | ||
693 | /* | |
694 | * Called if REQ_F_LINK is set, and we fail the head request | |
695 | */ | |
696 | static void io_fail_links(struct io_kiocb *req) | |
697 | { | |
698 | struct io_kiocb *link; | |
699 | ||
700 | while (!list_empty(&req->link_list)) { | |
701 | link = list_first_entry(&req->link_list, struct io_kiocb, list); | |
702 | list_del(&link->list); | |
703 | ||
704 | io_cqring_add_event(req->ctx, link->user_data, -ECANCELED); | |
705 | __io_free_req(link); | |
706 | } | |
707 | } | |
708 | ||
709 | static void io_free_req(struct io_kiocb *req) | |
710 | { | |
711 | /* | |
712 | * If LINK is set, we have dependent requests in this chain. If we | |
713 | * didn't fail this request, queue the first one up, moving any other | |
714 | * dependencies to the next request. In case of failure, fail the rest | |
715 | * of the chain. | |
716 | */ | |
717 | if (req->flags & REQ_F_LINK) { | |
718 | if (req->flags & REQ_F_FAIL_LINK) | |
719 | io_fail_links(req); | |
720 | else | |
721 | io_req_link_next(req); | |
722 | } | |
723 | ||
724 | __io_free_req(req); | |
725 | } | |
726 | ||
e65ef56d JA |
727 | static void io_put_req(struct io_kiocb *req) |
728 | { | |
729 | if (refcount_dec_and_test(&req->refs)) | |
730 | io_free_req(req); | |
2b188cc1 JA |
731 | } |
732 | ||
75b28aff | 733 | static unsigned io_cqring_events(struct io_rings *rings) |
a3a0e43f JA |
734 | { |
735 | /* See comment at the top of this file */ | |
736 | smp_rmb(); | |
75b28aff | 737 | return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head); |
a3a0e43f JA |
738 | } |
739 | ||
fb5ccc98 PB |
740 | static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) |
741 | { | |
742 | struct io_rings *rings = ctx->rings; | |
743 | ||
744 | /* make sure SQ entry isn't read before tail */ | |
745 | return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; | |
746 | } | |
747 | ||
def596e9 JA |
748 | /* |
749 | * Find and free completed poll iocbs | |
750 | */ | |
751 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
752 | struct list_head *done) | |
753 | { | |
754 | void *reqs[IO_IOPOLL_BATCH]; | |
755 | struct io_kiocb *req; | |
09bb8394 | 756 | int to_free; |
def596e9 | 757 | |
09bb8394 | 758 | to_free = 0; |
def596e9 JA |
759 | while (!list_empty(done)) { |
760 | req = list_first_entry(done, struct io_kiocb, list); | |
761 | list_del(&req->list); | |
762 | ||
9e645e11 | 763 | io_cqring_fill_event(ctx, req->user_data, req->result); |
def596e9 JA |
764 | (*nr_events)++; |
765 | ||
09bb8394 JA |
766 | if (refcount_dec_and_test(&req->refs)) { |
767 | /* If we're not using fixed files, we have to pair the | |
768 | * completion part with the file put. Use regular | |
769 | * completions for those, only batch free for fixed | |
9e645e11 | 770 | * file and non-linked commands. |
09bb8394 | 771 | */ |
9e645e11 JA |
772 | if ((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) == |
773 | REQ_F_FIXED_FILE) { | |
09bb8394 JA |
774 | reqs[to_free++] = req; |
775 | if (to_free == ARRAY_SIZE(reqs)) | |
776 | io_free_req_many(ctx, reqs, &to_free); | |
6b06314c | 777 | } else { |
09bb8394 | 778 | io_free_req(req); |
6b06314c | 779 | } |
9a56a232 | 780 | } |
def596e9 | 781 | } |
def596e9 | 782 | |
09bb8394 | 783 | io_commit_cqring(ctx); |
def596e9 JA |
784 | io_free_req_many(ctx, reqs, &to_free); |
785 | } | |
786 | ||
787 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
788 | long min) | |
789 | { | |
790 | struct io_kiocb *req, *tmp; | |
791 | LIST_HEAD(done); | |
792 | bool spin; | |
793 | int ret; | |
794 | ||
795 | /* | |
796 | * Only spin for completions if we don't have multiple devices hanging | |
797 | * off our complete list, and we're under the requested amount. | |
798 | */ | |
799 | spin = !ctx->poll_multi_file && *nr_events < min; | |
800 | ||
801 | ret = 0; | |
802 | list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) { | |
803 | struct kiocb *kiocb = &req->rw; | |
804 | ||
805 | /* | |
806 | * Move completed entries to our local list. If we find a | |
807 | * request that requires polling, break out and complete | |
808 | * the done list first, if we have entries there. | |
809 | */ | |
810 | if (req->flags & REQ_F_IOPOLL_COMPLETED) { | |
811 | list_move_tail(&req->list, &done); | |
812 | continue; | |
813 | } | |
814 | if (!list_empty(&done)) | |
815 | break; | |
816 | ||
817 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
818 | if (ret < 0) | |
819 | break; | |
820 | ||
821 | if (ret && spin) | |
822 | spin = false; | |
823 | ret = 0; | |
824 | } | |
825 | ||
826 | if (!list_empty(&done)) | |
827 | io_iopoll_complete(ctx, nr_events, &done); | |
828 | ||
829 | return ret; | |
830 | } | |
831 | ||
832 | /* | |
833 | * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a | |
834 | * non-spinning poll check - we'll still enter the driver poll loop, but only | |
835 | * as a non-spinning completion check. | |
836 | */ | |
837 | static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
838 | long min) | |
839 | { | |
08f5439f | 840 | while (!list_empty(&ctx->poll_list) && !need_resched()) { |
def596e9 JA |
841 | int ret; |
842 | ||
843 | ret = io_do_iopoll(ctx, nr_events, min); | |
844 | if (ret < 0) | |
845 | return ret; | |
846 | if (!min || *nr_events >= min) | |
847 | return 0; | |
848 | } | |
849 | ||
850 | return 1; | |
851 | } | |
852 | ||
853 | /* | |
854 | * We can't just wait for polled events to come to us, we have to actively | |
855 | * find and complete them. | |
856 | */ | |
857 | static void io_iopoll_reap_events(struct io_ring_ctx *ctx) | |
858 | { | |
859 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
860 | return; | |
861 | ||
862 | mutex_lock(&ctx->uring_lock); | |
863 | while (!list_empty(&ctx->poll_list)) { | |
864 | unsigned int nr_events = 0; | |
865 | ||
866 | io_iopoll_getevents(ctx, &nr_events, 1); | |
08f5439f JA |
867 | |
868 | /* | |
869 | * Ensure we allow local-to-the-cpu processing to take place, | |
870 | * in this case we need to ensure that we reap all events. | |
871 | */ | |
872 | cond_resched(); | |
def596e9 JA |
873 | } |
874 | mutex_unlock(&ctx->uring_lock); | |
875 | } | |
876 | ||
2b2ed975 JA |
877 | static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, |
878 | long min) | |
def596e9 | 879 | { |
2b2ed975 | 880 | int iters = 0, ret = 0; |
500f9fba | 881 | |
def596e9 JA |
882 | do { |
883 | int tmin = 0; | |
884 | ||
a3a0e43f JA |
885 | /* |
886 | * Don't enter poll loop if we already have events pending. | |
887 | * If we do, we can potentially be spinning for commands that | |
888 | * already triggered a CQE (eg in error). | |
889 | */ | |
75b28aff | 890 | if (io_cqring_events(ctx->rings)) |
a3a0e43f JA |
891 | break; |
892 | ||
500f9fba JA |
893 | /* |
894 | * If a submit got punted to a workqueue, we can have the | |
895 | * application entering polling for a command before it gets | |
896 | * issued. That app will hold the uring_lock for the duration | |
897 | * of the poll right here, so we need to take a breather every | |
898 | * now and then to ensure that the issue has a chance to add | |
899 | * the poll to the issued list. Otherwise we can spin here | |
900 | * forever, while the workqueue is stuck trying to acquire the | |
901 | * very same mutex. | |
902 | */ | |
903 | if (!(++iters & 7)) { | |
904 | mutex_unlock(&ctx->uring_lock); | |
905 | mutex_lock(&ctx->uring_lock); | |
906 | } | |
907 | ||
def596e9 JA |
908 | if (*nr_events < min) |
909 | tmin = min - *nr_events; | |
910 | ||
911 | ret = io_iopoll_getevents(ctx, nr_events, tmin); | |
912 | if (ret <= 0) | |
913 | break; | |
914 | ret = 0; | |
915 | } while (min && !*nr_events && !need_resched()); | |
916 | ||
2b2ed975 JA |
917 | return ret; |
918 | } | |
919 | ||
920 | static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, | |
921 | long min) | |
922 | { | |
923 | int ret; | |
924 | ||
925 | /* | |
926 | * We disallow the app entering submit/complete with polling, but we | |
927 | * still need to lock the ring to prevent racing with polled issue | |
928 | * that got punted to a workqueue. | |
929 | */ | |
930 | mutex_lock(&ctx->uring_lock); | |
931 | ret = __io_iopoll_check(ctx, nr_events, min); | |
500f9fba | 932 | mutex_unlock(&ctx->uring_lock); |
def596e9 JA |
933 | return ret; |
934 | } | |
935 | ||
491381ce | 936 | static void kiocb_end_write(struct io_kiocb *req) |
2b188cc1 | 937 | { |
491381ce JA |
938 | /* |
939 | * Tell lockdep we inherited freeze protection from submission | |
940 | * thread. | |
941 | */ | |
942 | if (req->flags & REQ_F_ISREG) { | |
943 | struct inode *inode = file_inode(req->file); | |
2b188cc1 | 944 | |
491381ce | 945 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); |
2b188cc1 | 946 | } |
491381ce | 947 | file_end_write(req->file); |
2b188cc1 JA |
948 | } |
949 | ||
950 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
951 | { | |
952 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
953 | ||
491381ce JA |
954 | if (kiocb->ki_flags & IOCB_WRITE) |
955 | kiocb_end_write(req); | |
2b188cc1 | 956 | |
9e645e11 JA |
957 | if ((req->flags & REQ_F_LINK) && res != req->result) |
958 | req->flags |= REQ_F_FAIL_LINK; | |
c71ffb67 | 959 | io_cqring_add_event(req->ctx, req->user_data, res); |
e65ef56d | 960 | io_put_req(req); |
2b188cc1 JA |
961 | } |
962 | ||
def596e9 JA |
963 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
964 | { | |
965 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
966 | ||
491381ce JA |
967 | if (kiocb->ki_flags & IOCB_WRITE) |
968 | kiocb_end_write(req); | |
def596e9 | 969 | |
9e645e11 JA |
970 | if ((req->flags & REQ_F_LINK) && res != req->result) |
971 | req->flags |= REQ_F_FAIL_LINK; | |
972 | req->result = res; | |
def596e9 JA |
973 | if (res != -EAGAIN) |
974 | req->flags |= REQ_F_IOPOLL_COMPLETED; | |
975 | } | |
976 | ||
977 | /* | |
978 | * After the iocb has been issued, it's safe to be found on the poll list. | |
979 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
980 | * find it from a io_iopoll_getevents() thread before the issuer is done | |
981 | * accessing the kiocb cookie. | |
982 | */ | |
983 | static void io_iopoll_req_issued(struct io_kiocb *req) | |
984 | { | |
985 | struct io_ring_ctx *ctx = req->ctx; | |
986 | ||
987 | /* | |
988 | * Track whether we have multiple files in our lists. This will impact | |
989 | * how we do polling eventually, not spinning if we're on potentially | |
990 | * different devices. | |
991 | */ | |
992 | if (list_empty(&ctx->poll_list)) { | |
993 | ctx->poll_multi_file = false; | |
994 | } else if (!ctx->poll_multi_file) { | |
995 | struct io_kiocb *list_req; | |
996 | ||
997 | list_req = list_first_entry(&ctx->poll_list, struct io_kiocb, | |
998 | list); | |
999 | if (list_req->rw.ki_filp != req->rw.ki_filp) | |
1000 | ctx->poll_multi_file = true; | |
1001 | } | |
1002 | ||
1003 | /* | |
1004 | * For fast devices, IO may have already completed. If it has, add | |
1005 | * it to the front so we find it first. | |
1006 | */ | |
1007 | if (req->flags & REQ_F_IOPOLL_COMPLETED) | |
1008 | list_add(&req->list, &ctx->poll_list); | |
1009 | else | |
1010 | list_add_tail(&req->list, &ctx->poll_list); | |
1011 | } | |
1012 | ||
3d6770fb | 1013 | static void io_file_put(struct io_submit_state *state) |
9a56a232 | 1014 | { |
3d6770fb | 1015 | if (state->file) { |
9a56a232 JA |
1016 | int diff = state->has_refs - state->used_refs; |
1017 | ||
1018 | if (diff) | |
1019 | fput_many(state->file, diff); | |
1020 | state->file = NULL; | |
1021 | } | |
1022 | } | |
1023 | ||
1024 | /* | |
1025 | * Get as many references to a file as we have IOs left in this submission, | |
1026 | * assuming most submissions are for one file, or at least that each file | |
1027 | * has more than one submission. | |
1028 | */ | |
1029 | static struct file *io_file_get(struct io_submit_state *state, int fd) | |
1030 | { | |
1031 | if (!state) | |
1032 | return fget(fd); | |
1033 | ||
1034 | if (state->file) { | |
1035 | if (state->fd == fd) { | |
1036 | state->used_refs++; | |
1037 | state->ios_left--; | |
1038 | return state->file; | |
1039 | } | |
3d6770fb | 1040 | io_file_put(state); |
9a56a232 JA |
1041 | } |
1042 | state->file = fget_many(fd, state->ios_left); | |
1043 | if (!state->file) | |
1044 | return NULL; | |
1045 | ||
1046 | state->fd = fd; | |
1047 | state->has_refs = state->ios_left; | |
1048 | state->used_refs = 1; | |
1049 | state->ios_left--; | |
1050 | return state->file; | |
1051 | } | |
1052 | ||
2b188cc1 JA |
1053 | /* |
1054 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1055 | * any file. For now, just ensure that anything potentially problematic is done | |
1056 | * inline. | |
1057 | */ | |
1058 | static bool io_file_supports_async(struct file *file) | |
1059 | { | |
1060 | umode_t mode = file_inode(file)->i_mode; | |
1061 | ||
1062 | if (S_ISBLK(mode) || S_ISCHR(mode)) | |
1063 | return true; | |
1064 | if (S_ISREG(mode) && file->f_op != &io_uring_fops) | |
1065 | return true; | |
1066 | ||
1067 | return false; | |
1068 | } | |
1069 | ||
6c271ce2 | 1070 | static int io_prep_rw(struct io_kiocb *req, const struct sqe_submit *s, |
8358e3a8 | 1071 | bool force_nonblock) |
2b188cc1 | 1072 | { |
6c271ce2 | 1073 | const struct io_uring_sqe *sqe = s->sqe; |
def596e9 | 1074 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 | 1075 | struct kiocb *kiocb = &req->rw; |
09bb8394 JA |
1076 | unsigned ioprio; |
1077 | int ret; | |
2b188cc1 | 1078 | |
09bb8394 JA |
1079 | if (!req->file) |
1080 | return -EBADF; | |
2b188cc1 | 1081 | |
491381ce JA |
1082 | if (S_ISREG(file_inode(req->file)->i_mode)) |
1083 | req->flags |= REQ_F_ISREG; | |
1084 | ||
1085 | /* | |
1086 | * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so | |
1087 | * we know to async punt it even if it was opened O_NONBLOCK | |
1088 | */ | |
1089 | if (force_nonblock && !io_file_supports_async(req->file)) { | |
1090 | req->flags |= REQ_F_MUST_PUNT; | |
1091 | return -EAGAIN; | |
1092 | } | |
6b06314c | 1093 | |
2b188cc1 JA |
1094 | kiocb->ki_pos = READ_ONCE(sqe->off); |
1095 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); | |
1096 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); | |
1097 | ||
1098 | ioprio = READ_ONCE(sqe->ioprio); | |
1099 | if (ioprio) { | |
1100 | ret = ioprio_check_cap(ioprio); | |
1101 | if (ret) | |
09bb8394 | 1102 | return ret; |
2b188cc1 JA |
1103 | |
1104 | kiocb->ki_ioprio = ioprio; | |
1105 | } else | |
1106 | kiocb->ki_ioprio = get_current_ioprio(); | |
1107 | ||
1108 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
