]>
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 | 73 | |
c826bd7a DD |
74 | #define CREATE_TRACE_POINTS |
75 | #include <trace/events/io_uring.h> | |
76 | ||
2b188cc1 JA |
77 | #include <uapi/linux/io_uring.h> |
78 | ||
79 | #include "internal.h" | |
80 | ||
5277deaa | 81 | #define IORING_MAX_ENTRIES 32768 |
33a107f0 | 82 | #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES) |
6b06314c | 83 | #define IORING_MAX_FIXED_FILES 1024 |
2b188cc1 JA |
84 | |
85 | struct io_uring { | |
86 | u32 head ____cacheline_aligned_in_smp; | |
87 | u32 tail ____cacheline_aligned_in_smp; | |
88 | }; | |
89 | ||
1e84b97b | 90 | /* |
75b28aff HV |
91 | * This data is shared with the application through the mmap at offsets |
92 | * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING. | |
1e84b97b SB |
93 | * |
94 | * The offsets to the member fields are published through struct | |
95 | * io_sqring_offsets when calling io_uring_setup. | |
96 | */ | |
75b28aff | 97 | struct io_rings { |
1e84b97b SB |
98 | /* |
99 | * Head and tail offsets into the ring; the offsets need to be | |
100 | * masked to get valid indices. | |
101 | * | |
75b28aff HV |
102 | * The kernel controls head of the sq ring and the tail of the cq ring, |
103 | * and the application controls tail of the sq ring and the head of the | |
104 | * cq ring. | |
1e84b97b | 105 | */ |
75b28aff | 106 | struct io_uring sq, cq; |
1e84b97b | 107 | /* |
75b28aff | 108 | * Bitmasks to apply to head and tail offsets (constant, equals |
1e84b97b SB |
109 | * ring_entries - 1) |
110 | */ | |
75b28aff HV |
111 | u32 sq_ring_mask, cq_ring_mask; |
112 | /* Ring sizes (constant, power of 2) */ | |
113 | u32 sq_ring_entries, cq_ring_entries; | |
1e84b97b SB |
114 | /* |
115 | * Number of invalid entries dropped by the kernel due to | |
116 | * invalid index stored in array | |
117 | * | |
118 | * Written by the kernel, shouldn't be modified by the | |
119 | * application (i.e. get number of "new events" by comparing to | |
120 | * cached value). | |
121 | * | |
122 | * After a new SQ head value was read by the application this | |
123 | * counter includes all submissions that were dropped reaching | |
124 | * the new SQ head (and possibly more). | |
125 | */ | |
75b28aff | 126 | u32 sq_dropped; |
1e84b97b SB |
127 | /* |
128 | * Runtime flags | |
129 | * | |
130 | * Written by the kernel, shouldn't be modified by the | |
131 | * application. | |
132 | * | |
133 | * The application needs a full memory barrier before checking | |
134 | * for IORING_SQ_NEED_WAKEUP after updating the sq tail. | |
135 | */ | |
75b28aff | 136 | u32 sq_flags; |
1e84b97b SB |
137 | /* |
138 | * Number of completion events lost because the queue was full; | |
139 | * this should be avoided by the application by making sure | |
140 | * there are not more requests pending thatn there is space in | |
141 | * the completion queue. | |
142 | * | |
143 | * Written by the kernel, shouldn't be modified by the | |
144 | * application (i.e. get number of "new events" by comparing to | |
145 | * cached value). | |
146 | * | |
147 | * As completion events come in out of order this counter is not | |
148 | * ordered with any other data. | |
149 | */ | |
75b28aff | 150 | u32 cq_overflow; |
1e84b97b SB |
151 | /* |
152 | * Ring buffer of completion events. | |
153 | * | |
154 | * The kernel writes completion events fresh every time they are | |
155 | * produced, so the application is allowed to modify pending | |
156 | * entries. | |
157 | */ | |
75b28aff | 158 | struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp; |
2b188cc1 JA |
159 | }; |
160 | ||
edafccee JA |
161 | struct io_mapped_ubuf { |
162 | u64 ubuf; | |
163 | size_t len; | |
164 | struct bio_vec *bvec; | |
165 | unsigned int nr_bvecs; | |
166 | }; | |
167 | ||
31b51510 JA |
168 | struct async_list { |
169 | spinlock_t lock; | |
170 | atomic_t cnt; | |
171 | struct list_head list; | |
172 | ||
173 | struct file *file; | |
6d5d5ac5 | 174 | off_t io_start; |
9310a7ba | 175 | size_t io_len; |
31b51510 JA |
176 | }; |
177 | ||
2b188cc1 JA |
178 | struct io_ring_ctx { |
179 | struct { | |
180 | struct percpu_ref refs; | |
181 | } ____cacheline_aligned_in_smp; | |
182 | ||
183 | struct { | |
184 | unsigned int flags; | |
185 | bool compat; | |
186 | bool account_mem; | |
187 | ||
75b28aff HV |
188 | /* |
189 | * Ring buffer of indices into array of io_uring_sqe, which is | |
190 | * mmapped by the application using the IORING_OFF_SQES offset. | |
191 | * | |
192 | * This indirection could e.g. be used to assign fixed | |
193 | * io_uring_sqe entries to operations and only submit them to | |
194 | * the queue when needed. | |
195 | * | |
196 | * The kernel modifies neither the indices array nor the entries | |
197 | * array. | |
198 | */ | |
199 | u32 *sq_array; | |
2b188cc1 JA |
200 | unsigned cached_sq_head; |
201 | unsigned sq_entries; | |
202 | unsigned sq_mask; | |
6c271ce2 | 203 | unsigned sq_thread_idle; |
498ccd9e | 204 | unsigned cached_sq_dropped; |
2b188cc1 | 205 | struct io_uring_sqe *sq_sqes; |
de0617e4 JA |
206 | |
207 | struct list_head defer_list; | |
5262f567 | 208 | struct list_head timeout_list; |
2b188cc1 JA |
209 | } ____cacheline_aligned_in_smp; |
210 | ||
211 | /* IO offload */ | |
54a91f3b | 212 | struct workqueue_struct *sqo_wq[2]; |
6c271ce2 | 213 | struct task_struct *sqo_thread; /* if using sq thread polling */ |
2b188cc1 | 214 | struct mm_struct *sqo_mm; |
6c271ce2 | 215 | wait_queue_head_t sqo_wait; |
a4c0b3de | 216 | struct completion sqo_thread_started; |
2b188cc1 JA |
217 | |
218 | struct { | |
2b188cc1 | 219 | unsigned cached_cq_tail; |
498ccd9e | 220 | atomic_t cached_cq_overflow; |
2b188cc1 JA |
221 | unsigned cq_entries; |
222 | unsigned cq_mask; | |
223 | struct wait_queue_head cq_wait; | |
224 | struct fasync_struct *cq_fasync; | |
9b402849 | 225 | struct eventfd_ctx *cq_ev_fd; |
5262f567 | 226 | atomic_t cq_timeouts; |
2b188cc1 JA |
227 | } ____cacheline_aligned_in_smp; |
228 | ||
75b28aff HV |
229 | struct io_rings *rings; |
230 | ||
6b06314c JA |
231 | /* |
232 | * If used, fixed file set. Writers must ensure that ->refs is dead, | |
233 | * readers must ensure that ->refs is alive as long as the file* is | |
234 | * used. Only updated through io_uring_register(2). | |
235 | */ | |
236 | struct file **user_files; | |
237 | unsigned nr_user_files; | |
238 | ||
edafccee JA |
239 | /* if used, fixed mapped user buffers */ |
240 | unsigned nr_user_bufs; | |
241 | struct io_mapped_ubuf *user_bufs; | |
242 | ||
2b188cc1 JA |
243 | struct user_struct *user; |
244 | ||
245 | struct completion ctx_done; | |
246 | ||
247 | struct { | |
248 | struct mutex uring_lock; | |
249 | wait_queue_head_t wait; | |
250 | } ____cacheline_aligned_in_smp; | |
251 | ||
252 | struct { | |
253 | spinlock_t completion_lock; | |
def596e9 JA |
254 | bool poll_multi_file; |
255 | /* | |
256 | * ->poll_list is protected by the ctx->uring_lock for | |
257 | * io_uring instances that don't use IORING_SETUP_SQPOLL. | |
258 | * For SQPOLL, only the single threaded io_sq_thread() will | |
259 | * manipulate the list, hence no extra locking is needed there. | |
260 | */ | |
261 | struct list_head poll_list; | |
221c5eb2 | 262 | struct list_head cancel_list; |
2b188cc1 JA |
263 | } ____cacheline_aligned_in_smp; |
264 | ||
31b51510 JA |
265 | struct async_list pending_async[2]; |
266 | ||
2b188cc1 JA |
267 | #if defined(CONFIG_UNIX) |
268 | struct socket *ring_sock; | |
269 | #endif | |
270 | }; | |
271 | ||
272 | struct sqe_submit { | |
273 | const struct io_uring_sqe *sqe; | |
8776f3fa | 274 | u32 sequence; |
2b188cc1 | 275 | bool has_user; |
ba5290cc | 276 | bool in_async; |
6c271ce2 | 277 | bool needs_fixed_file; |
2b188cc1 JA |
278 | }; |
279 | ||
09bb8394 JA |
280 | /* |
281 | * First field must be the file pointer in all the | |
282 | * iocb unions! See also 'struct kiocb' in <linux/fs.h> | |
283 | */ | |
221c5eb2 JA |
284 | struct io_poll_iocb { |
285 | struct file *file; | |
286 | struct wait_queue_head *head; | |
287 | __poll_t events; | |
8c838788 | 288 | bool done; |
221c5eb2 JA |
289 | bool canceled; |
290 | struct wait_queue_entry wait; | |
291 | }; | |
292 | ||
5262f567 JA |
293 | struct io_timeout { |
294 | struct file *file; | |
295 | struct hrtimer timer; | |
296 | }; | |
297 | ||
09bb8394 JA |
298 | /* |
299 | * NOTE! Each of the iocb union members has the file pointer | |
300 | * as the first entry in their struct definition. So you can | |
301 | * access the file pointer through any of the sub-structs, | |
302 | * or directly as just 'ki_filp' in this struct. | |
303 | */ | |
2b188cc1 | 304 | struct io_kiocb { |
221c5eb2 | 305 | union { |
09bb8394 | 306 | struct file *file; |
221c5eb2 JA |
307 | struct kiocb rw; |
308 | struct io_poll_iocb poll; | |
5262f567 | 309 | struct io_timeout timeout; |
221c5eb2 | 310 | }; |
2b188cc1 JA |
311 | |
312 | struct sqe_submit submit; | |
313 | ||
314 | struct io_ring_ctx *ctx; | |
315 | struct list_head list; | |
9e645e11 | 316 | struct list_head link_list; |
2b188cc1 | 317 | unsigned int flags; |
c16361c1 | 318 | refcount_t refs; |
8449eeda | 319 | #define REQ_F_NOWAIT 1 /* must not punt to workers */ |
def596e9 | 320 | #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */ |
6b06314c | 321 | #define REQ_F_FIXED_FILE 4 /* ctx owns file */ |
31b51510 | 322 | #define REQ_F_SEQ_PREV 8 /* sequential with previous */ |
e2033e33 SB |
323 | #define REQ_F_IO_DRAIN 16 /* drain existing IO first */ |
324 | #define REQ_F_IO_DRAINED 32 /* drain done */ | |
9e645e11 | 325 | #define REQ_F_LINK 64 /* linked sqes */ |
f7b76ac9 ZL |
326 | #define REQ_F_LINK_DONE 128 /* linked sqes done */ |
327 | #define REQ_F_FAIL_LINK 256 /* fail rest of links */ | |
4fe2c963 | 328 | #define REQ_F_SHADOW_DRAIN 512 /* link-drain shadow req */ |
5262f567 | 329 | #define REQ_F_TIMEOUT 1024 /* timeout request */ |
491381ce JA |
330 | #define REQ_F_ISREG 2048 /* regular file */ |
331 | #define REQ_F_MUST_PUNT 4096 /* must be punted even for NONBLOCK */ | |
2b188cc1 | 332 | u64 user_data; |
9e645e11 | 333 | u32 result; |
de0617e4 | 334 | u32 sequence; |
2b188cc1 JA |
335 | |
336 | struct work_struct work; | |
337 | }; | |
338 | ||
339 | #define IO_PLUG_THRESHOLD 2 | |
def596e9 | 340 | #define IO_IOPOLL_BATCH 8 |
2b188cc1 | 341 | |
9a56a232 JA |
342 | struct io_submit_state { |
343 | struct blk_plug plug; | |
344 | ||
2579f913 JA |
345 | /* |
346 | * io_kiocb alloc cache | |
347 | */ | |
348 | void *reqs[IO_IOPOLL_BATCH]; | |
349 | unsigned int free_reqs; | |
350 | unsigned int cur_req; | |
351 | ||
9a56a232 JA |
352 | /* |
353 | * File reference cache | |
354 | */ | |
355 | struct file *file; | |
356 | unsigned int fd; | |
357 | unsigned int has_refs; | |
358 | unsigned int used_refs; | |
359 | unsigned int ios_left; | |
360 | }; | |
361 | ||
de0617e4 | 362 | static void io_sq_wq_submit_work(struct work_struct *work); |
5262f567 JA |
363 | static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data, |
364 | long res); | |
4fe2c963 | 365 | static void __io_free_req(struct io_kiocb *req); |
de0617e4 | 366 | |
2b188cc1 JA |
367 | static struct kmem_cache *req_cachep; |
368 | ||
369 | static const struct file_operations io_uring_fops; | |
370 | ||
371 | struct sock *io_uring_get_socket(struct file *file) | |
372 | { | |
373 | #if defined(CONFIG_UNIX) | |
374 | if (file->f_op == &io_uring_fops) { | |
375 | struct io_ring_ctx *ctx = file->private_data; | |
376 | ||
377 | return ctx->ring_sock->sk; | |
378 | } | |
379 | #endif | |
380 | return NULL; | |
381 | } | |
382 | EXPORT_SYMBOL(io_uring_get_socket); | |
383 | ||
384 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) | |
385 | { | |
386 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
387 | ||
388 | complete(&ctx->ctx_done); | |
389 | } | |
390 | ||
391 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) | |
392 | { | |
393 | struct io_ring_ctx *ctx; | |
31b51510 | 394 | int i; |
2b188cc1 JA |
395 | |
396 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
397 | if (!ctx) | |
398 | return NULL; | |
399 | ||
21482896 RG |
400 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
401 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) { | |
2b188cc1 JA |
402 | kfree(ctx); |
403 | return NULL; | |
404 | } | |
405 | ||
406 | ctx->flags = p->flags; | |
407 | init_waitqueue_head(&ctx->cq_wait); | |
408 | init_completion(&ctx->ctx_done); | |
a4c0b3de | 409 | init_completion(&ctx->sqo_thread_started); |
2b188cc1 JA |
410 | mutex_init(&ctx->uring_lock); |
411 | init_waitqueue_head(&ctx->wait); | |
31b51510 JA |
412 | for (i = 0; i < ARRAY_SIZE(ctx->pending_async); i++) { |
413 | spin_lock_init(&ctx->pending_async[i].lock); | |
414 | INIT_LIST_HEAD(&ctx->pending_async[i].list); | |
415 | atomic_set(&ctx->pending_async[i].cnt, 0); | |
416 | } | |
2b188cc1 | 417 | spin_lock_init(&ctx->completion_lock); |
def596e9 | 418 | INIT_LIST_HEAD(&ctx->poll_list); |
221c5eb2 | 419 | INIT_LIST_HEAD(&ctx->cancel_list); |
de0617e4 | 420 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 421 | INIT_LIST_HEAD(&ctx->timeout_list); |
2b188cc1 JA |
422 | return ctx; |
423 | } | |
424 | ||
7adf4eaf JA |
425 | static inline bool __io_sequence_defer(struct io_ring_ctx *ctx, |
426 | struct io_kiocb *req) | |
427 | { | |
498ccd9e JA |
428 | return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped |
429 | + atomic_read(&ctx->cached_cq_overflow); | |
7adf4eaf JA |
430 | } |
431 | ||
de0617e4 JA |
432 | static inline bool io_sequence_defer(struct io_ring_ctx *ctx, |
433 | struct io_kiocb *req) | |
434 | { | |
7adf4eaf | 435 | if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) != REQ_F_IO_DRAIN) |
de0617e4 JA |
436 | return false; |
437 | ||
7adf4eaf | 438 | return __io_sequence_defer(ctx, req); |
de0617e4 JA |
439 | } |
440 | ||
7adf4eaf | 441 | static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx) |
de0617e4 JA |
442 | { |
443 | struct io_kiocb *req; | |
444 | ||
7adf4eaf JA |
445 | req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list); |
446 | if (req && !io_sequence_defer(ctx, req)) { | |
de0617e4 JA |
447 | list_del_init(&req->list); |
448 | return req; | |
449 | } | |
450 | ||
451 | return NULL; | |
452 | } | |
453 | ||
5262f567 JA |
454 | static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx) |
455 | { | |
7adf4eaf JA |
456 | struct io_kiocb *req; |
457 | ||
458 | req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list); | |
459 | if (req && !__io_sequence_defer(ctx, req)) { | |
460 | list_del_init(&req->list); | |
461 | return req; | |
462 | } | |
463 | ||
464 | return NULL; | |
5262f567 JA |
465 | } |
466 | ||
de0617e4 | 467 | static void __io_commit_cqring(struct io_ring_ctx *ctx) |
2b188cc1 | 468 | { |
75b28aff | 469 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 470 | |
75b28aff | 471 | if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) { |
2b188cc1 | 472 | /* order cqe stores with ring update */ |
75b28aff | 473 | smp_store_release(&rings->cq.tail, ctx->cached_cq_tail); |
2b188cc1 | 474 | |
2b188cc1 JA |
475 | if (wq_has_sleeper(&ctx->cq_wait)) { |
476 | wake_up_interruptible(&ctx->cq_wait); | |
477 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
478 | } | |
479 | } | |
480 | } | |
481 | ||
18d9be1a JA |
482 | static inline void io_queue_async_work(struct io_ring_ctx *ctx, |
483 | struct io_kiocb *req) | |
484 | { | |
6cc47d1d | 485 | int rw = 0; |
54a91f3b | 486 | |
6cc47d1d JA |
487 | if (req->submit.sqe) { |
488 | switch (req->submit.sqe->opcode) { | |
489 | case IORING_OP_WRITEV: | |
490 | case IORING_OP_WRITE_FIXED: | |
491 | rw = !(req->rw.ki_flags & IOCB_DIRECT); | |
492 | break; | |
493 | } | |
54a91f3b JA |
494 | } |
495 | ||
c826bd7a | 496 | trace_io_uring_queue_async_work(ctx, rw, req, &req->work, req->flags); |
54a91f3b | 497 | queue_work(ctx->sqo_wq[rw], &req->work); |
18d9be1a JA |
498 | } |
499 | ||
5262f567 JA |
500 | static void io_kill_timeout(struct io_kiocb *req) |
501 | { | |
502 | int ret; | |
503 | ||
504 | ret = hrtimer_try_to_cancel(&req->timeout.timer); | |
505 | if (ret != -1) { | |
506 | atomic_inc(&req->ctx->cq_timeouts); | |
507 | list_del(&req->list); | |
508 | io_cqring_fill_event(req->ctx, req->user_data, 0); | |
509 | __io_free_req(req); | |
510 | } | |
511 | } | |
512 | ||
513 | static void io_kill_timeouts(struct io_ring_ctx *ctx) | |
514 | { | |
515 | struct io_kiocb *req, *tmp; | |
516 | ||
517 | spin_lock_irq(&ctx->completion_lock); | |
518 | list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list) | |
519 | io_kill_timeout(req); | |
520 | spin_unlock_irq(&ctx->completion_lock); | |
521 | } | |
522 | ||
de0617e4 JA |
523 | static void io_commit_cqring(struct io_ring_ctx *ctx) |
524 | { | |
525 | struct io_kiocb *req; | |
526 | ||
5262f567 JA |
527 | while ((req = io_get_timeout_req(ctx)) != NULL) |
528 | io_kill_timeout(req); | |
529 | ||
de0617e4 JA |
530 | __io_commit_cqring(ctx); |
531 | ||
532 | while ((req = io_get_deferred_req(ctx)) != NULL) { | |
4fe2c963 JL |
533 | if (req->flags & REQ_F_SHADOW_DRAIN) { |
534 | /* Just for drain, free it. */ | |
535 | __io_free_req(req); | |
536 | continue; | |
537 | } | |
de0617e4 | 538 | req->flags |= REQ_F_IO_DRAINED; |
18d9be1a | 539 | io_queue_async_work(ctx, req); |
de0617e4 JA |
540 | } |
541 | } | |
542 | ||
2b188cc1 JA |
543 | static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx) |
544 | { | |
75b28aff | 545 | struct io_rings *rings = ctx->rings; |
2b188cc1 JA |
546 | unsigned tail; |
547 | ||
548 | tail = ctx->cached_cq_tail; | |
115e12e5 SB |
549 | /* |
550 | * writes to the cq entry need to come after reading head; the | |
551 | * control dependency is enough as we're using WRITE_ONCE to | |
552 | * fill the cq entry | |
553 | */ | |
75b28aff | 554 | if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries) |
2b188cc1 JA |
555 | return NULL; |
556 | ||
557 | ctx->cached_cq_tail++; | |
75b28aff | 558 | return &rings->cqes[tail & ctx->cq_mask]; |
2b188cc1 JA |
559 | } |
560 | ||
561 | static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data, | |
c71ffb67 | 562 | long res) |
2b188cc1 JA |
563 | { |
564 | struct io_uring_cqe *cqe; | |
565 | ||
566 | /* | |
567 | * If we can't get a cq entry, userspace overflowed the | |
568 | * submission (by quite a lot). Increment the overflow count in | |
569 | * the ring. | |
570 | */ | |
571 | cqe = io_get_cqring(ctx); | |
572 | if (cqe) { | |
573 | WRITE_ONCE(cqe->user_data, ki_user_data); | |
574 | WRITE_ONCE(cqe->res, res); | |
c71ffb67 | 575 | WRITE_ONCE(cqe->flags, 0); |
2b188cc1 | 576 | } else { |
498ccd9e JA |
577 | WRITE_ONCE(ctx->rings->cq_overflow, |
578 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
2b188cc1 JA |
579 | } |
580 | } | |
581 | ||
8c838788 JA |
582 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
583 | { | |
584 | if (waitqueue_active(&ctx->wait)) | |
585 | wake_up(&ctx->wait); | |
586 | if (waitqueue_active(&ctx->sqo_wait)) | |
587 | wake_up(&ctx->sqo_wait); | |
