]>
Commit | Line | Data |
---|---|---|
2b188cc1 JA |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* | |
3 | * Shared application/kernel submission and completion ring pairs, for | |
4 | * supporting fast/efficient IO. | |
5 | * | |
6 | * A note on the read/write ordering memory barriers that are matched between | |
1e84b97b SB |
7 | * the application and kernel side. |
8 | * | |
9 | * After the application reads the CQ ring tail, it must use an | |
10 | * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses | |
11 | * before writing the tail (using smp_load_acquire to read the tail will | |
12 | * do). It also needs a smp_mb() before updating CQ head (ordering the | |
13 | * entry load(s) with the head store), pairing with an implicit barrier | |
14 | * through a control-dependency in io_get_cqring (smp_store_release to | |
15 | * store head will do). Failure to do so could lead to reading invalid | |
16 | * CQ entries. | |
17 | * | |
18 | * Likewise, the application must use an appropriate smp_wmb() before | |
19 | * writing the SQ tail (ordering SQ entry stores with the tail store), | |
20 | * which pairs with smp_load_acquire in io_get_sqring (smp_store_release | |
21 | * to store the tail will do). And it needs a barrier ordering the SQ | |
22 | * head load before writing new SQ entries (smp_load_acquire to read | |
23 | * head will do). | |
24 | * | |
25 | * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application | |
26 | * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after* | |
27 | * updating the SQ tail; a full memory barrier smp_mb() is needed | |
28 | * between. | |
2b188cc1 JA |
29 | * |
30 | * Also see the examples in the liburing library: | |
31 | * | |
32 | * git://git.kernel.dk/liburing | |
33 | * | |
34 | * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens | |
35 | * from data shared between the kernel and application. This is done both | |
36 | * for ordering purposes, but also to ensure that once a value is loaded from | |
37 | * data that the application could potentially modify, it remains stable. | |
38 | * | |
39 | * Copyright (C) 2018-2019 Jens Axboe | |
c992fe29 | 40 | * Copyright (c) 2018-2019 Christoph Hellwig |
2b188cc1 JA |
41 | */ |
42 | #include <linux/kernel.h> | |
43 | #include <linux/init.h> | |
44 | #include <linux/errno.h> | |
45 | #include <linux/syscalls.h> | |
46 | #include <linux/compat.h> | |
47 | #include <linux/refcount.h> | |
48 | #include <linux/uio.h> | |
49 | ||
50 | #include <linux/sched/signal.h> | |
51 | #include <linux/fs.h> | |
52 | #include <linux/file.h> | |
53 | #include <linux/fdtable.h> | |
54 | #include <linux/mm.h> | |
55 | #include <linux/mman.h> | |
56 | #include <linux/mmu_context.h> | |
57 | #include <linux/percpu.h> | |
58 | #include <linux/slab.h> | |
59 | #include <linux/workqueue.h> | |
6c271ce2 | 60 | #include <linux/kthread.h> |
2b188cc1 | 61 | #include <linux/blkdev.h> |
edafccee | 62 | #include <linux/bvec.h> |
2b188cc1 JA |
63 | #include <linux/net.h> |
64 | #include <net/sock.h> | |
65 | #include <net/af_unix.h> | |
6b06314c | 66 | #include <net/scm.h> |
2b188cc1 JA |
67 | #include <linux/anon_inodes.h> |
68 | #include <linux/sched/mm.h> | |
69 | #include <linux/uaccess.h> | |
70 | #include <linux/nospec.h> | |
edafccee JA |
71 | #include <linux/sizes.h> |
72 | #include <linux/hugetlb.h> | |
2b188cc1 JA |
73 | |
74 | #include <uapi/linux/io_uring.h> | |
75 | ||
76 | #include "internal.h" | |
77 | ||
78 | #define IORING_MAX_ENTRIES 4096 | |
6b06314c | 79 | #define IORING_MAX_FIXED_FILES 1024 |
2b188cc1 JA |
80 | |
81 | struct io_uring { | |
82 | u32 head ____cacheline_aligned_in_smp; | |
83 | u32 tail ____cacheline_aligned_in_smp; | |
84 | }; | |
85 | ||
1e84b97b SB |
86 | /* |
87 | * This data is shared with the application through the mmap at offset | |
88 | * IORING_OFF_SQ_RING. | |
89 | * | |
90 | * The offsets to the member fields are published through struct | |
91 | * io_sqring_offsets when calling io_uring_setup. | |
92 | */ | |
2b188cc1 | 93 | struct io_sq_ring { |
1e84b97b SB |
94 | /* |
95 | * Head and tail offsets into the ring; the offsets need to be | |
96 | * masked to get valid indices. | |
97 | * | |
98 | * The kernel controls head and the application controls tail. | |
99 | */ | |