1109 | if (unlikely(ret)) | |
09bb8394 | 1110 | return ret; |
8449eeda SB |
1111 | |
1112 | /* don't allow async punt if RWF_NOWAIT was requested */ | |
491381ce JA |
1113 | if ((kiocb->ki_flags & IOCB_NOWAIT) || |
1114 | (req->file->f_flags & O_NONBLOCK)) | |
8449eeda SB |
1115 | req->flags |= REQ_F_NOWAIT; |
1116 | ||
1117 | if (force_nonblock) | |
2b188cc1 | 1118 | kiocb->ki_flags |= IOCB_NOWAIT; |
8449eeda | 1119 | |
def596e9 | 1120 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
1121 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
1122 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 1123 | return -EOPNOTSUPP; |
2b188cc1 | 1124 | |
def596e9 JA |
1125 | kiocb->ki_flags |= IOCB_HIPRI; |
1126 | kiocb->ki_complete = io_complete_rw_iopoll; | |
6873e0bd | 1127 | req->result = 0; |
def596e9 | 1128 | } else { |
09bb8394 JA |
1129 | if (kiocb->ki_flags & IOCB_HIPRI) |
1130 | return -EINVAL; | |
def596e9 JA |
1131 | kiocb->ki_complete = io_complete_rw; |
1132 | } | |
2b188cc1 | 1133 | return 0; |
2b188cc1 JA |
1134 | } |
1135 | ||
1136 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
1137 | { | |
1138 | switch (ret) { | |
1139 | case -EIOCBQUEUED: | |
1140 | break; | |
1141 | case -ERESTARTSYS: | |
1142 | case -ERESTARTNOINTR: | |
1143 | case -ERESTARTNOHAND: | |
1144 | case -ERESTART_RESTARTBLOCK: | |
1145 | /* | |
1146 | * We can't just restart the syscall, since previously | |
1147 | * submitted sqes may already be in progress. Just fail this | |
1148 | * IO with EINTR. | |
1149 | */ | |
1150 | ret = -EINTR; | |
1151 | /* fall through */ | |
1152 | default: | |
1153 | kiocb->ki_complete(kiocb, ret, 0); | |
1154 | } | |
1155 | } | |
1156 | ||
edafccee JA |
1157 | static int io_import_fixed(struct io_ring_ctx *ctx, int rw, |
1158 | const struct io_uring_sqe *sqe, | |
1159 | struct iov_iter *iter) | |
1160 | { | |
1161 | size_t len = READ_ONCE(sqe->len); | |
1162 | struct io_mapped_ubuf *imu; | |
1163 | unsigned index, buf_index; | |
1164 | size_t offset; | |
1165 | u64 buf_addr; | |
1166 | ||
1167 | /* attempt to use fixed buffers without having provided iovecs */ | |
1168 | if (unlikely(!ctx->user_bufs)) | |
1169 | return -EFAULT; | |
1170 | ||
1171 | buf_index = READ_ONCE(sqe->buf_index); | |
1172 | if (unlikely(buf_index >= ctx->nr_user_bufs)) | |
1173 | return -EFAULT; | |
1174 | ||
1175 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); | |
1176 | imu = &ctx->user_bufs[index]; | |
1177 | buf_addr = READ_ONCE(sqe->addr); | |
1178 | ||
1179 | /* overflow */ | |
1180 | if (buf_addr + len < buf_addr) | |
1181 | return -EFAULT; | |
1182 | /* not inside the mapped region */ | |
1183 | if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len) | |
1184 | return -EFAULT; | |
1185 | ||
1186 | /* | |
1187 | * May not be a start of buffer, set size appropriately | |
1188 | * and advance us to the beginning. | |
1189 | */ | |
1190 | offset = buf_addr - imu->ubuf; | |
1191 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
bd11b3a3 JA |
1192 | |
1193 | if (offset) { | |
1194 | /* | |
1195 | * Don't use iov_iter_advance() here, as it's really slow for | |
1196 | * using the latter parts of a big fixed buffer - it iterates | |
1197 | * over each segment manually. We can cheat a bit here, because | |
1198 | * we know that: | |
1199 | * | |
1200 | * 1) it's a BVEC iter, we set it up | |
1201 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
1202 | * first and last bvec | |
1203 | * | |
1204 | * So just find our index, and adjust the iterator afterwards. | |
1205 | * If the offset is within the first bvec (or the whole first | |
1206 | * bvec, just use iov_iter_advance(). This makes it easier | |
1207 | * since we can just skip the first segment, which may not | |
1208 | * be PAGE_SIZE aligned. | |
1209 | */ | |
1210 | const struct bio_vec *bvec = imu->bvec; | |
1211 | ||
1212 | if (offset <= bvec->bv_len) { | |
1213 | iov_iter_advance(iter, offset); | |
1214 | } else { | |
1215 | unsigned long seg_skip; | |
1216 | ||
1217 | /* skip first vec */ | |
1218 | offset -= bvec->bv_len; | |
1219 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
1220 | ||
1221 | iter->bvec = bvec + seg_skip; | |
1222 | iter->nr_segs -= seg_skip; | |
99c79f66 | 1223 | iter->count -= bvec->bv_len + offset; |
bd11b3a3 | 1224 | iter->iov_offset = offset & ~PAGE_MASK; |
bd11b3a3 JA |
1225 | } |
1226 | } | |
1227 | ||
edafccee JA |
1228 | return 0; |
1229 | } | |
1230 | ||
87e5e6da JA |
1231 | static ssize_t io_import_iovec(struct io_ring_ctx *ctx, int rw, |
1232 | const struct sqe_submit *s, struct iovec **iovec, | |
1233 | struct iov_iter *iter) | |
2b188cc1 JA |
1234 | { |
1235 | const struct io_uring_sqe *sqe = s->sqe; | |
1236 | void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
1237 | size_t sqe_len = READ_ONCE(sqe->len); | |
edafccee JA |
1238 | u8 opcode; |
1239 | ||
1240 | /* | |
1241 | * We're reading ->opcode for the second time, but the first read | |
1242 | * doesn't care whether it's _FIXED or not, so it doesn't matter | |
1243 | * whether ->opcode changes concurrently. The first read does care | |
1244 | * about whether it is a READ or a WRITE, so we don't trust this read | |
1245 | * for that purpose and instead let the caller pass in the read/write | |
1246 | * flag. | |
1247 | */ | |
1248 | opcode = READ_ONCE(sqe->opcode); | |
1249 | if (opcode == IORING_OP_READ_FIXED || | |
1250 | opcode == IORING_OP_WRITE_FIXED) { | |
87e5e6da | 1251 | ssize_t ret = io_import_fixed(ctx, rw, sqe, iter); |
edafccee JA |
1252 | *iovec = NULL; |
1253 | return ret; | |
1254 | } | |
2b188cc1 JA |
1255 | |
1256 | if (!s->has_user) | |
1257 | return -EFAULT; | |
1258 | ||
1259 | #ifdef CONFIG_COMPAT | |
1260 | if (ctx->compat) | |
1261 | return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV, | |
1262 | iovec, iter); | |
1263 | #endif | |
1264 | ||
1265 | return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter); | |
1266 | } | |
1267 | ||
6d5d5ac5 JA |
1268 | static inline bool io_should_merge(struct async_list *al, struct kiocb *kiocb) |
1269 | { | |
1270 | if (al->file == kiocb->ki_filp) { | |
1271 | off_t start, end; | |
1272 | ||
1273 | /* | |
1274 | * Allow merging if we're anywhere in the range of the same | |
1275 | * page. Generally this happens for sub-page reads or writes, | |
1276 | * and it's beneficial to allow the first worker to bring the | |
1277 | * page in and the piggy backed work can then work on the | |
1278 | * cached page. | |
1279 | */ | |
1280 | start = al->io_start & PAGE_MASK; | |
1281 | end = (al->io_start + al->io_len + PAGE_SIZE - 1) & PAGE_MASK; | |
1282 | if (kiocb->ki_pos >= start && kiocb->ki_pos <= end) | |
1283 | return true; | |
1284 | } | |
1285 | ||
1286 | al->file = NULL; | |
1287 | return false; | |
1288 | } | |
1289 | ||
31b51510 JA |
1290 | /* |
1291 | * Make a note of the last file/offset/direction we punted to async | |
1292 | * context. We'll use this information to see if we can piggy back a | |
1293 | * sequential request onto the previous one, if it's still hasn't been | |
1294 | * completed by the async worker. | |
1295 | */ | |
1296 | static void io_async_list_note(int rw, struct io_kiocb *req, size_t len) | |
1297 | { | |
1298 | struct async_list *async_list = &req->ctx->pending_async[rw]; | |
1299 | struct kiocb *kiocb = &req->rw; | |
1300 | struct file *filp = kiocb->ki_filp; | |
31b51510 | 1301 | |
6d5d5ac5 | 1302 | if (io_should_merge(async_list, kiocb)) { |
9310a7ba | 1303 | unsigned long max_bytes; |
31b51510 JA |
1304 | |
1305 | /* Use 8x RA size as a decent limiter for both reads/writes */ | |
9310a7ba ZL |
1306 | max_bytes = filp->f_ra.ra_pages << (PAGE_SHIFT + 3); |
1307 | if (!max_bytes) | |
1308 | max_bytes = VM_READAHEAD_PAGES << (PAGE_SHIFT + 3); | |
1309 | ||
1310 | /* If max len are exceeded, reset the state */ | |
1311 | if (async_list->io_len + len <= max_bytes) { | |
31b51510 | 1312 | req->flags |= REQ_F_SEQ_PREV; |
9310a7ba | 1313 | async_list->io_len += len; |
31b51510 | 1314 | } else { |
6d5d5ac5 | 1315 | async_list->file = NULL; |
31b51510 JA |
1316 | } |
1317 | } | |
1318 | ||
1319 | /* New file? Reset state. */ | |
1320 | if (async_list->file != filp) { | |
6d5d5ac5 JA |
1321 | async_list->io_start = kiocb->ki_pos; |
1322 | async_list->io_len = len; | |
31b51510 JA |
1323 | async_list->file = filp; |
1324 | } | |
31b51510 JA |
1325 | } |
1326 | ||
32960613 JA |
1327 | /* |
1328 | * For files that don't have ->read_iter() and ->write_iter(), handle them | |
1329 | * by looping over ->read() or ->write() manually. | |
1330 | */ | |
1331 | static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb, | |
1332 | struct iov_iter *iter) | |
1333 | { | |
1334 | ssize_t ret = 0; | |
1335 | ||
1336 | /* | |
1337 | * Don't support polled IO through this interface, and we can't | |
1338 | * support non-blocking either. For the latter, this just causes | |
1339 | * the kiocb to be handled from an async context. | |
1340 | */ | |
1341 | if (kiocb->ki_flags & IOCB_HIPRI) | |
1342 | return -EOPNOTSUPP; | |
1343 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
1344 | return -EAGAIN; | |
1345 | ||
1346 | while (iov_iter_count(iter)) { | |
1347 | struct iovec iovec = iov_iter_iovec(iter); | |
1348 | ssize_t nr; | |
1349 | ||
1350 | if (rw == READ) { | |
1351 | nr = file->f_op->read(file, iovec.iov_base, | |
1352 | iovec.iov_len, &kiocb->ki_pos); | |
1353 | } else { | |
1354 | nr = file->f_op->write(file, iovec.iov_base, | |
1355 | iovec.iov_len, &kiocb->ki_pos); | |
1356 | } | |
1357 | ||
1358 | if (nr < 0) { | |
1359 | if (!ret) | |
1360 | ret = nr; | |
1361 | break; | |
1362 | } | |
1363 | ret += nr; | |
1364 | if (nr != iovec.iov_len) | |
1365 | break; | |
1366 | iov_iter_advance(iter, nr); | |
1367 | } | |
1368 | ||
1369 | return ret; | |
1370 | } | |
1371 | ||
e0c5c576 | 1372 | static int io_read(struct io_kiocb *req, const struct sqe_submit *s, |
8358e3a8 | 1373 | bool force_nonblock) |
2b188cc1 JA |
1374 | { |
1375 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1376 | struct kiocb *kiocb = &req->rw; | |
1377 | struct iov_iter iter; | |
1378 | struct file *file; | |
31b51510 | 1379 | size_t iov_count; |
9d93a3f5 | 1380 | ssize_t read_size, ret; |
2b188cc1 | 1381 | |
8358e3a8 | 1382 | ret = io_prep_rw(req, s, force_nonblock); |
2b188cc1 JA |
1383 | if (ret) |
1384 | return ret; | |
1385 | file = kiocb->ki_filp; | |
1386 | ||
2b188cc1 | 1387 | if (unlikely(!(file->f_mode & FMODE_READ))) |
09bb8394 | 1388 | return -EBADF; |
2b188cc1 JA |
1389 | |
1390 | ret = io_import_iovec(req->ctx, READ, s, &iovec, &iter); | |
87e5e6da | 1391 | if (ret < 0) |
09bb8394 | 1392 | return ret; |
2b188cc1 | 1393 | |
9d93a3f5 | 1394 | read_size = ret; |
9e645e11 JA |
1395 | if (req->flags & REQ_F_LINK) |
1396 | req->result = read_size; | |
1397 | ||
31b51510 JA |
1398 | iov_count = iov_iter_count(&iter); |
1399 | ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 JA |
1400 | if (!ret) { |
1401 | ssize_t ret2; | |
1402 | ||
32960613 JA |
1403 | if (file->f_op->read_iter) |
1404 | ret2 = call_read_iter(file, kiocb, &iter); | |
1405 | else | |
1406 | ret2 = loop_rw_iter(READ, file, kiocb, &iter); | |
1407 | ||
9d93a3f5 JA |
1408 | /* |
1409 | * In case of a short read, punt to async. This can happen | |
1410 | * if we have data partially cached. Alternatively we can | |
1411 | * return the short read, in which case the application will | |
1412 | * need to issue another SQE and wait for it. That SQE will | |
1413 | * need async punt anyway, so it's more efficient to do it | |
1414 | * here. | |
1415 | */ | |
491381ce JA |
1416 | if (force_nonblock && !(req->flags & REQ_F_NOWAIT) && |
1417 | (req->flags & REQ_F_ISREG) && | |
1418 | ret2 > 0 && ret2 < read_size) | |
9d93a3f5 JA |
1419 | ret2 = -EAGAIN; |
1420 | /* Catch -EAGAIN return for forced non-blocking submission */ | |
31b51510 | 1421 | if (!force_nonblock || ret2 != -EAGAIN) { |
2b188cc1 | 1422 | io_rw_done(kiocb, ret2); |
31b51510 JA |
1423 | } else { |
1424 | /* | |
1425 | * If ->needs_lock is true, we're already in async | |
1426 | * context. | |
1427 | */ | |
1428 | if (!s->needs_lock) | |
1429 | io_async_list_note(READ, req, iov_count); | |
2b188cc1 | 1430 | ret = -EAGAIN; |
31b51510 | 1431 | } |
2b188cc1 JA |
1432 | } |
1433 | kfree(iovec); | |
2b188cc1 JA |
1434 | return ret; |
1435 | } | |
1436 | ||
e0c5c576 | 1437 | static int io_write(struct io_kiocb *req, const struct sqe_submit *s, |
8358e3a8 | 1438 | bool force_nonblock) |
2b188cc1 JA |
1439 | { |
1440 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1441 | struct kiocb *kiocb = &req->rw; | |
1442 | struct iov_iter iter; | |
1443 | struct file *file; | |
31b51510 | 1444 | size_t iov_count; |
87e5e6da | 1445 | ssize_t ret; |
2b188cc1 | 1446 | |
8358e3a8 | 1447 | ret = io_prep_rw(req, s, force_nonblock); |
2b188cc1 JA |
1448 | if (ret) |
1449 | return ret; | |
2b188cc1 | 1450 | |
2b188cc1 JA |
1451 | file = kiocb->ki_filp; |
1452 | if (unlikely(!(file->f_mode & FMODE_WRITE))) | |
09bb8394 | 1453 | return -EBADF; |
2b188cc1 JA |
1454 | |
1455 | ret = io_import_iovec(req->ctx, WRITE, s, &iovec, &iter); | |
87e5e6da | 1456 | if (ret < 0) |
09bb8394 | 1457 | return ret; |
2b188cc1 | 1458 | |
9e645e11 JA |
1459 | if (req->flags & REQ_F_LINK) |
1460 | req->result = ret; | |
1461 | ||
31b51510 JA |
1462 | iov_count = iov_iter_count(&iter); |
1463 | ||
1464 | ret = -EAGAIN; | |
1465 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT)) { | |
1466 | /* If ->needs_lock is true, we're already in async context. */ | |
1467 | if (!s->needs_lock) | |
1468 | io_async_list_note(WRITE, req, iov_count); | |
1469 | goto out_free; | |
1470 | } | |
1471 | ||
1472 | ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 | 1473 | if (!ret) { |
9bf7933f RP |
1474 | ssize_t ret2; |
1475 | ||
2b188cc1 JA |
1476 | /* |
1477 | * Open-code file_start_write here to grab freeze protection, | |
1478 | * which will be released by another thread in | |
1479 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
1480 | * released so that it doesn't complain about the held lock when | |
1481 | * we return to userspace. | |
1482 | */ | |
491381ce | 1483 | if (req->flags & REQ_F_ISREG) { |
2b188cc1 JA |
1484 | __sb_start_write(file_inode(file)->i_sb, |
1485 | SB_FREEZE_WRITE, true); | |
1486 | __sb_writers_release(file_inode(file)->i_sb, | |
1487 | SB_FREEZE_WRITE); | |
1488 | } | |
1489 | kiocb->ki_flags |= IOCB_WRITE; | |
9bf7933f | 1490 | |
32960613 JA |
1491 | if (file->f_op->write_iter) |