9b402849 JA |
588 | if (ctx->cq_ev_fd) |
589 | eventfd_signal(ctx->cq_ev_fd, 1); | |
8c838788 JA |
590 | } |
591 | ||
592 | static void io_cqring_add_event(struct io_ring_ctx *ctx, u64 user_data, | |
c71ffb67 | 593 | long res) |
2b188cc1 JA |
594 | { |
595 | unsigned long flags; | |
596 | ||
597 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
c71ffb67 | 598 | io_cqring_fill_event(ctx, user_data, res); |
2b188cc1 JA |
599 | io_commit_cqring(ctx); |
600 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
601 | ||
8c838788 | 602 | io_cqring_ev_posted(ctx); |
2b188cc1 JA |
603 | } |
604 | ||
2579f913 JA |
605 | static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx, |
606 | struct io_submit_state *state) | |
2b188cc1 | 607 | { |
fd6fab2c | 608 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
2b188cc1 JA |
609 | struct io_kiocb *req; |
610 | ||
611 | if (!percpu_ref_tryget(&ctx->refs)) | |
612 | return NULL; | |
613 | ||
2579f913 | 614 | if (!state) { |
fd6fab2c | 615 | req = kmem_cache_alloc(req_cachep, gfp); |
2579f913 JA |
616 | if (unlikely(!req)) |
617 | goto out; | |
618 | } else if (!state->free_reqs) { | |
619 | size_t sz; | |
620 | int ret; | |
621 | ||
622 | sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs)); | |
fd6fab2c JA |
623 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs); |
624 | ||
625 | /* | |
626 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
627 | * retry single alloc to be on the safe side. | |
628 | */ | |
629 | if (unlikely(ret <= 0)) { | |
630 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
631 | if (!state->reqs[0]) | |
632 | goto out; | |
633 | ret = 1; | |
634 | } | |
2579f913 JA |
635 | state->free_reqs = ret - 1; |
636 | state->cur_req = 1; | |
637 | req = state->reqs[0]; | |
638 | } else { | |
639 | req = state->reqs[state->cur_req]; | |
640 | state->free_reqs--; | |
641 | state->cur_req++; | |
2b188cc1 JA |
642 | } |
643 | ||
60c112b0 | 644 | req->file = NULL; |
2579f913 JA |
645 | req->ctx = ctx; |
646 | req->flags = 0; | |
e65ef56d JA |
647 | /* one is dropped after submission, the other at completion */ |
648 | refcount_set(&req->refs, 2); | |
9e645e11 | 649 | req->result = 0; |
2579f913 JA |
650 | return req; |
651 | out: | |
6805b32e | 652 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
653 | return NULL; |
654 | } | |
655 | ||
def596e9 JA |
656 | static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr) |
657 | { | |
658 | if (*nr) { | |
659 | kmem_cache_free_bulk(req_cachep, *nr, reqs); | |
6805b32e | 660 | percpu_ref_put_many(&ctx->refs, *nr); |
def596e9 JA |
661 | *nr = 0; |
662 | } | |
663 | } | |
664 | ||
9e645e11 | 665 | static void __io_free_req(struct io_kiocb *req) |
2b188cc1 | 666 | { |
09bb8394 JA |
667 | if (req->file && !(req->flags & REQ_F_FIXED_FILE)) |
668 | fput(req->file); | |
6805b32e | 669 | percpu_ref_put(&req->ctx->refs); |
e65ef56d JA |
670 | kmem_cache_free(req_cachep, req); |
671 | } | |
672 | ||
ba816ad6 | 673 | static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr) |
9e645e11 JA |
674 | { |
675 | struct io_kiocb *nxt; | |
676 | ||
677 | /* | |
678 | * The list should never be empty when we are called here. But could | |
679 | * potentially happen if the chain is messed up, check to be on the | |
680 | * safe side. | |
681 | */ | |
682 | nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list); | |
683 | if (nxt) { | |
684 | list_del(&nxt->list); | |
685 | if (!list_empty(&req->link_list)) { | |
686 | INIT_LIST_HEAD(&nxt->link_list); | |
687 | list_splice(&req->link_list, &nxt->link_list); | |
688 | nxt->flags |= REQ_F_LINK; | |
689 | } | |
690 | ||
f7b76ac9 | 691 | nxt->flags |= REQ_F_LINK_DONE; |
ba816ad6 JA |
692 | /* |
693 | * If we're in async work, we can continue processing the chain | |
694 | * in this context instead of having to queue up new async work. | |
695 | */ | |
696 | if (nxtptr && current_work()) { | |
697 | *nxtptr = nxt; | |
698 | } else { | |
699 | INIT_WORK(&nxt->work, io_sq_wq_submit_work); | |
700 | io_queue_async_work(req->ctx, nxt); | |
701 | } | |
9e645e11 JA |
702 | } |
703 | } | |
704 | ||
705 | /* | |
706 | * Called if REQ_F_LINK is set, and we fail the head request | |
707 | */ | |
708 | static void io_fail_links(struct io_kiocb *req) | |
709 | { | |
710 | struct io_kiocb *link; | |
711 | ||
712 | while (!list_empty(&req->link_list)) { | |
713 | link = list_first_entry(&req->link_list, struct io_kiocb, list); | |
714 | list_del(&link->list); | |
715 | ||
c826bd7a | 716 | trace_io_uring_fail_link(req, link); |
9e645e11 JA |
717 | io_cqring_add_event(req->ctx, link->user_data, -ECANCELED); |
718 | __io_free_req(link); | |
719 | } | |
720 | } | |
721 | ||
ba816ad6 | 722 | static void io_free_req(struct io_kiocb *req, struct io_kiocb **nxt) |
9e645e11 JA |
723 | { |
724 | /* | |
725 | * If LINK is set, we have dependent requests in this chain. If we | |
726 | * didn't fail this request, queue the first one up, moving any other | |
727 | * dependencies to the next request. In case of failure, fail the rest | |
728 | * of the chain. | |
729 | */ | |
730 | if (req->flags & REQ_F_LINK) { | |
731 | if (req->flags & REQ_F_FAIL_LINK) | |
732 | io_fail_links(req); | |
733 | else | |
ba816ad6 | 734 | io_req_link_next(req, nxt); |
9e645e11 JA |
735 | } |
736 | ||
737 | __io_free_req(req); | |
738 | } | |
739 | ||
ba816ad6 JA |
740 | /* |
741 | * Drop reference to request, return next in chain (if there is one) if this | |
742 | * was the last reference to this request. | |
743 | */ | |
744 | static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req) | |
e65ef56d | 745 | { |
ba816ad6 JA |
746 | struct io_kiocb *nxt = NULL; |
747 | ||
e65ef56d | 748 | if (refcount_dec_and_test(&req->refs)) |
ba816ad6 JA |
749 | io_free_req(req, &nxt); |
750 | ||
751 | return nxt; | |
752 | } | |
753 | ||
754 | static void io_put_req(struct io_kiocb *req, struct io_kiocb **nxtptr) | |
755 | { | |
756 | struct io_kiocb *nxt; | |
757 | ||
758 | nxt = io_put_req_find_next(req); | |
759 | if (nxt) { | |
760 | if (nxtptr) { | |
761 | *nxtptr = nxt; | |
762 | } else { | |
763 | INIT_WORK(&nxt->work, io_sq_wq_submit_work); | |
764 | io_queue_async_work(nxt->ctx, nxt); | |
765 | } | |
766 | } | |
2b188cc1 JA |
767 | } |
768 | ||
75b28aff | 769 | static unsigned io_cqring_events(struct io_rings *rings) |
a3a0e43f JA |
770 | { |
771 | /* See comment at the top of this file */ | |
772 | smp_rmb(); | |
75b28aff | 773 | return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head); |
a3a0e43f JA |
774 | } |
775 | ||
fb5ccc98 PB |
776 | static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) |
777 | { | |
778 | struct io_rings *rings = ctx->rings; | |
779 | ||
780 | /* make sure SQ entry isn't read before tail */ | |
781 | return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; | |
782 | } | |
783 | ||
def596e9 JA |
784 | /* |
785 | * Find and free completed poll iocbs | |
786 | */ | |
787 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
788 | struct list_head *done) | |
789 | { | |
790 | void *reqs[IO_IOPOLL_BATCH]; | |
791 | struct io_kiocb *req; | |
09bb8394 | 792 | int to_free; |
def596e9 | 793 | |
09bb8394 | 794 | to_free = 0; |
def596e9 JA |
795 | while (!list_empty(done)) { |
796 | req = list_first_entry(done, struct io_kiocb, list); | |
797 | list_del(&req->list); | |
798 | ||
9e645e11 | 799 | io_cqring_fill_event(ctx, req->user_data, req->result); |
def596e9 JA |
800 | (*nr_events)++; |
801 | ||
09bb8394 JA |
802 | if (refcount_dec_and_test(&req->refs)) { |
803 | /* If we're not using fixed files, we have to pair the | |
804 | * completion part with the file put. Use regular | |
805 | * completions for those, only batch free for fixed | |
9e645e11 | 806 | * file and non-linked commands. |
09bb8394 | 807 | */ |
9e645e11 JA |
808 | if ((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) == |
809 | REQ_F_FIXED_FILE) { | |
09bb8394 JA |
810 | reqs[to_free++] = req; |
811 | if (to_free == ARRAY_SIZE(reqs)) | |
812 | io_free_req_many(ctx, reqs, &to_free); | |
6b06314c | 813 | } else { |
ba816ad6 | 814 | io_free_req(req, NULL); |
6b06314c | 815 | } |
9a56a232 | 816 | } |
def596e9 | 817 | } |
def596e9 | 818 | |
09bb8394 | 819 | io_commit_cqring(ctx); |
def596e9 JA |
820 | io_free_req_many(ctx, reqs, &to_free); |
821 | } | |
822 | ||
823 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
824 | long min) | |
825 | { | |
826 | struct io_kiocb *req, *tmp; | |
827 | LIST_HEAD(done); | |
828 | bool spin; | |
829 | int ret; | |
830 | ||
831 | /* | |
832 | * Only spin for completions if we don't have multiple devices hanging | |
833 | * off our complete list, and we're under the requested amount. | |
834 | */ | |
835 | spin = !ctx->poll_multi_file && *nr_events < min; | |
836 | ||
837 | ret = 0; | |
838 | list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) { | |
839 | struct kiocb *kiocb = &req->rw; | |
840 | ||
841 | /* | |
842 | * Move completed entries to our local list. If we find a | |
843 | * request that requires polling, break out and complete | |
844 | * the done list first, if we have entries there. | |
845 | */ | |
846 | if (req->flags & REQ_F_IOPOLL_COMPLETED) { | |
847 | list_move_tail(&req->list, &done); | |
848 | continue; | |
849 | } | |
850 | if (!list_empty(&done)) | |
851 | break; | |
852 | ||
853 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
854 | if (ret < 0) | |
855 | break; | |
856 | ||
857 | if (ret && spin) | |
858 | spin = false; | |
859 | ret = 0; | |
860 | } | |
861 | ||
862 | if (!list_empty(&done)) | |
863 | io_iopoll_complete(ctx, nr_events, &done); | |
864 | ||
865 | return ret; | |
866 | } | |
867 | ||
868 | /* | |
869 | * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a | |
870 | * non-spinning poll check - we'll still enter the driver poll loop, but only | |
871 | * as a non-spinning completion check. | |
872 | */ | |
873 | static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
874 | long min) | |
875 | { | |
08f5439f | 876 | while (!list_empty(&ctx->poll_list) && !need_resched()) { |
def596e9 JA |
877 | int ret; |
878 | ||
879 | ret = io_do_iopoll(ctx, nr_events, min); | |
880 | if (ret < 0) | |
881 | return ret; | |
882 | if (!min || *nr_events >= min) | |
883 | return 0; | |
884 | } | |
885 | ||
886 | return 1; | |
887 | } | |
888 | ||
889 | /* | |
890 | * We can't just wait for polled events to come to us, we have to actively | |
891 | * find and complete them. | |
892 | */ | |
893 | static void io_iopoll_reap_events(struct io_ring_ctx *ctx) | |
894 | { | |
895 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
896 | return; | |
897 | ||
898 | mutex_lock(&ctx->uring_lock); | |
899 | while (!list_empty(&ctx->poll_list)) { | |
900 | unsigned int nr_events = 0; | |
901 | ||
902 | io_iopoll_getevents(ctx, &nr_events, 1); | |
08f5439f JA |
903 | |
904 | /* | |
905 | * Ensure we allow local-to-the-cpu processing to take place, | |
906 | * in this case we need to ensure that we reap all events. | |
907 | */ | |
908 | cond_resched(); | |
def596e9 JA |
909 | } |
910 | mutex_unlock(&ctx->uring_lock); | |
911 | } | |
912 | ||
2b2ed975 JA |
913 | static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, |
914 | long min) | |
def596e9 | 915 | { |
2b2ed975 | 916 | int iters = 0, ret = 0; |
500f9fba | 917 | |
def596e9 JA |
918 | do { |
919 | int tmin = 0; | |
920 | ||
a3a0e43f JA |
921 | /* |
922 | * Don't enter poll loop if we already have events pending. | |
923 | * If we do, we can potentially be spinning for commands that | |
924 | * already triggered a CQE (eg in error). | |
925 | */ | |
75b28aff | 926 | if (io_cqring_events(ctx->rings)) |
a3a0e43f JA |
927 | break; |
928 | ||
500f9fba JA |
929 | /* |
930 | * If a submit got punted to a workqueue, we can have the | |
931 | * application entering polling for a command before it gets | |
932 | * issued. That app will hold the uring_lock for the duration | |
933 | * of the poll right here, so we need to take a breather every | |
934 | * now and then to ensure that the issue has a chance to add | |
935 | * the poll to the issued list. Otherwise we can spin here | |
936 | * forever, while the workqueue is stuck trying to acquire the | |
937 | * very same mutex. | |
938 | */ | |
939 | if (!(++iters & 7)) { | |
940 | mutex_unlock(&ctx->uring_lock); | |
941 | mutex_lock(&ctx->uring_lock); | |
942 | } | |
943 | ||
def596e9 JA |
944 | if (*nr_events < min) |
945 | tmin = min - *nr_events; | |
946 | ||
947 | ret = io_iopoll_getevents(ctx, nr_events, tmin); | |
948 | if (ret <= 0) | |
949 | break; | |
950 | ret = 0; | |
951 | } while (min && !*nr_events && !need_resched()); | |
952 | ||
2b2ed975 JA |
953 | return ret; |
954 | } | |
955 | ||
956 | static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, | |
957 | long min) | |
958 | { | |
959 | int ret; | |
960 | ||
961 | /* | |
962 | * We disallow the app entering submit/complete with polling, but we | |
963 | * still need to lock the ring to prevent racing with polled issue | |
964 | * that got punted to a workqueue. | |
965 | */ | |
966 | mutex_lock(&ctx->uring_lock); | |
967 | ret = __io_iopoll_check(ctx, nr_events, min); | |
500f9fba | 968 | mutex_unlock(&ctx->uring_lock); |
def596e9 JA |
969 | return ret; |
970 | } | |
971 | ||
491381ce | 972 | static void kiocb_end_write(struct io_kiocb *req) |
2b188cc1 | 973 | { |
491381ce JA |
974 | /* |
975 | * Tell lockdep we inherited freeze protection from submission | |
976 | * thread. | |
977 | */ | |
978 | if (req->flags & REQ_F_ISREG) { | |
979 | struct inode *inode = file_inode(req->file); | |
2b188cc1 | 980 | |
491381ce | 981 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); |
2b188cc1 | 982 | } |
491381ce | 983 | file_end_write(req->file); |
2b188cc1 JA |
984 | } |
985 | ||
ba816ad6 | 986 | static void io_complete_rw_common(struct kiocb *kiocb, long res) |
2b188cc1 JA |
987 | { |
988 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
989 | ||
491381ce JA |
990 | if (kiocb->ki_flags & IOCB_WRITE) |
991 | kiocb_end_write(req); | |
2b188cc1 | 992 | |
9e645e11 JA |
993 | if ((req->flags & REQ_F_LINK) && res != req->result) |
994 | req->flags |= REQ_F_FAIL_LINK; | |
c71ffb67 | 995 | io_cqring_add_event(req->ctx, req->user_data, res); |
ba816ad6 JA |
996 | } |
997 | ||
998 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
999 | { | |
1000 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
1001 | ||
1002 | io_complete_rw_common(kiocb, res); | |
1003 | io_put_req(req, NULL); | |
1004 | } | |
1005 | ||
1006 | static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res) | |
1007 | { | |
1008 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
1009 | ||
1010 | io_complete_rw_common(kiocb, res); | |
1011 | return io_put_req_find_next(req); | |
2b188cc1 JA |
1012 | } |
1013 | ||
def596e9 JA |
1014 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
1015 | { | |
1016 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
1017 | ||
491381ce JA |
1018 | if (kiocb->ki_flags & IOCB_WRITE) |
1019 | kiocb_end_write(req); | |
def596e9 | 1020 | |
9e645e11 JA |
1021 | if ((req->flags & REQ_F_LINK) && res != req->result) |
1022 | req->flags |= REQ_F_FAIL_LINK; | |
1023 | req->result = res; | |
def596e9 JA |
1024 | if (res != -EAGAIN) |
1025 | req->flags |= REQ_F_IOPOLL_COMPLETED; | |
1026 | } | |
1027 | ||
1028 | /* | |
1029 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1030 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
1031 | * find it from a io_iopoll_getevents() thread before the issuer is done | |
1032 | * accessing the kiocb cookie. | |
1033 | */ | |
1034 | static void io_iopoll_req_issued(struct io_kiocb *req) | |
1035 | { | |
1036 | struct io_ring_ctx *ctx = req->ctx; | |
1037 | ||
1038 | /* | |
1039 | * Track whether we have multiple files in our lists. This will impact | |
1040 | * how we do polling eventually, not spinning if we're on potentially | |
1041 | * different devices. | |
1042 | */ | |
1043 | if (list_empty(&ctx->poll_list)) { | |
1044 | ctx->poll_multi_file = false; | |
1045 | } else if (!ctx->poll_multi_file) { | |
1046 | struct io_kiocb *list_req; | |
1047 | ||
1048 | list_req = list_first_entry(&ctx->poll_list, struct io_kiocb, | |
1049 | list); | |
1050 | if (list_req->rw.ki_filp != req->rw.ki_filp) | |
1051 | ctx->poll_multi_file = true; | |
1052 | } | |
1053 | ||
1054 | /* | |
1055 | * For fast devices, IO may have already completed. If it has, add | |
1056 | * it to the front so we find it first. | |
1057 | */ | |
1058 | if (req->flags & REQ_F_IOPOLL_COMPLETED) | |
1059 | list_add(&req->list, &ctx->poll_list); | |
1060 | else | |
1061 | list_add_tail(&req->list, &ctx->poll_list); | |
1062 | } | |
1063 | ||
3d6770fb | 1064 | static void io_file_put(struct io_submit_state *state) |
9a56a232 | 1065 | { |
3d6770fb | 1066 | if (state->file) { |
9a56a232 JA |
1067 | int diff = state->has_refs - state->used_refs; |
1068 | ||
1069 | if (diff) | |
1070 | fput_many(state->file, diff); | |
1071 | state->file = NULL; | |
1072 | } | |
1073 | } | |
1074 | ||
1075 | /* | |
1076 | * Get as many references to a file as we have IOs left in this submission, | |
1077 | * assuming most submissions are for one file, or at least that each file | |
1078 | * has more than one submission. | |
1079 | */ | |
1080 | static struct file *io_file_get(struct io_submit_state *state, int fd) | |
1081 | { | |
1082 | if (!state) | |
1083 | return fget(fd); | |
1084 | ||
1085 | if (state->file) { | |
1086 | if (state->fd == fd) { | |
1087 | state->used_refs++; | |
1088 | state->ios_left--; | |
1089 | return state->file; | |
1090 | } | |
3d6770fb | 1091 | io_file_put(state); |
9a56a232 JA |
1092 | } |
1093 | state->file = fget_many(fd, state->ios_left); | |
1094 | if (!state->file) | |
1095 | return NULL; | |
1096 | ||
1097 | state->fd = fd; | |
1098 | state->has_refs = state->ios_left; | |
1099 | state->used_refs = 1; | |
1100 | state->ios_left--; | |
1101 | return state->file; | |
1102 | } | |
1103 | ||
2b188cc1 JA |
1104 | /* |
1105 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1106 | * any file. For now, just ensure that anything potentially problematic is done | |
1107 | * inline. | |
1108 | */ | |
1109 | static bool io_file_supports_async(struct file *file) | |
1110 | { | |
1111 | umode_t mode = file_inode(file)->i_mode; | |
1112 | ||
1113 | if (S_ISBLK(mode) || S_ISCHR(mode)) | |
1114 | return true; | |
1115 | if (S_ISREG(mode) && file->f_op != &io_uring_fops) | |
1116 | return true; | |
1117 | ||
1118 | return false; | |
1119 | } | |
1120 | ||
6c271ce2 | 1121 | static int io_prep_rw(struct io_kiocb *req, const struct sqe_submit *s, |
8358e3a8 | 1122 | bool force_nonblock) |
2b188cc1 | 1123 | { |
6c271ce2 | 1124 | const struct io_uring_sqe *sqe = s->sqe; |
def596e9 | 1125 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 | 1126 | struct kiocb *kiocb = &req->rw; |
09bb8394 JA |
1127 | unsigned ioprio; |
1128 | int ret; | |
2b188cc1 | 1129 | |
09bb8394 JA |