2b188cc1 | 100 | struct io_uring r; |
1e84b97b SB |
101 | /* |
102 | * Bitmask to apply to head and tail offsets (constant, equals | |
103 | * ring_entries - 1) | |
104 | */ | |
2b188cc1 | 105 | u32 ring_mask; |
1e84b97b | 106 | /* Ring size (constant, power of 2) */ |
2b188cc1 | 107 | u32 ring_entries; |
1e84b97b SB |
108 | /* |
109 | * Number of invalid entries dropped by the kernel due to | |
110 | * invalid index stored in array | |
111 | * | |
112 | * Written by the kernel, shouldn't be modified by the | |
113 | * application (i.e. get number of "new events" by comparing to | |
114 | * cached value). | |
115 | * | |
116 | * After a new SQ head value was read by the application this | |
117 | * counter includes all submissions that were dropped reaching | |
118 | * the new SQ head (and possibly more). | |
119 | */ | |
2b188cc1 | 120 | u32 dropped; |
1e84b97b SB |
121 | /* |
122 | * Runtime flags | |
123 | * | |
124 | * Written by the kernel, shouldn't be modified by the | |
125 | * application. | |
126 | * | |
127 | * The application needs a full memory barrier before checking | |
128 | * for IORING_SQ_NEED_WAKEUP after updating the sq tail. | |
129 | */ | |
2b188cc1 | 130 | u32 flags; |
1e84b97b SB |
131 | /* |
132 | * Ring buffer of indices into array of io_uring_sqe, which is | |
133 | * mmapped by the application using the IORING_OFF_SQES offset. | |
134 | * | |
135 | * This indirection could e.g. be used to assign fixed | |
136 | * io_uring_sqe entries to operations and only submit them to | |
137 | * the queue when needed. | |
138 | * | |
139 | * The kernel modifies neither the indices array nor the entries | |
140 | * array. | |
141 | */ | |
2b188cc1 JA |
142 | u32 array[]; |
143 | }; | |
144 | ||
1e84b97b SB |
145 | /* |
146 | * This data is shared with the application through the mmap at offset | |
147 | * IORING_OFF_CQ_RING. | |
148 | * | |
149 | * The offsets to the member fields are published through struct | |
150 | * io_cqring_offsets when calling io_uring_setup. | |
151 | */ | |
2b188cc1 | 152 | struct io_cq_ring { |
1e84b97b SB |
153 | /* |
154 | * Head and tail offsets into the ring; the offsets need to be | |
155 | * masked to get valid indices. | |
156 | * | |
157 | * The application controls head and the kernel tail. | |
158 | */ | |
2b188cc1 | 159 | struct io_uring r; |
1e84b97b SB |
160 | /* |
161 | * Bitmask to apply to head and tail offsets (constant, equals | |
162 | * ring_entries - 1) | |
163 | */ | |
2b188cc1 | 164 | u32 ring_mask; |
1e84b97b | 165 | /* Ring size (constant, power of 2) */ |
2b188cc1 | 166 | u32 ring_entries; |
1e84b97b SB |
167 | /* |
168 | * Number of completion events lost because the queue was full; | |
169 | * this should be avoided by the application by making sure | |
170 | * there are not more requests pending thatn there is space in | |
171 | * the completion queue. | |
172 | * | |
173 | * Written by the kernel, shouldn't be modified by the | |
174 | * application (i.e. get number of "new events" by comparing to | |
175 | * cached value). | |
176 | * | |
177 | * As completion events come in out of order this counter is not | |
178 | * ordered with any other data. | |
179 | */ | |
2b188cc1 | 180 | u32 overflow; |
1e84b97b SB |
181 | /* |
182 | * Ring buffer of completion events. | |
183 | * | |
184 | * The kernel writes completion events fresh every time they are | |
185 | * produced, so the application is allowed to modify pending | |
186 | * entries. | |
187 | */ | |
2b188cc1 JA |
188 | struct io_uring_cqe cqes[]; |
189 | }; | |
190 | ||
edafccee JA |
191 | struct io_mapped_ubuf { |
192 | u64 ubuf; | |
193 | size_t len; | |
194 | struct bio_vec *bvec; | |
195 | unsigned int nr_bvecs; | |
196 | }; | |
197 | ||
31b51510 JA |
198 | struct async_list { |
199 | spinlock_t lock; | |
200 | atomic_t cnt; | |
201 | struct list_head list; | |
202 | ||
203 | struct file *file; | |
204 | off_t io_end; | |
9310a7ba | 205 | size_t io_len; |
31b51510 JA |
206 | }; |
207 | ||
2b188cc1 JA |
208 | struct io_ring_ctx { |
209 | struct { | |
210 | struct percpu_ref refs; | |
211 | } ____cacheline_aligned_in_smp; | |
212 | ||
213 | struct { | |
214 | unsigned int flags; | |
215 | bool compat; | |
216 | bool account_mem; | |
217 | ||
218 | /* SQ ring */ | |
219 | struct io_sq_ring *sq_ring; | |