1492 | ret2 = call_write_iter(file, kiocb, &iter); | |
1493 | else | |
1494 | ret2 = loop_rw_iter(WRITE, file, kiocb, &iter); | |
9bf7933f RP |
1495 | if (!force_nonblock || ret2 != -EAGAIN) { |
1496 | io_rw_done(kiocb, ret2); | |
1497 | } else { | |
1498 | /* | |
1499 | * If ->needs_lock is true, we're already in async | |
1500 | * context. | |
1501 | */ | |
1502 | if (!s->needs_lock) | |
1503 | io_async_list_note(WRITE, req, iov_count); | |
1504 | ret = -EAGAIN; | |
1505 | } | |
2b188cc1 | 1506 | } |
31b51510 | 1507 | out_free: |
2b188cc1 | 1508 | kfree(iovec); |
2b188cc1 JA |
1509 | return ret; |
1510 | } | |
1511 | ||
1512 | /* | |
1513 | * IORING_OP_NOP just posts a completion event, nothing else. | |
1514 | */ | |
1515 | static int io_nop(struct io_kiocb *req, u64 user_data) | |
1516 | { | |
1517 | struct io_ring_ctx *ctx = req->ctx; | |
1518 | long err = 0; | |
1519 | ||
def596e9 JA |
1520 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
1521 | return -EINVAL; | |
1522 | ||
c71ffb67 | 1523 | io_cqring_add_event(ctx, user_data, err); |
e65ef56d | 1524 | io_put_req(req); |
2b188cc1 JA |
1525 | return 0; |
1526 | } | |
1527 | ||
c992fe29 CH |
1528 | static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1529 | { | |
6b06314c | 1530 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 1531 | |
09bb8394 JA |
1532 | if (!req->file) |
1533 | return -EBADF; | |
c992fe29 | 1534 | |
6b06314c | 1535 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 1536 | return -EINVAL; |
edafccee | 1537 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
1538 | return -EINVAL; |
1539 | ||
c992fe29 CH |
1540 | return 0; |
1541 | } | |
1542 | ||
1543 | static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
1544 | bool force_nonblock) | |
1545 | { | |
1546 | loff_t sqe_off = READ_ONCE(sqe->off); | |
1547 | loff_t sqe_len = READ_ONCE(sqe->len); | |
1548 | loff_t end = sqe_off + sqe_len; | |
1549 | unsigned fsync_flags; | |
1550 | int ret; | |
1551 | ||
1552 | fsync_flags = READ_ONCE(sqe->fsync_flags); | |
1553 | if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC)) | |
1554 | return -EINVAL; | |
1555 | ||
1556 | ret = io_prep_fsync(req, sqe); | |
1557 | if (ret) | |
1558 | return ret; | |
1559 | ||
1560 | /* fsync always requires a blocking context */ | |
1561 | if (force_nonblock) | |
1562 | return -EAGAIN; | |
1563 | ||
1564 | ret = vfs_fsync_range(req->rw.ki_filp, sqe_off, | |
1565 | end > 0 ? end : LLONG_MAX, | |
1566 | fsync_flags & IORING_FSYNC_DATASYNC); | |
1567 | ||
9e645e11 JA |
1568 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1569 | req->flags |= REQ_F_FAIL_LINK; | |
c71ffb67 | 1570 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
e65ef56d | 1571 | io_put_req(req); |
c992fe29 CH |
1572 | return 0; |
1573 | } | |
1574 | ||
5d17b4a4 JA |
1575 | static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1576 | { | |
1577 | struct io_ring_ctx *ctx = req->ctx; | |
1578 | int ret = 0; | |
1579 | ||
1580 | if (!req->file) | |
1581 | return -EBADF; | |
5d17b4a4 JA |
1582 | |
1583 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
1584 | return -EINVAL; | |
1585 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) | |
1586 | return -EINVAL; | |
1587 | ||
5d17b4a4 JA |
1588 | return ret; |
1589 | } | |
1590 | ||
1591 | static int io_sync_file_range(struct io_kiocb *req, | |
1592 | const struct io_uring_sqe *sqe, | |
1593 | bool force_nonblock) | |
1594 | { | |
1595 | loff_t sqe_off; | |
1596 | loff_t sqe_len; | |
1597 | unsigned flags; | |
1598 | int ret; | |
1599 | ||
1600 | ret = io_prep_sfr(req, sqe); | |
1601 | if (ret) | |
1602 | return ret; | |
1603 | ||
1604 | /* sync_file_range always requires a blocking context */ | |
1605 | if (force_nonblock) | |
1606 | return -EAGAIN; | |
1607 | ||
1608 | sqe_off = READ_ONCE(sqe->off); | |
1609 | sqe_len = READ_ONCE(sqe->len); | |
1610 | flags = READ_ONCE(sqe->sync_range_flags); | |
1611 | ||
1612 | ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags); | |
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); |
5d17b4a4 JA |
1617 | io_put_req(req); |
1618 | return 0; | |
1619 | } | |
1620 | ||
0fa03c62 | 1621 | #if defined(CONFIG_NET) |
aa1fa28f JA |
1622 | static int io_send_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
1623 | bool force_nonblock, | |
1624 | long (*fn)(struct socket *, struct user_msghdr __user *, | |
1625 | unsigned int)) | |
1626 | { | |
0fa03c62 JA |
1627 | struct socket *sock; |
1628 | int ret; | |
1629 | ||
1630 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1631 | return -EINVAL; | |
1632 | ||
1633 | sock = sock_from_file(req->file, &ret); | |
1634 | if (sock) { | |
1635 | struct user_msghdr __user *msg; | |
1636 | unsigned flags; | |
1637 | ||
1638 | flags = READ_ONCE(sqe->msg_flags); | |
1639 | if (flags & MSG_DONTWAIT) | |
1640 | req->flags |= REQ_F_NOWAIT; | |
1641 | else if (force_nonblock) | |
1642 | flags |= MSG_DONTWAIT; | |
1643 | ||
1644 | msg = (struct user_msghdr __user *) (unsigned long) | |
1645 | READ_ONCE(sqe->addr); | |
1646 | ||
aa1fa28f | 1647 | ret = fn(sock, msg, flags); |
0fa03c62 JA |
1648 | if (force_nonblock && ret == -EAGAIN) |
1649 | return ret; | |
1650 | } | |
1651 | ||
c71ffb67 | 1652 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
5d17b4a4 JA |
1653 | io_put_req(req); |
1654 | return 0; | |
1655 | } | |
aa1fa28f JA |
1656 | #endif |
1657 | ||
1658 | static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
1659 | bool force_nonblock) | |
1660 | { | |
1661 | #if defined(CONFIG_NET) | |
1662 | return io_send_recvmsg(req, sqe, force_nonblock, __sys_sendmsg_sock); | |
1663 | #else | |
1664 | return -EOPNOTSUPP; | |
1665 | #endif | |
1666 | } | |
1667 | ||
1668 | static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
1669 | bool force_nonblock) | |
1670 | { | |
1671 | #if defined(CONFIG_NET) | |
1672 | return io_send_recvmsg(req, sqe, force_nonblock, __sys_recvmsg_sock); | |
0fa03c62 JA |
1673 | #else |
1674 | return -EOPNOTSUPP; | |
1675 | #endif | |
1676 | } | |
5d17b4a4 | 1677 | |
221c5eb2 JA |
1678 | static void io_poll_remove_one(struct io_kiocb *req) |
1679 | { | |
1680 | struct io_poll_iocb *poll = &req->poll; | |
1681 | ||
1682 | spin_lock(&poll->head->lock); | |
1683 | WRITE_ONCE(poll->canceled, true); | |
1684 | if (!list_empty(&poll->wait.entry)) { | |
1685 | list_del_init(&poll->wait.entry); | |
18d9be1a | 1686 | io_queue_async_work(req->ctx, req); |
221c5eb2 JA |
1687 | } |
1688 | spin_unlock(&poll->head->lock); | |
1689 | ||
1690 | list_del_init(&req->list); | |
1691 | } | |
1692 | ||
1693 | static void io_poll_remove_all(struct io_ring_ctx *ctx) | |
1694 | { | |
1695 | struct io_kiocb *req; | |
1696 | ||
1697 | spin_lock_irq(&ctx->completion_lock); | |
1698 | while (!list_empty(&ctx->cancel_list)) { | |
1699 | req = list_first_entry(&ctx->cancel_list, struct io_kiocb,list); | |
1700 | io_poll_remove_one(req); | |
1701 | } | |
1702 | spin_unlock_irq(&ctx->completion_lock); | |
1703 | } | |
1704 | ||
1705 | /* | |
1706 | * Find a running poll command that matches one specified in sqe->addr, | |
1707 | * and remove it if found. | |
1708 | */ | |
1709 | static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
1710 | { | |
1711 | struct io_ring_ctx *ctx = req->ctx; | |
1712 | struct io_kiocb *poll_req, *next; | |
1713 | int ret = -ENOENT; | |
1714 | ||
1715 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1716 | return -EINVAL; | |
1717 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index || | |
1718 | sqe->poll_events) | |
1719 | return -EINVAL; | |
1720 | ||
1721 | spin_lock_irq(&ctx->completion_lock); | |
1722 | list_for_each_entry_safe(poll_req, next, &ctx->cancel_list, list) { | |
1723 | if (READ_ONCE(sqe->addr) == poll_req->user_data) { | |
1724 | io_poll_remove_one(poll_req); | |
1725 | ret = 0; | |
1726 | break; | |
1727 | } | |
1728 | } | |
1729 | spin_unlock_irq(&ctx->completion_lock); | |
1730 | ||
c71ffb67 | 1731 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
e65ef56d | 1732 | io_put_req(req); |
221c5eb2 JA |
1733 | return 0; |
1734 | } | |
1735 | ||
8c838788 JA |
1736 | static void io_poll_complete(struct io_ring_ctx *ctx, struct io_kiocb *req, |
1737 | __poll_t mask) | |
221c5eb2 | 1738 | { |
8c838788 | 1739 | req->poll.done = true; |
c71ffb67 | 1740 | io_cqring_fill_event(ctx, req->user_data, mangle_poll(mask)); |
8c838788 | 1741 | io_commit_cqring(ctx); |
221c5eb2 JA |
1742 | } |
1743 | ||
1744 | static void io_poll_complete_work(struct work_struct *work) | |
1745 | { | |
1746 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
1747 | struct io_poll_iocb *poll = &req->poll; | |
1748 | struct poll_table_struct pt = { ._key = poll->events }; | |
1749 | struct io_ring_ctx *ctx = req->ctx; | |
1750 | __poll_t mask = 0; | |
1751 | ||
1752 | if (!READ_ONCE(poll->canceled)) | |
1753 | mask = vfs_poll(poll->file, &pt) & poll->events; | |
1754 | ||
1755 | /* | |
1756 | * Note that ->ki_cancel callers also delete iocb from active_reqs after | |
1757 | * calling ->ki_cancel. We need the ctx_lock roundtrip here to | |
1758 | * synchronize with them. In the cancellation case the list_del_init | |
1759 | * itself is not actually needed, but harmless so we keep it in to | |
1760 | * avoid further branches in the fast path. | |
1761 | */ | |
1762 | spin_lock_irq(&ctx->completion_lock); | |
1763 | if (!mask && !READ_ONCE(poll->canceled)) { | |
1764 | add_wait_queue(poll->head, &poll->wait); | |
1765 | spin_unlock_irq(&ctx->completion_lock); | |
1766 | return; | |
1767 | } | |
1768 | list_del_init(&req->list); | |
8c838788 | 1769 | io_poll_complete(ctx, req, mask); |
221c5eb2 JA |
1770 | spin_unlock_irq(&ctx->completion_lock); |
1771 | ||
8c838788 JA |
1772 | io_cqring_ev_posted(ctx); |
1773 | io_put_req(req); | |
221c5eb2 JA |
1774 | } |
1775 | ||
1776 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
1777 | void *key) | |
1778 | { | |
1779 | struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb, | |
1780 | wait); | |
1781 | struct io_kiocb *req = container_of(poll, struct io_kiocb, poll); | |
1782 | struct io_ring_ctx *ctx = req->ctx; | |
1783 | __poll_t mask = key_to_poll(key); | |
8c838788 | 1784 | unsigned long flags; |
221c5eb2 JA |
1785 | |
1786 | /* for instances that support it check for an event match first: */ | |
8c838788 JA |
1787 | if (mask && !(mask & poll->events)) |
1788 | return 0; | |
221c5eb2 | 1789 | |
8c838788 | 1790 | list_del_init(&poll->wait.entry); |
221c5eb2 | 1791 | |
8c838788 JA |
1792 | if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) { |
1793 | list_del(&req->list); | |
1794 | io_poll_complete(ctx, req, mask); | |
1795 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
221c5eb2 | 1796 | |
8c838788 JA |
1797 | io_cqring_ev_posted(ctx); |
1798 | io_put_req(req); | |
1799 | } else { | |
18d9be1a | 1800 | io_queue_async_work(ctx, req); |
221c5eb2 JA |
1801 | } |
1802 | ||
221c5eb2 JA |
1803 | return 1; |
1804 | } | |
1805 | ||
1806 | struct io_poll_table { | |
1807 | struct poll_table_struct pt; | |
1808 | struct io_kiocb *req; | |
1809 | int error; | |
1810 | }; | |
1811 | ||
1812 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
1813 | struct poll_table_struct *p) | |
1814 | { | |
1815 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
1816 | ||
1817 | if (unlikely(pt->req->poll.head)) { | |
1818 | pt->error = -EINVAL; | |
1819 | return; | |
1820 | } | |
1821 | ||
1822 | pt->error = 0; | |
1823 | pt->req->poll.head = head; | |
1824 | add_wait_queue(head, &pt->req->poll.wait); | |
1825 | } | |
1826 | ||
1827 | static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
1828 | { | |
1829 | struct io_poll_iocb *poll = &req->poll; | |
1830 | struct io_ring_ctx *ctx = req->ctx; | |
1831 | struct io_poll_table ipt; | |
8c838788 | 1832 | bool cancel = false; |
221c5eb2 JA |
1833 | __poll_t mask; |
1834 | u16 events; | |
221c5eb2 JA |
1835 | |
1836 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1837 | return -EINVAL; | |
1838 | if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
1839 | return -EINVAL; | |
09bb8394 JA |
1840 | if (!poll->file) |
1841 | return -EBADF; | |
221c5eb2 | 1842 | |
6cc47d1d | 1843 | req->submit.sqe = NULL; |
221c5eb2 JA |
1844 | INIT_WORK(&req->work, io_poll_complete_work); |
1845 | events = READ_ONCE(sqe->poll_events); | |
1846 | poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP; | |
1847 | ||
221c5eb2 | 1848 | poll->head = NULL; |
8c838788 | 1849 | poll->done = false; |
221c5eb2 JA |
1850 | poll->canceled = false; |
1851 | ||
1852 | ipt.pt._qproc = io_poll_queue_proc; | |
1853 | ipt.pt._key = poll->events; | |
1854 | ipt.req = req; | |
1855 | ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */ | |
1856 | ||
1857 | /* initialized the list so that we can do list_empty checks */ | |
1858 | INIT_LIST_HEAD(&poll->wait.entry); | |
1859 | init_waitqueue_func_entry(&poll->wait, io_poll_wake); | |
1860 | ||
36703247 JA |
1861 | INIT_LIST_HEAD(&req->list); |
1862 | ||
221c5eb2 | 1863 | mask = vfs_poll(poll->file, &ipt.pt) & poll->events; |
221c5eb2 JA |
1864 | |
1865 | spin_lock_irq(&ctx->completion_lock); | |
8c838788 JA |
1866 | if (likely(poll->head)) { |
1867 | spin_lock(&poll->head->lock); | |
1868 | if (unlikely(list_empty(&poll->wait.entry))) { | |
1869 | if (ipt.error) | |
1870 | cancel = true; | |
1871 | ipt.error = 0; | |
1872 | mask = 0; | |
1873 | } | |
1874 | if (mask || ipt.error) | |
1875 | list_del_init(&poll->wait.entry); | |
1876 | else if (cancel) | |
1877 | WRITE_ONCE(poll->canceled, true); | |
1878 | else if (!poll->done) /* actually waiting for an event */ | |
1879 | list_add_tail(&req->list, &ctx->cancel_list); | |
1880 | spin_unlock(&poll->head->lock); | |
1881 | } | |
1882 | if (mask) { /* no async, we'd stolen it */ | |
221c5eb2 | 1883 | ipt.error = 0; |
8c838788 | 1884 | io_poll_complete(ctx, req, mask); |
221c5eb2 | 1885 | } |
221c5eb2 JA |
1886 | spin_unlock_irq(&ctx->completion_lock); |
1887 | ||
8c838788 JA |
1888 | if (mask) { |
1889 | io_cqring_ev_posted(ctx); | |
e65ef56d | 1890 | io_put_req(req); |
221c5eb2 | 1891 | } |
8c838788 | 1892 | return ipt.error; |
221c5eb2 JA |
1893 | } |
1894 | ||
5262f567 JA |
1895 | static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer) |
1896 | { | |
1897 | struct io_ring_ctx *ctx; | |
ef03681a | 1898 | struct io_kiocb *req, *prev; |
5262f567 JA |
1899 | unsigned long flags; |
1900 | ||
1901 | req = container_of(timer, struct io_kiocb, timeout.timer); | |
1902 | ctx = req->ctx; | |