1130 | if (!req->file) |
1131 | return -EBADF; | |
2b188cc1 | 1132 | |
491381ce JA |
1133 | if (S_ISREG(file_inode(req->file)->i_mode)) |
1134 | req->flags |= REQ_F_ISREG; | |
1135 | ||
1136 | /* | |
1137 | * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so | |
1138 | * we know to async punt it even if it was opened O_NONBLOCK | |
1139 | */ | |
1140 | if (force_nonblock && !io_file_supports_async(req->file)) { | |
1141 | req->flags |= REQ_F_MUST_PUNT; | |
1142 | return -EAGAIN; | |
1143 | } | |
6b06314c | 1144 | |
2b188cc1 JA |
1145 | kiocb->ki_pos = READ_ONCE(sqe->off); |
1146 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); | |
1147 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); | |
1148 | ||
1149 | ioprio = READ_ONCE(sqe->ioprio); | |
1150 | if (ioprio) { | |
1151 | ret = ioprio_check_cap(ioprio); | |
1152 | if (ret) | |
09bb8394 | 1153 | return ret; |
2b188cc1 JA |
1154 | |
1155 | kiocb->ki_ioprio = ioprio; | |
1156 | } else | |
1157 | kiocb->ki_ioprio = get_current_ioprio(); | |
1158 | ||
1159 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
1160 | if (unlikely(ret)) | |
09bb8394 | 1161 | return ret; |
8449eeda SB |
1162 | |
1163 | /* don't allow async punt if RWF_NOWAIT was requested */ | |
491381ce JA |
1164 | if ((kiocb->ki_flags & IOCB_NOWAIT) || |
1165 | (req->file->f_flags & O_NONBLOCK)) | |
8449eeda SB |
1166 | req->flags |= REQ_F_NOWAIT; |
1167 | ||
1168 | if (force_nonblock) | |
2b188cc1 | 1169 | kiocb->ki_flags |= IOCB_NOWAIT; |
8449eeda | 1170 | |
def596e9 | 1171 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
1172 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
1173 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 1174 | return -EOPNOTSUPP; |
2b188cc1 | 1175 | |
def596e9 JA |
1176 | kiocb->ki_flags |= IOCB_HIPRI; |
1177 | kiocb->ki_complete = io_complete_rw_iopoll; | |
1178 | } else { | |
09bb8394 JA |
1179 | if (kiocb->ki_flags & IOCB_HIPRI) |
1180 | return -EINVAL; | |
def596e9 JA |
1181 | kiocb->ki_complete = io_complete_rw; |
1182 | } | |
2b188cc1 | 1183 | return 0; |
2b188cc1 JA |
1184 | } |
1185 | ||
1186 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
1187 | { | |
1188 | switch (ret) { | |
1189 | case -EIOCBQUEUED: | |
1190 | break; | |
1191 | case -ERESTARTSYS: | |
1192 | case -ERESTARTNOINTR: | |
1193 | case -ERESTARTNOHAND: | |
1194 | case -ERESTART_RESTARTBLOCK: | |
1195 | /* | |
1196 | * We can't just restart the syscall, since previously | |
1197 | * submitted sqes may already be in progress. Just fail this | |
1198 | * IO with EINTR. | |
1199 | */ | |
1200 | ret = -EINTR; | |
1201 | /* fall through */ | |
1202 | default: | |
1203 | kiocb->ki_complete(kiocb, ret, 0); | |
1204 | } | |
1205 | } | |
1206 | ||
ba816ad6 JA |
1207 | static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt, |
1208 | bool in_async) | |
1209 | { | |
1210 | if (in_async && ret >= 0 && nxt && kiocb->ki_complete == io_complete_rw) | |
1211 | *nxt = __io_complete_rw(kiocb, ret); | |
1212 | else | |
1213 | io_rw_done(kiocb, ret); | |
1214 | } | |
1215 | ||
edafccee JA |
1216 | static int io_import_fixed(struct io_ring_ctx *ctx, int rw, |
1217 | const struct io_uring_sqe *sqe, | |
1218 | struct iov_iter *iter) | |
1219 | { | |
1220 | size_t len = READ_ONCE(sqe->len); | |
1221 | struct io_mapped_ubuf *imu; | |
1222 | unsigned index, buf_index; | |
1223 | size_t offset; | |
1224 | u64 buf_addr; | |
1225 | ||
1226 | /* attempt to use fixed buffers without having provided iovecs */ | |
1227 | if (unlikely(!ctx->user_bufs)) | |
1228 | return -EFAULT; | |
1229 | ||
1230 | buf_index = READ_ONCE(sqe->buf_index); | |
1231 | if (unlikely(buf_index >= ctx->nr_user_bufs)) | |
1232 | return -EFAULT; | |
1233 | ||
1234 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); | |
1235 | imu = &ctx->user_bufs[index]; | |
1236 | buf_addr = READ_ONCE(sqe->addr); | |
1237 | ||
1238 | /* overflow */ | |
1239 | if (buf_addr + len < buf_addr) | |
1240 | return -EFAULT; | |
1241 | /* not inside the mapped region */ | |
1242 | if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len) | |
1243 | return -EFAULT; | |
1244 | ||
1245 | /* | |
1246 | * May not be a start of buffer, set size appropriately | |
1247 | * and advance us to the beginning. | |
1248 | */ | |
1249 | offset = buf_addr - imu->ubuf; | |
1250 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
bd11b3a3 JA |
1251 | |
1252 | if (offset) { | |
1253 | /* | |
1254 | * Don't use iov_iter_advance() here, as it's really slow for | |
1255 | * using the latter parts of a big fixed buffer - it iterates | |
1256 | * over each segment manually. We can cheat a bit here, because | |
1257 | * we know that: | |
1258 | * | |
1259 | * 1) it's a BVEC iter, we set it up | |
1260 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
1261 | * first and last bvec | |
1262 | * | |
1263 | * So just find our index, and adjust the iterator afterwards. | |
1264 | * If the offset is within the first bvec (or the whole first | |
1265 | * bvec, just use iov_iter_advance(). This makes it easier | |
1266 | * since we can just skip the first segment, which may not | |
1267 | * be PAGE_SIZE aligned. | |
1268 | */ | |
1269 | const struct bio_vec *bvec = imu->bvec; | |
1270 | ||
1271 | if (offset <= bvec->bv_len) { | |
1272 | iov_iter_advance(iter, offset); | |
1273 | } else { | |
1274 | unsigned long seg_skip; | |
1275 | ||
1276 | /* skip first vec */ | |
1277 | offset -= bvec->bv_len; | |
1278 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
1279 | ||
1280 | iter->bvec = bvec + seg_skip; | |
1281 | iter->nr_segs -= seg_skip; | |
99c79f66 | 1282 | iter->count -= bvec->bv_len + offset; |
bd11b3a3 | 1283 | iter->iov_offset = offset & ~PAGE_MASK; |
bd11b3a3 JA |
1284 | } |
1285 | } | |
1286 | ||
edafccee JA |
1287 | return 0; |
1288 | } | |
1289 | ||
87e5e6da JA |
1290 | static ssize_t io_import_iovec(struct io_ring_ctx *ctx, int rw, |
1291 | const struct sqe_submit *s, struct iovec **iovec, | |
1292 | struct iov_iter *iter) | |
2b188cc1 JA |
1293 | { |
1294 | const struct io_uring_sqe *sqe = s->sqe; | |
1295 | void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
1296 | size_t sqe_len = READ_ONCE(sqe->len); | |
edafccee JA |
1297 | u8 opcode; |
1298 | ||
1299 | /* | |
1300 | * We're reading ->opcode for the second time, but the first read | |
1301 | * doesn't care whether it's _FIXED or not, so it doesn't matter | |
1302 | * whether ->opcode changes concurrently. The first read does care | |
1303 | * about whether it is a READ or a WRITE, so we don't trust this read | |
1304 | * for that purpose and instead let the caller pass in the read/write | |
1305 | * flag. | |
1306 | */ | |
1307 | opcode = READ_ONCE(sqe->opcode); | |
1308 | if (opcode == IORING_OP_READ_FIXED || | |
1309 | opcode == IORING_OP_WRITE_FIXED) { | |
87e5e6da | 1310 | ssize_t ret = io_import_fixed(ctx, rw, sqe, iter); |
edafccee JA |
1311 | *iovec = NULL; |
1312 | return ret; | |
1313 | } | |
2b188cc1 JA |
1314 | |
1315 | if (!s->has_user) | |
1316 | return -EFAULT; | |
1317 | ||
1318 | #ifdef CONFIG_COMPAT | |
1319 | if (ctx->compat) | |
1320 | return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV, | |
1321 | iovec, iter); | |
1322 | #endif | |
1323 | ||
1324 | return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter); | |
1325 | } | |
1326 | ||
6d5d5ac5 JA |
1327 | static inline bool io_should_merge(struct async_list *al, struct kiocb *kiocb) |
1328 | { | |
1329 | if (al->file == kiocb->ki_filp) { | |
1330 | off_t start, end; | |
1331 | ||
1332 | /* | |
1333 | * Allow merging if we're anywhere in the range of the same | |
1334 | * page. Generally this happens for sub-page reads or writes, | |
1335 | * and it's beneficial to allow the first worker to bring the | |
1336 | * page in and the piggy backed work can then work on the | |
1337 | * cached page. | |
1338 | */ | |
1339 | start = al->io_start & PAGE_MASK; | |
1340 | end = (al->io_start + al->io_len + PAGE_SIZE - 1) & PAGE_MASK; | |
1341 | if (kiocb->ki_pos >= start && kiocb->ki_pos <= end) | |
1342 | return true; | |
1343 | } | |
1344 | ||
1345 | al->file = NULL; | |
1346 | return false; | |
1347 | } | |
1348 | ||
31b51510 JA |
1349 | /* |
1350 | * Make a note of the last file/offset/direction we punted to async | |
1351 | * context. We'll use this information to see if we can piggy back a | |
1352 | * sequential request onto the previous one, if it's still hasn't been | |
1353 | * completed by the async worker. | |
1354 | */ | |
1355 | static void io_async_list_note(int rw, struct io_kiocb *req, size_t len) | |
1356 | { | |
1357 | struct async_list *async_list = &req->ctx->pending_async[rw]; | |
1358 | struct kiocb *kiocb = &req->rw; | |
1359 | struct file *filp = kiocb->ki_filp; | |
31b51510 | 1360 | |
6d5d5ac5 | 1361 | if (io_should_merge(async_list, kiocb)) { |
9310a7ba | 1362 | unsigned long max_bytes; |
31b51510 JA |
1363 | |
1364 | /* Use 8x RA size as a decent limiter for both reads/writes */ | |
9310a7ba ZL |
1365 | max_bytes = filp->f_ra.ra_pages << (PAGE_SHIFT + 3); |
1366 | if (!max_bytes) | |
1367 | max_bytes = VM_READAHEAD_PAGES << (PAGE_SHIFT + 3); | |
1368 | ||
1369 | /* If max len are exceeded, reset the state */ | |
1370 | if (async_list->io_len + len <= max_bytes) { | |
31b51510 | 1371 | req->flags |= REQ_F_SEQ_PREV; |
9310a7ba | 1372 | async_list->io_len += len; |
31b51510 | 1373 | } else { |
6d5d5ac5 | 1374 | async_list->file = NULL; |
31b51510 JA |
1375 | } |
1376 | } | |
1377 | ||
1378 | /* New file? Reset state. */ | |
1379 | if (async_list->file != filp) { | |
6d5d5ac5 JA |
1380 | async_list->io_start = kiocb->ki_pos; |
1381 | async_list->io_len = len; | |
31b51510 JA |
1382 | async_list->file = filp; |
1383 | } | |
31b51510 JA |
1384 | } |
1385 | ||
32960613 JA |
1386 | /* |
1387 | * For files that don't have ->read_iter() and ->write_iter(), handle them | |
1388 | * by looping over ->read() or ->write() manually. | |
1389 | */ | |
1390 | static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb, | |
1391 | struct iov_iter *iter) | |
1392 | { | |
1393 | ssize_t ret = 0; | |
1394 | ||
1395 | /* | |
1396 | * Don't support polled IO through this interface, and we can't | |
1397 | * support non-blocking either. For the latter, this just causes | |
1398 | * the kiocb to be handled from an async context. | |
1399 | */ | |
1400 | if (kiocb->ki_flags & IOCB_HIPRI) | |
1401 | return -EOPNOTSUPP; | |
1402 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
1403 | return -EAGAIN; | |
1404 | ||
1405 | while (iov_iter_count(iter)) { | |
1406 | struct iovec iovec = iov_iter_iovec(iter); | |
1407 | ssize_t nr; | |
1408 | ||
1409 | if (rw == READ) { | |
1410 | nr = file->f_op->read(file, iovec.iov_base, | |
1411 | iovec.iov_len, &kiocb->ki_pos); | |
1412 | } else { | |
1413 | nr = file->f_op->write(file, iovec.iov_base, | |
1414 | iovec.iov_len, &kiocb->ki_pos); | |
1415 | } | |
1416 | ||
1417 | if (nr < 0) { | |
1418 | if (!ret) | |
1419 | ret = nr; | |
1420 | break; | |
1421 | } | |
1422 | ret += nr; | |
1423 | if (nr != iovec.iov_len) | |
1424 | break; | |
1425 | iov_iter_advance(iter, nr); | |
1426 | } | |
1427 | ||
1428 | return ret; | |
1429 | } | |
1430 | ||
e0c5c576 | 1431 | static int io_read(struct io_kiocb *req, const struct sqe_submit *s, |
ba816ad6 | 1432 | struct io_kiocb **nxt, bool force_nonblock) |
2b188cc1 JA |
1433 | { |
1434 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1435 | struct kiocb *kiocb = &req->rw; | |
1436 | struct iov_iter iter; | |
1437 | struct file *file; | |
31b51510 | 1438 | size_t iov_count; |
9d93a3f5 | 1439 | ssize_t read_size, ret; |
2b188cc1 | 1440 | |
8358e3a8 | 1441 | ret = io_prep_rw(req, s, force_nonblock); |
2b188cc1 JA |
1442 | if (ret) |
1443 | return ret; | |
1444 | file = kiocb->ki_filp; | |
1445 | ||
2b188cc1 | 1446 | if (unlikely(!(file->f_mode & FMODE_READ))) |
09bb8394 | 1447 | return -EBADF; |
2b188cc1 JA |
1448 | |
1449 | ret = io_import_iovec(req->ctx, READ, s, &iovec, &iter); | |
87e5e6da | 1450 | if (ret < 0) |
09bb8394 | 1451 | return ret; |
2b188cc1 | 1452 | |
9d93a3f5 | 1453 | read_size = ret; |
9e645e11 JA |
1454 | if (req->flags & REQ_F_LINK) |
1455 | req->result = read_size; | |
1456 | ||
31b51510 JA |
1457 | iov_count = iov_iter_count(&iter); |
1458 | ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 JA |
1459 | if (!ret) { |
1460 | ssize_t ret2; | |
1461 | ||
32960613 JA |
1462 | if (file->f_op->read_iter) |
1463 | ret2 = call_read_iter(file, kiocb, &iter); | |
1464 | else | |
1465 | ret2 = loop_rw_iter(READ, file, kiocb, &iter); | |
1466 | ||
9d93a3f5 JA |
1467 | /* |
1468 | * In case of a short read, punt to async. This can happen | |
1469 | * if we have data partially cached. Alternatively we can | |
1470 | * return the short read, in which case the application will | |
1471 | * need to issue another SQE and wait for it. That SQE will | |
1472 | * need async punt anyway, so it's more efficient to do it | |
1473 | * here. | |
1474 | */ | |
491381ce JA |
1475 | if (force_nonblock && !(req->flags & REQ_F_NOWAIT) && |
1476 | (req->flags & REQ_F_ISREG) && | |
1477 | ret2 > 0 && ret2 < read_size) | |
9d93a3f5 JA |
1478 | ret2 = -EAGAIN; |
1479 | /* Catch -EAGAIN return for forced non-blocking submission */ | |
31b51510 | 1480 | if (!force_nonblock || ret2 != -EAGAIN) { |
ba5290cc | 1481 | kiocb_done(kiocb, ret2, nxt, s->in_async); |
31b51510 | 1482 | } else { |
ba5290cc | 1483 | if (!s->in_async) |
31b51510 | 1484 | io_async_list_note(READ, req, iov_count); |
2b188cc1 | 1485 | ret = -EAGAIN; |
31b51510 | 1486 | } |
2b188cc1 JA |
1487 | } |
1488 | kfree(iovec); | |
2b188cc1 JA |
1489 | return ret; |
1490 | } | |
1491 | ||
e0c5c576 | 1492 | static int io_write(struct io_kiocb *req, const struct sqe_submit *s, |
ba816ad6 | 1493 | struct io_kiocb **nxt, bool force_nonblock) |
2b188cc1 JA |
1494 | { |
1495 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1496 | struct kiocb *kiocb = &req->rw; | |
1497 | struct iov_iter iter; | |
1498 | struct file *file; | |
31b51510 | 1499 | size_t iov_count; |
87e5e6da | 1500 | ssize_t ret; |
2b188cc1 | 1501 | |
8358e3a8 | 1502 | ret = io_prep_rw(req, s, force_nonblock); |
2b188cc1 JA |
1503 | if (ret) |
1504 | return ret; | |
2b188cc1 | 1505 | |
2b188cc1 JA |
1506 | file = kiocb->ki_filp; |
1507 | if (unlikely(!(file->f_mode & FMODE_WRITE))) | |
09bb8394 | 1508 | return -EBADF; |
2b188cc1 JA |
1509 | |
1510 | ret = io_import_iovec(req->ctx, WRITE, s, &iovec, &iter); | |
87e5e6da | 1511 | if (ret < 0) |
09bb8394 | 1512 | return ret; |
2b188cc1 | 1513 | |
9e645e11 JA |
1514 | if (req->flags & REQ_F_LINK) |
1515 | req->result = ret; | |
1516 | ||
31b51510 JA |
1517 | iov_count = iov_iter_count(&iter); |
1518 | ||
1519 | ret = -EAGAIN; | |
1520 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT)) { | |
ba5290cc | 1521 | if (!s->in_async) |
31b51510 JA |
1522 | io_async_list_note(WRITE, req, iov_count); |
1523 | goto out_free; | |
1524 | } | |
1525 | ||
1526 | ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 | 1527 | if (!ret) { |
9bf7933f RP |
1528 | ssize_t ret2; |
1529 | ||
2b188cc1 JA |
1530 | /* |
1531 | * Open-code file_start_write here to grab freeze protection, | |
1532 | * which will be released by another thread in | |
1533 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
1534 | * released so that it doesn't complain about the held lock when | |
1535 | * we return to userspace. | |
1536 | */ | |
491381ce | 1537 | if (req->flags & REQ_F_ISREG) { |
2b188cc1 JA |
1538 | __sb_start_write(file_inode(file)->i_sb, |
1539 | SB_FREEZE_WRITE, true); | |
1540 | __sb_writers_release(file_inode(file)->i_sb, | |
1541 | SB_FREEZE_WRITE); | |
1542 | } | |
1543 | kiocb->ki_flags |= IOCB_WRITE; | |
9bf7933f | 1544 | |
32960613 JA |
1545 | if (file->f_op->write_iter) |
1546 | ret2 = call_write_iter(file, kiocb, &iter); | |
1547 | else | |
1548 | ret2 = loop_rw_iter(WRITE, file, kiocb, &iter); | |
9bf7933f | 1549 | if (!force_nonblock || ret2 != -EAGAIN) { |
ba5290cc | 1550 | kiocb_done(kiocb, ret2, nxt, s->in_async); |
9bf7933f | 1551 | } else { |
ba5290cc | 1552 | if (!s->in_async) |
9bf7933f RP |
1553 | io_async_list_note(WRITE, req, iov_count); |
1554 | ret = -EAGAIN; | |
1555 | } | |
2b188cc1 | 1556 | } |
31b51510 | 1557 | out_free: |
2b188cc1 | 1558 | kfree(iovec); |
2b188cc1 JA |
1559 | return ret; |
1560 | } | |
1561 | ||
1562 | /* | |
1563 | * IORING_OP_NOP just posts a completion event, nothing else. | |
1564 | */ | |
1565 | static int io_nop(struct io_kiocb *req, u64 user_data) | |
1566 | { | |
1567 | struct io_ring_ctx *ctx = req->ctx; | |
1568 | long err = 0; | |
1569 | ||
def596e9 JA |
1570 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
1571 | return -EINVAL; | |
1572 | ||
c71ffb67 | 1573 | io_cqring_add_event(ctx, user_data, err); |
ba816ad6 | 1574 | io_put_req(req, NULL); |
2b188cc1 JA |
1575 | return 0; |
1576 | } | |
1577 | ||
c992fe29 CH |
1578 | static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1579 | { | |
6b06314c | 1580 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 1581 | |
09bb8394 JA |
1582 | if (!req->file) |
1583 | return -EBADF; | |
c992fe29 | 1584 | |
6b06314c | 1585 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 1586 | return -EINVAL; |
edafccee | 1587 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
1588 | return -EINVAL; |
1589 | ||
c992fe29 CH |
1590 | return 0; |
1591 | } | |
1592 | ||
1593 | static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
ba816ad6 | 1594 | struct io_kiocb **nxt, bool force_nonblock) |
c992fe29 CH |
1595 | { |
1596 | loff_t sqe_off = READ_ONCE(sqe->off); | |
1597 | loff_t sqe_len = READ_ONCE(sqe->len); | |
1598 | loff_t end = sqe_off + sqe_len; | |
1599 | unsigned fsync_flags; | |
1600 | int ret; | |
1601 | ||
1602 | fsync_flags = READ_ONCE(sqe->fsync_flags); | |
1603 | if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC)) | |
1604 | return -EINVAL; | |
1605 | ||
1606 | ret = io_prep_fsync(req, sqe); | |
1607 | if (ret) | |
1608 | return ret; | |
1609 | ||
1610 | /* fsync always requires a blocking context */ | |
1611 | if (force_nonblock) | |
1612 | return -EAGAIN; | |
1613 | ||
1614 | ret = vfs_fsync_range(req->rw.ki_filp, sqe_off, | |
1615 | end > 0 ? end : LLONG_MAX, | |
1616 | fsync_flags & IORING_FSYNC_DATASYNC); | |
1617 | ||
9e645e11 JA |
1618 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1619 | req->flags |= REQ_F_FAIL_LINK; | |
c71ffb67 | 1620 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
ba816ad6 | 1621 | io_put_req(req, nxt); |
c992fe29 CH |
1622 | return 0; |
1623 | } | |
1624 | ||
5d17b4a4 JA |
1625 | static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1626 | { | |
1627 | struct io_ring_ctx *ctx = req->ctx; | |
1628 | int ret = 0; | |
1629 | ||
1630 | if (!req->file) | |
1631 | return -EBADF; | |
5d17b4a4 JA |
1632 | |
1633 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
1634 | return -EINVAL; | |
1635 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) | |