220 | unsigned cached_sq_head; | |
221 | unsigned sq_entries; | |
222 | unsigned sq_mask; | |
6c271ce2 | 223 | unsigned sq_thread_idle; |
2b188cc1 | 224 | struct io_uring_sqe *sq_sqes; |
de0617e4 JA |
225 | |
226 | struct list_head defer_list; | |
2b188cc1 JA |
227 | } ____cacheline_aligned_in_smp; |
228 | ||
229 | /* IO offload */ | |
230 | struct workqueue_struct *sqo_wq; | |
6c271ce2 | 231 | struct task_struct *sqo_thread; /* if using sq thread polling */ |
2b188cc1 | 232 | struct mm_struct *sqo_mm; |
6c271ce2 | 233 | wait_queue_head_t sqo_wait; |
a4c0b3de | 234 | struct completion sqo_thread_started; |
2b188cc1 JA |
235 | |
236 | struct { | |
237 | /* CQ ring */ | |
238 | struct io_cq_ring *cq_ring; | |
239 | unsigned cached_cq_tail; | |
240 | unsigned cq_entries; | |
241 | unsigned cq_mask; | |
242 | struct wait_queue_head cq_wait; | |
243 | struct fasync_struct *cq_fasync; | |
9b402849 | 244 | struct eventfd_ctx *cq_ev_fd; |
2b188cc1 JA |
245 | } ____cacheline_aligned_in_smp; |
246 | ||
6b06314c JA |
247 | /* |
248 | * If used, fixed file set. Writers must ensure that ->refs is dead, | |
249 | * readers must ensure that ->refs is alive as long as the file* is | |
250 | * used. Only updated through io_uring_register(2). | |
251 | */ | |
252 | struct file **user_files; | |
253 | unsigned nr_user_files; | |
254 | ||
edafccee JA |
255 | /* if used, fixed mapped user buffers */ |
256 | unsigned nr_user_bufs; | |
257 | struct io_mapped_ubuf *user_bufs; | |
258 | ||
2b188cc1 JA |
259 | struct user_struct *user; |
260 | ||
261 | struct completion ctx_done; | |
262 | ||
263 | struct { | |
264 | struct mutex uring_lock; | |
265 | wait_queue_head_t wait; | |
266 | } ____cacheline_aligned_in_smp; | |
267 | ||
268 | struct { | |
269 | spinlock_t completion_lock; | |
def596e9 JA |
270 | bool poll_multi_file; |
271 | /* | |
272 | * ->poll_list is protected by the ctx->uring_lock for | |
273 | * io_uring instances that don't use IORING_SETUP_SQPOLL. | |
274 | * For SQPOLL, only the single threaded io_sq_thread() will | |
275 | * manipulate the list, hence no extra locking is needed there. | |
276 | */ | |
277 | struct list_head poll_list; | |
221c5eb2 | 278 | struct list_head cancel_list; |
2b188cc1 JA |
279 | } ____cacheline_aligned_in_smp; |
280 | ||
31b51510 JA |
281 | struct async_list pending_async[2]; |
282 | ||
2b188cc1 JA |
283 | #if defined(CONFIG_UNIX) |
284 | struct socket *ring_sock; | |
285 | #endif | |
286 | }; | |
287 | ||
288 | struct sqe_submit { | |
289 | const struct io_uring_sqe *sqe; | |
290 | unsigned short index; | |
291 | bool has_user; | |
def596e9 | 292 | bool needs_lock; |
6c271ce2 | 293 | bool needs_fixed_file; |
2b188cc1 JA |
294 | }; |
295 | ||
09bb8394 JA |
296 | /* |
297 | * First field must be the file pointer in all the | |
298 | * iocb unions! See also 'struct kiocb' in <linux/fs.h> | |
299 | */ | |
221c5eb2 JA |
300 | struct io_poll_iocb { |
301 | struct file *file; | |
302 | struct wait_queue_head *head; | |
303 | __poll_t events; | |
8c838788 | 304 | bool done; |
221c5eb2 JA |
305 | bool canceled; |
306 | struct wait_queue_entry wait; | |
307 | }; | |
308 | ||
09bb8394 JA |
309 | /* |
310 | * NOTE! Each of the iocb union members has the file pointer | |
311 | * as the first entry in their struct definition. So you can | |
312 | * access the file pointer through any of the sub-structs, | |
313 | * or directly as just 'ki_filp' in this struct. | |
314 | */ | |
2b188cc1 | 315 | struct io_kiocb { |
221c5eb2 | 316 | union { |
09bb8394 | 317 | struct file *file; |
221c5eb2 JA |
318 | struct kiocb rw; |
319 | struct io_poll_iocb poll; | |
320 | }; | |
2b188cc1 JA |
321 | |
322 | struct sqe_submit submit; | |
323 | ||
324 | struct io_ring_ctx *ctx; | |
325 | struct list_head list; | |
9e645e11 | 326 | struct list_head link_list; |
2b188cc1 | 327 | unsigned int flags; |
c16361c1 | 328 | refcount_t refs; |
8449eeda | 329 | #define REQ_F_NOWAIT 1 /* must not punt to workers */ |
def596e9 | 330 | #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */ |
6b06314c | 331 | #define REQ_F_FIXED_FILE 4 /* ctx owns file */ |
31b51510 | 332 | #define REQ_F_SEQ_PREV 8 /* sequential with previous */ |
e2033e33 SB |
333 | #define REQ_F_IO_DRAIN 16 /* drain existing IO first */ |