1903 | atomic_inc(&ctx->cq_timeouts); | |
1904 | ||
1905 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
ef03681a | 1906 | /* |
1907 | * Adjust the reqs sequence before the current one because it | |
1908 | * will consume a slot in the cq_ring and the the cq_tail pointer | |
1909 | * will be increased, otherwise other timeout reqs may return in | |
1910 | * advance without waiting for enough wait_nr. | |
1911 | */ | |
1912 | prev = req; | |
1913 | list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list) | |
1914 | prev->sequence++; | |
5262f567 JA |
1915 | list_del(&req->list); |
1916 | ||
1917 | io_cqring_fill_event(ctx, req->user_data, -ETIME); | |
1918 | io_commit_cqring(ctx); | |
1919 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1920 | ||
1921 | io_cqring_ev_posted(ctx); | |
1922 | ||
1923 | io_put_req(req); | |
1924 | return HRTIMER_NORESTART; | |
1925 | } | |
1926 | ||
1927 | static int io_timeout(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
1928 | { | |
5da0fb1a | 1929 | unsigned count; |
5262f567 JA |
1930 | struct io_ring_ctx *ctx = req->ctx; |
1931 | struct list_head *entry; | |
bdf20073 | 1932 | struct timespec64 ts; |
a1f58ba4 | 1933 | unsigned span = 0; |
5262f567 JA |
1934 | |
1935 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
1936 | return -EINVAL; | |
1937 | if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->timeout_flags || | |
1938 | sqe->len != 1) | |
1939 | return -EINVAL; | |
bdf20073 AB |
1940 | |
1941 | if (get_timespec64(&ts, u64_to_user_ptr(sqe->addr))) | |
5262f567 JA |
1942 | return -EFAULT; |
1943 | ||
1944 | /* | |
1945 | * sqe->off holds how many events that need to occur for this | |
1946 | * timeout event to be satisfied. | |
1947 | */ | |
1948 | count = READ_ONCE(sqe->off); | |
1949 | if (!count) | |
1950 | count = 1; | |
1951 | ||
1952 | req->sequence = ctx->cached_sq_head + count - 1; | |
5da0fb1a | 1953 | /* reuse it to store the count */ |
1954 | req->submit.sequence = count; | |
5262f567 JA |
1955 | req->flags |= REQ_F_TIMEOUT; |
1956 | ||
1957 | /* | |
1958 | * Insertion sort, ensuring the first entry in the list is always | |
1959 | * the one we need first. | |
1960 | */ | |
5262f567 JA |
1961 | spin_lock_irq(&ctx->completion_lock); |
1962 | list_for_each_prev(entry, &ctx->timeout_list) { | |
1963 | struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list); | |
5da0fb1a | 1964 | unsigned nxt_sq_head; |
1965 | long long tmp, tmp_nxt; | |
5262f567 | 1966 | |
5da0fb1a | 1967 | /* |
1968 | * Since cached_sq_head + count - 1 can overflow, use type long | |
1969 | * long to store it. | |
1970 | */ | |
1971 | tmp = (long long)ctx->cached_sq_head + count - 1; | |
1972 | nxt_sq_head = nxt->sequence - nxt->submit.sequence + 1; | |
1973 | tmp_nxt = (long long)nxt_sq_head + nxt->submit.sequence - 1; | |
1974 | ||
1975 | /* | |
1976 | * cached_sq_head may overflow, and it will never overflow twice | |
1977 | * once there is some timeout req still be valid. | |
1978 | */ | |
1979 | if (ctx->cached_sq_head < nxt_sq_head) | |
8b07a65a | 1980 | tmp += UINT_MAX; |
5da0fb1a | 1981 | |
a1f58ba4 | 1982 | if (tmp > tmp_nxt) |
5262f567 | 1983 | break; |
a1f58ba4 | 1984 | |
1985 | /* | |
1986 | * Sequence of reqs after the insert one and itself should | |
1987 | * be adjusted because each timeout req consumes a slot. | |
1988 | */ | |
1989 | span++; | |
1990 | nxt->sequence++; | |
5262f567 | 1991 | } |
a1f58ba4 | 1992 | req->sequence -= span; |
5262f567 JA |
1993 | list_add(&req->list, entry); |
1994 | spin_unlock_irq(&ctx->completion_lock); | |
1995 | ||
1996 | hrtimer_init(&req->timeout.timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
1997 | req->timeout.timer.function = io_timeout_fn; | |
bdf20073 | 1998 | hrtimer_start(&req->timeout.timer, timespec64_to_ktime(ts), |
5262f567 JA |
1999 | HRTIMER_MODE_REL); |
2000 | return 0; | |
2001 | } | |
2002 | ||
de0617e4 JA |
2003 | static int io_req_defer(struct io_ring_ctx *ctx, struct io_kiocb *req, |
2004 | const struct io_uring_sqe *sqe) | |
2005 | { | |
2006 | struct io_uring_sqe *sqe_copy; | |
2007 | ||
2008 | if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) | |
2009 | return 0; | |
2010 | ||
2011 | sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL); | |
2012 | if (!sqe_copy) | |
2013 | return -EAGAIN; | |
2014 | ||
2015 | spin_lock_irq(&ctx->completion_lock); | |
2016 | if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) { | |
2017 | spin_unlock_irq(&ctx->completion_lock); | |
2018 | kfree(sqe_copy); | |
2019 | return 0; | |
2020 | } | |
2021 | ||
2022 | memcpy(sqe_copy, sqe, sizeof(*sqe_copy)); | |
2023 | req->submit.sqe = sqe_copy; | |
2024 | ||
2025 | INIT_WORK(&req->work, io_sq_wq_submit_work); | |
2026 | list_add_tail(&req->list, &ctx->defer_list); | |
2027 | spin_unlock_irq(&ctx->completion_lock); | |
2028 | return -EIOCBQUEUED; | |
2029 | } | |
2030 | ||
2b188cc1 | 2031 | static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
8358e3a8 | 2032 | const struct sqe_submit *s, bool force_nonblock) |
2b188cc1 | 2033 | { |
e0c5c576 | 2034 | int ret, opcode; |
2b188cc1 | 2035 | |
9e645e11 JA |
2036 | req->user_data = READ_ONCE(s->sqe->user_data); |
2037 | ||
2b188cc1 JA |
2038 | if (unlikely(s->index >= ctx->sq_entries)) |
2039 | return -EINVAL; | |
2b188cc1 JA |
2040 | |
2041 | opcode = READ_ONCE(s->sqe->opcode); | |
2042 | switch (opcode) { | |
2043 | case IORING_OP_NOP: | |
2044 | ret = io_nop(req, req->user_data); | |
2045 | break; | |
2046 | case IORING_OP_READV: | |
edafccee JA |
2047 | if (unlikely(s->sqe->buf_index)) |
2048 | return -EINVAL; | |
8358e3a8 | 2049 | ret = io_read(req, s, force_nonblock); |
2b188cc1 JA |
2050 | break; |
2051 | case IORING_OP_WRITEV: | |
edafccee JA |
2052 | if (unlikely(s->sqe->buf_index)) |
2053 | return -EINVAL; | |
8358e3a8 | 2054 | ret = io_write(req, s, force_nonblock); |
edafccee JA |
2055 | break; |
2056 | case IORING_OP_READ_FIXED: | |
8358e3a8 | 2057 | ret = io_read(req, s, force_nonblock); |
edafccee JA |
2058 | break; |
2059 | case IORING_OP_WRITE_FIXED: | |
8358e3a8 | 2060 | ret = io_write(req, s, force_nonblock); |
2b188cc1 | 2061 | break; |
c992fe29 CH |
2062 | case IORING_OP_FSYNC: |
2063 | ret = io_fsync(req, s->sqe, force_nonblock); | |
2064 | break; | |
221c5eb2 JA |
2065 | case IORING_OP_POLL_ADD: |
2066 | ret = io_poll_add(req, s->sqe); | |
2067 | break; | |
2068 | case IORING_OP_POLL_REMOVE: | |
2069 | ret = io_poll_remove(req, s->sqe); | |
2070 | break; | |
5d17b4a4 JA |
2071 | case IORING_OP_SYNC_FILE_RANGE: |
2072 | ret = io_sync_file_range(req, s->sqe, force_nonblock); | |
2073 | break; | |
0fa03c62 JA |
2074 | case IORING_OP_SENDMSG: |
2075 | ret = io_sendmsg(req, s->sqe, force_nonblock); | |
2076 | break; | |
aa1fa28f JA |
2077 | case IORING_OP_RECVMSG: |
2078 | ret = io_recvmsg(req, s->sqe, force_nonblock); | |
2079 | break; | |
5262f567 JA |
2080 | case IORING_OP_TIMEOUT: |
2081 | ret = io_timeout(req, s->sqe); | |
2082 | break; | |
2b188cc1 JA |
2083 | default: |
2084 | ret = -EINVAL; | |
2085 | break; | |
2086 | } | |
2087 | ||
def596e9 JA |
2088 | if (ret) |
2089 | return ret; | |
2090 | ||
2091 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
9e645e11 | 2092 | if (req->result == -EAGAIN) |
def596e9 JA |
2093 | return -EAGAIN; |
2094 | ||
2095 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
2096 | if (s->needs_lock) | |
2097 | mutex_lock(&ctx->uring_lock); | |
2098 | io_iopoll_req_issued(req); | |
2099 | if (s->needs_lock) | |
2100 | mutex_unlock(&ctx->uring_lock); | |
2101 | } | |
2102 | ||
2103 | return 0; | |
2b188cc1 JA |
2104 | } |
2105 | ||
31b51510 JA |
2106 | static struct async_list *io_async_list_from_sqe(struct io_ring_ctx *ctx, |
2107 | const struct io_uring_sqe *sqe) | |
2108 | { | |
2109 | switch (sqe->opcode) { | |
2110 | case IORING_OP_READV: | |
2111 | case IORING_OP_READ_FIXED: | |
2112 | return &ctx->pending_async[READ]; | |
2113 | case IORING_OP_WRITEV: | |
2114 | case IORING_OP_WRITE_FIXED: | |
2115 | return &ctx->pending_async[WRITE]; | |
2116 | default: | |
2117 | return NULL; | |
2118 | } | |
2119 | } | |
2120 | ||
edafccee JA |
2121 | static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe) |
2122 | { | |
2123 | u8 opcode = READ_ONCE(sqe->opcode); | |
2124 | ||
2125 | return !(opcode == IORING_OP_READ_FIXED || | |
2126 | opcode == IORING_OP_WRITE_FIXED); | |
2127 | } | |
2128 | ||
2b188cc1 JA |
2129 | static void io_sq_wq_submit_work(struct work_struct *work) |
2130 | { | |
2131 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
2b188cc1 | 2132 | struct io_ring_ctx *ctx = req->ctx; |
31b51510 JA |
2133 | struct mm_struct *cur_mm = NULL; |
2134 | struct async_list *async_list; | |
2135 | LIST_HEAD(req_list); | |
edafccee | 2136 | mm_segment_t old_fs; |
2b188cc1 JA |
2137 | int ret; |
2138 | ||
31b51510 JA |
2139 | async_list = io_async_list_from_sqe(ctx, req->submit.sqe); |
2140 | restart: | |
2141 | do { | |
2142 | struct sqe_submit *s = &req->submit; | |
2143 | const struct io_uring_sqe *sqe = s->sqe; | |
d0ee8791 | 2144 | unsigned int flags = req->flags; |
2b188cc1 | 2145 | |
8449eeda | 2146 | /* Ensure we clear previously set non-block flag */ |
31b51510 JA |
2147 | req->rw.ki_flags &= ~IOCB_NOWAIT; |
2148 | ||
2149 | ret = 0; | |
2150 | if (io_sqe_needs_user(sqe) && !cur_mm) { | |
2151 | if (!mmget_not_zero(ctx->sqo_mm)) { | |
2152 | ret = -EFAULT; | |
2153 | } else { | |
2154 | cur_mm = ctx->sqo_mm; | |
2155 | use_mm(cur_mm); | |
2156 | old_fs = get_fs(); | |
2157 | set_fs(USER_DS); | |
2158 | } | |
2159 | } | |
2160 | ||
2161 | if (!ret) { | |
2162 | s->has_user = cur_mm != NULL; | |
2163 | s->needs_lock = true; | |
2164 | do { | |
8358e3a8 | 2165 | ret = __io_submit_sqe(ctx, req, s, false); |
31b51510 JA |
2166 | /* |
2167 | * We can get EAGAIN for polled IO even though | |
2168 | * we're forcing a sync submission from here, | |
2169 | * since we can't wait for request slots on the | |
2170 | * block side. | |
2171 | */ | |
2172 | if (ret != -EAGAIN) | |
2173 | break; | |
2174 | cond_resched(); | |
2175 | } while (1); | |
2176 | } | |
817869d2 JA |
2177 | |
2178 | /* drop submission reference */ | |
2179 | io_put_req(req); | |
2180 | ||
31b51510 | 2181 | if (ret) { |
c71ffb67 | 2182 | io_cqring_add_event(ctx, sqe->user_data, ret); |
e65ef56d | 2183 | io_put_req(req); |
31b51510 JA |
2184 | } |
2185 | ||
2186 | /* async context always use a copy of the sqe */ | |
2187 | kfree(sqe); | |
2188 | ||
f7b76ac9 | 2189 | /* req from defer and link list needn't decrease async cnt */ |
d0ee8791 | 2190 | if (flags & (REQ_F_IO_DRAINED | REQ_F_LINK_DONE)) |
f7b76ac9 ZL |
2191 | goto out; |
2192 | ||
31b51510 JA |
2193 | if (!async_list) |
2194 | break; | |
2195 | if (!list_empty(&req_list)) { | |
2196 | req = list_first_entry(&req_list, struct io_kiocb, | |
2197 | list); | |
2198 | list_del(&req->list); | |
2199 | continue; | |
2200 | } | |
2201 | if (list_empty(&async_list->list)) | |
2202 | break; | |
2203 | ||
2204 | req = NULL; | |
2205 | spin_lock(&async_list->lock); | |
2206 | if (list_empty(&async_list->list)) { | |
2207 | spin_unlock(&async_list->lock); | |
2208 | break; | |
2209 | } | |
2210 | list_splice_init(&async_list->list, &req_list); | |
2211 | spin_unlock(&async_list->lock); | |
2212 | ||
2213 | req = list_first_entry(&req_list, struct io_kiocb, list); | |
2214 | list_del(&req->list); | |
2215 | } while (req); | |
edafccee JA |
2216 | |
2217 | /* | |
31b51510 JA |
2218 | * Rare case of racing with a submitter. If we find the count has |
2219 | * dropped to zero AND we have pending work items, then restart | |
2220 | * the processing. This is a tiny race window. | |
edafccee | 2221 | */ |
31b51510 JA |
2222 | if (async_list) { |
2223 | ret = atomic_dec_return(&async_list->cnt); | |
2224 | while (!ret && !list_empty(&async_list->list)) { | |
2225 | spin_lock(&async_list->lock); | |
2226 | atomic_inc(&async_list->cnt); | |
2227 | list_splice_init(&async_list->list, &req_list); | |
2228 | spin_unlock(&async_list->lock); | |
2229 | ||
2230 | if (!list_empty(&req_list)) { | |
2231 | req = list_first_entry(&req_list, | |
2232 | struct io_kiocb, list); | |
2233 | list_del(&req->list); | |
2234 | goto restart; | |
2235 | } | |
2236 | ret = atomic_dec_return(&async_list->cnt); | |
edafccee | 2237 | } |
edafccee | 2238 | } |
2b188cc1 | 2239 | |
f7b76ac9 | 2240 | out: |
31b51510 | 2241 | if (cur_mm) { |
edafccee | 2242 | set_fs(old_fs); |
31b51510 JA |
2243 | unuse_mm(cur_mm); |
2244 | mmput(cur_mm); | |
2b188cc1 | 2245 | } |
31b51510 | 2246 | } |
2b188cc1 | 2247 | |
31b51510 JA |
2248 | /* |
2249 | * See if we can piggy back onto previously submitted work, that is still | |
2250 | * running. We currently only allow this if the new request is sequential | |
2251 | * to the previous one we punted. | |
2252 | */ | |
2253 | static bool io_add_to_prev_work(struct async_list *list, struct io_kiocb *req) | |
2254 | { | |
6d5d5ac5 | 2255 | bool ret; |
31b51510 JA |
2256 | |
2257 | if (!list) | |
2258 | return false; | |
2259 | if (!(req->flags & REQ_F_SEQ_PREV)) | |
2260 | return false; | |
2261 | if (!atomic_read(&list->cnt)) | |
2262 | return false; | |
2263 | ||
2264 | ret = true; | |
2265 | spin_lock(&list->lock); | |
2266 | list_add_tail(&req->list, &list->list); | |
c0e48f9d ZL |
2267 | /* |
2268 | * Ensure we see a simultaneous modification from io_sq_wq_submit_work() | |
2269 | */ | |
2270 | smp_mb(); | |
31b51510 JA |
2271 | if (!atomic_read(&list->cnt)) { |
2272 | list_del_init(&req->list); | |
2273 | ret = false; | |
2274 | } | |
2275 | spin_unlock(&list->lock); | |
2276 | return ret; | |
2b188cc1 JA |
2277 | } |
2278 | ||
09bb8394 JA |
2279 | static bool io_op_needs_file(const struct io_uring_sqe *sqe) |
2280 | { | |
2281 | int op = READ_ONCE(sqe->opcode); | |
2282 | ||
2283 | switch (op) { | |
2284 | case IORING_OP_NOP: | |
2285 | case IORING_OP_POLL_REMOVE: | |
2286 | return false; | |
2287 | default: | |
2288 | return true; | |
2289 | } | |
2290 | } | |
2291 | ||
2292 | static int io_req_set_file(struct io_ring_ctx *ctx, const struct sqe_submit *s, | |
2293 | struct io_submit_state *state, struct io_kiocb *req) | |
2294 | { | |
2295 | unsigned flags; | |
2296 | int fd; | |
2297 | ||
2298 | flags = READ_ONCE(s->sqe->flags); | |
2299 | fd = READ_ONCE(s->sqe->fd); | |
2300 | ||
4fe2c963 | 2301 | if (flags & IOSQE_IO_DRAIN) |
de0617e4 | 2302 | req->flags |= REQ_F_IO_DRAIN; |
4fe2c963 JL |
2303 | /* |
2304 | * All io need record the previous position, if LINK vs DARIN, | |
2305 | * it can be used to mark the position of the first IO in the | |