1636 | return -EINVAL; | |
1637 | ||
5d17b4a4 JA |
1638 | return ret; |
1639 | } | |
1640 | ||
1641 | static int io_sync_file_range(struct io_kiocb *req, | |
1642 | const struct io_uring_sqe *sqe, | |
ba816ad6 | 1643 | struct io_kiocb **nxt, |
5d17b4a4 JA |
1644 | bool force_nonblock) |
1645 | { | |
1646 | loff_t sqe_off; | |
1647 | loff_t sqe_len; | |
1648 | unsigned flags; | |
1649 | int ret; | |
1650 | ||
1651 | ret = io_prep_sfr(req, sqe); | |
1652 | if (ret) | |
1653 | return ret; | |
1654 | ||
1655 | /* sync_file_range always requires a blocking context */ | |
1656 | if (force_nonblock) | |
1657 | return -EAGAIN; | |
1658 | ||
1659 | sqe_off = READ_ONCE(sqe->off); | |
1660 | sqe_len = READ_ONCE(sqe->len); | |
1661 | flags = READ_ONCE(sqe->sync_range_flags); | |
1662 | ||
1663 | ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags); | |
1664 | ||
9e645e11 JA |
1665 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1666 | req->flags |= REQ_F_FAIL_LINK; | |
c71ffb67 | 1667 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
ba816ad6 | 1668 | io_put_req(req, nxt); |
5d17b4a4 JA |
1669 | return 0; |
1670 | } | |
1671 | ||
0fa03c62 | 1672 | #if defined(CONFIG_NET) |
aa1fa28f | 1673 | static int io_send_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
ba816ad6 | 1674 | struct io_kiocb **nxt, bool force_nonblock, |
aa1fa28f JA |
1675 | long (*fn)(struct socket *, struct user_msghdr __user *, |
1676 | unsigned int)) | |
1677 | { | |
0fa03c62 JA |
1678 | struct socket *sock; |
1679 | int ret; | |
1680 | ||
1681 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1682 | return -EINVAL; | |
1683 | ||
1684 | sock = sock_from_file(req->file, &ret); | |
1685 | if (sock) { | |
1686 | struct user_msghdr __user *msg; | |
1687 | unsigned flags; | |
1688 | ||
1689 | flags = READ_ONCE(sqe->msg_flags); | |
1690 | if (flags & MSG_DONTWAIT) | |
1691 | req->flags |= REQ_F_NOWAIT; | |
1692 | else if (force_nonblock) | |
1693 | flags |= MSG_DONTWAIT; | |
1694 | ||
1695 | msg = (struct user_msghdr __user *) (unsigned long) | |
1696 | READ_ONCE(sqe->addr); | |
1697 | ||
aa1fa28f | 1698 | ret = fn(sock, msg, flags); |
0fa03c62 JA |
1699 | if (force_nonblock && ret == -EAGAIN) |
1700 | return ret; | |
1701 | } | |
1702 | ||
c71ffb67 | 1703 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
ba816ad6 | 1704 | io_put_req(req, nxt); |
5d17b4a4 JA |
1705 | return 0; |
1706 | } | |
aa1fa28f JA |
1707 | #endif |
1708 | ||
1709 | static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
ba816ad6 | 1710 | struct io_kiocb **nxt, bool force_nonblock) |
aa1fa28f JA |
1711 | { |
1712 | #if defined(CONFIG_NET) | |
ba816ad6 JA |
1713 | return io_send_recvmsg(req, sqe, nxt, force_nonblock, |
1714 | __sys_sendmsg_sock); | |
aa1fa28f JA |
1715 | #else |
1716 | return -EOPNOTSUPP; | |
1717 | #endif | |
1718 | } | |
1719 | ||
1720 | static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
ba816ad6 | 1721 | struct io_kiocb **nxt, bool force_nonblock) |
aa1fa28f JA |
1722 | { |
1723 | #if defined(CONFIG_NET) | |
ba816ad6 JA |
1724 | return io_send_recvmsg(req, sqe, nxt, force_nonblock, |
1725 | __sys_recvmsg_sock); | |
0fa03c62 JA |
1726 | #else |
1727 | return -EOPNOTSUPP; | |
1728 | #endif | |
1729 | } | |
5d17b4a4 | 1730 | |
221c5eb2 JA |
1731 | static void io_poll_remove_one(struct io_kiocb *req) |
1732 | { | |
1733 | struct io_poll_iocb *poll = &req->poll; | |
1734 | ||
1735 | spin_lock(&poll->head->lock); | |
1736 | WRITE_ONCE(poll->canceled, true); | |
1737 | if (!list_empty(&poll->wait.entry)) { | |
1738 | list_del_init(&poll->wait.entry); | |
18d9be1a | 1739 | io_queue_async_work(req->ctx, req); |
221c5eb2 JA |
1740 | } |
1741 | spin_unlock(&poll->head->lock); | |
1742 | ||
1743 | list_del_init(&req->list); | |
1744 | } | |
1745 | ||
1746 | static void io_poll_remove_all(struct io_ring_ctx *ctx) | |
1747 | { | |
1748 | struct io_kiocb *req; | |
1749 | ||
1750 | spin_lock_irq(&ctx->completion_lock); | |
1751 | while (!list_empty(&ctx->cancel_list)) { | |
1752 | req = list_first_entry(&ctx->cancel_list, struct io_kiocb,list); | |
1753 | io_poll_remove_one(req); | |
1754 | } | |
1755 | spin_unlock_irq(&ctx->completion_lock); | |
1756 | } | |
1757 | ||
1758 | /* | |
1759 | * Find a running poll command that matches one specified in sqe->addr, | |
1760 | * and remove it if found. | |
1761 | */ | |
1762 | static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
1763 | { | |
1764 | struct io_ring_ctx *ctx = req->ctx; | |
1765 | struct io_kiocb *poll_req, *next; | |
1766 | int ret = -ENOENT; | |
1767 | ||
1768 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1769 | return -EINVAL; | |
1770 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index || | |
1771 | sqe->poll_events) | |
1772 | return -EINVAL; | |
1773 | ||
1774 | spin_lock_irq(&ctx->completion_lock); | |
1775 | list_for_each_entry_safe(poll_req, next, &ctx->cancel_list, list) { | |
1776 | if (READ_ONCE(sqe->addr) == poll_req->user_data) { | |
1777 | io_poll_remove_one(poll_req); | |
1778 | ret = 0; | |
1779 | break; | |
1780 | } | |
1781 | } | |
1782 | spin_unlock_irq(&ctx->completion_lock); | |
1783 | ||
c71ffb67 | 1784 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
ba816ad6 | 1785 | io_put_req(req, NULL); |
221c5eb2 JA |
1786 | return 0; |
1787 | } | |
1788 | ||
8c838788 JA |
1789 | static void io_poll_complete(struct io_ring_ctx *ctx, struct io_kiocb *req, |
1790 | __poll_t mask) | |
221c5eb2 | 1791 | { |
8c838788 | 1792 | req->poll.done = true; |
c71ffb67 | 1793 | io_cqring_fill_event(ctx, req->user_data, mangle_poll(mask)); |
8c838788 | 1794 | io_commit_cqring(ctx); |
221c5eb2 JA |
1795 | } |
1796 | ||
1797 | static void io_poll_complete_work(struct work_struct *work) | |
1798 | { | |
1799 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
1800 | struct io_poll_iocb *poll = &req->poll; | |
1801 | struct poll_table_struct pt = { ._key = poll->events }; | |
1802 | struct io_ring_ctx *ctx = req->ctx; | |
1803 | __poll_t mask = 0; | |
1804 | ||
1805 | if (!READ_ONCE(poll->canceled)) | |
1806 | mask = vfs_poll(poll->file, &pt) & poll->events; | |
1807 | ||
1808 | /* | |
1809 | * Note that ->ki_cancel callers also delete iocb from active_reqs after | |
1810 | * calling ->ki_cancel. We need the ctx_lock roundtrip here to | |
1811 | * synchronize with them. In the cancellation case the list_del_init | |
1812 | * itself is not actually needed, but harmless so we keep it in to | |
1813 | * avoid further branches in the fast path. | |
1814 | */ | |
1815 | spin_lock_irq(&ctx->completion_lock); | |
1816 | if (!mask && !READ_ONCE(poll->canceled)) { | |
1817 | add_wait_queue(poll->head, &poll->wait); | |
1818 | spin_unlock_irq(&ctx->completion_lock); | |
1819 | return; | |
1820 | } | |
1821 | list_del_init(&req->list); | |
8c838788 | 1822 | io_poll_complete(ctx, req, mask); |
221c5eb2 JA |
1823 | spin_unlock_irq(&ctx->completion_lock); |
1824 | ||
8c838788 | 1825 | io_cqring_ev_posted(ctx); |
ba816ad6 | 1826 | io_put_req(req, NULL); |
221c5eb2 JA |
1827 | } |
1828 | ||
1829 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
1830 | void *key) | |
1831 | { | |
1832 | struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb, | |
1833 | wait); | |
1834 | struct io_kiocb *req = container_of(poll, struct io_kiocb, poll); | |
1835 | struct io_ring_ctx *ctx = req->ctx; | |
1836 | __poll_t mask = key_to_poll(key); | |
8c838788 | 1837 | unsigned long flags; |
221c5eb2 JA |
1838 | |
1839 | /* for instances that support it check for an event match first: */ | |
8c838788 JA |
1840 | if (mask && !(mask & poll->events)) |
1841 | return 0; | |
221c5eb2 | 1842 | |
8c838788 | 1843 | list_del_init(&poll->wait.entry); |
221c5eb2 | 1844 | |
8c838788 JA |
1845 | if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) { |
1846 | list_del(&req->list); | |
1847 | io_poll_complete(ctx, req, mask); | |
1848 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
221c5eb2 | 1849 | |
8c838788 | 1850 | io_cqring_ev_posted(ctx); |
ba816ad6 | 1851 | io_put_req(req, NULL); |
8c838788 | 1852 | } else { |
18d9be1a | 1853 | io_queue_async_work(ctx, req); |
221c5eb2 JA |
1854 | } |
1855 | ||
221c5eb2 JA |
1856 | return 1; |
1857 | } | |
1858 | ||
1859 | struct io_poll_table { | |
1860 | struct poll_table_struct pt; | |
1861 | struct io_kiocb *req; | |
1862 | int error; | |
1863 | }; | |
1864 | ||
1865 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
1866 | struct poll_table_struct *p) | |
1867 | { | |
1868 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
1869 | ||
1870 | if (unlikely(pt->req->poll.head)) { | |
1871 | pt->error = -EINVAL; | |
1872 | return; | |
1873 | } | |
1874 | ||
1875 | pt->error = 0; | |
1876 | pt->req->poll.head = head; | |
1877 | add_wait_queue(head, &pt->req->poll.wait); | |
1878 | } | |
1879 | ||
1880 | static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
1881 | { | |
1882 | struct io_poll_iocb *poll = &req->poll; | |
1883 | struct io_ring_ctx *ctx = req->ctx; | |
1884 | struct io_poll_table ipt; | |
8c838788 | 1885 | bool cancel = false; |
221c5eb2 JA |
1886 | __poll_t mask; |
1887 | u16 events; | |
221c5eb2 JA |
1888 | |
1889 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1890 | return -EINVAL; | |
1891 | if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
1892 | return -EINVAL; | |
09bb8394 JA |
1893 | if (!poll->file) |
1894 | return -EBADF; | |
221c5eb2 | 1895 | |
6cc47d1d | 1896 | req->submit.sqe = NULL; |
221c5eb2 JA |
1897 | INIT_WORK(&req->work, io_poll_complete_work); |
1898 | events = READ_ONCE(sqe->poll_events); | |
1899 | poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP; | |
1900 | ||
221c5eb2 | 1901 | poll->head = NULL; |
8c838788 | 1902 | poll->done = false; |
221c5eb2 JA |
1903 | poll->canceled = false; |
1904 | ||
1905 | ipt.pt._qproc = io_poll_queue_proc; | |
1906 | ipt.pt._key = poll->events; | |
1907 | ipt.req = req; | |
1908 | ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */ | |
1909 | ||
1910 | /* initialized the list so that we can do list_empty checks */ | |
1911 | INIT_LIST_HEAD(&poll->wait.entry); | |
1912 | init_waitqueue_func_entry(&poll->wait, io_poll_wake); | |
1913 | ||
36703247 JA |
1914 | INIT_LIST_HEAD(&req->list); |
1915 | ||
221c5eb2 | 1916 | mask = vfs_poll(poll->file, &ipt.pt) & poll->events; |
221c5eb2 JA |
1917 | |
1918 | spin_lock_irq(&ctx->completion_lock); | |
8c838788 JA |
1919 | if (likely(poll->head)) { |
1920 | spin_lock(&poll->head->lock); | |
1921 | if (unlikely(list_empty(&poll->wait.entry))) { | |
1922 | if (ipt.error) | |
1923 | cancel = true; | |
1924 | ipt.error = 0; | |
1925 | mask = 0; | |
1926 | } | |
1927 | if (mask || ipt.error) | |
1928 | list_del_init(&poll->wait.entry); | |
1929 | else if (cancel) | |
1930 | WRITE_ONCE(poll->canceled, true); | |
1931 | else if (!poll->done) /* actually waiting for an event */ | |
1932 | list_add_tail(&req->list, &ctx->cancel_list); | |
1933 | spin_unlock(&poll->head->lock); | |
1934 | } | |
1935 | if (mask) { /* no async, we'd stolen it */ | |
221c5eb2 | 1936 | ipt.error = 0; |
8c838788 | 1937 | io_poll_complete(ctx, req, mask); |
221c5eb2 | 1938 | } |
221c5eb2 JA |
1939 | spin_unlock_irq(&ctx->completion_lock); |
1940 | ||
8c838788 JA |
1941 | if (mask) { |
1942 | io_cqring_ev_posted(ctx); | |
ba816ad6 | 1943 | io_put_req(req, NULL); |
221c5eb2 | 1944 | } |
8c838788 | 1945 | return ipt.error; |
221c5eb2 JA |
1946 | } |
1947 | ||
5262f567 JA |
1948 | static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer) |
1949 | { | |
1950 | struct io_ring_ctx *ctx; | |
11365043 | 1951 | struct io_kiocb *req; |
5262f567 | 1952 | unsigned long flags; |
11365043 | 1953 | bool comp; |
5262f567 JA |
1954 | |
1955 | req = container_of(timer, struct io_kiocb, timeout.timer); | |
1956 | ctx = req->ctx; | |
1957 | atomic_inc(&ctx->cq_timeouts); | |
1958 | ||
1959 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
ef03681a | 1960 | /* |
11365043 JA |
1961 | * We could be racing with timeout deletion. If the list is empty, |
1962 | * then timeout lookup already found it and will be handling it. | |
ef03681a | 1963 | */ |
11365043 JA |
1964 | comp = !list_empty(&req->list); |
1965 | if (comp) { | |
1966 | struct io_kiocb *prev; | |
5262f567 | 1967 | |
11365043 JA |
1968 | /* |
1969 | * Adjust the reqs sequence before the current one because it | |
1970 | * will consume a slot in the cq_ring and the the cq_tail | |
1971 | * pointer will be increased, otherwise other timeout reqs may | |
1972 | * return in advance without waiting for enough wait_nr. | |
1973 | */ | |
1974 | prev = req; | |
1975 | list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list) | |
1976 | prev->sequence++; | |
1977 | ||
1978 | list_del_init(&req->list); | |
1979 | io_cqring_fill_event(ctx, req->user_data, -ETIME); | |
1980 | io_commit_cqring(ctx); | |
1981 | } | |
5262f567 JA |
1982 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
1983 | ||
11365043 JA |
1984 | if (comp) { |
1985 | io_cqring_ev_posted(ctx); | |
1986 | io_put_req(req, NULL); | |
1987 | } | |
1988 | return HRTIMER_NORESTART; | |
1989 | } | |
1990 | ||
1991 | /* | |
1992 | * Remove or update an existing timeout command | |
1993 | */ | |
1994 | static int io_timeout_remove(struct io_kiocb *req, | |
1995 | const struct io_uring_sqe *sqe) | |
1996 | { | |
1997 | struct io_ring_ctx *ctx = req->ctx; | |
1998 | struct io_kiocb *treq; | |
1999 | int ret = -ENOENT; | |
2000 | __u64 user_data; | |
2001 | unsigned flags; | |
2002 | ||
2003 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
2004 | return -EINVAL; | |
2005 | if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len) | |
2006 | return -EINVAL; | |
2007 | flags = READ_ONCE(sqe->timeout_flags); | |
2008 | if (flags) | |
2009 | return -EINVAL; | |
2010 | ||
2011 | user_data = READ_ONCE(sqe->addr); | |
2012 | spin_lock_irq(&ctx->completion_lock); | |
2013 | list_for_each_entry(treq, &ctx->timeout_list, list) { | |
2014 | if (user_data == treq->user_data) { | |
2015 | list_del_init(&treq->list); | |
2016 | ret = 0; | |
2017 | break; | |
2018 | } | |
2019 | } | |
2020 | ||
2021 | /* didn't find timeout */ | |
2022 | if (ret) { | |
2023 | fill_ev: | |
2024 | io_cqring_fill_event(ctx, req->user_data, ret); | |
2025 | io_commit_cqring(ctx); | |
2026 | spin_unlock_irq(&ctx->completion_lock); | |
2027 | io_cqring_ev_posted(ctx); | |
2028 | io_put_req(req, NULL); | |
2029 | return 0; | |
2030 | } | |
2031 | ||
2032 | ret = hrtimer_try_to_cancel(&treq->timeout.timer); | |
2033 | if (ret == -1) { | |
2034 | ret = -EBUSY; | |
2035 | goto fill_ev; | |
2036 | } | |
2037 | ||
2038 | io_cqring_fill_event(ctx, req->user_data, 0); | |
2039 | io_cqring_fill_event(ctx, treq->user_data, -ECANCELED); | |
2040 | io_commit_cqring(ctx); | |
2041 | spin_unlock_irq(&ctx->completion_lock); | |
5262f567 JA |
2042 | io_cqring_ev_posted(ctx); |
2043 | ||
11365043 | 2044 | io_put_req(treq, NULL); |
ba816ad6 | 2045 | io_put_req(req, NULL); |
11365043 | 2046 | return 0; |
5262f567 JA |
2047 | } |
2048 | ||
2049 | static int io_timeout(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
2050 | { | |
5da0fb1a | 2051 | unsigned count; |
5262f567 JA |
2052 | struct io_ring_ctx *ctx = req->ctx; |
2053 | struct list_head *entry; | |
a41525ab | 2054 | enum hrtimer_mode mode; |
bdf20073 | 2055 | struct timespec64 ts; |
a1f58ba4 | 2056 | unsigned span = 0; |
a41525ab | 2057 | unsigned flags; |
5262f567 JA |
2058 | |
2059 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
2060 | return -EINVAL; | |
a41525ab JA |
2061 | if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len != 1) |
2062 | return -EINVAL; | |
2063 | flags = READ_ONCE(sqe->timeout_flags); | |
2064 | if (flags & ~IORING_TIMEOUT_ABS) | |
5262f567 | 2065 | return -EINVAL; |
bdf20073 AB |
2066 | |
2067 | if (get_timespec64(&ts, u64_to_user_ptr(sqe->addr))) | |
5262f567 JA |
2068 | return -EFAULT; |
2069 | ||
11365043 JA |
2070 | if (flags & IORING_TIMEOUT_ABS) |
2071 | mode = HRTIMER_MODE_ABS; | |
2072 | else | |
2073 | mode = HRTIMER_MODE_REL; | |
2074 | ||
2075 | hrtimer_init(&req->timeout.timer, CLOCK_MONOTONIC, mode); | |
2076 | ||
5262f567 JA |
2077 | /* |
2078 | * sqe->off holds how many events that need to occur for this | |
2079 | * timeout event to be satisfied. | |
2080 | */ | |
2081 | count = READ_ONCE(sqe->off); | |
2082 | if (!count) | |
2083 | count = 1; | |
2084 | ||
2085 | req->sequence = ctx->cached_sq_head + count - 1; | |
5da0fb1a | 2086 | /* reuse it to store the count */ |
2087 | req->submit.sequence = count; | |
5262f567 JA |
2088 | req->flags |= REQ_F_TIMEOUT; |
2089 | ||
2090 | /* | |
2091 | * Insertion sort, ensuring the first entry in the list is always | |
2092 | * the one we need first. | |
2093 | */ | |
5262f567 JA |
2094 | spin_lock_irq(&ctx->completion_lock); |
2095 | list_for_each_prev(entry, &ctx->timeout_list) { | |
2096 | struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list); | |
5da0fb1a | 2097 | unsigned nxt_sq_head; |
2098 | long long tmp, tmp_nxt; | |
5262f567 | 2099 | |
5da0fb1a | 2100 | /* |
2101 | * Since cached_sq_head + count - 1 can overflow, use type long | |
2102 | * long to store it. | |
2103 | */ | |
2104 | tmp = (long long)ctx->cached_sq_head + count - 1; | |
2105 | nxt_sq_head = nxt->sequence - nxt->submit.sequence + 1; | |
2106 | tmp_nxt = (long long)nxt_sq_head + nxt->submit.sequence - 1; | |
2107 | ||
2108 | /* | |
2109 | * cached_sq_head may overflow, and it will never overflow twice | |
2110 | * once there is some timeout req still be valid. | |
2111 | */ | |
2112 | if (ctx->cached_sq_head < nxt_sq_head) | |
8b07a65a | 2113 | tmp += UINT_MAX; |
5da0fb1a | 2114 | |
a1f58ba4 | 2115 | if (tmp > tmp_nxt) |
5262f567 | 2116 | break; |
a1f58ba4 | 2117 | |
2118 | /* | |
2119 | * Sequence of reqs after the insert one and itself should | |
2120 | * be adjusted because each timeout req consumes a slot. | |
2121 | */ | |
2122 | span++; | |
2123 | nxt->sequence++; | |
5262f567 | 2124 | } |
a1f58ba4 | 2125 | req->sequence -= span; |
5262f567 JA |
2126 | list_add(&req->list, entry); |
2127 | spin_unlock_irq(&ctx->completion_lock); | |
5262f567 | 2128 | req->timeout.timer.function = io_timeout_fn; |
a41525ab | 2129 | hrtimer_start(&req->timeout.timer, timespec64_to_ktime(ts), mode); |
5262f567 JA |
2130 | return 0; |
2131 | } | |
2132 | ||
de0617e4 JA |
2133 | static int io_req_defer(struct io_ring_ctx *ctx, struct io_kiocb *req, |
2134 | const struct io_uring_sqe *sqe) | |
2135 | { | |
2136 | struct io_uring_sqe *sqe_copy; | |
2137 | ||
2138 | if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) | |
2139 | return 0; | |
2140 | ||
2141 | sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL); | |
2142 | if (!sqe_copy) | |
2143 | return -EAGAIN; | |
2144 | ||
2145 | spin_lock_irq(&ctx->completion_lock); | |
2146 | if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) { | |
2147 | spin_unlock_irq(&ctx->completion_lock); | |
2148 | kfree(sqe_copy); | |
2149 | return 0; | |