334 | #define REQ_F_IO_DRAINED 32 /* drain done */ | |
9e645e11 | 335 | #define REQ_F_LINK 64 /* linked sqes */ |
f7b76ac9 ZL |
336 | #define REQ_F_LINK_DONE 128 /* linked sqes done */ |
337 | #define REQ_F_FAIL_LINK 256 /* fail rest of links */ | |
2b188cc1 | 338 | u64 user_data; |
9e645e11 | 339 | u32 result; |
de0617e4 | 340 | u32 sequence; |
2b188cc1 JA |
341 | |
342 | struct work_struct work; | |
343 | }; | |
344 | ||
345 | #define IO_PLUG_THRESHOLD 2 | |
def596e9 | 346 | #define IO_IOPOLL_BATCH 8 |
2b188cc1 | 347 | |
9a56a232 JA |
348 | struct io_submit_state { |
349 | struct blk_plug plug; | |
350 | ||
2579f913 JA |
351 | /* |
352 | * io_kiocb alloc cache | |
353 | */ | |
354 | void *reqs[IO_IOPOLL_BATCH]; | |
355 | unsigned int free_reqs; | |
356 | unsigned int cur_req; | |
357 | ||
9a56a232 JA |
358 | /* |
359 | * File reference cache | |
360 | */ | |
361 | struct file *file; | |
362 | unsigned int fd; | |
363 | unsigned int has_refs; | |
364 | unsigned int used_refs; | |
365 | unsigned int ios_left; | |
366 | }; | |
367 | ||
de0617e4 JA |
368 | static void io_sq_wq_submit_work(struct work_struct *work); |
369 | ||
2b188cc1 JA |
370 | static struct kmem_cache *req_cachep; |
371 | ||
372 | static const struct file_operations io_uring_fops; | |
373 | ||
374 | struct sock *io_uring_get_socket(struct file *file) | |
375 | { | |
376 | #if defined(CONFIG_UNIX) | |
377 | if (file->f_op == &io_uring_fops) { | |
378 | struct io_ring_ctx *ctx = file->private_data; | |
379 | ||
380 | return ctx->ring_sock->sk; | |
381 | } | |
382 | #endif | |
383 | return NULL; | |
384 | } | |
385 | EXPORT_SYMBOL(io_uring_get_socket); | |
386 | ||
387 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) | |
388 | { | |
389 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
390 | ||
391 | complete(&ctx->ctx_done); | |
392 | } | |
393 | ||
394 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) | |
395 | { | |
396 | struct io_ring_ctx *ctx; | |
31b51510 | 397 | int i; |
2b188cc1 JA |
398 | |
399 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
400 | if (!ctx) | |
401 | return NULL; | |
402 | ||
21482896 RG |
403 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
404 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) { | |
2b188cc1 JA |
405 | kfree(ctx); |
406 | return NULL; | |
407 | } | |
408 | ||
409 | ctx->flags = p->flags; | |
410 | init_waitqueue_head(&ctx->cq_wait); | |
411 | init_completion(&ctx->ctx_done); | |
a4c0b3de | 412 | init_completion(&ctx->sqo_thread_started); |
2b188cc1 JA |
413 | mutex_init(&ctx->uring_lock); |
414 | init_waitqueue_head(&ctx->wait); | |
31b51510 JA |
415 | for (i = 0; i < ARRAY_SIZE(ctx->pending_async); i++) { |
416 | spin_lock_init(&ctx->pending_async[i].lock); | |
417 | INIT_LIST_HEAD(&ctx->pending_async[i].list); | |
418 | atomic_set(&ctx->pending_async[i].cnt, 0); | |
419 | } | |
2b188cc1 | 420 | spin_lock_init(&ctx->completion_lock); |
def596e9 | 421 | INIT_LIST_HEAD(&ctx->poll_list); |
221c5eb2 | 422 | INIT_LIST_HEAD(&ctx->cancel_list); |
de0617e4 | 423 | INIT_LIST_HEAD(&ctx->defer_list); |
2b188cc1 JA |
424 | return ctx; |
425 | } | |
426 | ||
de0617e4 JA |
427 | static inline bool io_sequence_defer(struct io_ring_ctx *ctx, |
428 | struct io_kiocb *req) | |
429 | { | |
430 | if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) != REQ_F_IO_DRAIN) | |
431 | return false; | |
432 | ||
dbd0f6d6 | 433 | return req->sequence != ctx->cached_cq_tail + ctx->sq_ring->dropped; |
de0617e4 JA |
434 | } |
435 | ||
436 | static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx) | |
437 | { | |
438 | struct io_kiocb *req; | |
439 | ||
440 | if (list_empty(&ctx->defer_list)) | |
441 | return NULL; | |
442 | ||
443 | req = list_first_entry(&ctx->defer_list, struct io_kiocb, list); | |
444 | if (!io_sequence_defer(ctx, req)) { | |
445 | list_del_init(&req->list); | |
446 | return req; | |
447 | } | |
448 | ||
449 | return NULL; | |
450 | } | |
451 | ||
452 | static void __io_commit_cqring(struct io_ring_ctx *ctx) | |
2b188cc1 JA |
453 | { |
454 | struct io_cq_ring *ring = ctx->cq_ring; | |
455 | ||
456 | if (ctx->cached_cq_tail != READ_ONCE(ring->r.tail)) { | |
457 | /* order cqe stores with ring update */ | |
458 | smp_store_release(&ring->r.tail, ctx->cached_cq_tail); | |
459 | ||
2b188cc1 JA |