2306 | * link list. | |
2307 | */ | |
2308 | req->sequence = s->sequence; | |
de0617e4 | 2309 | |
60c112b0 | 2310 | if (!io_op_needs_file(s->sqe)) |
09bb8394 | 2311 | return 0; |
09bb8394 JA |
2312 | |
2313 | if (flags & IOSQE_FIXED_FILE) { | |
2314 | if (unlikely(!ctx->user_files || | |
2315 | (unsigned) fd >= ctx->nr_user_files)) | |
2316 | return -EBADF; | |
2317 | req->file = ctx->user_files[fd]; | |
2318 | req->flags |= REQ_F_FIXED_FILE; | |
2319 | } else { | |
2320 | if (s->needs_fixed_file) | |
2321 | return -EBADF; | |
2322 | req->file = io_file_get(state, fd); | |
2323 | if (unlikely(!req->file)) | |
2324 | return -EBADF; | |
2325 | } | |
2326 | ||
2327 | return 0; | |
2328 | } | |
2329 | ||
4fe2c963 | 2330 | static int __io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
bc808bce | 2331 | struct sqe_submit *s) |
2b188cc1 | 2332 | { |
e0c5c576 | 2333 | int ret; |
2b188cc1 | 2334 | |
bc808bce | 2335 | ret = __io_submit_sqe(ctx, req, s, true); |
491381ce JA |
2336 | |
2337 | /* | |
2338 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
2339 | * doesn't support non-blocking read/write attempts | |
2340 | */ | |
2341 | if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) || | |
2342 | (req->flags & REQ_F_MUST_PUNT))) { | |
2b188cc1 JA |
2343 | struct io_uring_sqe *sqe_copy; |
2344 | ||
954dab19 | 2345 | sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL); |
2b188cc1 | 2346 | if (sqe_copy) { |
31b51510 JA |
2347 | struct async_list *list; |
2348 | ||
2b188cc1 | 2349 | s->sqe = sqe_copy; |
2b188cc1 | 2350 | memcpy(&req->submit, s, sizeof(*s)); |
31b51510 JA |
2351 | list = io_async_list_from_sqe(ctx, s->sqe); |
2352 | if (!io_add_to_prev_work(list, req)) { | |
2353 | if (list) | |
2354 | atomic_inc(&list->cnt); | |
2355 | INIT_WORK(&req->work, io_sq_wq_submit_work); | |
18d9be1a | 2356 | io_queue_async_work(ctx, req); |
31b51510 | 2357 | } |
e65ef56d JA |
2358 | |
2359 | /* | |
2360 | * Queued up for async execution, worker will release | |
9e645e11 | 2361 | * submit reference when the iocb is actually submitted. |
e65ef56d JA |
2362 | */ |
2363 | return 0; | |
2b188cc1 JA |
2364 | } |
2365 | } | |
e65ef56d JA |
2366 | |
2367 | /* drop submission reference */ | |
2368 | io_put_req(req); | |
2369 | ||
2370 | /* and drop final reference, if we failed */ | |
9e645e11 JA |
2371 | if (ret) { |
2372 | io_cqring_add_event(ctx, req->user_data, ret); | |
2373 | if (req->flags & REQ_F_LINK) | |
2374 | req->flags |= REQ_F_FAIL_LINK; | |
e65ef56d | 2375 | io_put_req(req); |
9e645e11 | 2376 | } |
2b188cc1 JA |
2377 | |
2378 | return ret; | |
2379 | } | |
2380 | ||
4fe2c963 | 2381 | static int io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
bc808bce | 2382 | struct sqe_submit *s) |
4fe2c963 JL |
2383 | { |
2384 | int ret; | |
2385 | ||
2386 | ret = io_req_defer(ctx, req, s->sqe); | |
2387 | if (ret) { | |
2388 | if (ret != -EIOCBQUEUED) { | |
2389 | io_free_req(req); | |
2390 | io_cqring_add_event(ctx, s->sqe->user_data, ret); | |
2391 | } | |
2392 | return 0; | |
2393 | } | |
2394 | ||
bc808bce | 2395 | return __io_queue_sqe(ctx, req, s); |
4fe2c963 JL |
2396 | } |
2397 | ||
2398 | static int io_queue_link_head(struct io_ring_ctx *ctx, struct io_kiocb *req, | |
bc808bce | 2399 | struct sqe_submit *s, struct io_kiocb *shadow) |
4fe2c963 JL |
2400 | { |
2401 | int ret; | |
2402 | int need_submit = false; | |
2403 | ||
2404 | if (!shadow) | |
bc808bce | 2405 | return io_queue_sqe(ctx, req, s); |
4fe2c963 JL |
2406 | |
2407 | /* | |
2408 | * Mark the first IO in link list as DRAIN, let all the following | |
2409 | * IOs enter the defer list. all IO needs to be completed before link | |
2410 | * list. | |
2411 | */ | |
2412 | req->flags |= REQ_F_IO_DRAIN; | |
2413 | ret = io_req_defer(ctx, req, s->sqe); | |
2414 | if (ret) { | |
2415 | if (ret != -EIOCBQUEUED) { | |
2416 | io_free_req(req); | |
7b20238d | 2417 | __io_free_req(shadow); |
4fe2c963 JL |
2418 | io_cqring_add_event(ctx, s->sqe->user_data, ret); |
2419 | return 0; | |
2420 | } | |
2421 | } else { | |
2422 | /* | |
2423 | * If ret == 0 means that all IOs in front of link io are | |
2424 | * running done. let's queue link head. | |
2425 | */ | |
2426 | need_submit = true; | |
2427 | } | |
2428 | ||
2429 | /* Insert shadow req to defer_list, blocking next IOs */ | |
2430 | spin_lock_irq(&ctx->completion_lock); | |
2431 | list_add_tail(&shadow->list, &ctx->defer_list); | |
2432 | spin_unlock_irq(&ctx->completion_lock); | |
2433 | ||
2434 | if (need_submit) | |
bc808bce | 2435 | return __io_queue_sqe(ctx, req, s); |
4fe2c963 JL |
2436 | |
2437 | return 0; | |
2438 | } | |
2439 | ||
9e645e11 JA |
2440 | #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK) |
2441 | ||
2442 | static void io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s, | |
bc808bce | 2443 | struct io_submit_state *state, struct io_kiocb **link) |
9e645e11 JA |
2444 | { |
2445 | struct io_uring_sqe *sqe_copy; | |
2446 | struct io_kiocb *req; | |
2447 | int ret; | |
2448 | ||
2449 | /* enforce forwards compatibility on users */ | |
2450 | if (unlikely(s->sqe->flags & ~SQE_VALID_FLAGS)) { | |
2451 | ret = -EINVAL; | |
2452 | goto err; | |
2453 | } | |
2454 | ||
2455 | req = io_get_req(ctx, state); | |
2456 | if (unlikely(!req)) { | |
2457 | ret = -EAGAIN; | |
2458 | goto err; | |
2459 | } | |
2460 | ||
2461 | ret = io_req_set_file(ctx, s, state, req); | |
2462 | if (unlikely(ret)) { | |
2463 | err_req: | |
2464 | io_free_req(req); | |
2465 | err: | |
2466 | io_cqring_add_event(ctx, s->sqe->user_data, ret); | |
2467 | return; | |
2468 | } | |
2469 | ||
84d55dc5 PB |
2470 | req->user_data = s->sqe->user_data; |
2471 | ||
9e645e11 JA |
2472 | /* |
2473 | * If we already have a head request, queue this one for async | |
2474 | * submittal once the head completes. If we don't have a head but | |
2475 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
2476 | * submitted sync once the chain is complete. If none of those | |
2477 | * conditions are true (normal request), then just queue it. | |
2478 | */ | |
2479 | if (*link) { | |
2480 | struct io_kiocb *prev = *link; | |
2481 | ||
2482 | sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL); | |
2483 | if (!sqe_copy) { | |
2484 | ret = -EAGAIN; | |
2485 | goto err_req; | |
2486 | } | |
2487 | ||
2488 | s->sqe = sqe_copy; | |
2489 | memcpy(&req->submit, s, sizeof(*s)); | |
2490 | list_add_tail(&req->list, &prev->link_list); | |
2491 | } else if (s->sqe->flags & IOSQE_IO_LINK) { | |
2492 | req->flags |= REQ_F_LINK; | |
2493 | ||
2494 | memcpy(&req->submit, s, sizeof(*s)); | |
2495 | INIT_LIST_HEAD(&req->link_list); | |
2496 | *link = req; | |
2497 | } else { | |
bc808bce | 2498 | io_queue_sqe(ctx, req, s); |
9e645e11 JA |
2499 | } |
2500 | } | |
2501 | ||
9a56a232 JA |
2502 | /* |
2503 | * Batched submission is done, ensure local IO is flushed out. | |
2504 | */ | |
2505 | static void io_submit_state_end(struct io_submit_state *state) | |
2506 | { | |
2507 | blk_finish_plug(&state->plug); | |
3d6770fb | 2508 | io_file_put(state); |
2579f913 JA |
2509 | if (state->free_reqs) |
2510 | kmem_cache_free_bulk(req_cachep, state->free_reqs, | |
2511 | &state->reqs[state->cur_req]); | |
9a56a232 JA |
2512 | } |
2513 | ||
2514 | /* | |
2515 | * Start submission side cache. | |
2516 | */ | |
2517 | static void io_submit_state_start(struct io_submit_state *state, | |
2518 | struct io_ring_ctx *ctx, unsigned max_ios) | |
2519 | { | |
2520 | blk_start_plug(&state->plug); | |
2579f913 | 2521 | state->free_reqs = 0; |
9a56a232 JA |
2522 | state->file = NULL; |
2523 | state->ios_left = max_ios; | |
2524 | } | |
2525 | ||
2b188cc1 JA |
2526 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
2527 | { | |
75b28aff | 2528 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 2529 | |
75b28aff | 2530 | if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) { |
2b188cc1 JA |
2531 | /* |
2532 | * Ensure any loads from the SQEs are done at this point, | |
2533 | * since once we write the new head, the application could | |
2534 | * write new data to them. | |
2535 | */ | |
75b28aff | 2536 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); |
2b188cc1 JA |
2537 | } |
2538 | } | |
2539 | ||
2b188cc1 JA |
2540 | /* |
2541 | * Fetch an sqe, if one is available. Note that s->sqe will point to memory | |
2542 | * that is mapped by userspace. This means that care needs to be taken to | |
2543 | * ensure that reads are stable, as we cannot rely on userspace always | |
2544 | * being a good citizen. If members of the sqe are validated and then later | |
2545 | * used, it's important that those reads are done through READ_ONCE() to | |
2546 | * prevent a re-load down the line. | |
2547 | */ | |
2548 | static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s) | |
2549 | { | |
75b28aff HV |
2550 | struct io_rings *rings = ctx->rings; |
2551 | u32 *sq_array = ctx->sq_array; | |
2b188cc1 JA |
2552 | unsigned head; |
2553 | ||
2554 | /* | |
2555 | * The cached sq head (or cq tail) serves two purposes: | |
2556 | * | |
2557 | * 1) allows us to batch the cost of updating the user visible | |
2558 | * head updates. | |
2559 | * 2) allows the kernel side to track the head on its own, even | |
2560 | * though the application is the one updating it. | |
2561 | */ | |
2562 | head = ctx->cached_sq_head; | |
e523a29c | 2563 | /* make sure SQ entry isn't read before tail */ |
75b28aff | 2564 | if (head == smp_load_acquire(&rings->sq.tail)) |
2b188cc1 JA |
2565 | return false; |
2566 | ||
75b28aff | 2567 | head = READ_ONCE(sq_array[head & ctx->sq_mask]); |
2b188cc1 JA |
2568 | if (head < ctx->sq_entries) { |
2569 | s->index = head; | |
2570 | s->sqe = &ctx->sq_sqes[head]; | |
8776f3fa | 2571 | s->sequence = ctx->cached_sq_head; |
2b188cc1 JA |
2572 | ctx->cached_sq_head++; |
2573 | return true; | |
2574 | } | |
2575 | ||
2576 | /* drop invalid entries */ | |
2577 | ctx->cached_sq_head++; | |
498ccd9e JA |
2578 | ctx->cached_sq_dropped++; |
2579 | WRITE_ONCE(rings->sq_dropped, ctx->cached_sq_dropped); | |
2b188cc1 JA |
2580 | return false; |
2581 | } | |
2582 | ||
fb5ccc98 PB |
2583 | static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr, |
2584 | bool has_user, bool mm_fault) | |
6c271ce2 JA |
2585 | { |
2586 | struct io_submit_state state, *statep = NULL; | |
9e645e11 | 2587 | struct io_kiocb *link = NULL; |
4fe2c963 | 2588 | struct io_kiocb *shadow_req = NULL; |
9e645e11 JA |
2589 | bool prev_was_link = false; |
2590 | int i, submitted = 0; | |
6c271ce2 JA |
2591 | |
2592 | if (nr > IO_PLUG_THRESHOLD) { | |
2593 | io_submit_state_start(&state, ctx, nr); | |
2594 | statep = &state; | |
2595 | } | |
2596 | ||
2597 | for (i = 0; i < nr; i++) { | |
fb5ccc98 PB |
2598 | struct sqe_submit s; |
2599 | ||
2600 | if (!io_get_sqring(ctx, &s)) | |
2601 | break; | |
2602 | ||
9e645e11 JA |
2603 | /* |
2604 | * If previous wasn't linked and we have a linked command, | |
2605 | * that's the end of the chain. Submit the previous link. | |
2606 | */ | |
2607 | if (!prev_was_link && link) { | |
bc808bce | 2608 | io_queue_link_head(ctx, link, &link->submit, shadow_req); |
9e645e11 | 2609 | link = NULL; |
5f5ad9ce | 2610 | shadow_req = NULL; |
9e645e11 | 2611 | } |
fb5ccc98 | 2612 | prev_was_link = (s.sqe->flags & IOSQE_IO_LINK) != 0; |
9e645e11 | 2613 | |
fb5ccc98 | 2614 | if (link && (s.sqe->flags & IOSQE_IO_DRAIN)) { |
4fe2c963 JL |
2615 | if (!shadow_req) { |
2616 | shadow_req = io_get_req(ctx, NULL); | |
a1041c27 JL |
2617 | if (unlikely(!shadow_req)) |
2618 | goto out; | |
4fe2c963 JL |
2619 | shadow_req->flags |= (REQ_F_IO_DRAIN | REQ_F_SHADOW_DRAIN); |
2620 | refcount_dec(&shadow_req->refs); | |
2621 | } | |
fb5ccc98 | 2622 | shadow_req->sequence = s.sequence; |
4fe2c963 JL |
2623 | } |
2624 | ||
a1041c27 | 2625 | out: |
6c271ce2 | 2626 | if (unlikely(mm_fault)) { |
fb5ccc98 | 2627 | io_cqring_add_event(ctx, s.sqe->user_data, |
9e645e11 | 2628 | -EFAULT); |
6c271ce2 | 2629 | } else { |
fb5ccc98 PB |
2630 | s.has_user = has_user; |
2631 | s.needs_lock = true; | |
2632 | s.needs_fixed_file = true; | |
2633 | io_submit_sqe(ctx, &s, statep, &link); | |
6c271ce2 | 2634 | submitted++; |
6c271ce2 | 2635 | } |
6c271ce2 JA |
2636 | } |
2637 | ||
9e645e11 | 2638 | if (link) |
bc808bce | 2639 | io_queue_link_head(ctx, link, &link->submit, shadow_req); |
6c271ce2 JA |
2640 | if (statep) |
2641 | io_submit_state_end(&state); | |
2642 | ||
2643 | return submitted; | |
2644 | } | |
2645 | ||
2646 | static int io_sq_thread(void *data) | |
2647 | { | |
6c271ce2 JA |
2648 | struct io_ring_ctx *ctx = data; |
2649 | struct mm_struct *cur_mm = NULL; | |
2650 | mm_segment_t old_fs; | |
2651 | DEFINE_WAIT(wait); | |
2652 | unsigned inflight; | |
2653 | unsigned long timeout; | |
2654 | ||
a4c0b3de JL |
2655 | complete(&ctx->sqo_thread_started); |
2656 | ||
6c271ce2 JA |
2657 | old_fs = get_fs(); |
2658 | set_fs(USER_DS); | |
2659 | ||
2660 | timeout = inflight = 0; | |
2bbcd6d3 | 2661 | while (!kthread_should_park()) { |
fb5ccc98 PB |
2662 | bool mm_fault = false; |
2663 | unsigned int to_submit; | |
6c271ce2 JA |
2664 | |
2665 | if (inflight) { | |
2666 | unsigned nr_events = 0; | |
2667 | ||
2668 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
2b2ed975 JA |
2669 | /* |
2670 | * inflight is the count of the maximum possible | |
2671 | * entries we submitted, but it can be smaller | |
2672 | * if we dropped some of them. If we don't have | |
2673 | * poll entries available, then we know that we | |
2674 | * have nothing left to poll for. Reset the | |
2675 | * inflight count to zero in that case. | |
2676 | */ | |
2677 | mutex_lock(&ctx->uring_lock); | |
2678 | if (!list_empty(&ctx->poll_list)) | |
2679 | __io_iopoll_check(ctx, &nr_events, 0); | |
2680 | else | |
2681 | inflight = 0; | |
2682 | mutex_unlock(&ctx->uring_lock); | |
6c271ce2 JA |
2683 | } else { |
2684 | /* | |
2685 | * Normal IO, just pretend everything completed. | |
2686 | * We don't have to poll completions for that. | |
2687 | */ | |
2688 | nr_events = inflight; | |
2689 | } | |
2690 | ||
2691 | inflight -= nr_events; | |
2692 | if (!inflight) | |
2693 | timeout = jiffies + ctx->sq_thread_idle; | |
2694 | } | |
2695 | ||
fb5ccc98 PB |
2696 | to_submit = io_sqring_entries(ctx); |
2697 | if (!to_submit) { | |
6c271ce2 JA |
2698 | /* |
2699 | * We're polling. If we're within the defined idle | |
2700 | * period, then let us spin without work before going | |
2701 | * to sleep. | |
2702 | */ | |