2150 | } | |
2151 | ||
2152 | memcpy(sqe_copy, sqe, sizeof(*sqe_copy)); | |
2153 | req->submit.sqe = sqe_copy; | |
2154 | ||
2155 | INIT_WORK(&req->work, io_sq_wq_submit_work); | |
c826bd7a | 2156 | trace_io_uring_defer(ctx, req, false); |
de0617e4 JA |
2157 | list_add_tail(&req->list, &ctx->defer_list); |
2158 | spin_unlock_irq(&ctx->completion_lock); | |
2159 | return -EIOCBQUEUED; | |
2160 | } | |
2161 | ||
2b188cc1 | 2162 | static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
ba816ad6 JA |
2163 | const struct sqe_submit *s, struct io_kiocb **nxt, |
2164 | bool force_nonblock) | |
2b188cc1 | 2165 | { |
e0c5c576 | 2166 | int ret, opcode; |
2b188cc1 | 2167 | |
9e645e11 JA |
2168 | req->user_data = READ_ONCE(s->sqe->user_data); |
2169 | ||
2b188cc1 JA |
2170 | opcode = READ_ONCE(s->sqe->opcode); |
2171 | switch (opcode) { | |
2172 | case IORING_OP_NOP: | |
2173 | ret = io_nop(req, req->user_data); | |
2174 | break; | |
2175 | case IORING_OP_READV: | |
edafccee JA |
2176 | if (unlikely(s->sqe->buf_index)) |
2177 | return -EINVAL; | |
ba816ad6 | 2178 | ret = io_read(req, s, nxt, force_nonblock); |
2b188cc1 JA |
2179 | break; |
2180 | case IORING_OP_WRITEV: | |
edafccee JA |
2181 | if (unlikely(s->sqe->buf_index)) |
2182 | return -EINVAL; | |
ba816ad6 | 2183 | ret = io_write(req, s, nxt, force_nonblock); |
edafccee JA |
2184 | break; |
2185 | case IORING_OP_READ_FIXED: | |
ba816ad6 | 2186 | ret = io_read(req, s, nxt, force_nonblock); |
edafccee JA |
2187 | break; |
2188 | case IORING_OP_WRITE_FIXED: | |
ba816ad6 | 2189 | ret = io_write(req, s, nxt, force_nonblock); |
2b188cc1 | 2190 | break; |
c992fe29 | 2191 | case IORING_OP_FSYNC: |
ba816ad6 | 2192 | ret = io_fsync(req, s->sqe, nxt, force_nonblock); |
c992fe29 | 2193 | break; |
221c5eb2 JA |
2194 | case IORING_OP_POLL_ADD: |
2195 | ret = io_poll_add(req, s->sqe); | |
2196 | break; | |
2197 | case IORING_OP_POLL_REMOVE: | |
2198 | ret = io_poll_remove(req, s->sqe); | |
2199 | break; | |
5d17b4a4 | 2200 | case IORING_OP_SYNC_FILE_RANGE: |
ba816ad6 | 2201 | ret = io_sync_file_range(req, s->sqe, nxt, force_nonblock); |
5d17b4a4 | 2202 | break; |
0fa03c62 | 2203 | case IORING_OP_SENDMSG: |
ba816ad6 | 2204 | ret = io_sendmsg(req, s->sqe, nxt, force_nonblock); |
0fa03c62 | 2205 | break; |
aa1fa28f | 2206 | case IORING_OP_RECVMSG: |
ba816ad6 | 2207 | ret = io_recvmsg(req, s->sqe, nxt, force_nonblock); |
aa1fa28f | 2208 | break; |
5262f567 JA |
2209 | case IORING_OP_TIMEOUT: |
2210 | ret = io_timeout(req, s->sqe); | |
2211 | break; | |
11365043 JA |
2212 | case IORING_OP_TIMEOUT_REMOVE: |
2213 | ret = io_timeout_remove(req, s->sqe); | |
2214 | break; | |
2b188cc1 JA |
2215 | default: |
2216 | ret = -EINVAL; | |
2217 | break; | |
2218 | } | |
2219 | ||
def596e9 JA |
2220 | if (ret) |
2221 | return ret; | |
2222 | ||
2223 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
9e645e11 | 2224 | if (req->result == -EAGAIN) |
def596e9 JA |
2225 | return -EAGAIN; |
2226 | ||
2227 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
ba5290cc | 2228 | if (s->in_async) |
def596e9 JA |
2229 | mutex_lock(&ctx->uring_lock); |
2230 | io_iopoll_req_issued(req); | |
ba5290cc | 2231 | if (s->in_async) |
def596e9 JA |
2232 | mutex_unlock(&ctx->uring_lock); |
2233 | } | |
2234 | ||
2235 | return 0; | |
2b188cc1 JA |
2236 | } |
2237 | ||
31b51510 JA |
2238 | static struct async_list *io_async_list_from_sqe(struct io_ring_ctx *ctx, |
2239 | const struct io_uring_sqe *sqe) | |
2240 | { | |
2241 | switch (sqe->opcode) { | |
2242 | case IORING_OP_READV: | |
2243 | case IORING_OP_READ_FIXED: | |
2244 | return &ctx->pending_async[READ]; | |
2245 | case IORING_OP_WRITEV: | |
2246 | case IORING_OP_WRITE_FIXED: | |
2247 | return &ctx->pending_async[WRITE]; | |
2248 | default: | |
2249 | return NULL; | |
2250 | } | |
2251 | } | |
2252 | ||
edafccee JA |
2253 | static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe) |
2254 | { | |
2255 | u8 opcode = READ_ONCE(sqe->opcode); | |
2256 | ||
2257 | return !(opcode == IORING_OP_READ_FIXED || | |
2258 | opcode == IORING_OP_WRITE_FIXED); | |
2259 | } | |
2260 | ||
2b188cc1 JA |
2261 | static void io_sq_wq_submit_work(struct work_struct *work) |
2262 | { | |
2263 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
2b188cc1 | 2264 | struct io_ring_ctx *ctx = req->ctx; |
31b51510 JA |
2265 | struct mm_struct *cur_mm = NULL; |
2266 | struct async_list *async_list; | |
2267 | LIST_HEAD(req_list); | |
edafccee | 2268 | mm_segment_t old_fs; |
2b188cc1 JA |
2269 | int ret; |
2270 | ||
31b51510 JA |
2271 | async_list = io_async_list_from_sqe(ctx, req->submit.sqe); |
2272 | restart: | |
2273 | do { | |
2274 | struct sqe_submit *s = &req->submit; | |
2275 | const struct io_uring_sqe *sqe = s->sqe; | |
d0ee8791 | 2276 | unsigned int flags = req->flags; |
ba816ad6 | 2277 | struct io_kiocb *nxt = NULL; |
2b188cc1 | 2278 | |
8449eeda | 2279 | /* Ensure we clear previously set non-block flag */ |
31b51510 JA |
2280 | req->rw.ki_flags &= ~IOCB_NOWAIT; |
2281 | ||
2282 | ret = 0; | |
2283 | if (io_sqe_needs_user(sqe) && !cur_mm) { | |
2284 | if (!mmget_not_zero(ctx->sqo_mm)) { | |
2285 | ret = -EFAULT; | |
2286 | } else { | |
2287 | cur_mm = ctx->sqo_mm; | |
2288 | use_mm(cur_mm); | |
2289 | old_fs = get_fs(); | |
2290 | set_fs(USER_DS); | |
2291 | } | |
2292 | } | |
2293 | ||
2294 | if (!ret) { | |
2295 | s->has_user = cur_mm != NULL; | |
ba5290cc | 2296 | s->in_async = true; |
31b51510 | 2297 | do { |
ba816ad6 | 2298 | ret = __io_submit_sqe(ctx, req, s, &nxt, false); |
31b51510 JA |
2299 | /* |
2300 | * We can get EAGAIN for polled IO even though | |
2301 | * we're forcing a sync submission from here, | |
2302 | * since we can't wait for request slots on the | |
2303 | * block side. | |
2304 | */ | |
2305 | if (ret != -EAGAIN) | |
2306 | break; | |
2307 | cond_resched(); | |
2308 | } while (1); | |
2309 | } | |
817869d2 JA |
2310 | |
2311 | /* drop submission reference */ | |
ba816ad6 | 2312 | io_put_req(req, NULL); |
817869d2 | 2313 | |
31b51510 | 2314 | if (ret) { |
c71ffb67 | 2315 | io_cqring_add_event(ctx, sqe->user_data, ret); |
ba816ad6 | 2316 | io_put_req(req, NULL); |
31b51510 JA |
2317 | } |
2318 | ||
2319 | /* async context always use a copy of the sqe */ | |
2320 | kfree(sqe); | |
2321 | ||
ba816ad6 JA |
2322 | /* if a dependent link is ready, do that as the next one */ |
2323 | if (!ret && nxt) { | |
2324 | req = nxt; | |
2325 | continue; | |
2326 | } | |
2327 | ||
f7b76ac9 | 2328 | /* req from defer and link list needn't decrease async cnt */ |
d0ee8791 | 2329 | if (flags & (REQ_F_IO_DRAINED | REQ_F_LINK_DONE)) |
f7b76ac9 ZL |
2330 | goto out; |
2331 | ||
31b51510 JA |
2332 | if (!async_list) |
2333 | break; | |
2334 | if (!list_empty(&req_list)) { | |
2335 | req = list_first_entry(&req_list, struct io_kiocb, | |
2336 | list); | |
2337 | list_del(&req->list); | |
2338 | continue; | |
2339 | } | |
2340 | if (list_empty(&async_list->list)) | |
2341 | break; | |
2342 | ||
2343 | req = NULL; | |
2344 | spin_lock(&async_list->lock); | |
2345 | if (list_empty(&async_list->list)) { | |
2346 | spin_unlock(&async_list->lock); | |
2347 | break; | |
2348 | } | |
2349 | list_splice_init(&async_list->list, &req_list); | |
2350 | spin_unlock(&async_list->lock); | |
2351 | ||
2352 | req = list_first_entry(&req_list, struct io_kiocb, list); | |
2353 | list_del(&req->list); | |
2354 | } while (req); | |
edafccee JA |
2355 | |
2356 | /* | |
31b51510 JA |
2357 | * Rare case of racing with a submitter. If we find the count has |
2358 | * dropped to zero AND we have pending work items, then restart | |
2359 | * the processing. This is a tiny race window. | |
edafccee | 2360 | */ |
31b51510 JA |
2361 | if (async_list) { |
2362 | ret = atomic_dec_return(&async_list->cnt); | |
2363 | while (!ret && !list_empty(&async_list->list)) { | |
2364 | spin_lock(&async_list->lock); | |
2365 | atomic_inc(&async_list->cnt); | |
2366 | list_splice_init(&async_list->list, &req_list); | |
2367 | spin_unlock(&async_list->lock); | |
2368 | ||
2369 | if (!list_empty(&req_list)) { | |
2370 | req = list_first_entry(&req_list, | |
2371 | struct io_kiocb, list); | |
2372 | list_del(&req->list); | |
2373 | goto restart; | |
2374 | } | |
2375 | ret = atomic_dec_return(&async_list->cnt); | |
edafccee | 2376 | } |
edafccee | 2377 | } |
2b188cc1 | 2378 | |
f7b76ac9 | 2379 | out: |
31b51510 | 2380 | if (cur_mm) { |
edafccee | 2381 | set_fs(old_fs); |
31b51510 JA |
2382 | unuse_mm(cur_mm); |
2383 | mmput(cur_mm); | |
2b188cc1 | 2384 | } |
31b51510 | 2385 | } |
2b188cc1 | 2386 | |
31b51510 JA |
2387 | /* |
2388 | * See if we can piggy back onto previously submitted work, that is still | |
2389 | * running. We currently only allow this if the new request is sequential | |
2390 | * to the previous one we punted. | |
2391 | */ | |
2392 | static bool io_add_to_prev_work(struct async_list *list, struct io_kiocb *req) | |
2393 | { | |
6d5d5ac5 | 2394 | bool ret; |
31b51510 JA |
2395 | |
2396 | if (!list) | |
2397 | return false; | |
2398 | if (!(req->flags & REQ_F_SEQ_PREV)) | |
2399 | return false; | |
2400 | if (!atomic_read(&list->cnt)) | |
2401 | return false; | |
2402 | ||
2403 | ret = true; | |
2404 | spin_lock(&list->lock); | |
2405 | list_add_tail(&req->list, &list->list); | |
c0e48f9d ZL |
2406 | /* |
2407 | * Ensure we see a simultaneous modification from io_sq_wq_submit_work() | |
2408 | */ | |
2409 | smp_mb(); | |
31b51510 JA |
2410 | if (!atomic_read(&list->cnt)) { |
2411 | list_del_init(&req->list); | |
2412 | ret = false; | |
2413 | } | |
2414 | spin_unlock(&list->lock); | |
c826bd7a DD |
2415 | |
2416 | trace_io_uring_add_to_prev(req, ret); | |
31b51510 | 2417 | return ret; |
2b188cc1 JA |
2418 | } |
2419 | ||
09bb8394 JA |
2420 | static bool io_op_needs_file(const struct io_uring_sqe *sqe) |
2421 | { | |
2422 | int op = READ_ONCE(sqe->opcode); | |
2423 | ||
2424 | switch (op) { | |
2425 | case IORING_OP_NOP: | |
2426 | case IORING_OP_POLL_REMOVE: | |
2427 | return false; | |
2428 | default: | |
2429 | return true; | |
2430 | } | |
2431 | } | |
2432 | ||
2433 | static int io_req_set_file(struct io_ring_ctx *ctx, const struct sqe_submit *s, | |
2434 | struct io_submit_state *state, struct io_kiocb *req) | |
2435 | { | |
2436 | unsigned flags; | |
2437 | int fd; | |
2438 | ||
2439 | flags = READ_ONCE(s->sqe->flags); | |
2440 | fd = READ_ONCE(s->sqe->fd); | |
2441 | ||
4fe2c963 | 2442 | if (flags & IOSQE_IO_DRAIN) |
de0617e4 | 2443 | req->flags |= REQ_F_IO_DRAIN; |
4fe2c963 JL |
2444 | /* |
2445 | * All io need record the previous position, if LINK vs DARIN, | |
2446 | * it can be used to mark the position of the first IO in the | |
2447 | * link list. | |
2448 | */ | |
2449 | req->sequence = s->sequence; | |
de0617e4 | 2450 | |
60c112b0 | 2451 | if (!io_op_needs_file(s->sqe)) |
09bb8394 | 2452 | return 0; |
09bb8394 JA |
2453 | |
2454 | if (flags & IOSQE_FIXED_FILE) { | |
2455 | if (unlikely(!ctx->user_files || | |
2456 | (unsigned) fd >= ctx->nr_user_files)) | |
2457 | return -EBADF; | |
08a45173 JA |
2458 | if (!ctx->user_files[fd]) |
2459 | return -EBADF; | |
09bb8394 JA |
2460 | req->file = ctx->user_files[fd]; |
2461 | req->flags |= REQ_F_FIXED_FILE; | |
2462 | } else { | |
2463 | if (s->needs_fixed_file) | |
2464 | return -EBADF; | |
c826bd7a | 2465 | trace_io_uring_file_get(ctx, fd); |
09bb8394 JA |
2466 | req->file = io_file_get(state, fd); |
2467 | if (unlikely(!req->file)) | |
2468 | return -EBADF; | |
2469 | } | |
2470 | ||
2471 | return 0; | |
2472 | } | |
2473 | ||
4fe2c963 | 2474 | static int __io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
bc808bce | 2475 | struct sqe_submit *s) |
2b188cc1 | 2476 | { |
e0c5c576 | 2477 | int ret; |
2b188cc1 | 2478 | |
ba816ad6 | 2479 | ret = __io_submit_sqe(ctx, req, s, NULL, true); |
491381ce JA |
2480 | |
2481 | /* | |
2482 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
2483 | * doesn't support non-blocking read/write attempts | |
2484 | */ | |
2485 | if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) || | |
2486 | (req->flags & REQ_F_MUST_PUNT))) { | |
2b188cc1 JA |
2487 | struct io_uring_sqe *sqe_copy; |
2488 | ||
954dab19 | 2489 | sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL); |
2b188cc1 | 2490 | if (sqe_copy) { |
31b51510 JA |
2491 | struct async_list *list; |
2492 | ||
2b188cc1 | 2493 | s->sqe = sqe_copy; |
2b188cc1 | 2494 | memcpy(&req->submit, s, sizeof(*s)); |
31b51510 JA |
2495 | list = io_async_list_from_sqe(ctx, s->sqe); |
2496 | if (!io_add_to_prev_work(list, req)) { | |
2497 | if (list) | |
2498 | atomic_inc(&list->cnt); | |
2499 | INIT_WORK(&req->work, io_sq_wq_submit_work); | |
18d9be1a | 2500 | io_queue_async_work(ctx, req); |
31b51510 | 2501 | } |
e65ef56d JA |
2502 | |
2503 | /* | |
2504 | * Queued up for async execution, worker will release | |
9e645e11 | 2505 | * submit reference when the iocb is actually submitted. |
e65ef56d JA |
2506 | */ |
2507 | return 0; | |
2b188cc1 JA |
2508 | } |
2509 | } | |
e65ef56d JA |
2510 | |
2511 | /* drop submission reference */ | |
ba816ad6 | 2512 | io_put_req(req, NULL); |
e65ef56d JA |
2513 | |
2514 | /* and drop final reference, if we failed */ | |
9e645e11 JA |
2515 | if (ret) { |
2516 | io_cqring_add_event(ctx, req->user_data, ret); | |
2517 | if (req->flags & REQ_F_LINK) | |
2518 | req->flags |= REQ_F_FAIL_LINK; | |
ba816ad6 | 2519 | io_put_req(req, NULL); |
9e645e11 | 2520 | } |
2b188cc1 JA |
2521 | |
2522 | return ret; | |
2523 | } | |
2524 | ||
4fe2c963 | 2525 | static int io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
bc808bce | 2526 | struct sqe_submit *s) |
4fe2c963 JL |
2527 | { |
2528 | int ret; | |
2529 | ||
2530 | ret = io_req_defer(ctx, req, s->sqe); | |
2531 | if (ret) { | |
2532 | if (ret != -EIOCBQUEUED) { | |
ba816ad6 | 2533 | io_free_req(req, NULL); |
4fe2c963 JL |
2534 | io_cqring_add_event(ctx, s->sqe->user_data, ret); |
2535 | } | |
2536 | return 0; | |
2537 | } | |
2538 | ||
bc808bce | 2539 | return __io_queue_sqe(ctx, req, s); |
4fe2c963 JL |
2540 | } |
2541 | ||
2542 | static int io_queue_link_head(struct io_ring_ctx *ctx, struct io_kiocb *req, | |
bc808bce | 2543 | struct sqe_submit *s, struct io_kiocb *shadow) |
4fe2c963 JL |
2544 | { |
2545 | int ret; | |
2546 | int need_submit = false; | |
2547 | ||
2548 | if (!shadow) | |
bc808bce | 2549 | return io_queue_sqe(ctx, req, s); |
4fe2c963 JL |
2550 | |
2551 | /* | |
2552 | * Mark the first IO in link list as DRAIN, let all the following | |
2553 | * IOs enter the defer list. all IO needs to be completed before link | |
2554 | * list. | |
2555 | */ | |
2556 | req->flags |= REQ_F_IO_DRAIN; | |
2557 | ret = io_req_defer(ctx, req, s->sqe); | |
2558 | if (ret) { | |
2559 | if (ret != -EIOCBQUEUED) { | |
ba816ad6 | 2560 | io_free_req(req, NULL); |
7b20238d | 2561 | __io_free_req(shadow); |
4fe2c963 JL |
2562 | io_cqring_add_event(ctx, s->sqe->user_data, ret); |
2563 | return 0; | |
2564 | } | |
2565 | } else { | |
2566 | /* | |
2567 | * If ret == 0 means that all IOs in front of link io are | |
2568 | * running done. let's queue link head. | |
2569 | */ | |
2570 | need_submit = true; | |
2571 | } | |
2572 | ||
2573 | /* Insert shadow req to defer_list, blocking next IOs */ | |
2574 | spin_lock_irq(&ctx->completion_lock); | |
c826bd7a | 2575 | trace_io_uring_defer(ctx, shadow, true); |
4fe2c963 JL |
2576 | list_add_tail(&shadow->list, &ctx->defer_list); |
2577 | spin_unlock_irq(&ctx->completion_lock); | |
2578 | ||
2579 | if (need_submit) | |
bc808bce | 2580 | return __io_queue_sqe(ctx, req, s); |
4fe2c963 JL |
2581 | |
2582 | return 0; | |
2583 | } | |
2584 | ||
9e645e11 JA |
2585 | #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK) |
2586 | ||
2587 | static void io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s, | |
bc808bce | 2588 | struct io_submit_state *state, struct io_kiocb **link) |
9e645e11 JA |
2589 | { |
2590 | struct io_uring_sqe *sqe_copy; | |
2591 | struct io_kiocb *req; | |
2592 | int ret; | |
2593 | ||
2594 | /* enforce forwards compatibility on users */ | |
2595 | if (unlikely(s->sqe->flags & ~SQE_VALID_FLAGS)) { | |
2596 | ret = -EINVAL; | |
2597 | goto err; | |
2598 | } | |
2599 | ||
2600 | req = io_get_req(ctx, state); | |
2601 | if (unlikely(!req)) { | |
2602 | ret = -EAGAIN; | |
2603 | goto err; | |
2604 | } | |
2605 | ||
2606 | ret = io_req_set_file(ctx, s, state, req); | |
2607 | if (unlikely(ret)) { | |
2608 | err_req: | |
ba816ad6 | 2609 | io_free_req(req, NULL); |
9e645e11 JA |
2610 | err: |
2611 | io_cqring_add_event(ctx, s->sqe->user_data, ret); | |
2612 | return; | |
2613 | } | |
2614 | ||
84d55dc5 PB |
2615 | req->user_data = s->sqe->user_data; |
2616 | ||
9e645e11 JA |
2617 | /* |
2618 | * If we already have a head request, queue this one for async | |
2619 | * submittal once the head completes. If we don't have a head but | |
2620 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
2621 | * submitted sync once the chain is complete. If none of those | |
2622 | * conditions are true (normal request), then just queue it. | |
2623 | */ | |
2624 | if (*link) { | |
2625 | struct io_kiocb *prev = *link; | |
2626 | ||
2627 | sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL); | |
2628 | if (!sqe_copy) { | |
2629 | ret = -EAGAIN; | |
2630 | goto err_req; | |
2631 | } | |
2632 | ||
2633 | s->sqe = sqe_copy; | |
2634 | memcpy(&req->submit, s, sizeof(*s)); | |
c826bd7a | 2635 | trace_io_uring_link(ctx, req, prev); |
9e645e11 JA |
2636 | list_add_tail(&req->list, &prev->link_list); |
2637 | } else if (s->sqe->flags & IOSQE_IO_LINK) { | |
2638 | req->flags |= REQ_F_LINK; | |
2639 | ||
2640 | memcpy(&req->submit, s, sizeof(*s)); | |
2641 | INIT_LIST_HEAD(&req->link_list); | |
2642 | *link = req; | |
2643 | } else { | |
bc808bce | 2644 | io_queue_sqe(ctx, req, s); |
9e645e11 JA |
2645 | } |
2646 | } | |
2647 | ||
9a56a232 JA |
2648 | /* |
2649 | * Batched submission is done, ensure local IO is flushed out. | |
2650 | */ | |
2651 | static void io_submit_state_end(struct io_submit_state *state) | |
2652 | { | |
2653 | blk_finish_plug(&state->plug); | |
3d6770fb | 2654 | io_file_put(state); |
2579f913 JA |
2655 | if (state->free_reqs) |
2656 | kmem_cache_free_bulk(req_cachep, state->free_reqs, | |
2657 | &state->reqs[state->cur_req]); | |
9a56a232 JA |
2658 | } |
2659 | ||
2660 | /* | |
2661 | * Start submission side cache. | |
2662 | */ | |
2663 | static void io_submit_state_start(struct io_submit_state *state, | |
2664 | struct io_ring_ctx *ctx, unsigned max_ios) | |
2665 | { | |
2666 | blk_start_plug(&state->plug); | |
2579f913 | 2667 | state->free_reqs = 0; |
9a56a232 JA |
2668 | state->file = NULL; |
2669 | state->ios_left = max_ios; | |
2670 | } | |
2671 | ||
2b188cc1 JA |
2672 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
2673 | { | |
75b28aff | 2674 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 2675 | |
75b28aff | 2676 | if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) { |
2b188cc1 JA |
2677 | /* |