460 | if (wq_has_sleeper(&ctx->cq_wait)) { |
461 | wake_up_interruptible(&ctx->cq_wait); | |
462 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
463 | } | |
464 | } | |
465 | } | |
466 | ||
de0617e4 JA |
467 | static void io_commit_cqring(struct io_ring_ctx *ctx) |
468 | { | |
469 | struct io_kiocb *req; | |
470 | ||
471 | __io_commit_cqring(ctx); | |
472 | ||
473 | while ((req = io_get_deferred_req(ctx)) != NULL) { | |
474 | req->flags |= REQ_F_IO_DRAINED; | |
475 | queue_work(ctx->sqo_wq, &req->work); | |
476 | } | |
477 | } | |
478 | ||
2b188cc1 JA |
479 | static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx) |
480 | { | |
481 | struct io_cq_ring *ring = ctx->cq_ring; | |
482 | unsigned tail; | |
483 | ||
484 | tail = ctx->cached_cq_tail; | |
115e12e5 SB |
485 | /* |
486 | * writes to the cq entry need to come after reading head; the | |
487 | * control dependency is enough as we're using WRITE_ONCE to | |
488 | * fill the cq entry | |
489 | */ | |
74f464e9 | 490 | if (tail - READ_ONCE(ring->r.head) == ring->ring_entries) |
2b188cc1 JA |
491 | return NULL; |
492 | ||
493 | ctx->cached_cq_tail++; | |
494 | return &ring->cqes[tail & ctx->cq_mask]; | |
495 | } | |
496 | ||
497 | static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data, | |
c71ffb67 | 498 | long res) |
2b188cc1 JA |
499 | { |
500 | struct io_uring_cqe *cqe; | |
501 | ||
502 | /* | |
503 | * If we can't get a cq entry, userspace overflowed the | |
504 | * submission (by quite a lot). Increment the overflow count in | |
505 | * the ring. | |
506 | */ | |
507 | cqe = io_get_cqring(ctx); | |
508 | if (cqe) { | |
509 | WRITE_ONCE(cqe->user_data, ki_user_data); | |
510 | WRITE_ONCE(cqe->res, res); | |
c71ffb67 | 511 | WRITE_ONCE(cqe->flags, 0); |
2b188cc1 JA |
512 | } else { |
513 | unsigned overflow = READ_ONCE(ctx->cq_ring->overflow); | |
514 | ||
515 | WRITE_ONCE(ctx->cq_ring->overflow, overflow + 1); | |
516 | } | |
517 | } | |
518 | ||
8c838788 JA |
519 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
520 | { | |
521 | if (waitqueue_active(&ctx->wait)) | |
522 | wake_up(&ctx->wait); | |
523 | if (waitqueue_active(&ctx->sqo_wait)) | |
524 | wake_up(&ctx->sqo_wait); | |
9b402849 JA |
525 | if (ctx->cq_ev_fd) |
526 | eventfd_signal(ctx->cq_ev_fd, 1); | |
8c838788 JA |
527 | } |
528 | ||
529 | static void io_cqring_add_event(struct io_ring_ctx *ctx, u64 user_data, | |
c71ffb67 | 530 | long res) |
2b188cc1 JA |
531 | { |
532 | unsigned long flags; | |
533 | ||
534 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
c71ffb67 | 535 | io_cqring_fill_event(ctx, user_data, res); |
2b188cc1 JA |
536 | io_commit_cqring(ctx); |
537 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
538 | ||
8c838788 | 539 | io_cqring_ev_posted(ctx); |
2b188cc1 JA |
540 | } |
541 | ||
542 | static void io_ring_drop_ctx_refs(struct io_ring_ctx *ctx, unsigned refs) | |
543 | { | |
544 | percpu_ref_put_many(&ctx->refs, refs); | |
545 | ||
546 | if (waitqueue_active(&ctx->wait)) | |
547 | wake_up(&ctx->wait); | |
548 | } | |
549 | ||
2579f913 JA |
550 | static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx, |
551 | struct io_submit_state *state) | |
2b188cc1 | 552 | { |
fd6fab2c | 553 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
2b188cc1 JA |
554 | struct io_kiocb *req; |
555 | ||
556 | if (!percpu_ref_tryget(&ctx->refs)) | |
557 | return NULL; | |
558 | ||
2579f913 | 559 | if (!state) { |
fd6fab2c | 560 | req = kmem_cache_alloc(req_cachep, gfp); |
2579f913 JA |
561 | if (unlikely(!req)) |
562 | goto out; | |
563 | } else if (!state->free_reqs) { | |
564 | size_t sz; | |
565 | int ret; | |
566 | ||
567 | sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs)); | |
fd6fab2c JA |
568 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs); |
569 | ||
570 | /* | |
571 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
572 | * retry single alloc to be on the safe side. | |
573 | */ | |
574 | if (unlikely(ret <= 0)) { | |
575 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
576 | if (!state->reqs[0]) | |
577 | goto out; | |
578 | ret = 1; | |
579 | } | |
2579f913 JA |
580 | state->free_reqs = ret - 1; |
581 | state->cur_req = 1; | |
582 | req = state->reqs[0]; | |
583 | } else { | |
584 | req = state->reqs[state->cur_req]; | |