2703 | if (inflight || !time_after(jiffies, timeout)) { | |
9831a90c | 2704 | cond_resched(); |
6c271ce2 JA |
2705 | continue; |
2706 | } | |
2707 | ||
2708 | /* | |
2709 | * Drop cur_mm before scheduling, we can't hold it for | |
2710 | * long periods (or over schedule()). Do this before | |
2711 | * adding ourselves to the waitqueue, as the unuse/drop | |
2712 | * may sleep. | |
2713 | */ | |
2714 | if (cur_mm) { | |
2715 | unuse_mm(cur_mm); | |
2716 | mmput(cur_mm); | |
2717 | cur_mm = NULL; | |
2718 | } | |
2719 | ||
2720 | prepare_to_wait(&ctx->sqo_wait, &wait, | |
2721 | TASK_INTERRUPTIBLE); | |
2722 | ||
2723 | /* Tell userspace we may need a wakeup call */ | |
75b28aff | 2724 | ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP; |
0d7bae69 SB |
2725 | /* make sure to read SQ tail after writing flags */ |
2726 | smp_mb(); | |
6c271ce2 | 2727 | |
fb5ccc98 PB |
2728 | to_submit = io_sqring_entries(ctx); |
2729 | if (!to_submit) { | |
2bbcd6d3 | 2730 | if (kthread_should_park()) { |
6c271ce2 JA |
2731 | finish_wait(&ctx->sqo_wait, &wait); |
2732 | break; | |
2733 | } | |
2734 | if (signal_pending(current)) | |
2735 | flush_signals(current); | |
2736 | schedule(); | |
2737 | finish_wait(&ctx->sqo_wait, &wait); | |
2738 | ||
75b28aff | 2739 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
2740 | continue; |
2741 | } | |
2742 | finish_wait(&ctx->sqo_wait, &wait); | |
2743 | ||
75b28aff | 2744 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
2745 | } |
2746 | ||
6c271ce2 | 2747 | /* Unless all new commands are FIXED regions, grab mm */ |
fb5ccc98 | 2748 | if (!cur_mm) { |
6c271ce2 JA |
2749 | mm_fault = !mmget_not_zero(ctx->sqo_mm); |
2750 | if (!mm_fault) { | |
2751 | use_mm(ctx->sqo_mm); | |
2752 | cur_mm = ctx->sqo_mm; | |
2753 | } | |
2754 | } | |
2755 | ||
fb5ccc98 PB |
2756 | to_submit = min(to_submit, ctx->sq_entries); |
2757 | inflight += io_submit_sqes(ctx, to_submit, cur_mm != NULL, | |
2758 | mm_fault); | |
6c271ce2 JA |
2759 | |
2760 | /* Commit SQ ring head once we've consumed all SQEs */ | |
2761 | io_commit_sqring(ctx); | |
2762 | } | |
2763 | ||
2764 | set_fs(old_fs); | |
2765 | if (cur_mm) { | |
2766 | unuse_mm(cur_mm); | |
2767 | mmput(cur_mm); | |
2768 | } | |
06058632 | 2769 | |
2bbcd6d3 | 2770 | kthread_parkme(); |
06058632 | 2771 | |
6c271ce2 JA |
2772 | return 0; |
2773 | } | |
2774 | ||
bc808bce | 2775 | static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit) |
2b188cc1 | 2776 | { |
9a56a232 | 2777 | struct io_submit_state state, *statep = NULL; |
9e645e11 | 2778 | struct io_kiocb *link = NULL; |
4fe2c963 | 2779 | struct io_kiocb *shadow_req = NULL; |
9e645e11 | 2780 | bool prev_was_link = false; |
5c8b0b54 | 2781 | int i, submit = 0; |
2b188cc1 | 2782 | |
9a56a232 JA |
2783 | if (to_submit > IO_PLUG_THRESHOLD) { |
2784 | io_submit_state_start(&state, ctx, to_submit); | |
2785 | statep = &state; | |
2786 | } | |
2b188cc1 JA |
2787 | |
2788 | for (i = 0; i < to_submit; i++) { | |
2789 | struct sqe_submit s; | |
2790 | ||
2791 | if (!io_get_sqring(ctx, &s)) | |
2792 | break; | |
2793 | ||
9e645e11 JA |
2794 | /* |
2795 | * If previous wasn't linked and we have a linked command, | |
2796 | * that's the end of the chain. Submit the previous link. | |
2797 | */ | |
2798 | if (!prev_was_link && link) { | |
bc808bce | 2799 | io_queue_link_head(ctx, link, &link->submit, shadow_req); |
9e645e11 | 2800 | link = NULL; |
5f5ad9ce | 2801 | shadow_req = NULL; |
9e645e11 JA |
2802 | } |
2803 | prev_was_link = (s.sqe->flags & IOSQE_IO_LINK) != 0; | |
2804 | ||
4fe2c963 JL |
2805 | if (link && (s.sqe->flags & IOSQE_IO_DRAIN)) { |
2806 | if (!shadow_req) { | |
2807 | shadow_req = io_get_req(ctx, NULL); | |
a1041c27 JL |
2808 | if (unlikely(!shadow_req)) |
2809 | goto out; | |
4fe2c963 JL |
2810 | shadow_req->flags |= (REQ_F_IO_DRAIN | REQ_F_SHADOW_DRAIN); |
2811 | refcount_dec(&shadow_req->refs); | |
2812 | } | |
2813 | shadow_req->sequence = s.sequence; | |
2814 | } | |
2815 | ||
a1041c27 | 2816 | out: |
2b188cc1 | 2817 | s.has_user = true; |
def596e9 | 2818 | s.needs_lock = false; |
6c271ce2 | 2819 | s.needs_fixed_file = false; |
5c8b0b54 | 2820 | submit++; |
bc808bce | 2821 | io_submit_sqe(ctx, &s, statep, &link); |
2b188cc1 | 2822 | } |
2b188cc1 | 2823 | |
9e645e11 | 2824 | if (link) |
bc808bce | 2825 | io_queue_link_head(ctx, link, &link->submit, shadow_req); |
9a56a232 JA |
2826 | if (statep) |
2827 | io_submit_state_end(statep); | |
2b188cc1 | 2828 | |
935d1e45 PB |
2829 | io_commit_sqring(ctx); |
2830 | ||
5c8b0b54 | 2831 | return submit; |
2b188cc1 JA |
2832 | } |
2833 | ||
bda52162 JA |
2834 | struct io_wait_queue { |
2835 | struct wait_queue_entry wq; | |
2836 | struct io_ring_ctx *ctx; | |
2837 | unsigned to_wait; | |
2838 | unsigned nr_timeouts; | |
2839 | }; | |
2840 | ||
2841 | static inline bool io_should_wake(struct io_wait_queue *iowq) | |
2842 | { | |
2843 | struct io_ring_ctx *ctx = iowq->ctx; | |
2844 | ||
2845 | /* | |
2846 | * Wake up if we have enough events, or if a timeout occured since we | |
2847 | * started waiting. For timeouts, we always want to return to userspace, | |
2848 | * regardless of event count. | |
2849 | */ | |
2850 | return io_cqring_events(ctx->rings) >= iowq->to_wait || | |
2851 | atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; | |
2852 | } | |
2853 | ||
2854 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
2855 | int wake_flags, void *key) | |
2856 | { | |
2857 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
2858 | wq); | |
2859 | ||
2860 | if (!io_should_wake(iowq)) | |
2861 | return -1; | |
2862 | ||
2863 | return autoremove_wake_function(curr, mode, wake_flags, key); | |
2864 | } | |
2865 | ||
2b188cc1 JA |
2866 | /* |
2867 | * Wait until events become available, if we don't already have some. The | |
2868 | * application must reap them itself, as they reside on the shared cq ring. | |
2869 | */ | |
2870 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
2871 | const sigset_t __user *sig, size_t sigsz) | |
2872 | { | |
bda52162 JA |
2873 | struct io_wait_queue iowq = { |
2874 | .wq = { | |
2875 | .private = current, | |
2876 | .func = io_wake_function, | |
2877 | .entry = LIST_HEAD_INIT(iowq.wq.entry), | |
2878 | }, | |
2879 | .ctx = ctx, | |
2880 | .to_wait = min_events, | |
2881 | }; | |
75b28aff | 2882 | struct io_rings *rings = ctx->rings; |
2b188cc1 JA |
2883 | int ret; |
2884 | ||
75b28aff | 2885 | if (io_cqring_events(rings) >= min_events) |
2b188cc1 JA |
2886 | return 0; |
2887 | ||
2888 | if (sig) { | |
9e75ad5d AB |
2889 | #ifdef CONFIG_COMPAT |
2890 | if (in_compat_syscall()) | |
2891 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 2892 | sigsz); |
9e75ad5d AB |
2893 | else |
2894 | #endif | |
b772434b | 2895 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 2896 | |
2b188cc1 JA |
2897 | if (ret) |
2898 | return ret; | |
2899 | } | |
2900 | ||
bda52162 JA |
2901 | ret = 0; |
2902 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); | |
2903 | do { | |
2904 | prepare_to_wait_exclusive(&ctx->wait, &iowq.wq, | |
2905 | TASK_INTERRUPTIBLE); | |
2906 | if (io_should_wake(&iowq)) | |
2907 | break; | |
2908 | schedule(); | |
2909 | if (signal_pending(current)) { | |
2910 | ret = -ERESTARTSYS; | |
2911 | break; | |
2912 | } | |
2913 | } while (1); | |
2914 | finish_wait(&ctx->wait, &iowq.wq); | |
2915 | ||
b772434b | 2916 | restore_saved_sigmask_unless(ret == -ERESTARTSYS); |
97abc889 ON |
2917 | if (ret == -ERESTARTSYS) |
2918 | ret = -EINTR; | |
2b188cc1 | 2919 | |
75b28aff | 2920 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
2921 | } |
2922 | ||
6b06314c JA |
2923 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
2924 | { | |
2925 | #if defined(CONFIG_UNIX) | |
2926 | if (ctx->ring_sock) { | |
2927 | struct sock *sock = ctx->ring_sock->sk; | |
2928 | struct sk_buff *skb; | |
2929 | ||
2930 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
2931 | kfree_skb(skb); | |
2932 | } | |
2933 | #else | |
2934 | int i; | |
2935 | ||
2936 | for (i = 0; i < ctx->nr_user_files; i++) | |
2937 | fput(ctx->user_files[i]); | |
2938 | #endif | |
2939 | } | |
2940 | ||
2941 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) | |
2942 | { | |
2943 | if (!ctx->user_files) | |
2944 | return -ENXIO; | |
2945 | ||
2946 | __io_sqe_files_unregister(ctx); | |
2947 | kfree(ctx->user_files); | |
2948 | ctx->user_files = NULL; | |
2949 | ctx->nr_user_files = 0; | |
2950 | return 0; | |
2951 | } | |
2952 | ||
6c271ce2 JA |
2953 | static void io_sq_thread_stop(struct io_ring_ctx *ctx) |
2954 | { | |
2955 | if (ctx->sqo_thread) { | |
a4c0b3de | 2956 | wait_for_completion(&ctx->sqo_thread_started); |
2bbcd6d3 RP |
2957 | /* |
2958 | * The park is a bit of a work-around, without it we get | |
2959 | * warning spews on shutdown with SQPOLL set and affinity | |
2960 | * set to a single CPU. | |
2961 | */ | |
06058632 | 2962 | kthread_park(ctx->sqo_thread); |
6c271ce2 JA |
2963 | kthread_stop(ctx->sqo_thread); |
2964 | ctx->sqo_thread = NULL; | |
2965 | } | |
2966 | } | |
2967 | ||
6b06314c JA |
2968 | static void io_finish_async(struct io_ring_ctx *ctx) |
2969 | { | |
54a91f3b JA |
2970 | int i; |
2971 | ||
6c271ce2 JA |
2972 | io_sq_thread_stop(ctx); |
2973 | ||
54a91f3b JA |
2974 | for (i = 0; i < ARRAY_SIZE(ctx->sqo_wq); i++) { |
2975 | if (ctx->sqo_wq[i]) { | |
2976 | destroy_workqueue(ctx->sqo_wq[i]); | |
2977 | ctx->sqo_wq[i] = NULL; | |
2978 | } | |
6b06314c JA |
2979 | } |
2980 | } | |
2981 | ||
2982 | #if defined(CONFIG_UNIX) | |
2983 | static void io_destruct_skb(struct sk_buff *skb) | |
2984 | { | |
2985 | struct io_ring_ctx *ctx = skb->sk->sk_user_data; | |
8a997340 JA |
2986 | int i; |
2987 | ||
2988 | for (i = 0; i < ARRAY_SIZE(ctx->sqo_wq); i++) | |
2989 | if (ctx->sqo_wq[i]) | |
2990 | flush_workqueue(ctx->sqo_wq[i]); | |
6b06314c | 2991 | |
6b06314c JA |
2992 | unix_destruct_scm(skb); |
2993 | } | |
2994 | ||
2995 | /* | |
2996 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
2997 | * the io_uring can be safely unregistered on process exit, even if we have | |
2998 | * loops in the file referencing. | |
2999 | */ | |
3000 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
3001 | { | |
3002 | struct sock *sk = ctx->ring_sock->sk; | |
3003 | struct scm_fp_list *fpl; | |
3004 | struct sk_buff *skb; | |
3005 | int i; | |
3006 | ||
3007 | if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) { | |
3008 | unsigned long inflight = ctx->user->unix_inflight + nr; | |
3009 | ||
3010 | if (inflight > task_rlimit(current, RLIMIT_NOFILE)) | |
3011 | return -EMFILE; | |
3012 | } | |
3013 | ||
3014 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); | |
3015 | if (!fpl) | |
3016 | return -ENOMEM; | |
3017 | ||
3018 | skb = alloc_skb(0, GFP_KERNEL); | |
3019 | if (!skb) { | |
3020 | kfree(fpl); | |
3021 | return -ENOMEM; | |
3022 | } | |
3023 | ||
3024 | skb->sk = sk; | |
3025 | skb->destructor = io_destruct_skb; | |
3026 | ||
3027 | fpl->user = get_uid(ctx->user); | |
3028 | for (i = 0; i < nr; i++) { | |
3029 | fpl->fp[i] = get_file(ctx->user_files[i + offset]); | |
3030 | unix_inflight(fpl->user, fpl->fp[i]); | |
3031 | } | |
3032 | ||
3033 | fpl->max = fpl->count = nr; | |
3034 | UNIXCB(skb).fp = fpl; | |
3035 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); | |
3036 | skb_queue_head(&sk->sk_receive_queue, skb); | |
3037 | ||
3038 | for (i = 0; i < nr; i++) | |
3039 | fput(fpl->fp[i]); | |
3040 | ||
3041 | return 0; | |
3042 | } | |
3043 | ||
3044 | /* | |
3045 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
3046 | * causes regular reference counting to break down. We rely on the UNIX | |
3047 | * garbage collection to take care of this problem for us. | |
3048 | */ | |
3049 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
3050 | { | |
3051 | unsigned left, total; | |
3052 | int ret = 0; | |
3053 | ||
3054 | total = 0; | |
3055 | left = ctx->nr_user_files; | |
3056 | while (left) { | |
3057 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
3058 | |
3059 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
3060 | if (ret) | |
3061 | break; | |
3062 | left -= this_files; | |
3063 | total += this_files; | |
3064 | } | |
3065 | ||
3066 | if (!ret) | |
3067 | return 0; | |
3068 | ||
3069 | while (total < ctx->nr_user_files) { | |
3070 | fput(ctx->user_files[total]); | |
3071 | total++; | |
3072 | } | |
3073 | ||
3074 | return ret; | |
3075 | } | |
3076 | #else | |
3077 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
3078 | { | |
3079 | return 0; | |
3080 | } | |
3081 | #endif | |
3082 | ||
3083 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, | |
3084 | unsigned nr_args) | |
3085 | { | |
3086 | __s32 __user *fds = (__s32 __user *) arg; | |
3087 | int fd, ret = 0; | |
3088 | unsigned i; | |
3089 | ||
3090 | if (ctx->user_files) | |
3091 | return -EBUSY; | |
3092 | if (!nr_args) | |
3093 | return -EINVAL; | |
3094 | if (nr_args > IORING_MAX_FIXED_FILES) | |
3095 | return -EMFILE; | |
3096 | ||
3097 | ctx->user_files = kcalloc(nr_args, sizeof(struct file *), GFP_KERNEL); | |
3098 | if (!ctx->user_files) | |
3099 | return -ENOMEM; | |
3100 | ||
3101 | for (i = 0; i < nr_args; i++) { | |
3102 | ret = -EFAULT; | |
3103 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) | |
3104 | break; | |
3105 | ||
3106 | ctx->user_files[i] = fget(fd); | |
3107 | ||
3108 | ret = -EBADF; | |
3109 | if (!ctx->user_files[i]) | |
3110 | break; | |
3111 | /* | |
3112 | * Don't allow io_uring instances to be registered. If UNIX | |
3113 | * isn't enabled, then this causes a reference cycle and this | |
3114 | * instance can never get freed. If UNIX is enabled we'll | |
3115 | * handle it just fine, but there's still no point in allowing | |
3116 | * a ring fd as it doesn't support regular read/write anyway. | |
3117 | */ | |
3118 | if (ctx->user_files[i]->f_op == &io_uring_fops) { | |
3119 | fput(ctx->user_files[i]); | |
3120 | break; | |
3121 | } | |
3122 | ctx->nr_user_files++; | |
3123 | ret = 0; | |
3124 | } | |
3125 | ||
3126 | if (ret) { | |
3127 | for (i = 0; i < ctx->nr_user_files; i++) | |
3128 | fput(ctx->user_files[i]); | |
3129 | ||
3130 | kfree(ctx->user_files); | |
25adf50f | 3131 | ctx->user_files = NULL; |
6b06314c JA |
3132 | ctx->nr_user_files = 0; |
3133 | return ret; | |
3134 | } | |
3135 | ||
3136 | ret = io_sqe_files_scm(ctx); | |
3137 | if (ret) | |
3138 | io_sqe_files_unregister(ctx); | |
3139 | ||
3140 | return ret; | |
3141 | } | |
3142 | ||
6c271ce2 JA |
3143 | static int io_sq_offload_start(struct io_ring_ctx *ctx, |