2678 | * Ensure any loads from the SQEs are done at this point, | |
2679 | * since once we write the new head, the application could | |
2680 | * write new data to them. | |
2681 | */ | |
75b28aff | 2682 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); |
2b188cc1 JA |
2683 | } |
2684 | } | |
2685 | ||
2b188cc1 JA |
2686 | /* |
2687 | * Fetch an sqe, if one is available. Note that s->sqe will point to memory | |
2688 | * that is mapped by userspace. This means that care needs to be taken to | |
2689 | * ensure that reads are stable, as we cannot rely on userspace always | |
2690 | * being a good citizen. If members of the sqe are validated and then later | |
2691 | * used, it's important that those reads are done through READ_ONCE() to | |
2692 | * prevent a re-load down the line. | |
2693 | */ | |
2694 | static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s) | |
2695 | { | |
75b28aff HV |
2696 | struct io_rings *rings = ctx->rings; |
2697 | u32 *sq_array = ctx->sq_array; | |
2b188cc1 JA |
2698 | unsigned head; |
2699 | ||
2700 | /* | |
2701 | * The cached sq head (or cq tail) serves two purposes: | |
2702 | * | |
2703 | * 1) allows us to batch the cost of updating the user visible | |
2704 | * head updates. | |
2705 | * 2) allows the kernel side to track the head on its own, even | |
2706 | * though the application is the one updating it. | |
2707 | */ | |
2708 | head = ctx->cached_sq_head; | |
e523a29c | 2709 | /* make sure SQ entry isn't read before tail */ |
75b28aff | 2710 | if (head == smp_load_acquire(&rings->sq.tail)) |
2b188cc1 JA |
2711 | return false; |
2712 | ||
75b28aff | 2713 | head = READ_ONCE(sq_array[head & ctx->sq_mask]); |
2b188cc1 | 2714 | if (head < ctx->sq_entries) { |
2b188cc1 | 2715 | s->sqe = &ctx->sq_sqes[head]; |
8776f3fa | 2716 | s->sequence = ctx->cached_sq_head; |
2b188cc1 JA |
2717 | ctx->cached_sq_head++; |
2718 | return true; | |
2719 | } | |
2720 | ||
2721 | /* drop invalid entries */ | |
2722 | ctx->cached_sq_head++; | |
498ccd9e JA |
2723 | ctx->cached_sq_dropped++; |
2724 | WRITE_ONCE(rings->sq_dropped, ctx->cached_sq_dropped); | |
2b188cc1 JA |
2725 | return false; |
2726 | } | |
2727 | ||
fb5ccc98 | 2728 | static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr, |
95a1b3ff | 2729 | struct mm_struct **mm) |
6c271ce2 JA |
2730 | { |
2731 | struct io_submit_state state, *statep = NULL; | |
9e645e11 | 2732 | struct io_kiocb *link = NULL; |
4fe2c963 | 2733 | struct io_kiocb *shadow_req = NULL; |
9e645e11 JA |
2734 | bool prev_was_link = false; |
2735 | int i, submitted = 0; | |
95a1b3ff | 2736 | bool mm_fault = false; |
6c271ce2 JA |
2737 | |
2738 | if (nr > IO_PLUG_THRESHOLD) { | |
2739 | io_submit_state_start(&state, ctx, nr); | |
2740 | statep = &state; | |
2741 | } | |
2742 | ||
2743 | for (i = 0; i < nr; i++) { | |
fb5ccc98 PB |
2744 | struct sqe_submit s; |
2745 | ||
2746 | if (!io_get_sqring(ctx, &s)) | |
2747 | break; | |
2748 | ||
95a1b3ff PB |
2749 | if (io_sqe_needs_user(s.sqe) && !*mm) { |
2750 | mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm); | |
2751 | if (!mm_fault) { | |
2752 | use_mm(ctx->sqo_mm); | |
2753 | *mm = ctx->sqo_mm; | |
2754 | } | |
2755 | } | |
2756 | ||
9e645e11 JA |
2757 | /* |
2758 | * If previous wasn't linked and we have a linked command, | |
2759 | * that's the end of the chain. Submit the previous link. | |
2760 | */ | |
2761 | if (!prev_was_link && link) { | |
bc808bce | 2762 | io_queue_link_head(ctx, link, &link->submit, shadow_req); |
9e645e11 | 2763 | link = NULL; |
5f5ad9ce | 2764 | shadow_req = NULL; |
9e645e11 | 2765 | } |
fb5ccc98 | 2766 | prev_was_link = (s.sqe->flags & IOSQE_IO_LINK) != 0; |
9e645e11 | 2767 | |
fb5ccc98 | 2768 | if (link && (s.sqe->flags & IOSQE_IO_DRAIN)) { |
4fe2c963 JL |
2769 | if (!shadow_req) { |
2770 | shadow_req = io_get_req(ctx, NULL); | |
a1041c27 JL |
2771 | if (unlikely(!shadow_req)) |
2772 | goto out; | |
4fe2c963 JL |
2773 | shadow_req->flags |= (REQ_F_IO_DRAIN | REQ_F_SHADOW_DRAIN); |
2774 | refcount_dec(&shadow_req->refs); | |
2775 | } | |
fb5ccc98 | 2776 | shadow_req->sequence = s.sequence; |
4fe2c963 JL |
2777 | } |
2778 | ||
a1041c27 | 2779 | out: |
95a1b3ff PB |
2780 | s.has_user = *mm != NULL; |
2781 | s.in_async = true; | |
2782 | s.needs_fixed_file = true; | |
2783 | trace_io_uring_submit_sqe(ctx, true, true); | |
2784 | io_submit_sqe(ctx, &s, statep, &link); | |
2785 | submitted++; | |
6c271ce2 JA |
2786 | } |
2787 | ||
9e645e11 | 2788 | if (link) |
bc808bce | 2789 | io_queue_link_head(ctx, link, &link->submit, shadow_req); |
6c271ce2 JA |
2790 | if (statep) |
2791 | io_submit_state_end(&state); | |
2792 | ||
2793 | return submitted; | |
2794 | } | |
2795 | ||
2796 | static int io_sq_thread(void *data) | |
2797 | { | |
6c271ce2 JA |
2798 | struct io_ring_ctx *ctx = data; |
2799 | struct mm_struct *cur_mm = NULL; | |
2800 | mm_segment_t old_fs; | |
2801 | DEFINE_WAIT(wait); | |
2802 | unsigned inflight; | |
2803 | unsigned long timeout; | |
2804 | ||
a4c0b3de JL |
2805 | complete(&ctx->sqo_thread_started); |
2806 | ||
6c271ce2 JA |
2807 | old_fs = get_fs(); |
2808 | set_fs(USER_DS); | |
2809 | ||
2810 | timeout = inflight = 0; | |
2bbcd6d3 | 2811 | while (!kthread_should_park()) { |
fb5ccc98 | 2812 | unsigned int to_submit; |
6c271ce2 JA |
2813 | |
2814 | if (inflight) { | |
2815 | unsigned nr_events = 0; | |
2816 | ||
2817 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
2b2ed975 JA |
2818 | /* |
2819 | * inflight is the count of the maximum possible | |
2820 | * entries we submitted, but it can be smaller | |
2821 | * if we dropped some of them. If we don't have | |
2822 | * poll entries available, then we know that we | |
2823 | * have nothing left to poll for. Reset the | |
2824 | * inflight count to zero in that case. | |
2825 | */ | |
2826 | mutex_lock(&ctx->uring_lock); | |
2827 | if (!list_empty(&ctx->poll_list)) | |
2828 | __io_iopoll_check(ctx, &nr_events, 0); | |
2829 | else | |
2830 | inflight = 0; | |
2831 | mutex_unlock(&ctx->uring_lock); | |
6c271ce2 JA |
2832 | } else { |
2833 | /* | |
2834 | * Normal IO, just pretend everything completed. | |
2835 | * We don't have to poll completions for that. | |
2836 | */ | |
2837 | nr_events = inflight; | |
2838 | } | |
2839 | ||
2840 | inflight -= nr_events; | |
2841 | if (!inflight) | |
2842 | timeout = jiffies + ctx->sq_thread_idle; | |
2843 | } | |
2844 | ||
fb5ccc98 PB |
2845 | to_submit = io_sqring_entries(ctx); |
2846 | if (!to_submit) { | |
6c271ce2 JA |
2847 | /* |
2848 | * We're polling. If we're within the defined idle | |
2849 | * period, then let us spin without work before going | |
2850 | * to sleep. | |
2851 | */ | |
2852 | if (inflight || !time_after(jiffies, timeout)) { | |
9831a90c | 2853 | cond_resched(); |
6c271ce2 JA |
2854 | continue; |
2855 | } | |
2856 | ||
2857 | /* | |
2858 | * Drop cur_mm before scheduling, we can't hold it for | |
2859 | * long periods (or over schedule()). Do this before | |
2860 | * adding ourselves to the waitqueue, as the unuse/drop | |
2861 | * may sleep. | |
2862 | */ | |
2863 | if (cur_mm) { | |
2864 | unuse_mm(cur_mm); | |
2865 | mmput(cur_mm); | |
2866 | cur_mm = NULL; | |
2867 | } | |
2868 | ||
2869 | prepare_to_wait(&ctx->sqo_wait, &wait, | |
2870 | TASK_INTERRUPTIBLE); | |
2871 | ||
2872 | /* Tell userspace we may need a wakeup call */ | |
75b28aff | 2873 | ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP; |
0d7bae69 SB |
2874 | /* make sure to read SQ tail after writing flags */ |
2875 | smp_mb(); | |
6c271ce2 | 2876 | |
fb5ccc98 PB |
2877 | to_submit = io_sqring_entries(ctx); |
2878 | if (!to_submit) { | |
2bbcd6d3 | 2879 | if (kthread_should_park()) { |
6c271ce2 JA |
2880 | finish_wait(&ctx->sqo_wait, &wait); |
2881 | break; | |
2882 | } | |
2883 | if (signal_pending(current)) | |
2884 | flush_signals(current); | |
2885 | schedule(); | |
2886 | finish_wait(&ctx->sqo_wait, &wait); | |
2887 | ||
75b28aff | 2888 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
2889 | continue; |
2890 | } | |
2891 | finish_wait(&ctx->sqo_wait, &wait); | |
2892 | ||
75b28aff | 2893 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
2894 | } |
2895 | ||
fb5ccc98 | 2896 | to_submit = min(to_submit, ctx->sq_entries); |
95a1b3ff | 2897 | inflight += io_submit_sqes(ctx, to_submit, &cur_mm); |
6c271ce2 JA |
2898 | |
2899 | /* Commit SQ ring head once we've consumed all SQEs */ | |
2900 | io_commit_sqring(ctx); | |
2901 | } | |
2902 | ||
2903 | set_fs(old_fs); | |
2904 | if (cur_mm) { | |
2905 | unuse_mm(cur_mm); | |
2906 | mmput(cur_mm); | |
2907 | } | |
06058632 | 2908 | |
2bbcd6d3 | 2909 | kthread_parkme(); |
06058632 | 2910 | |
6c271ce2 JA |
2911 | return 0; |
2912 | } | |
2913 | ||
bc808bce | 2914 | static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit) |
2b188cc1 | 2915 | { |
9a56a232 | 2916 | struct io_submit_state state, *statep = NULL; |
9e645e11 | 2917 | struct io_kiocb *link = NULL; |
4fe2c963 | 2918 | struct io_kiocb *shadow_req = NULL; |
9e645e11 | 2919 | bool prev_was_link = false; |
5c8b0b54 | 2920 | int i, submit = 0; |
2b188cc1 | 2921 | |
9a56a232 JA |
2922 | if (to_submit > IO_PLUG_THRESHOLD) { |
2923 | io_submit_state_start(&state, ctx, to_submit); | |
2924 | statep = &state; | |
2925 | } | |
2b188cc1 JA |
2926 | |
2927 | for (i = 0; i < to_submit; i++) { | |
2928 | struct sqe_submit s; | |
2929 | ||
2930 | if (!io_get_sqring(ctx, &s)) | |
2931 | break; | |
2932 | ||
9e645e11 JA |
2933 | /* |
2934 | * If previous wasn't linked and we have a linked command, | |
2935 | * that's the end of the chain. Submit the previous link. | |
2936 | */ | |
2937 | if (!prev_was_link && link) { | |
bc808bce | 2938 | io_queue_link_head(ctx, link, &link->submit, shadow_req); |
9e645e11 | 2939 | link = NULL; |
5f5ad9ce | 2940 | shadow_req = NULL; |
9e645e11 JA |
2941 | } |
2942 | prev_was_link = (s.sqe->flags & IOSQE_IO_LINK) != 0; | |
2943 | ||
4fe2c963 JL |
2944 | if (link && (s.sqe->flags & IOSQE_IO_DRAIN)) { |
2945 | if (!shadow_req) { | |
2946 | shadow_req = io_get_req(ctx, NULL); | |
a1041c27 JL |
2947 | if (unlikely(!shadow_req)) |
2948 | goto out; | |
4fe2c963 JL |
2949 | shadow_req->flags |= (REQ_F_IO_DRAIN | REQ_F_SHADOW_DRAIN); |
2950 | refcount_dec(&shadow_req->refs); | |
2951 | } | |
2952 | shadow_req->sequence = s.sequence; | |
2953 | } | |
2954 | ||
a1041c27 | 2955 | out: |
2b188cc1 | 2956 | s.has_user = true; |
ba5290cc | 2957 | s.in_async = false; |
6c271ce2 | 2958 | s.needs_fixed_file = false; |
5c8b0b54 | 2959 | submit++; |
c826bd7a | 2960 | trace_io_uring_submit_sqe(ctx, true, false); |
bc808bce | 2961 | io_submit_sqe(ctx, &s, statep, &link); |
2b188cc1 | 2962 | } |
2b188cc1 | 2963 | |
9e645e11 | 2964 | if (link) |
bc808bce | 2965 | io_queue_link_head(ctx, link, &link->submit, shadow_req); |
9a56a232 JA |
2966 | if (statep) |
2967 | io_submit_state_end(statep); | |
2b188cc1 | 2968 | |
935d1e45 PB |
2969 | io_commit_sqring(ctx); |
2970 | ||
5c8b0b54 | 2971 | return submit; |
2b188cc1 JA |
2972 | } |
2973 | ||
bda52162 JA |
2974 | struct io_wait_queue { |
2975 | struct wait_queue_entry wq; | |
2976 | struct io_ring_ctx *ctx; | |
2977 | unsigned to_wait; | |
2978 | unsigned nr_timeouts; | |
2979 | }; | |
2980 | ||
2981 | static inline bool io_should_wake(struct io_wait_queue *iowq) | |
2982 | { | |
2983 | struct io_ring_ctx *ctx = iowq->ctx; | |
2984 | ||
2985 | /* | |
2986 | * Wake up if we have enough events, or if a timeout occured since we | |
2987 | * started waiting. For timeouts, we always want to return to userspace, | |
2988 | * regardless of event count. | |
2989 | */ | |
2990 | return io_cqring_events(ctx->rings) >= iowq->to_wait || | |
2991 | atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; | |
2992 | } | |
2993 | ||
2994 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
2995 | int wake_flags, void *key) | |
2996 | { | |
2997 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
2998 | wq); | |
2999 | ||
3000 | if (!io_should_wake(iowq)) | |
3001 | return -1; | |
3002 | ||
3003 | return autoremove_wake_function(curr, mode, wake_flags, key); | |
3004 | } | |
3005 | ||
2b188cc1 JA |
3006 | /* |
3007 | * Wait until events become available, if we don't already have some. The | |
3008 | * application must reap them itself, as they reside on the shared cq ring. | |
3009 | */ | |
3010 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
3011 | const sigset_t __user *sig, size_t sigsz) | |
3012 | { | |
bda52162 JA |
3013 | struct io_wait_queue iowq = { |
3014 | .wq = { | |
3015 | .private = current, | |
3016 | .func = io_wake_function, | |
3017 | .entry = LIST_HEAD_INIT(iowq.wq.entry), | |
3018 | }, | |
3019 | .ctx = ctx, | |
3020 | .to_wait = min_events, | |
3021 | }; | |
75b28aff | 3022 | struct io_rings *rings = ctx->rings; |
2b188cc1 JA |
3023 | int ret; |
3024 | ||
75b28aff | 3025 | if (io_cqring_events(rings) >= min_events) |
2b188cc1 JA |
3026 | return 0; |
3027 | ||
3028 | if (sig) { | |
9e75ad5d AB |
3029 | #ifdef CONFIG_COMPAT |
3030 | if (in_compat_syscall()) | |
3031 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 3032 | sigsz); |
9e75ad5d AB |
3033 | else |
3034 | #endif | |
b772434b | 3035 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 3036 | |
2b188cc1 JA |
3037 | if (ret) |
3038 | return ret; | |
3039 | } | |
3040 | ||
bda52162 JA |
3041 | ret = 0; |
3042 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); | |
c826bd7a | 3043 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 JA |
3044 | do { |
3045 | prepare_to_wait_exclusive(&ctx->wait, &iowq.wq, | |
3046 | TASK_INTERRUPTIBLE); | |
3047 | if (io_should_wake(&iowq)) | |
3048 | break; | |
3049 | schedule(); | |
3050 | if (signal_pending(current)) { | |
3051 | ret = -ERESTARTSYS; | |
3052 | break; | |
3053 | } | |
3054 | } while (1); | |
3055 | finish_wait(&ctx->wait, &iowq.wq); | |
3056 | ||
b772434b | 3057 | restore_saved_sigmask_unless(ret == -ERESTARTSYS); |
97abc889 ON |
3058 | if (ret == -ERESTARTSYS) |
3059 | ret = -EINTR; | |
2b188cc1 | 3060 | |
75b28aff | 3061 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
3062 | } |
3063 | ||
6b06314c JA |
3064 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
3065 | { | |
3066 | #if defined(CONFIG_UNIX) | |
3067 | if (ctx->ring_sock) { | |
3068 | struct sock *sock = ctx->ring_sock->sk; | |
3069 | struct sk_buff *skb; | |
3070 | ||
3071 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
3072 | kfree_skb(skb); | |
3073 | } | |
3074 | #else | |
3075 | int i; | |
3076 | ||
3077 | for (i = 0; i < ctx->nr_user_files; i++) | |
08a45173 JA |
3078 | if (ctx->user_files[i]) |
3079 | fput(ctx->user_files[i]); | |
6b06314c JA |
3080 | #endif |
3081 | } | |
3082 | ||
3083 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) | |
3084 | { | |
3085 | if (!ctx->user_files) | |
3086 | return -ENXIO; | |
3087 | ||
3088 | __io_sqe_files_unregister(ctx); | |
3089 | kfree(ctx->user_files); | |
3090 | ctx->user_files = NULL; | |
3091 | ctx->nr_user_files = 0; | |
3092 | return 0; | |
3093 | } | |
3094 | ||
6c271ce2 JA |
3095 | static void io_sq_thread_stop(struct io_ring_ctx *ctx) |
3096 | { | |
3097 | if (ctx->sqo_thread) { | |
a4c0b3de | 3098 | wait_for_completion(&ctx->sqo_thread_started); |
2bbcd6d3 RP |
3099 | /* |
3100 | * The park is a bit of a work-around, without it we get | |
3101 | * warning spews on shutdown with SQPOLL set and affinity | |
3102 | * set to a single CPU. | |
3103 | */ | |
06058632 | 3104 | kthread_park(ctx->sqo_thread); |
6c271ce2 JA |
3105 | kthread_stop(ctx->sqo_thread); |
3106 | ctx->sqo_thread = NULL; | |
3107 | } | |
3108 | } | |
3109 | ||
6b06314c JA |
3110 | static void io_finish_async(struct io_ring_ctx *ctx) |
3111 | { | |
54a91f3b JA |
3112 | int i; |
3113 | ||
6c271ce2 JA |
3114 | io_sq_thread_stop(ctx); |
3115 | ||
54a91f3b JA |
3116 | for (i = 0; i < ARRAY_SIZE(ctx->sqo_wq); i++) { |
3117 | if (ctx->sqo_wq[i]) { | |
3118 | destroy_workqueue(ctx->sqo_wq[i]); | |
3119 | ctx->sqo_wq[i] = NULL; | |
3120 | } | |
6b06314c JA |
3121 | } |
3122 | } | |
3123 | ||
3124 | #if defined(CONFIG_UNIX) | |
3125 | static void io_destruct_skb(struct sk_buff *skb) | |
3126 | { | |
3127 | struct io_ring_ctx *ctx = skb->sk->sk_user_data; | |
8a997340 JA |
3128 | int i; |
3129 | ||
3130 | for (i = 0; i < ARRAY_SIZE(ctx->sqo_wq); i++) | |
3131 | if (ctx->sqo_wq[i]) | |
3132 | flush_workqueue(ctx->sqo_wq[i]); | |
6b06314c | 3133 | |
6b06314c JA |
3134 | unix_destruct_scm(skb); |
3135 | } | |
3136 | ||
3137 | /* | |
3138 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
3139 | * the io_uring can be safely unregistered on process exit, even if we have | |
3140 | * loops in the file referencing. | |
3141 | */ | |
3142 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
3143 | { | |
3144 | struct sock *sk = ctx->ring_sock->sk; | |
3145 | struct scm_fp_list *fpl; | |
3146 | struct sk_buff *skb; | |
08a45173 | 3147 | int i, nr_files; |
6b06314c JA |
3148 | |
3149 | if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) { | |
3150 | unsigned long inflight = ctx->user->unix_inflight + nr; | |
3151 | ||
3152 | if (inflight > task_rlimit(current, RLIMIT_NOFILE)) | |
3153 | return -EMFILE; | |
3154 | } | |
3155 | ||
3156 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); | |
3157 | if (!fpl) | |
3158 | return -ENOMEM; | |
3159 | ||
3160 | skb = alloc_skb(0, GFP_KERNEL); | |
3161 | if (!skb) { | |
3162 | kfree(fpl); | |
3163 | return -ENOMEM; | |
3164 | } | |
3165 | ||
3166 | skb->sk = sk; | |
6b06314c | 3167 | |
08a45173 | 3168 | nr_files = 0; |
6b06314c JA |
3169 | fpl->user = get_uid(ctx->user); |
3170 | for (i = 0; i < nr; i++) { | |
08a45173 JA |
3171 | if (!ctx->user_files[i + offset]) |
3172 | continue; | |
3173 | fpl->fp[nr_files] = get_file(ctx->user_files[i + offset]); | |
3174 | unix_inflight(fpl->user, fpl->fp[nr_files]); | |
3175 | nr_files++; | |
6b06314c JA |
3176 | } |
3177 | ||
08a45173 JA |
3178 | if (nr_files) { |
3179 | fpl->max = SCM_MAX_FD; | |
3180 | fpl->count = nr_files; | |
3181 | UNIXCB(skb).fp = fpl; | |
3182 | skb->destructor = io_destruct_skb; | |
3183 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); | |
3184 | skb_queue_head(&sk->sk_receive_queue, skb); | |
6b06314c | 3185 | |
08a45173 JA |
3186 | for (i = 0; i < nr_files; i++) |
3187 | fput(fpl->fp[i]); | |
3188 | } else { | |
3189 | kfree_skb(skb); | |
3190 | kfree(fpl); | |
3191 | } | |
6b06314c JA |
3192 | |
3193 | return 0; | |
3194 | } | |
3195 | ||
3196 | /* | |
3197 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
3198 | * causes regular reference counting to break down. We rely on the UNIX | |
3199 | * garbage collection to take care of this problem for us. | |
3200 | */ | |
3201 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
3202 | { | |
3203 | unsigned left, total; | |
3204 | int ret = 0; | |
3205 | ||
3206 | total = 0; | |
3207 | left = ctx->nr_user_files; | |
3208 | while (left) { | |
3209 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
3210 | |
3211 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