585 | state->free_reqs--; | |
586 | state->cur_req++; | |
2b188cc1 JA |
587 | } |
588 | ||
60c112b0 | 589 | req->file = NULL; |
2579f913 JA |
590 | req->ctx = ctx; |
591 | req->flags = 0; | |
e65ef56d JA |
592 | /* one is dropped after submission, the other at completion */ |
593 | refcount_set(&req->refs, 2); | |
9e645e11 | 594 | req->result = 0; |
2579f913 JA |
595 | return req; |
596 | out: | |
2b188cc1 JA |
597 | io_ring_drop_ctx_refs(ctx, 1); |
598 | return NULL; | |
599 | } | |
600 | ||
def596e9 JA |
601 | static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr) |
602 | { | |
603 | if (*nr) { | |
604 | kmem_cache_free_bulk(req_cachep, *nr, reqs); | |
605 | io_ring_drop_ctx_refs(ctx, *nr); | |
606 | *nr = 0; | |
607 | } | |
608 | } | |
609 | ||
9e645e11 | 610 | static void __io_free_req(struct io_kiocb *req) |
2b188cc1 | 611 | { |
09bb8394 JA |
612 | if (req->file && !(req->flags & REQ_F_FIXED_FILE)) |
613 | fput(req->file); | |
e65ef56d JA |
614 | io_ring_drop_ctx_refs(req->ctx, 1); |
615 | kmem_cache_free(req_cachep, req); | |
616 | } | |
617 | ||
9e645e11 JA |
618 | static void io_req_link_next(struct io_kiocb *req) |
619 | { | |
620 | struct io_kiocb *nxt; | |
621 | ||
622 | /* | |
623 | * The list should never be empty when we are called here. But could | |
624 | * potentially happen if the chain is messed up, check to be on the | |
625 | * safe side. | |
626 | */ | |
627 | nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list); | |
628 | if (nxt) { | |
629 | list_del(&nxt->list); | |
630 | if (!list_empty(&req->link_list)) { | |
631 | INIT_LIST_HEAD(&nxt->link_list); | |
632 | list_splice(&req->link_list, &nxt->link_list); | |
633 | nxt->flags |= REQ_F_LINK; | |
634 | } | |
635 | ||
f7b76ac9 | 636 | nxt->flags |= REQ_F_LINK_DONE; |
9e645e11 JA |
637 | INIT_WORK(&nxt->work, io_sq_wq_submit_work); |
638 | queue_work(req->ctx->sqo_wq, &nxt->work); | |
639 | } | |
640 | } | |
641 | ||
642 | /* | |
643 | * Called if REQ_F_LINK is set, and we fail the head request | |
644 | */ | |
645 | static void io_fail_links(struct io_kiocb *req) | |
646 | { | |
647 | struct io_kiocb *link; | |
648 | ||
649 | while (!list_empty(&req->link_list)) { | |
650 | link = list_first_entry(&req->link_list, struct io_kiocb, list); | |
651 | list_del(&link->list); | |
652 | ||
653 | io_cqring_add_event(req->ctx, link->user_data, -ECANCELED); | |
654 | __io_free_req(link); | |
655 | } | |
656 | } | |
657 | ||
658 | static void io_free_req(struct io_kiocb *req) | |
659 | { | |
660 | /* | |
661 | * If LINK is set, we have dependent requests in this chain. If we | |
662 | * didn't fail this request, queue the first one up, moving any other | |
663 | * dependencies to the next request. In case of failure, fail the rest | |
664 | * of the chain. | |
665 | */ | |
666 | if (req->flags & REQ_F_LINK) { | |
667 | if (req->flags & REQ_F_FAIL_LINK) | |
668 | io_fail_links(req); | |
669 | else | |
670 | io_req_link_next(req); | |
671 | } | |
672 | ||
673 | __io_free_req(req); | |
674 | } | |
675 | ||
e65ef56d JA |
676 | static void io_put_req(struct io_kiocb *req) |
677 | { | |
678 | if (refcount_dec_and_test(&req->refs)) | |
679 | io_free_req(req); | |
2b188cc1 JA |
680 | } |
681 | ||
a3a0e43f JA |
682 | static unsigned io_cqring_events(struct io_cq_ring *ring) |
683 | { | |
684 | /* See comment at the top of this file */ | |
685 | smp_rmb(); | |
686 | return READ_ONCE(ring->r.tail) - READ_ONCE(ring->r.head); | |
687 | } | |
688 | ||
def596e9 JA |
689 | /* |
690 | * Find and free completed poll iocbs | |
691 | */ | |
692 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
693 | struct list_head *done) | |
694 | { | |
695 | void *reqs[IO_IOPOLL_BATCH]; | |
696 | struct io_kiocb *req; | |
09bb8394 | 697 | int to_free; |
def596e9 | 698 | |
09bb8394 | 699 | to_free = 0; |
def596e9 JA |
700 | while (!list_empty(done)) { |
701 | req = list_first_entry(done, struct io_kiocb, list); | |
702 | list_del(&req->list); | |
703 | ||
9e645e11 | 704 | io_cqring_fill_event(ctx, req->user_data, req->result); |
def596e9 JA |
705 | (*nr_events)++; |
706 | ||
09bb8394 JA |
707 | if (refcount_dec_and_test(&req->refs)) { |
708 | /* If we're not using fixed files, we have to pair the | |
709 | * completion part with the file put. Use regular | |