3144 | struct io_uring_params *p) | |
2b188cc1 JA |
3145 | { |
3146 | int ret; | |
3147 | ||
6c271ce2 | 3148 | init_waitqueue_head(&ctx->sqo_wait); |
2b188cc1 JA |
3149 | mmgrab(current->mm); |
3150 | ctx->sqo_mm = current->mm; | |
3151 | ||
6c271ce2 | 3152 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3ec482d1 JA |
3153 | ret = -EPERM; |
3154 | if (!capable(CAP_SYS_ADMIN)) | |
3155 | goto err; | |
3156 | ||
917257da JA |
3157 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
3158 | if (!ctx->sq_thread_idle) | |
3159 | ctx->sq_thread_idle = HZ; | |
3160 | ||
6c271ce2 | 3161 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 3162 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 3163 | |
917257da | 3164 | ret = -EINVAL; |
44a9bd18 JA |
3165 | if (cpu >= nr_cpu_ids) |
3166 | goto err; | |
7889f44d | 3167 | if (!cpu_online(cpu)) |
917257da JA |
3168 | goto err; |
3169 | ||
6c271ce2 JA |
3170 | ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread, |
3171 | ctx, cpu, | |
3172 | "io_uring-sq"); | |
3173 | } else { | |
3174 | ctx->sqo_thread = kthread_create(io_sq_thread, ctx, | |
3175 | "io_uring-sq"); | |
3176 | } | |
3177 | if (IS_ERR(ctx->sqo_thread)) { | |
3178 | ret = PTR_ERR(ctx->sqo_thread); | |
3179 | ctx->sqo_thread = NULL; | |
3180 | goto err; | |
3181 | } | |
3182 | wake_up_process(ctx->sqo_thread); | |
3183 | } else if (p->flags & IORING_SETUP_SQ_AFF) { | |
3184 | /* Can't have SQ_AFF without SQPOLL */ | |
3185 | ret = -EINVAL; | |
3186 | goto err; | |
3187 | } | |
3188 | ||
2b188cc1 | 3189 | /* Do QD, or 2 * CPUS, whatever is smallest */ |
54a91f3b JA |
3190 | ctx->sqo_wq[0] = alloc_workqueue("io_ring-wq", |
3191 | WQ_UNBOUND | WQ_FREEZABLE, | |
2b188cc1 | 3192 | min(ctx->sq_entries - 1, 2 * num_online_cpus())); |
54a91f3b JA |
3193 | if (!ctx->sqo_wq[0]) { |
3194 | ret = -ENOMEM; | |
3195 | goto err; | |
3196 | } | |
3197 | ||
3198 | /* | |
3199 | * This is for buffered writes, where we want to limit the parallelism | |
3200 | * due to file locking in file systems. As "normal" buffered writes | |
3201 | * should parellelize on writeout quite nicely, limit us to having 2 | |
3202 | * pending. This avoids massive contention on the inode when doing | |
3203 | * buffered async writes. | |
3204 | */ | |
3205 | ctx->sqo_wq[1] = alloc_workqueue("io_ring-write-wq", | |
3206 | WQ_UNBOUND | WQ_FREEZABLE, 2); | |
3207 | if (!ctx->sqo_wq[1]) { | |
2b188cc1 JA |
3208 | ret = -ENOMEM; |
3209 | goto err; | |
3210 | } | |
3211 | ||
3212 | return 0; | |
3213 | err: | |
54a91f3b | 3214 | io_finish_async(ctx); |
2b188cc1 JA |
3215 | mmdrop(ctx->sqo_mm); |
3216 | ctx->sqo_mm = NULL; | |
3217 | return ret; | |
3218 | } | |
3219 | ||
3220 | static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages) | |
3221 | { | |
3222 | atomic_long_sub(nr_pages, &user->locked_vm); | |
3223 | } | |
3224 | ||
3225 | static int io_account_mem(struct user_struct *user, unsigned long nr_pages) | |
3226 | { | |
3227 | unsigned long page_limit, cur_pages, new_pages; | |
3228 | ||
3229 | /* Don't allow more pages than we can safely lock */ | |
3230 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
3231 | ||
3232 | do { | |
3233 | cur_pages = atomic_long_read(&user->locked_vm); | |
3234 | new_pages = cur_pages + nr_pages; | |
3235 | if (new_pages > page_limit) | |
3236 | return -ENOMEM; | |
3237 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
3238 | new_pages) != cur_pages); | |
3239 | ||
3240 | return 0; | |
3241 | } | |
3242 | ||
3243 | static void io_mem_free(void *ptr) | |
3244 | { | |
52e04ef4 MR |
3245 | struct page *page; |
3246 | ||
3247 | if (!ptr) | |
3248 | return; | |
2b188cc1 | 3249 | |
52e04ef4 | 3250 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
3251 | if (put_page_testzero(page)) |
3252 | free_compound_page(page); | |
3253 | } | |
3254 | ||
3255 | static void *io_mem_alloc(size_t size) | |
3256 | { | |
3257 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
3258 | __GFP_NORETRY; | |
3259 | ||
3260 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
3261 | } | |
3262 | ||
75b28aff HV |
3263 | static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries, |
3264 | size_t *sq_offset) | |
3265 | { | |
3266 | struct io_rings *rings; | |
3267 | size_t off, sq_array_size; | |
3268 | ||
3269 | off = struct_size(rings, cqes, cq_entries); | |
3270 | if (off == SIZE_MAX) | |
3271 | return SIZE_MAX; | |
3272 | ||
3273 | #ifdef CONFIG_SMP | |
3274 | off = ALIGN(off, SMP_CACHE_BYTES); | |
3275 | if (off == 0) | |
3276 | return SIZE_MAX; | |
3277 | #endif | |
3278 | ||
3279 | sq_array_size = array_size(sizeof(u32), sq_entries); | |
3280 | if (sq_array_size == SIZE_MAX) | |
3281 | return SIZE_MAX; | |
3282 | ||
3283 | if (check_add_overflow(off, sq_array_size, &off)) | |
3284 | return SIZE_MAX; | |
3285 | ||
3286 | if (sq_offset) | |
3287 | *sq_offset = off; | |
3288 | ||
3289 | return off; | |
3290 | } | |
3291 | ||
2b188cc1 JA |
3292 | static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries) |
3293 | { | |
75b28aff | 3294 | size_t pages; |
2b188cc1 | 3295 | |
75b28aff HV |
3296 | pages = (size_t)1 << get_order( |
3297 | rings_size(sq_entries, cq_entries, NULL)); | |
3298 | pages += (size_t)1 << get_order( | |
3299 | array_size(sizeof(struct io_uring_sqe), sq_entries)); | |
2b188cc1 | 3300 | |
75b28aff | 3301 | return pages; |
2b188cc1 JA |
3302 | } |
3303 | ||
edafccee JA |
3304 | static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx) |
3305 | { | |
3306 | int i, j; | |
3307 | ||
3308 | if (!ctx->user_bufs) | |
3309 | return -ENXIO; | |
3310 | ||
3311 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
3312 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
3313 | ||
3314 | for (j = 0; j < imu->nr_bvecs; j++) | |
27c4d3a3 | 3315 | put_user_page(imu->bvec[j].bv_page); |
edafccee JA |
3316 | |
3317 | if (ctx->account_mem) | |
3318 | io_unaccount_mem(ctx->user, imu->nr_bvecs); | |
d4ef6475 | 3319 | kvfree(imu->bvec); |
edafccee JA |
3320 | imu->nr_bvecs = 0; |
3321 | } | |
3322 | ||
3323 | kfree(ctx->user_bufs); | |
3324 | ctx->user_bufs = NULL; | |
3325 | ctx->nr_user_bufs = 0; | |
3326 | return 0; | |
3327 | } | |
3328 | ||
3329 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
3330 | void __user *arg, unsigned index) | |
3331 | { | |
3332 | struct iovec __user *src; | |
3333 | ||
3334 | #ifdef CONFIG_COMPAT | |
3335 | if (ctx->compat) { | |
3336 | struct compat_iovec __user *ciovs; | |
3337 | struct compat_iovec ciov; | |
3338 | ||
3339 | ciovs = (struct compat_iovec __user *) arg; | |
3340 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
3341 | return -EFAULT; | |
3342 | ||
3343 | dst->iov_base = (void __user *) (unsigned long) ciov.iov_base; | |
3344 | dst->iov_len = ciov.iov_len; | |
3345 | return 0; | |
3346 | } | |
3347 | #endif | |
3348 | src = (struct iovec __user *) arg; | |
3349 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
3350 | return -EFAULT; | |
3351 | return 0; | |
3352 | } | |
3353 | ||
3354 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, | |
3355 | unsigned nr_args) | |
3356 | { | |
3357 | struct vm_area_struct **vmas = NULL; | |
3358 | struct page **pages = NULL; | |
3359 | int i, j, got_pages = 0; | |
3360 | int ret = -EINVAL; | |
3361 | ||
3362 | if (ctx->user_bufs) | |
3363 | return -EBUSY; | |
3364 | if (!nr_args || nr_args > UIO_MAXIOV) | |
3365 | return -EINVAL; | |
3366 | ||
3367 | ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf), | |
3368 | GFP_KERNEL); | |
3369 | if (!ctx->user_bufs) | |
3370 | return -ENOMEM; | |
3371 | ||
3372 | for (i = 0; i < nr_args; i++) { | |
3373 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
3374 | unsigned long off, start, end, ubuf; | |
3375 | int pret, nr_pages; | |
3376 | struct iovec iov; | |
3377 | size_t size; | |
3378 | ||
3379 | ret = io_copy_iov(ctx, &iov, arg, i); | |
3380 | if (ret) | |
a278682d | 3381 | goto err; |
edafccee JA |
3382 | |
3383 | /* | |
3384 | * Don't impose further limits on the size and buffer | |
3385 | * constraints here, we'll -EINVAL later when IO is | |
3386 | * submitted if they are wrong. | |
3387 | */ | |
3388 | ret = -EFAULT; | |
3389 | if (!iov.iov_base || !iov.iov_len) | |
3390 | goto err; | |
3391 | ||
3392 | /* arbitrary limit, but we need something */ | |
3393 | if (iov.iov_len > SZ_1G) | |
3394 | goto err; | |
3395 | ||
3396 | ubuf = (unsigned long) iov.iov_base; | |
3397 | end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
3398 | start = ubuf >> PAGE_SHIFT; | |
3399 | nr_pages = end - start; | |
3400 | ||
3401 | if (ctx->account_mem) { | |
3402 | ret = io_account_mem(ctx->user, nr_pages); | |
3403 | if (ret) | |
3404 | goto err; | |
3405 | } | |
3406 | ||
3407 | ret = 0; | |
3408 | if (!pages || nr_pages > got_pages) { | |
3409 | kfree(vmas); | |
3410 | kfree(pages); | |
d4ef6475 | 3411 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), |
edafccee | 3412 | GFP_KERNEL); |
d4ef6475 | 3413 | vmas = kvmalloc_array(nr_pages, |
edafccee JA |
3414 | sizeof(struct vm_area_struct *), |
3415 | GFP_KERNEL); | |
3416 | if (!pages || !vmas) { | |
3417 | ret = -ENOMEM; | |
3418 | if (ctx->account_mem) | |
3419 | io_unaccount_mem(ctx->user, nr_pages); | |
3420 | goto err; | |
3421 | } | |
3422 | got_pages = nr_pages; | |
3423 | } | |
3424 | ||
d4ef6475 | 3425 | imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec), |
edafccee JA |
3426 | GFP_KERNEL); |
3427 | ret = -ENOMEM; | |
3428 | if (!imu->bvec) { | |
3429 | if (ctx->account_mem) | |
3430 | io_unaccount_mem(ctx->user, nr_pages); | |
3431 | goto err; | |
3432 | } | |
3433 | ||
3434 | ret = 0; | |
3435 | down_read(¤t->mm->mmap_sem); | |
932f4a63 IW |
3436 | pret = get_user_pages(ubuf, nr_pages, |
3437 | FOLL_WRITE | FOLL_LONGTERM, | |
3438 | pages, vmas); | |
edafccee JA |
3439 | if (pret == nr_pages) { |
3440 | /* don't support file backed memory */ | |
3441 | for (j = 0; j < nr_pages; j++) { | |
3442 | struct vm_area_struct *vma = vmas[j]; | |
3443 | ||
3444 | if (vma->vm_file && | |
3445 | !is_file_hugepages(vma->vm_file)) { | |
3446 | ret = -EOPNOTSUPP; | |
3447 | break; | |
3448 | } | |
3449 | } | |
3450 | } else { | |
3451 | ret = pret < 0 ? pret : -EFAULT; | |
3452 | } | |
3453 | up_read(¤t->mm->mmap_sem); | |
3454 | if (ret) { | |
3455 | /* | |
3456 | * if we did partial map, or found file backed vmas, | |
3457 | * release any pages we did get | |
3458 | */ | |
27c4d3a3 JH |
3459 | if (pret > 0) |
3460 | put_user_pages(pages, pret); | |
edafccee JA |
3461 | if (ctx->account_mem) |
3462 | io_unaccount_mem(ctx->user, nr_pages); | |
d4ef6475 | 3463 | kvfree(imu->bvec); |
edafccee JA |
3464 | goto err; |
3465 | } | |
3466 | ||
3467 | off = ubuf & ~PAGE_MASK; | |
3468 | size = iov.iov_len; | |
3469 | for (j = 0; j < nr_pages; j++) { | |
3470 | size_t vec_len; | |
3471 | ||
3472 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
3473 | imu->bvec[j].bv_page = pages[j]; | |
3474 | imu->bvec[j].bv_len = vec_len; | |
3475 | imu->bvec[j].bv_offset = off; | |
3476 | off = 0; | |
3477 | size -= vec_len; | |
3478 | } | |
3479 | /* store original address for later verification */ | |
3480 | imu->ubuf = ubuf; | |
3481 | imu->len = iov.iov_len; | |
3482 | imu->nr_bvecs = nr_pages; | |
3483 | ||
3484 | ctx->nr_user_bufs++; | |
3485 | } | |
d4ef6475 MR |
3486 | kvfree(pages); |
3487 | kvfree(vmas); | |
edafccee JA |
3488 | return 0; |
3489 | err: | |
d4ef6475 MR |
3490 | kvfree(pages); |
3491 | kvfree(vmas); | |
edafccee JA |
3492 | io_sqe_buffer_unregister(ctx); |
3493 | return ret; | |
3494 | } | |
3495 | ||
9b402849 JA |
3496 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
3497 | { | |
3498 | __s32 __user *fds = arg; | |
3499 | int fd; | |
3500 | ||
3501 | if (ctx->cq_ev_fd) | |
3502 | return -EBUSY; | |
3503 | ||
3504 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
3505 | return -EFAULT; | |
3506 | ||
3507 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
3508 | if (IS_ERR(ctx->cq_ev_fd)) { | |
3509 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
3510 | ctx->cq_ev_fd = NULL; | |
3511 | return ret; | |
3512 | } | |
3513 | ||
3514 | return 0; | |
3515 | } | |
3516 | ||
3517 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
3518 | { | |
3519 | if (ctx->cq_ev_fd) { | |
3520 | eventfd_ctx_put(ctx->cq_ev_fd); | |
3521 | ctx->cq_ev_fd = NULL; | |
3522 | return 0; | |
3523 | } | |
3524 | ||
3525 | return -ENXIO; | |
3526 | } | |
3527 | ||
2b188cc1 JA |
3528 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
3529 | { | |
6b06314c | 3530 | io_finish_async(ctx); |
2b188cc1 JA |
3531 | if (ctx->sqo_mm) |
3532 | mmdrop(ctx->sqo_mm); | |
def596e9 JA |
3533 | |
3534 | io_iopoll_reap_events(ctx); | |
edafccee | 3535 | io_sqe_buffer_unregister(ctx); |
6b06314c | 3536 | io_sqe_files_unregister(ctx); |
9b402849 | 3537 | io_eventfd_unregister(ctx); |
def596e9 | 3538 | |
2b188cc1 | 3539 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
3540 | if (ctx->ring_sock) { |
3541 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 3542 | sock_release(ctx->ring_sock); |
355e8d26 | 3543 | } |
2b188cc1 JA |
3544 | #endif |
3545 | ||
75b28aff | 3546 | io_mem_free(ctx->rings); |
2b188cc1 | 3547 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
3548 | |
3549 | percpu_ref_exit(&ctx->refs); | |
3550 | if (ctx->account_mem) | |
3551 | io_unaccount_mem(ctx->user, | |
3552 | ring_pages(ctx->sq_entries, ctx->cq_entries)); | |
3553 | free_uid(ctx->user); | |
3554 | kfree(ctx); | |
3555 | } | |
3556 | ||
3557 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
3558 | { | |
3559 | struct io_ring_ctx *ctx = file->private_data; | |
3560 | __poll_t mask = 0; | |
3561 | ||
3562 | poll_wait(file, &ctx->cq_wait, wait); | |
4f7067c3 SB |
3563 | /* |
3564 | * synchronizes with barrier from wq_has_sleeper call in | |
3565 | * io_commit_cqring | |
3566 | */ | |
2b188cc1 | 3567 | smp_rmb(); |
75b28aff HV |
3568 | if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head != |
3569 | ctx->rings->sq_ring_entries) | |
2b188cc1 | 3570 | mask |= EPOLLOUT | EPOLLWRNORM; |
daa5de54 | 3571 | if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail) |
2b188cc1 JA |
3572 | mask |= EPOLLIN | EPOLLRDNORM; |
3573 | ||
3574 | return mask; | |
3575 | } | |
3576 | ||
3577 | static int io_uring_fasync(int fd, struct file *file, int on) | |
3578 | { | |
3579 | struct io_ring_ctx *ctx = file->private_data; | |
3580 | ||
3581 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
3582 | } | |
3583 | ||
3584 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) | |