3212 | if (ret) | |
3213 | break; | |
3214 | left -= this_files; | |
3215 | total += this_files; | |
3216 | } | |
3217 | ||
3218 | if (!ret) | |
3219 | return 0; | |
3220 | ||
3221 | while (total < ctx->nr_user_files) { | |
08a45173 JA |
3222 | if (ctx->user_files[total]) |
3223 | fput(ctx->user_files[total]); | |
6b06314c JA |
3224 | total++; |
3225 | } | |
3226 | ||
3227 | return ret; | |
3228 | } | |
3229 | #else | |
3230 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
3231 | { | |
3232 | return 0; | |
3233 | } | |
3234 | #endif | |
3235 | ||
3236 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, | |
3237 | unsigned nr_args) | |
3238 | { | |
3239 | __s32 __user *fds = (__s32 __user *) arg; | |
3240 | int fd, ret = 0; | |
3241 | unsigned i; | |
3242 | ||
3243 | if (ctx->user_files) | |
3244 | return -EBUSY; | |
3245 | if (!nr_args) | |
3246 | return -EINVAL; | |
3247 | if (nr_args > IORING_MAX_FIXED_FILES) | |
3248 | return -EMFILE; | |
3249 | ||
3250 | ctx->user_files = kcalloc(nr_args, sizeof(struct file *), GFP_KERNEL); | |
3251 | if (!ctx->user_files) | |
3252 | return -ENOMEM; | |
3253 | ||
08a45173 | 3254 | for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { |
6b06314c JA |
3255 | ret = -EFAULT; |
3256 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) | |
3257 | break; | |
08a45173 JA |
3258 | /* allow sparse sets */ |
3259 | if (fd == -1) { | |
3260 | ret = 0; | |
3261 | continue; | |
3262 | } | |
6b06314c JA |
3263 | |
3264 | ctx->user_files[i] = fget(fd); | |
3265 | ||
3266 | ret = -EBADF; | |
3267 | if (!ctx->user_files[i]) | |
3268 | break; | |
3269 | /* | |
3270 | * Don't allow io_uring instances to be registered. If UNIX | |
3271 | * isn't enabled, then this causes a reference cycle and this | |
3272 | * instance can never get freed. If UNIX is enabled we'll | |
3273 | * handle it just fine, but there's still no point in allowing | |
3274 | * a ring fd as it doesn't support regular read/write anyway. | |
3275 | */ | |
3276 | if (ctx->user_files[i]->f_op == &io_uring_fops) { | |
3277 | fput(ctx->user_files[i]); | |
3278 | break; | |
3279 | } | |
6b06314c JA |
3280 | ret = 0; |
3281 | } | |
3282 | ||
3283 | if (ret) { | |
3284 | for (i = 0; i < ctx->nr_user_files; i++) | |
08a45173 JA |
3285 | if (ctx->user_files[i]) |
3286 | fput(ctx->user_files[i]); | |
6b06314c JA |
3287 | |
3288 | kfree(ctx->user_files); | |
25adf50f | 3289 | ctx->user_files = NULL; |
6b06314c JA |
3290 | ctx->nr_user_files = 0; |
3291 | return ret; | |
3292 | } | |
3293 | ||
3294 | ret = io_sqe_files_scm(ctx); | |
3295 | if (ret) | |
3296 | io_sqe_files_unregister(ctx); | |
3297 | ||
3298 | return ret; | |
3299 | } | |
3300 | ||
c3a31e60 JA |
3301 | static void io_sqe_file_unregister(struct io_ring_ctx *ctx, int index) |
3302 | { | |
3303 | #if defined(CONFIG_UNIX) | |
3304 | struct file *file = ctx->user_files[index]; | |
3305 | struct sock *sock = ctx->ring_sock->sk; | |
3306 | struct sk_buff_head list, *head = &sock->sk_receive_queue; | |
3307 | struct sk_buff *skb; | |
3308 | int i; | |
3309 | ||
3310 | __skb_queue_head_init(&list); | |
3311 | ||
3312 | /* | |
3313 | * Find the skb that holds this file in its SCM_RIGHTS. When found, | |
3314 | * remove this entry and rearrange the file array. | |
3315 | */ | |
3316 | skb = skb_dequeue(head); | |
3317 | while (skb) { | |
3318 | struct scm_fp_list *fp; | |
3319 | ||
3320 | fp = UNIXCB(skb).fp; | |
3321 | for (i = 0; i < fp->count; i++) { | |
3322 | int left; | |
3323 | ||
3324 | if (fp->fp[i] != file) | |
3325 | continue; | |
3326 | ||
3327 | unix_notinflight(fp->user, fp->fp[i]); | |
3328 | left = fp->count - 1 - i; | |
3329 | if (left) { | |
3330 | memmove(&fp->fp[i], &fp->fp[i + 1], | |
3331 | left * sizeof(struct file *)); | |
3332 | } | |
3333 | fp->count--; | |
3334 | if (!fp->count) { | |
3335 | kfree_skb(skb); | |
3336 | skb = NULL; | |
3337 | } else { | |
3338 | __skb_queue_tail(&list, skb); | |
3339 | } | |
3340 | fput(file); | |
3341 | file = NULL; | |
3342 | break; | |
3343 | } | |
3344 | ||
3345 | if (!file) | |
3346 | break; | |
3347 | ||
3348 | __skb_queue_tail(&list, skb); | |
3349 | ||
3350 | skb = skb_dequeue(head); | |
3351 | } | |
3352 | ||
3353 | if (skb_peek(&list)) { | |
3354 | spin_lock_irq(&head->lock); | |
3355 | while ((skb = __skb_dequeue(&list)) != NULL) | |
3356 | __skb_queue_tail(head, skb); | |
3357 | spin_unlock_irq(&head->lock); | |
3358 | } | |
3359 | #else | |
3360 | fput(ctx->user_files[index]); | |
3361 | #endif | |
3362 | } | |
3363 | ||
3364 | static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file, | |
3365 | int index) | |
3366 | { | |
3367 | #if defined(CONFIG_UNIX) | |
3368 | struct sock *sock = ctx->ring_sock->sk; | |
3369 | struct sk_buff_head *head = &sock->sk_receive_queue; | |
3370 | struct sk_buff *skb; | |
3371 | ||
3372 | /* | |
3373 | * See if we can merge this file into an existing skb SCM_RIGHTS | |
3374 | * file set. If there's no room, fall back to allocating a new skb | |
3375 | * and filling it in. | |
3376 | */ | |
3377 | spin_lock_irq(&head->lock); | |
3378 | skb = skb_peek(head); | |
3379 | if (skb) { | |
3380 | struct scm_fp_list *fpl = UNIXCB(skb).fp; | |
3381 | ||
3382 | if (fpl->count < SCM_MAX_FD) { | |
3383 | __skb_unlink(skb, head); | |
3384 | spin_unlock_irq(&head->lock); | |
3385 | fpl->fp[fpl->count] = get_file(file); | |
3386 | unix_inflight(fpl->user, fpl->fp[fpl->count]); | |
3387 | fpl->count++; | |
3388 | spin_lock_irq(&head->lock); | |
3389 | __skb_queue_head(head, skb); | |
3390 | } else { | |
3391 | skb = NULL; | |
3392 | } | |
3393 | } | |
3394 | spin_unlock_irq(&head->lock); | |
3395 | ||
3396 | if (skb) { | |
3397 | fput(file); | |
3398 | return 0; | |
3399 | } | |
3400 | ||
3401 | return __io_sqe_files_scm(ctx, 1, index); | |
3402 | #else | |
3403 | return 0; | |
3404 | #endif | |
3405 | } | |
3406 | ||
3407 | static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg, | |
3408 | unsigned nr_args) | |
3409 | { | |
3410 | struct io_uring_files_update up; | |
3411 | __s32 __user *fds; | |
3412 | int fd, i, err; | |
3413 | __u32 done; | |
3414 | ||
3415 | if (!ctx->user_files) | |
3416 | return -ENXIO; | |
3417 | if (!nr_args) | |
3418 | return -EINVAL; | |
3419 | if (copy_from_user(&up, arg, sizeof(up))) | |
3420 | return -EFAULT; | |
3421 | if (check_add_overflow(up.offset, nr_args, &done)) | |
3422 | return -EOVERFLOW; | |
3423 | if (done > ctx->nr_user_files) | |
3424 | return -EINVAL; | |
3425 | ||
3426 | done = 0; | |
3427 | fds = (__s32 __user *) up.fds; | |
3428 | while (nr_args) { | |
3429 | err = 0; | |
3430 | if (copy_from_user(&fd, &fds[done], sizeof(fd))) { | |
3431 | err = -EFAULT; | |
3432 | break; | |
3433 | } | |
3434 | i = array_index_nospec(up.offset, ctx->nr_user_files); | |
3435 | if (ctx->user_files[i]) { | |
3436 | io_sqe_file_unregister(ctx, i); | |
3437 | ctx->user_files[i] = NULL; | |
3438 | } | |
3439 | if (fd != -1) { | |
3440 | struct file *file; | |
3441 | ||
3442 | file = fget(fd); | |
3443 | if (!file) { | |
3444 | err = -EBADF; | |
3445 | break; | |
3446 | } | |
3447 | /* | |
3448 | * Don't allow io_uring instances to be registered. If | |
3449 | * UNIX isn't enabled, then this causes a reference | |
3450 | * cycle and this instance can never get freed. If UNIX | |
3451 | * is enabled we'll handle it just fine, but there's | |
3452 | * still no point in allowing a ring fd as it doesn't | |
3453 | * support regular read/write anyway. | |
3454 | */ | |
3455 | if (file->f_op == &io_uring_fops) { | |
3456 | fput(file); | |
3457 | err = -EBADF; | |
3458 | break; | |
3459 | } | |
3460 | ctx->user_files[i] = file; | |
3461 | err = io_sqe_file_register(ctx, file, i); | |
3462 | if (err) | |
3463 | break; | |
3464 | } | |
3465 | nr_args--; | |
3466 | done++; | |
3467 | up.offset++; | |
3468 | } | |
3469 | ||
3470 | return done ? done : err; | |
3471 | } | |
3472 | ||
6c271ce2 JA |
3473 | static int io_sq_offload_start(struct io_ring_ctx *ctx, |
3474 | struct io_uring_params *p) | |
2b188cc1 JA |
3475 | { |
3476 | int ret; | |
3477 | ||
6c271ce2 | 3478 | init_waitqueue_head(&ctx->sqo_wait); |
2b188cc1 JA |
3479 | mmgrab(current->mm); |
3480 | ctx->sqo_mm = current->mm; | |
3481 | ||
6c271ce2 | 3482 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3ec482d1 JA |
3483 | ret = -EPERM; |
3484 | if (!capable(CAP_SYS_ADMIN)) | |
3485 | goto err; | |
3486 | ||
917257da JA |
3487 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
3488 | if (!ctx->sq_thread_idle) | |
3489 | ctx->sq_thread_idle = HZ; | |
3490 | ||
6c271ce2 | 3491 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 3492 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 3493 | |
917257da | 3494 | ret = -EINVAL; |
44a9bd18 JA |
3495 | if (cpu >= nr_cpu_ids) |
3496 | goto err; | |
7889f44d | 3497 | if (!cpu_online(cpu)) |
917257da JA |
3498 | goto err; |
3499 | ||
6c271ce2 JA |
3500 | ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread, |
3501 | ctx, cpu, | |
3502 | "io_uring-sq"); | |
3503 | } else { | |
3504 | ctx->sqo_thread = kthread_create(io_sq_thread, ctx, | |
3505 | "io_uring-sq"); | |
3506 | } | |
3507 | if (IS_ERR(ctx->sqo_thread)) { | |
3508 | ret = PTR_ERR(ctx->sqo_thread); | |
3509 | ctx->sqo_thread = NULL; | |
3510 | goto err; | |
3511 | } | |
3512 | wake_up_process(ctx->sqo_thread); | |
3513 | } else if (p->flags & IORING_SETUP_SQ_AFF) { | |
3514 | /* Can't have SQ_AFF without SQPOLL */ | |
3515 | ret = -EINVAL; | |
3516 | goto err; | |
3517 | } | |
3518 | ||
2b188cc1 | 3519 | /* Do QD, or 2 * CPUS, whatever is smallest */ |
54a91f3b JA |
3520 | ctx->sqo_wq[0] = alloc_workqueue("io_ring-wq", |
3521 | WQ_UNBOUND | WQ_FREEZABLE, | |
2b188cc1 | 3522 | min(ctx->sq_entries - 1, 2 * num_online_cpus())); |
54a91f3b JA |
3523 | if (!ctx->sqo_wq[0]) { |
3524 | ret = -ENOMEM; | |
3525 | goto err; | |
3526 | } | |
3527 | ||
3528 | /* | |
3529 | * This is for buffered writes, where we want to limit the parallelism | |
3530 | * due to file locking in file systems. As "normal" buffered writes | |
3531 | * should parellelize on writeout quite nicely, limit us to having 2 | |
3532 | * pending. This avoids massive contention on the inode when doing | |
3533 | * buffered async writes. | |
3534 | */ | |
3535 | ctx->sqo_wq[1] = alloc_workqueue("io_ring-write-wq", | |
3536 | WQ_UNBOUND | WQ_FREEZABLE, 2); | |
3537 | if (!ctx->sqo_wq[1]) { | |
2b188cc1 JA |
3538 | ret = -ENOMEM; |
3539 | goto err; | |
3540 | } | |
3541 | ||
3542 | return 0; | |
3543 | err: | |
54a91f3b | 3544 | io_finish_async(ctx); |
2b188cc1 JA |
3545 | mmdrop(ctx->sqo_mm); |
3546 | ctx->sqo_mm = NULL; | |
3547 | return ret; | |
3548 | } | |
3549 | ||
3550 | static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages) | |
3551 | { | |
3552 | atomic_long_sub(nr_pages, &user->locked_vm); | |
3553 | } | |
3554 | ||
3555 | static int io_account_mem(struct user_struct *user, unsigned long nr_pages) | |
3556 | { | |
3557 | unsigned long page_limit, cur_pages, new_pages; | |
3558 | ||
3559 | /* Don't allow more pages than we can safely lock */ | |
3560 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
3561 | ||
3562 | do { | |
3563 | cur_pages = atomic_long_read(&user->locked_vm); | |
3564 | new_pages = cur_pages + nr_pages; | |
3565 | if (new_pages > page_limit) | |
3566 | return -ENOMEM; | |
3567 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
3568 | new_pages) != cur_pages); | |
3569 | ||
3570 | return 0; | |
3571 | } | |
3572 | ||
3573 | static void io_mem_free(void *ptr) | |
3574 | { | |
52e04ef4 MR |
3575 | struct page *page; |
3576 | ||
3577 | if (!ptr) | |
3578 | return; | |
2b188cc1 | 3579 | |
52e04ef4 | 3580 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
3581 | if (put_page_testzero(page)) |
3582 | free_compound_page(page); | |
3583 | } | |
3584 | ||
3585 | static void *io_mem_alloc(size_t size) | |
3586 | { | |
3587 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
3588 | __GFP_NORETRY; | |
3589 | ||
3590 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
3591 | } | |
3592 | ||
75b28aff HV |
3593 | static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries, |
3594 | size_t *sq_offset) | |
3595 | { | |
3596 | struct io_rings *rings; | |
3597 | size_t off, sq_array_size; | |
3598 | ||
3599 | off = struct_size(rings, cqes, cq_entries); | |
3600 | if (off == SIZE_MAX) | |
3601 | return SIZE_MAX; | |
3602 | ||
3603 | #ifdef CONFIG_SMP | |
3604 | off = ALIGN(off, SMP_CACHE_BYTES); | |
3605 | if (off == 0) | |
3606 | return SIZE_MAX; | |
3607 | #endif | |
3608 | ||
3609 | sq_array_size = array_size(sizeof(u32), sq_entries); | |
3610 | if (sq_array_size == SIZE_MAX) | |
3611 | return SIZE_MAX; | |
3612 | ||
3613 | if (check_add_overflow(off, sq_array_size, &off)) | |
3614 | return SIZE_MAX; | |
3615 | ||
3616 | if (sq_offset) | |
3617 | *sq_offset = off; | |
3618 | ||
3619 | return off; | |
3620 | } | |
3621 | ||
2b188cc1 JA |
3622 | static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries) |
3623 | { | |
75b28aff | 3624 | size_t pages; |
2b188cc1 | 3625 | |
75b28aff HV |
3626 | pages = (size_t)1 << get_order( |
3627 | rings_size(sq_entries, cq_entries, NULL)); | |
3628 | pages += (size_t)1 << get_order( | |
3629 | array_size(sizeof(struct io_uring_sqe), sq_entries)); | |
2b188cc1 | 3630 | |
75b28aff | 3631 | return pages; |
2b188cc1 JA |
3632 | } |
3633 | ||
edafccee JA |
3634 | static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx) |
3635 | { | |
3636 | int i, j; | |
3637 | ||
3638 | if (!ctx->user_bufs) | |
3639 | return -ENXIO; | |
3640 | ||
3641 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
3642 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
3643 | ||
3644 | for (j = 0; j < imu->nr_bvecs; j++) | |
27c4d3a3 | 3645 | put_user_page(imu->bvec[j].bv_page); |
edafccee JA |
3646 | |
3647 | if (ctx->account_mem) | |
3648 | io_unaccount_mem(ctx->user, imu->nr_bvecs); | |
d4ef6475 | 3649 | kvfree(imu->bvec); |
edafccee JA |
3650 | imu->nr_bvecs = 0; |
3651 | } | |
3652 | ||
3653 | kfree(ctx->user_bufs); | |
3654 | ctx->user_bufs = NULL; | |
3655 | ctx->nr_user_bufs = 0; | |
3656 | return 0; | |
3657 | } | |
3658 | ||
3659 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
3660 | void __user *arg, unsigned index) | |
3661 | { | |
3662 | struct iovec __user *src; | |
3663 | ||
3664 | #ifdef CONFIG_COMPAT | |
3665 | if (ctx->compat) { | |
3666 | struct compat_iovec __user *ciovs; | |
3667 | struct compat_iovec ciov; | |
3668 | ||
3669 | ciovs = (struct compat_iovec __user *) arg; | |
3670 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
3671 | return -EFAULT; | |
3672 | ||
3673 | dst->iov_base = (void __user *) (unsigned long) ciov.iov_base; | |
3674 | dst->iov_len = ciov.iov_len; | |
3675 | return 0; | |
3676 | } | |
3677 | #endif | |
3678 | src = (struct iovec __user *) arg; | |
3679 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
3680 | return -EFAULT; | |
3681 | return 0; | |
3682 | } | |
3683 | ||
3684 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, | |
3685 | unsigned nr_args) | |
3686 | { | |
3687 | struct vm_area_struct **vmas = NULL; | |
3688 | struct page **pages = NULL; | |
3689 | int i, j, got_pages = 0; | |
3690 | int ret = -EINVAL; | |
3691 | ||
3692 | if (ctx->user_bufs) | |
3693 | return -EBUSY; | |
3694 | if (!nr_args || nr_args > UIO_MAXIOV) | |
3695 | return -EINVAL; | |
3696 | ||
3697 | ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf), | |
3698 | GFP_KERNEL); | |
3699 | if (!ctx->user_bufs) | |
3700 | return -ENOMEM; | |
3701 | ||
3702 | for (i = 0; i < nr_args; i++) { | |
3703 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
3704 | unsigned long off, start, end, ubuf; | |
3705 | int pret, nr_pages; | |
3706 | struct iovec iov; | |
3707 | size_t size; | |
3708 | ||
3709 | ret = io_copy_iov(ctx, &iov, arg, i); | |
3710 | if (ret) | |
a278682d | 3711 | goto err; |
edafccee JA |
3712 | |
3713 | /* | |
3714 | * Don't impose further limits on the size and buffer | |
3715 | * constraints here, we'll -EINVAL later when IO is | |
3716 | * submitted if they are wrong. | |
3717 | */ | |
3718 | ret = -EFAULT; | |
3719 | if (!iov.iov_base || !iov.iov_len) | |
3720 | goto err; | |
3721 | ||
3722 | /* arbitrary limit, but we need something */ | |
3723 | if (iov.iov_len > SZ_1G) | |
3724 | goto err; | |
3725 | ||
3726 | ubuf = (unsigned long) iov.iov_base; | |
3727 | end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
3728 | start = ubuf >> PAGE_SHIFT; | |
3729 | nr_pages = end - start; | |
3730 | ||
3731 | if (ctx->account_mem) { | |
3732 | ret = io_account_mem(ctx->user, nr_pages); | |
3733 | if (ret) | |
3734 | goto err; | |
3735 | } | |
3736 | ||
3737 | ret = 0; | |
3738 | if (!pages || nr_pages > got_pages) { | |
3739 | kfree(vmas); | |
3740 | kfree(pages); | |
d4ef6475 | 3741 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), |
edafccee | 3742 | GFP_KERNEL); |
d4ef6475 | 3743 | vmas = kvmalloc_array(nr_pages, |
edafccee JA |
3744 | sizeof(struct vm_area_struct *), |
3745 | GFP_KERNEL); | |
3746 | if (!pages || !vmas) { | |
3747 | ret = -ENOMEM; | |
3748 | if (ctx->account_mem) | |
3749 | io_unaccount_mem(ctx->user, nr_pages); | |
3750 | goto err; | |
3751 | } | |
3752 | got_pages = nr_pages; | |
3753 | } | |
3754 | ||
d4ef6475 | 3755 | imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec), |
edafccee JA |
3756 | GFP_KERNEL); |
3757 | ret = -ENOMEM; | |
3758 | if (!imu->bvec) { | |
3759 | if (ctx->account_mem) | |
3760 | io_unaccount_mem(ctx->user, nr_pages); | |
3761 | goto err; | |
3762 | } | |
3763 | ||
3764 | ret = 0; | |
3765 | down_read(¤t->mm->mmap_sem); | |
932f4a63 IW |
3766 | pret = get_user_pages(ubuf, nr_pages, |
3767 | FOLL_WRITE | FOLL_LONGTERM, | |
3768 | pages, vmas); | |
edafccee JA |
3769 | if (pret == nr_pages) { |
3770 | /* don't support file backed memory */ | |
3771 | for (j = 0; j < nr_pages; j++) { | |
3772 | struct vm_area_struct *vma = vmas[j]; | |
3773 | ||
3774 | if (vma->vm_file && | |
3775 | !is_file_hugepages(vma->vm_file)) { | |
3776 | ret = -EOPNOTSUPP; | |
3777 | break; | |
3778 | } | |
3779 | } | |
3780 | } else { | |
3781 | ret = pret < 0 ? pret : -EFAULT; | |
3782 | } | |
3783 | up_read(¤t->mm->mmap_sem); | |
3784 | if (ret) { | |
3785 | /* | |
3786 | * if we did partial map, or found file backed vmas, | |
3787 | * release any pages we did get | |
3788 | */ | |
27c4d3a3 JH |
3789 | if (pret > 0) |
3790 | put_user_pages(pages, pret); | |
edafccee JA |
3791 | if (ctx->account_mem) |
3792 | io_unaccount_mem(ctx->user, nr_pages); | |
d4ef6475 | 3793 | kvfree(imu->bvec); |
edafccee JA |
3794 | goto err; |
3795 | } | |
3796 | ||
3797 | off = ubuf & ~PAGE_MASK; | |
3798 | size = iov.iov_len; | |
3799 | for (j = 0; j < nr_pages; j++) { | |
3800 | size_t vec_len; | |
3801 | ||
3802 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
3803 | imu->bvec[j].bv_page = pages[j]; | |
3804 | imu->bvec[j].bv_len = vec_len; | |
3805 | imu->bvec[j].bv_offset = off; | |
3806 | off = 0; | |