710 | * completions for those, only batch free for fixed | |
9e645e11 | 711 | * file and non-linked commands. |
09bb8394 | 712 | */ |
9e645e11 JA |
713 | if ((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) == |
714 | REQ_F_FIXED_FILE) { | |
09bb8394 JA |
715 | reqs[to_free++] = req; |
716 | if (to_free == ARRAY_SIZE(reqs)) | |
717 | io_free_req_many(ctx, reqs, &to_free); | |
6b06314c | 718 | } else { |
09bb8394 | 719 | io_free_req(req); |
6b06314c | 720 | } |
9a56a232 | 721 | } |
def596e9 | 722 | } |
def596e9 | 723 | |
09bb8394 | 724 | io_commit_cqring(ctx); |
def596e9 JA |
725 | io_free_req_many(ctx, reqs, &to_free); |
726 | } | |
727 | ||
728 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
729 | long min) | |
730 | { | |
731 | struct io_kiocb *req, *tmp; | |
732 | LIST_HEAD(done); | |
733 | bool spin; | |
734 | int ret; | |
735 | ||
736 | /* | |
737 | * Only spin for completions if we don't have multiple devices hanging | |
738 | * off our complete list, and we're under the requested amount. | |
739 | */ | |
740 | spin = !ctx->poll_multi_file && *nr_events < min; | |
741 | ||
742 | ret = 0; | |
743 | list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) { | |
744 | struct kiocb *kiocb = &req->rw; | |
745 | ||
746 | /* | |
747 | * Move completed entries to our local list. If we find a | |
748 | * request that requires polling, break out and complete | |
749 | * the done list first, if we have entries there. | |
750 | */ | |
751 | if (req->flags & REQ_F_IOPOLL_COMPLETED) { | |
752 | list_move_tail(&req->list, &done); | |
753 | continue; | |
754 | } | |
755 | if (!list_empty(&done)) | |
756 | break; | |
757 | ||
758 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
759 | if (ret < 0) | |
760 | break; | |
761 | ||
762 | if (ret && spin) | |
763 | spin = false; | |
764 | ret = 0; | |
765 | } | |
766 | ||
767 | if (!list_empty(&done)) | |
768 | io_iopoll_complete(ctx, nr_events, &done); | |
769 | ||
770 | return ret; | |
771 | } | |
772 | ||
773 | /* | |
774 | * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a | |
775 | * non-spinning poll check - we'll still enter the driver poll loop, but only | |
776 | * as a non-spinning completion check. | |
777 | */ | |
778 | static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
779 | long min) | |
780 | { | |
781 | while (!list_empty(&ctx->poll_list)) { | |
782 | int ret; | |
783 | ||
784 | ret = io_do_iopoll(ctx, nr_events, min); | |
785 | if (ret < 0) | |
786 | return ret; | |
787 | if (!min || *nr_events >= min) | |
788 | return 0; | |
789 | } | |
790 | ||
791 | return 1; | |
792 | } | |
793 | ||
794 | /* | |
795 | * We can't just wait for polled events to come to us, we have to actively | |
796 | * find and complete them. | |
797 | */ | |
798 | static void io_iopoll_reap_events(struct io_ring_ctx *ctx) | |
799 | { | |
800 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
801 | return; | |
802 | ||
803 | mutex_lock(&ctx->uring_lock); | |
804 | while (!list_empty(&ctx->poll_list)) { | |
805 | unsigned int nr_events = 0; | |
806 | ||
807 | io_iopoll_getevents(ctx, &nr_events, 1); | |
808 | } | |
809 | mutex_unlock(&ctx->uring_lock); | |
810 | } | |
811 | ||
812 | static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, | |
813 | long min) | |
814 | { | |
500f9fba JA |
815 | int iters, ret = 0; |
816 | ||
817 | /* | |
818 | * We disallow the app entering submit/complete with polling, but we | |
819 | * still need to lock the ring to prevent racing with polled issue | |
820 | * that got punted to a workqueue. | |
821 | */ | |
822 | mutex_lock(&ctx->uring_lock); | |
def596e9 | 823 | |
500f9fba | 824 | iters = 0; |
def596e9 JA |
825 | do { |
826 | int tmin = 0; | |
827 | ||
a3a0e43f JA |
828 | /* |
829 | * Don't enter poll loop if we already have events pending. | |
830 | * If we do, we can potentially be spinning for commands that | |
831 | * already triggered a CQE (eg in error). | |
832 | */ | |
833 | if (io_cqring_events(ctx->cq_ring)) | |
834 | break; | |
835 | ||
500f9fba JA |
836 | /* |
837 | * If a submit got punted to a workqueue, we can have the | |
838 | * application entering polling for a command before it gets | |
839 | * issued. That app will hold the uring_lock for the duration | |
840 | * of the poll right here, so we need to take a breather every | |