3585 | { | |
3586 | mutex_lock(&ctx->uring_lock); | |
3587 | percpu_ref_kill(&ctx->refs); | |
3588 | mutex_unlock(&ctx->uring_lock); | |
3589 | ||
5262f567 | 3590 | io_kill_timeouts(ctx); |
221c5eb2 | 3591 | io_poll_remove_all(ctx); |
def596e9 | 3592 | io_iopoll_reap_events(ctx); |
2b188cc1 JA |
3593 | wait_for_completion(&ctx->ctx_done); |
3594 | io_ring_ctx_free(ctx); | |
3595 | } | |
3596 | ||
3597 | static int io_uring_release(struct inode *inode, struct file *file) | |
3598 | { | |
3599 | struct io_ring_ctx *ctx = file->private_data; | |
3600 | ||
3601 | file->private_data = NULL; | |
3602 | io_ring_ctx_wait_and_kill(ctx); | |
3603 | return 0; | |
3604 | } | |
3605 | ||
3606 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
3607 | { | |
3608 | loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT; | |
3609 | unsigned long sz = vma->vm_end - vma->vm_start; | |
3610 | struct io_ring_ctx *ctx = file->private_data; | |
3611 | unsigned long pfn; | |
3612 | struct page *page; | |
3613 | void *ptr; | |
3614 | ||
3615 | switch (offset) { | |
3616 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
3617 | case IORING_OFF_CQ_RING: |
3618 | ptr = ctx->rings; | |
2b188cc1 JA |
3619 | break; |
3620 | case IORING_OFF_SQES: | |
3621 | ptr = ctx->sq_sqes; | |
3622 | break; | |
2b188cc1 JA |
3623 | default: |
3624 | return -EINVAL; | |
3625 | } | |
3626 | ||
3627 | page = virt_to_head_page(ptr); | |
a50b854e | 3628 | if (sz > page_size(page)) |
2b188cc1 JA |
3629 | return -EINVAL; |
3630 | ||
3631 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
3632 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
3633 | } | |
3634 | ||
3635 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, | |
3636 | u32, min_complete, u32, flags, const sigset_t __user *, sig, | |
3637 | size_t, sigsz) | |
3638 | { | |
3639 | struct io_ring_ctx *ctx; | |
3640 | long ret = -EBADF; | |
3641 | int submitted = 0; | |
3642 | struct fd f; | |
3643 | ||
6c271ce2 | 3644 | if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP)) |
2b188cc1 JA |
3645 | return -EINVAL; |
3646 | ||
3647 | f = fdget(fd); | |
3648 | if (!f.file) | |
3649 | return -EBADF; | |
3650 | ||
3651 | ret = -EOPNOTSUPP; | |
3652 | if (f.file->f_op != &io_uring_fops) | |
3653 | goto out_fput; | |
3654 | ||
3655 | ret = -ENXIO; | |
3656 | ctx = f.file->private_data; | |
3657 | if (!percpu_ref_tryget(&ctx->refs)) | |
3658 | goto out_fput; | |
3659 | ||
6c271ce2 JA |
3660 | /* |
3661 | * For SQ polling, the thread will do all submissions and completions. | |
3662 | * Just return the requested submit count, and wake the thread if | |
3663 | * we were asked to. | |
3664 | */ | |
b2a9eada | 3665 | ret = 0; |
6c271ce2 JA |
3666 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3667 | if (flags & IORING_ENTER_SQ_WAKEUP) | |
3668 | wake_up(&ctx->sqo_wait); | |
3669 | submitted = to_submit; | |
b2a9eada | 3670 | } else if (to_submit) { |
2b188cc1 JA |
3671 | to_submit = min(to_submit, ctx->sq_entries); |
3672 | ||
3673 | mutex_lock(&ctx->uring_lock); | |
bc808bce | 3674 | submitted = io_ring_submit(ctx, to_submit); |
2b188cc1 | 3675 | mutex_unlock(&ctx->uring_lock); |
2b188cc1 JA |
3676 | } |
3677 | if (flags & IORING_ENTER_GETEVENTS) { | |
def596e9 JA |
3678 | unsigned nr_events = 0; |
3679 | ||
2b188cc1 JA |
3680 | min_complete = min(min_complete, ctx->cq_entries); |
3681 | ||
def596e9 | 3682 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 | 3683 | ret = io_iopoll_check(ctx, &nr_events, min_complete); |
def596e9 JA |
3684 | } else { |
3685 | ret = io_cqring_wait(ctx, min_complete, sig, sigsz); | |
3686 | } | |
2b188cc1 JA |
3687 | } |
3688 | ||
6805b32e | 3689 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
3690 | out_fput: |
3691 | fdput(f); | |
3692 | return submitted ? submitted : ret; | |
3693 | } | |
3694 | ||
3695 | static const struct file_operations io_uring_fops = { | |
3696 | .release = io_uring_release, | |
3697 | .mmap = io_uring_mmap, | |
3698 | .poll = io_uring_poll, | |
3699 | .fasync = io_uring_fasync, | |
3700 | }; | |
3701 | ||
3702 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
3703 | struct io_uring_params *p) | |
3704 | { | |
75b28aff HV |
3705 | struct io_rings *rings; |
3706 | size_t size, sq_array_offset; | |
2b188cc1 | 3707 | |
75b28aff HV |
3708 | size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset); |
3709 | if (size == SIZE_MAX) | |
3710 | return -EOVERFLOW; | |
3711 | ||
3712 | rings = io_mem_alloc(size); | |
3713 | if (!rings) | |
2b188cc1 JA |
3714 | return -ENOMEM; |
3715 | ||
75b28aff HV |
3716 | ctx->rings = rings; |
3717 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
3718 | rings->sq_ring_mask = p->sq_entries - 1; | |
3719 | rings->cq_ring_mask = p->cq_entries - 1; | |
3720 | rings->sq_ring_entries = p->sq_entries; | |
3721 | rings->cq_ring_entries = p->cq_entries; | |
3722 | ctx->sq_mask = rings->sq_ring_mask; | |
3723 | ctx->cq_mask = rings->cq_ring_mask; | |
3724 | ctx->sq_entries = rings->sq_ring_entries; | |
3725 | ctx->cq_entries = rings->cq_ring_entries; | |
2b188cc1 JA |
3726 | |
3727 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
3728 | if (size == SIZE_MAX) | |
3729 | return -EOVERFLOW; | |
3730 | ||
3731 | ctx->sq_sqes = io_mem_alloc(size); | |
52e04ef4 | 3732 | if (!ctx->sq_sqes) |
2b188cc1 | 3733 | return -ENOMEM; |
2b188cc1 | 3734 | |
2b188cc1 JA |
3735 | return 0; |
3736 | } | |
3737 | ||
3738 | /* | |
3739 | * Allocate an anonymous fd, this is what constitutes the application | |
3740 | * visible backing of an io_uring instance. The application mmaps this | |
3741 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
3742 | * we have to tie this fd to a socket for file garbage collection purposes. | |
3743 | */ | |
3744 | static int io_uring_get_fd(struct io_ring_ctx *ctx) | |
3745 | { | |
3746 | struct file *file; | |
3747 | int ret; | |
3748 | ||
3749 | #if defined(CONFIG_UNIX) | |
3750 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, | |
3751 | &ctx->ring_sock); | |
3752 | if (ret) | |
3753 | return ret; | |
3754 | #endif | |
3755 | ||
3756 | ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
3757 | if (ret < 0) | |
3758 | goto err; | |
3759 | ||
3760 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, | |
3761 | O_RDWR | O_CLOEXEC); | |
3762 | if (IS_ERR(file)) { | |
3763 | put_unused_fd(ret); | |
3764 | ret = PTR_ERR(file); | |
3765 | goto err; | |
3766 | } | |
3767 | ||
3768 | #if defined(CONFIG_UNIX) | |
3769 | ctx->ring_sock->file = file; | |
6b06314c | 3770 | ctx->ring_sock->sk->sk_user_data = ctx; |
2b188cc1 JA |
3771 | #endif |
3772 | fd_install(ret, file); | |
3773 | return ret; | |
3774 | err: | |
3775 | #if defined(CONFIG_UNIX) | |
3776 | sock_release(ctx->ring_sock); | |
3777 | ctx->ring_sock = NULL; | |
3778 | #endif | |
3779 | return ret; | |
3780 | } | |
3781 | ||
3782 | static int io_uring_create(unsigned entries, struct io_uring_params *p) | |
3783 | { | |
3784 | struct user_struct *user = NULL; | |
3785 | struct io_ring_ctx *ctx; | |
3786 | bool account_mem; | |
3787 | int ret; | |
3788 | ||
3789 | if (!entries || entries > IORING_MAX_ENTRIES) | |
3790 | return -EINVAL; | |
3791 | ||
3792 | /* | |
3793 | * Use twice as many entries for the CQ ring. It's possible for the | |
3794 | * application to drive a higher depth than the size of the SQ ring, | |
3795 | * since the sqes are only used at submission time. This allows for | |
3796 | * some flexibility in overcommitting a bit. | |
3797 | */ | |
3798 | p->sq_entries = roundup_pow_of_two(entries); | |
3799 | p->cq_entries = 2 * p->sq_entries; | |
3800 | ||
3801 | user = get_uid(current_user()); | |
3802 | account_mem = !capable(CAP_IPC_LOCK); | |
3803 | ||
3804 | if (account_mem) { | |
3805 | ret = io_account_mem(user, | |
3806 | ring_pages(p->sq_entries, p->cq_entries)); | |
3807 | if (ret) { | |
3808 | free_uid(user); | |
3809 | return ret; | |
3810 | } | |
3811 | } | |
3812 | ||
3813 | ctx = io_ring_ctx_alloc(p); | |
3814 | if (!ctx) { | |
3815 | if (account_mem) | |
3816 | io_unaccount_mem(user, ring_pages(p->sq_entries, | |
3817 | p->cq_entries)); | |
3818 | free_uid(user); | |
3819 | return -ENOMEM; | |
3820 | } | |
3821 | ctx->compat = in_compat_syscall(); | |
3822 | ctx->account_mem = account_mem; | |
3823 | ctx->user = user; | |
3824 | ||
3825 | ret = io_allocate_scq_urings(ctx, p); | |
3826 | if (ret) | |
3827 | goto err; | |
3828 | ||
6c271ce2 | 3829 | ret = io_sq_offload_start(ctx, p); |
2b188cc1 JA |
3830 | if (ret) |
3831 | goto err; | |
3832 | ||
2b188cc1 | 3833 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
3834 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
3835 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
3836 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
3837 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
3838 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
3839 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
3840 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
3841 | |
3842 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
3843 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
3844 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
3845 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
3846 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
3847 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
3848 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
ac90f249 | 3849 | |
044c1ab3 JA |
3850 | /* |
3851 | * Install ring fd as the very last thing, so we don't risk someone | |
3852 | * having closed it before we finish setup | |
3853 | */ | |
3854 | ret = io_uring_get_fd(ctx); | |
3855 | if (ret < 0) | |
3856 | goto err; | |
3857 | ||
ac90f249 | 3858 | p->features = IORING_FEAT_SINGLE_MMAP; |
2b188cc1 JA |
3859 | return ret; |
3860 | err: | |
3861 | io_ring_ctx_wait_and_kill(ctx); | |
3862 | return ret; | |
3863 | } | |
3864 | ||
3865 | /* | |
3866 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
3867 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
3868 | * params structure passed in. | |
3869 | */ | |
3870 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
3871 | { | |
3872 | struct io_uring_params p; | |
3873 | long ret; | |
3874 | int i; | |
3875 | ||
3876 | if (copy_from_user(&p, params, sizeof(p))) | |
3877 | return -EFAULT; | |
3878 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
3879 | if (p.resv[i]) | |
3880 | return -EINVAL; | |
3881 | } | |
3882 | ||
6c271ce2 JA |
3883 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
3884 | IORING_SETUP_SQ_AFF)) | |
2b188cc1 JA |
3885 | return -EINVAL; |
3886 | ||
3887 | ret = io_uring_create(entries, &p); | |
3888 | if (ret < 0) | |
3889 | return ret; | |
3890 | ||
3891 | if (copy_to_user(params, &p, sizeof(p))) | |
3892 | return -EFAULT; | |
3893 | ||
3894 | return ret; | |
3895 | } | |
3896 | ||
3897 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
3898 | struct io_uring_params __user *, params) | |
3899 | { | |
3900 | return io_uring_setup(entries, params); | |
3901 | } | |
3902 | ||
edafccee JA |
3903 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
3904 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
3905 | __releases(ctx->uring_lock) |
3906 | __acquires(ctx->uring_lock) | |
edafccee JA |
3907 | { |
3908 | int ret; | |
3909 | ||
35fa71a0 JA |
3910 | /* |
3911 | * We're inside the ring mutex, if the ref is already dying, then | |
3912 | * someone else killed the ctx or is already going through | |
3913 | * io_uring_register(). | |
3914 | */ | |
3915 | if (percpu_ref_is_dying(&ctx->refs)) | |
3916 | return -ENXIO; | |
3917 | ||
edafccee | 3918 | percpu_ref_kill(&ctx->refs); |
b19062a5 JA |
3919 | |
3920 | /* | |
3921 | * Drop uring mutex before waiting for references to exit. If another | |
3922 | * thread is currently inside io_uring_enter() it might need to grab | |
3923 | * the uring_lock to make progress. If we hold it here across the drain | |
3924 | * wait, then we can deadlock. It's safe to drop the mutex here, since | |
3925 | * no new references will come in after we've killed the percpu ref. | |
3926 | */ | |
3927 | mutex_unlock(&ctx->uring_lock); | |
edafccee | 3928 | wait_for_completion(&ctx->ctx_done); |
b19062a5 | 3929 | mutex_lock(&ctx->uring_lock); |
edafccee JA |
3930 | |
3931 | switch (opcode) { | |
3932 | case IORING_REGISTER_BUFFERS: | |
3933 | ret = io_sqe_buffer_register(ctx, arg, nr_args); | |
3934 | break; | |
3935 | case IORING_UNREGISTER_BUFFERS: | |
3936 | ret = -EINVAL; | |
3937 | if (arg || nr_args) | |
3938 | break; | |
3939 | ret = io_sqe_buffer_unregister(ctx); | |
3940 | break; | |
6b06314c JA |
3941 | case IORING_REGISTER_FILES: |
3942 | ret = io_sqe_files_register(ctx, arg, nr_args); | |
3943 | break; | |
3944 | case IORING_UNREGISTER_FILES: | |
3945 | ret = -EINVAL; | |
3946 | if (arg || nr_args) | |
3947 | break; | |
3948 | ret = io_sqe_files_unregister(ctx); | |
3949 | break; | |
9b402849 JA |
3950 | case IORING_REGISTER_EVENTFD: |
3951 | ret = -EINVAL; | |
3952 | if (nr_args != 1) | |
3953 | break; | |
3954 | ret = io_eventfd_register(ctx, arg); | |
3955 | break; | |
3956 | case IORING_UNREGISTER_EVENTFD: | |
3957 | ret = -EINVAL; | |
3958 | if (arg || nr_args) | |
3959 | break; | |
3960 | ret = io_eventfd_unregister(ctx); | |
3961 | break; | |
edafccee JA |
3962 | default: |
3963 | ret = -EINVAL; | |
3964 | break; | |
3965 | } | |
3966 | ||
3967 | /* bring the ctx back to life */ | |
3968 | reinit_completion(&ctx->ctx_done); | |
3969 | percpu_ref_reinit(&ctx->refs); | |
3970 | return ret; | |
3971 | } | |
3972 | ||
3973 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
3974 | void __user *, arg, unsigned int, nr_args) | |
3975 | { | |
3976 | struct io_ring_ctx *ctx; | |
3977 | long ret = -EBADF; | |
3978 | struct fd f; | |
3979 | ||
3980 | f = fdget(fd); | |
3981 | if (!f.file) | |
3982 | return -EBADF; | |
3983 | ||
3984 | ret = -EOPNOTSUPP; | |
3985 | if (f.file->f_op != &io_uring_fops) | |
3986 | goto out_fput; | |
3987 | ||
3988 | ctx = f.file->private_data; | |
3989 | ||
3990 | mutex_lock(&ctx->uring_lock); | |
3991 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
3992 | mutex_unlock(&ctx->uring_lock); | |
3993 | out_fput: | |
3994 | fdput(f); | |
3995 | return ret; | |
3996 | } | |
3997 | ||
2b188cc1 JA |
3998 | static int __init io_uring_init(void) |
3999 | { | |
4000 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC); | |
4001 | return 0; | |
4002 | }; | |
4003 | __initcall(io_uring_init); |