3807 | size -= vec_len; | |
3808 | } | |
3809 | /* store original address for later verification */ | |
3810 | imu->ubuf = ubuf; | |
3811 | imu->len = iov.iov_len; | |
3812 | imu->nr_bvecs = nr_pages; | |
3813 | ||
3814 | ctx->nr_user_bufs++; | |
3815 | } | |
d4ef6475 MR |
3816 | kvfree(pages); |
3817 | kvfree(vmas); | |
edafccee JA |
3818 | return 0; |
3819 | err: | |
d4ef6475 MR |
3820 | kvfree(pages); |
3821 | kvfree(vmas); | |
edafccee JA |
3822 | io_sqe_buffer_unregister(ctx); |
3823 | return ret; | |
3824 | } | |
3825 | ||
9b402849 JA |
3826 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
3827 | { | |
3828 | __s32 __user *fds = arg; | |
3829 | int fd; | |
3830 | ||
3831 | if (ctx->cq_ev_fd) | |
3832 | return -EBUSY; | |
3833 | ||
3834 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
3835 | return -EFAULT; | |
3836 | ||
3837 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
3838 | if (IS_ERR(ctx->cq_ev_fd)) { | |
3839 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
3840 | ctx->cq_ev_fd = NULL; | |
3841 | return ret; | |
3842 | } | |
3843 | ||
3844 | return 0; | |
3845 | } | |
3846 | ||
3847 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
3848 | { | |
3849 | if (ctx->cq_ev_fd) { | |
3850 | eventfd_ctx_put(ctx->cq_ev_fd); | |
3851 | ctx->cq_ev_fd = NULL; | |
3852 | return 0; | |
3853 | } | |
3854 | ||
3855 | return -ENXIO; | |
3856 | } | |
3857 | ||
2b188cc1 JA |
3858 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
3859 | { | |
6b06314c | 3860 | io_finish_async(ctx); |
2b188cc1 JA |
3861 | if (ctx->sqo_mm) |
3862 | mmdrop(ctx->sqo_mm); | |
def596e9 JA |
3863 | |
3864 | io_iopoll_reap_events(ctx); | |
edafccee | 3865 | io_sqe_buffer_unregister(ctx); |
6b06314c | 3866 | io_sqe_files_unregister(ctx); |
9b402849 | 3867 | io_eventfd_unregister(ctx); |
def596e9 | 3868 | |
2b188cc1 | 3869 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
3870 | if (ctx->ring_sock) { |
3871 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 3872 | sock_release(ctx->ring_sock); |
355e8d26 | 3873 | } |
2b188cc1 JA |
3874 | #endif |
3875 | ||
75b28aff | 3876 | io_mem_free(ctx->rings); |
2b188cc1 | 3877 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
3878 | |
3879 | percpu_ref_exit(&ctx->refs); | |
3880 | if (ctx->account_mem) | |
3881 | io_unaccount_mem(ctx->user, | |
3882 | ring_pages(ctx->sq_entries, ctx->cq_entries)); | |
3883 | free_uid(ctx->user); | |
3884 | kfree(ctx); | |
3885 | } | |
3886 | ||
3887 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
3888 | { | |
3889 | struct io_ring_ctx *ctx = file->private_data; | |
3890 | __poll_t mask = 0; | |
3891 | ||
3892 | poll_wait(file, &ctx->cq_wait, wait); | |
4f7067c3 SB |
3893 | /* |
3894 | * synchronizes with barrier from wq_has_sleeper call in | |
3895 | * io_commit_cqring | |
3896 | */ | |
2b188cc1 | 3897 | smp_rmb(); |
75b28aff HV |
3898 | if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head != |
3899 | ctx->rings->sq_ring_entries) | |
2b188cc1 | 3900 | mask |= EPOLLOUT | EPOLLWRNORM; |
daa5de54 | 3901 | if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail) |
2b188cc1 JA |
3902 | mask |= EPOLLIN | EPOLLRDNORM; |
3903 | ||
3904 | return mask; | |
3905 | } | |
3906 | ||
3907 | static int io_uring_fasync(int fd, struct file *file, int on) | |
3908 | { | |
3909 | struct io_ring_ctx *ctx = file->private_data; | |
3910 | ||
3911 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
3912 | } | |
3913 | ||
3914 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) | |
3915 | { | |
3916 | mutex_lock(&ctx->uring_lock); | |
3917 | percpu_ref_kill(&ctx->refs); | |
3918 | mutex_unlock(&ctx->uring_lock); | |
3919 | ||
5262f567 | 3920 | io_kill_timeouts(ctx); |
221c5eb2 | 3921 | io_poll_remove_all(ctx); |
def596e9 | 3922 | io_iopoll_reap_events(ctx); |
2b188cc1 JA |
3923 | wait_for_completion(&ctx->ctx_done); |
3924 | io_ring_ctx_free(ctx); | |
3925 | } | |
3926 | ||
3927 | static int io_uring_release(struct inode *inode, struct file *file) | |
3928 | { | |
3929 | struct io_ring_ctx *ctx = file->private_data; | |
3930 | ||
3931 | file->private_data = NULL; | |
3932 | io_ring_ctx_wait_and_kill(ctx); | |
3933 | return 0; | |
3934 | } | |
3935 | ||
3936 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
3937 | { | |
3938 | loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT; | |
3939 | unsigned long sz = vma->vm_end - vma->vm_start; | |
3940 | struct io_ring_ctx *ctx = file->private_data; | |
3941 | unsigned long pfn; | |
3942 | struct page *page; | |
3943 | void *ptr; | |
3944 | ||
3945 | switch (offset) { | |
3946 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
3947 | case IORING_OFF_CQ_RING: |
3948 | ptr = ctx->rings; | |
2b188cc1 JA |
3949 | break; |
3950 | case IORING_OFF_SQES: | |
3951 | ptr = ctx->sq_sqes; | |
3952 | break; | |
2b188cc1 JA |
3953 | default: |
3954 | return -EINVAL; | |
3955 | } | |
3956 | ||
3957 | page = virt_to_head_page(ptr); | |
a50b854e | 3958 | if (sz > page_size(page)) |
2b188cc1 JA |
3959 | return -EINVAL; |
3960 | ||
3961 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
3962 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
3963 | } | |
3964 | ||
3965 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, | |
3966 | u32, min_complete, u32, flags, const sigset_t __user *, sig, | |
3967 | size_t, sigsz) | |
3968 | { | |
3969 | struct io_ring_ctx *ctx; | |
3970 | long ret = -EBADF; | |
3971 | int submitted = 0; | |
3972 | struct fd f; | |
3973 | ||
6c271ce2 | 3974 | if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP)) |
2b188cc1 JA |
3975 | return -EINVAL; |
3976 | ||
3977 | f = fdget(fd); | |
3978 | if (!f.file) | |
3979 | return -EBADF; | |
3980 | ||
3981 | ret = -EOPNOTSUPP; | |
3982 | if (f.file->f_op != &io_uring_fops) | |
3983 | goto out_fput; | |
3984 | ||
3985 | ret = -ENXIO; | |
3986 | ctx = f.file->private_data; | |
3987 | if (!percpu_ref_tryget(&ctx->refs)) | |
3988 | goto out_fput; | |
3989 | ||
6c271ce2 JA |
3990 | /* |
3991 | * For SQ polling, the thread will do all submissions and completions. | |
3992 | * Just return the requested submit count, and wake the thread if | |
3993 | * we were asked to. | |
3994 | */ | |
b2a9eada | 3995 | ret = 0; |
6c271ce2 JA |
3996 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3997 | if (flags & IORING_ENTER_SQ_WAKEUP) | |
3998 | wake_up(&ctx->sqo_wait); | |
3999 | submitted = to_submit; | |
b2a9eada | 4000 | } else if (to_submit) { |
2b188cc1 JA |
4001 | to_submit = min(to_submit, ctx->sq_entries); |
4002 | ||
4003 | mutex_lock(&ctx->uring_lock); | |
bc808bce | 4004 | submitted = io_ring_submit(ctx, to_submit); |
2b188cc1 | 4005 | mutex_unlock(&ctx->uring_lock); |
2b188cc1 JA |
4006 | } |
4007 | if (flags & IORING_ENTER_GETEVENTS) { | |
def596e9 JA |
4008 | unsigned nr_events = 0; |
4009 | ||
2b188cc1 JA |
4010 | min_complete = min(min_complete, ctx->cq_entries); |
4011 | ||
def596e9 | 4012 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 | 4013 | ret = io_iopoll_check(ctx, &nr_events, min_complete); |
def596e9 JA |
4014 | } else { |
4015 | ret = io_cqring_wait(ctx, min_complete, sig, sigsz); | |
4016 | } | |
2b188cc1 JA |
4017 | } |
4018 | ||
6805b32e | 4019 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
4020 | out_fput: |
4021 | fdput(f); | |
4022 | return submitted ? submitted : ret; | |
4023 | } | |
4024 | ||
4025 | static const struct file_operations io_uring_fops = { | |
4026 | .release = io_uring_release, | |
4027 | .mmap = io_uring_mmap, | |
4028 | .poll = io_uring_poll, | |
4029 | .fasync = io_uring_fasync, | |
4030 | }; | |
4031 | ||
4032 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
4033 | struct io_uring_params *p) | |
4034 | { | |
75b28aff HV |
4035 | struct io_rings *rings; |
4036 | size_t size, sq_array_offset; | |
2b188cc1 | 4037 | |
75b28aff HV |
4038 | size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset); |
4039 | if (size == SIZE_MAX) | |
4040 | return -EOVERFLOW; | |
4041 | ||
4042 | rings = io_mem_alloc(size); | |
4043 | if (!rings) | |
2b188cc1 JA |
4044 | return -ENOMEM; |
4045 | ||
75b28aff HV |
4046 | ctx->rings = rings; |
4047 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
4048 | rings->sq_ring_mask = p->sq_entries - 1; | |
4049 | rings->cq_ring_mask = p->cq_entries - 1; | |
4050 | rings->sq_ring_entries = p->sq_entries; | |
4051 | rings->cq_ring_entries = p->cq_entries; | |
4052 | ctx->sq_mask = rings->sq_ring_mask; | |
4053 | ctx->cq_mask = rings->cq_ring_mask; | |
4054 | ctx->sq_entries = rings->sq_ring_entries; | |
4055 | ctx->cq_entries = rings->cq_ring_entries; | |
2b188cc1 JA |
4056 | |
4057 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
4058 | if (size == SIZE_MAX) | |
4059 | return -EOVERFLOW; | |
4060 | ||
4061 | ctx->sq_sqes = io_mem_alloc(size); | |
52e04ef4 | 4062 | if (!ctx->sq_sqes) |
2b188cc1 | 4063 | return -ENOMEM; |
2b188cc1 | 4064 | |
2b188cc1 JA |
4065 | return 0; |
4066 | } | |
4067 | ||
4068 | /* | |
4069 | * Allocate an anonymous fd, this is what constitutes the application | |
4070 | * visible backing of an io_uring instance. The application mmaps this | |
4071 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
4072 | * we have to tie this fd to a socket for file garbage collection purposes. | |
4073 | */ | |
4074 | static int io_uring_get_fd(struct io_ring_ctx *ctx) | |
4075 | { | |
4076 | struct file *file; | |
4077 | int ret; | |
4078 | ||
4079 | #if defined(CONFIG_UNIX) | |
4080 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, | |
4081 | &ctx->ring_sock); | |
4082 | if (ret) | |
4083 | return ret; | |
4084 | #endif | |
4085 | ||
4086 | ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
4087 | if (ret < 0) | |
4088 | goto err; | |
4089 | ||
4090 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, | |
4091 | O_RDWR | O_CLOEXEC); | |
4092 | if (IS_ERR(file)) { | |
4093 | put_unused_fd(ret); | |
4094 | ret = PTR_ERR(file); | |
4095 | goto err; | |
4096 | } | |
4097 | ||
4098 | #if defined(CONFIG_UNIX) | |
4099 | ctx->ring_sock->file = file; | |
6b06314c | 4100 | ctx->ring_sock->sk->sk_user_data = ctx; |
2b188cc1 JA |
4101 | #endif |
4102 | fd_install(ret, file); | |
4103 | return ret; | |
4104 | err: | |
4105 | #if defined(CONFIG_UNIX) | |
4106 | sock_release(ctx->ring_sock); | |
4107 | ctx->ring_sock = NULL; | |
4108 | #endif | |
4109 | return ret; | |
4110 | } | |
4111 | ||
4112 | static int io_uring_create(unsigned entries, struct io_uring_params *p) | |
4113 | { | |
4114 | struct user_struct *user = NULL; | |
4115 | struct io_ring_ctx *ctx; | |
4116 | bool account_mem; | |
4117 | int ret; | |
4118 | ||
4119 | if (!entries || entries > IORING_MAX_ENTRIES) | |
4120 | return -EINVAL; | |
4121 | ||
4122 | /* | |
4123 | * Use twice as many entries for the CQ ring. It's possible for the | |
4124 | * application to drive a higher depth than the size of the SQ ring, | |
4125 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
4126 | * some flexibility in overcommitting a bit. If the application has |
4127 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
4128 | * of CQ ring entries manually. | |
2b188cc1 JA |
4129 | */ |
4130 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
4131 | if (p->flags & IORING_SETUP_CQSIZE) { |
4132 | /* | |
4133 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
4134 | * to a power-of-two, if it isn't already. We do NOT impose | |
4135 | * any cq vs sq ring sizing. | |
4136 | */ | |
4137 | if (p->cq_entries < p->sq_entries || p->cq_entries > IORING_MAX_CQ_ENTRIES) | |
4138 | return -EINVAL; | |
4139 | p->cq_entries = roundup_pow_of_two(p->cq_entries); | |
4140 | } else { | |
4141 | p->cq_entries = 2 * p->sq_entries; | |
4142 | } | |
2b188cc1 JA |
4143 | |
4144 | user = get_uid(current_user()); | |
4145 | account_mem = !capable(CAP_IPC_LOCK); | |
4146 | ||
4147 | if (account_mem) { | |
4148 | ret = io_account_mem(user, | |
4149 | ring_pages(p->sq_entries, p->cq_entries)); | |
4150 | if (ret) { | |
4151 | free_uid(user); | |
4152 | return ret; | |
4153 | } | |
4154 | } | |
4155 | ||
4156 | ctx = io_ring_ctx_alloc(p); | |
4157 | if (!ctx) { | |
4158 | if (account_mem) | |
4159 | io_unaccount_mem(user, ring_pages(p->sq_entries, | |
4160 | p->cq_entries)); | |
4161 | free_uid(user); | |
4162 | return -ENOMEM; | |
4163 | } | |
4164 | ctx->compat = in_compat_syscall(); | |
4165 | ctx->account_mem = account_mem; | |
4166 | ctx->user = user; | |
4167 | ||
4168 | ret = io_allocate_scq_urings(ctx, p); | |
4169 | if (ret) | |
4170 | goto err; | |
4171 | ||
6c271ce2 | 4172 | ret = io_sq_offload_start(ctx, p); |
2b188cc1 JA |
4173 | if (ret) |
4174 | goto err; | |
4175 | ||
2b188cc1 | 4176 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
4177 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
4178 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
4179 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
4180 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
4181 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
4182 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
4183 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
4184 | |
4185 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
4186 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
4187 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
4188 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
4189 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
4190 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
4191 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
ac90f249 | 4192 | |
044c1ab3 JA |
4193 | /* |
4194 | * Install ring fd as the very last thing, so we don't risk someone | |
4195 | * having closed it before we finish setup | |
4196 | */ | |
4197 | ret = io_uring_get_fd(ctx); | |
4198 | if (ret < 0) | |
4199 | goto err; | |
4200 | ||
ac90f249 | 4201 | p->features = IORING_FEAT_SINGLE_MMAP; |
c826bd7a | 4202 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
4203 | return ret; |
4204 | err: | |
4205 | io_ring_ctx_wait_and_kill(ctx); | |
4206 | return ret; | |
4207 | } | |
4208 | ||
4209 | /* | |
4210 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
4211 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
4212 | * params structure passed in. | |
4213 | */ | |
4214 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
4215 | { | |
4216 | struct io_uring_params p; | |
4217 | long ret; | |
4218 | int i; | |
4219 | ||
4220 | if (copy_from_user(&p, params, sizeof(p))) | |
4221 | return -EFAULT; | |
4222 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
4223 | if (p.resv[i]) | |
4224 | return -EINVAL; | |
4225 | } | |
4226 | ||
6c271ce2 | 4227 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
33a107f0 | 4228 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE)) |
2b188cc1 JA |
4229 | return -EINVAL; |
4230 | ||
4231 | ret = io_uring_create(entries, &p); | |
4232 | if (ret < 0) | |
4233 | return ret; | |
4234 | ||
4235 | if (copy_to_user(params, &p, sizeof(p))) | |
4236 | return -EFAULT; | |
4237 | ||
4238 | return ret; | |
4239 | } | |
4240 | ||
4241 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
4242 | struct io_uring_params __user *, params) | |
4243 | { | |
4244 | return io_uring_setup(entries, params); | |
4245 | } | |
4246 | ||
edafccee JA |
4247 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
4248 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
4249 | __releases(ctx->uring_lock) |
4250 | __acquires(ctx->uring_lock) | |
edafccee JA |
4251 | { |
4252 | int ret; | |
4253 | ||
35fa71a0 JA |
4254 | /* |
4255 | * We're inside the ring mutex, if the ref is already dying, then | |
4256 | * someone else killed the ctx or is already going through | |
4257 | * io_uring_register(). | |
4258 | */ | |
4259 | if (percpu_ref_is_dying(&ctx->refs)) | |
4260 | return -ENXIO; | |
4261 | ||
edafccee | 4262 | percpu_ref_kill(&ctx->refs); |
b19062a5 JA |
4263 | |
4264 | /* | |
4265 | * Drop uring mutex before waiting for references to exit. If another | |
4266 | * thread is currently inside io_uring_enter() it might need to grab | |
4267 | * the uring_lock to make progress. If we hold it here across the drain | |
4268 | * wait, then we can deadlock. It's safe to drop the mutex here, since | |
4269 | * no new references will come in after we've killed the percpu ref. | |
4270 | */ | |
4271 | mutex_unlock(&ctx->uring_lock); | |
edafccee | 4272 | wait_for_completion(&ctx->ctx_done); |
b19062a5 | 4273 | mutex_lock(&ctx->uring_lock); |
edafccee JA |
4274 | |
4275 | switch (opcode) { | |
4276 | case IORING_REGISTER_BUFFERS: | |
4277 | ret = io_sqe_buffer_register(ctx, arg, nr_args); | |
4278 | break; | |
4279 | case IORING_UNREGISTER_BUFFERS: | |
4280 | ret = -EINVAL; | |
4281 | if (arg || nr_args) | |
4282 | break; | |
4283 | ret = io_sqe_buffer_unregister(ctx); | |
4284 | break; | |
6b06314c JA |
4285 | case IORING_REGISTER_FILES: |
4286 | ret = io_sqe_files_register(ctx, arg, nr_args); | |
4287 | break; | |
4288 | case IORING_UNREGISTER_FILES: | |
4289 | ret = -EINVAL; | |
4290 | if (arg || nr_args) | |
4291 | break; | |
4292 | ret = io_sqe_files_unregister(ctx); | |
4293 | break; | |
c3a31e60 JA |
4294 | case IORING_REGISTER_FILES_UPDATE: |
4295 | ret = io_sqe_files_update(ctx, arg, nr_args); | |
4296 | break; | |
9b402849 JA |
4297 | case IORING_REGISTER_EVENTFD: |
4298 | ret = -EINVAL; | |
4299 | if (nr_args != 1) | |
4300 | break; | |
4301 | ret = io_eventfd_register(ctx, arg); | |
4302 | break; | |
4303 | case IORING_UNREGISTER_EVENTFD: | |
4304 | ret = -EINVAL; | |
4305 | if (arg || nr_args) | |
4306 | break; | |
4307 | ret = io_eventfd_unregister(ctx); | |
4308 | break; | |
edafccee JA |
4309 | default: |
4310 | ret = -EINVAL; | |
4311 | break; | |
4312 | } | |
4313 | ||
4314 | /* bring the ctx back to life */ | |
4315 | reinit_completion(&ctx->ctx_done); | |
4316 | percpu_ref_reinit(&ctx->refs); | |
4317 | return ret; | |
4318 | } | |
4319 | ||
4320 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
4321 | void __user *, arg, unsigned int, nr_args) | |
4322 | { | |
4323 | struct io_ring_ctx *ctx; | |
4324 | long ret = -EBADF; | |
4325 | struct fd f; | |
4326 | ||
4327 | f = fdget(fd); | |
4328 | if (!f.file) | |
4329 | return -EBADF; | |
4330 | ||
4331 | ret = -EOPNOTSUPP; | |
4332 | if (f.file->f_op != &io_uring_fops) | |
4333 | goto out_fput; | |
4334 | ||
4335 | ctx = f.file->private_data; | |
4336 | ||
4337 | mutex_lock(&ctx->uring_lock); | |
4338 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
4339 | mutex_unlock(&ctx->uring_lock); | |
c826bd7a DD |
4340 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, |
4341 | ctx->cq_ev_fd != NULL, ret); | |
edafccee JA |
4342 | out_fput: |
4343 | fdput(f); | |
4344 | return ret; | |
4345 | } | |
4346 | ||
2b188cc1 JA |
4347 | static int __init io_uring_init(void) |
4348 | { | |
4349 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC); | |
4350 | return 0; | |
4351 | }; | |
4352 | __initcall(io_uring_init); |