841 | * now and then to ensure that the issue has a chance to add | |
842 | * the poll to the issued list. Otherwise we can spin here | |
843 | * forever, while the workqueue is stuck trying to acquire the | |
844 | * very same mutex. | |
845 | */ | |
846 | if (!(++iters & 7)) { | |
847 | mutex_unlock(&ctx->uring_lock); | |
848 | mutex_lock(&ctx->uring_lock); | |
849 | } | |
850 | ||
def596e9 JA |
851 | if (*nr_events < min) |
852 | tmin = min - *nr_events; | |
853 | ||
854 | ret = io_iopoll_getevents(ctx, nr_events, tmin); | |
855 | if (ret <= 0) | |
856 | break; | |
857 | ret = 0; | |
858 | } while (min && !*nr_events && !need_resched()); | |
859 | ||
500f9fba | 860 | mutex_unlock(&ctx->uring_lock); |
def596e9 JA |
861 | return ret; |
862 | } | |
863 | ||
2b188cc1 JA |
864 | static void kiocb_end_write(struct kiocb *kiocb) |
865 | { | |
866 | if (kiocb->ki_flags & IOCB_WRITE) { | |
867 | struct inode *inode = file_inode(kiocb->ki_filp); | |
868 | ||
869 | /* | |
870 | * Tell lockdep we inherited freeze protection from submission | |
871 | * thread. | |
872 | */ | |
873 | if (S_ISREG(inode->i_mode)) | |
874 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); | |
875 | file_end_write(kiocb->ki_filp); | |
876 | } | |
877 | } | |
878 | ||
879 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
880 | { | |
881 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
882 | ||
883 | kiocb_end_write(kiocb); | |
884 | ||
9e645e11 JA |
885 | if ((req->flags & REQ_F_LINK) && res != req->result) |
886 | req->flags |= REQ_F_FAIL_LINK; | |
c71ffb67 | 887 | io_cqring_add_event(req->ctx, req->user_data, res); |
e65ef56d | 888 | io_put_req(req); |
2b188cc1 JA |
889 | } |
890 | ||
def596e9 JA |
891 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
892 | { | |
893 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
894 | ||
895 | kiocb_end_write(kiocb); | |
896 | ||
9e645e11 JA |
897 | if ((req->flags & REQ_F_LINK) && res != req->result) |
898 | req->flags |= REQ_F_FAIL_LINK; | |
899 | req->result = res; | |
def596e9 JA |
900 | if (res != -EAGAIN) |
901 | req->flags |= REQ_F_IOPOLL_COMPLETED; | |
902 | } | |
903 | ||
904 | /* | |
905 | * After the iocb has been issued, it's safe to be found on the poll list. | |
906 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
907 | * find it from a io_iopoll_getevents() thread before the issuer is done | |
908 | * accessing the kiocb cookie. | |
909 | */ | |
910 | static void io_iopoll_req_issued(struct io_kiocb *req) | |
911 | { | |
912 | struct io_ring_ctx *ctx = req->ctx; | |
913 | ||
914 | /* | |
915 | * Track whether we have multiple files in our lists. This will impact | |
916 | * how we do polling eventually, not spinning if we're on potentially | |
917 | * different devices. | |
918 | */ | |
919 | if (list_empty(&ctx->poll_list)) { | |
920 | ctx->poll_multi_file = false; | |
921 | } else if (!ctx->poll_multi_file) { | |
922 | struct io_kiocb *list_req; | |
923 | ||
924 | list_req = list_first_entry(&ctx->poll_list, struct io_kiocb, | |
925 | list); | |
926 | if (list_req->rw.ki_filp != req->rw.ki_filp) | |
927 | ctx->poll_multi_file = true; | |
928 | } | |
929 | ||
930 | /* | |
931 | * For fast devices, IO may have already completed. If it has, add | |
932 | * it to the front so we find it first. | |
933 | */ | |
934 | if (req->flags & REQ_F_IOPOLL_COMPLETED) | |
935 | list_add(&req->list, &ctx->poll_list); | |
936 | else | |
937 | list_add_tail(&req->list, &ctx->poll_list); | |
938 | } | |
939 | ||
3d6770fb | 940 | static void io_file_put(struct io_submit_state *state) |
9a56a232 | 941 | { |
3d6770fb | 942 | if (state->file) { |
9a56a232 JA |
943 | int diff = state->has_refs - state->used_refs; |
944 | ||
945 | if (diff) | |
946 | fput_many(state->file, diff); | |
947 | state->file = NULL; | |
948 | } | |
949 | } | |
950 | ||
951 | /* | |
952 | * Get as many references to a file as we have IOs left in this submission, | |
953 | * assuming most submissions are for one file, or at least that each file | |
954 | * has more than one submission. | |
955 | */ | |
956 | static struct file *io_file_get(struct io_submit_state *state, int fd) | |
957 | { | |
958 | if (!state) | |
959 | return fget(fd); | |
960 | ||
961 | if (state->file) { | |
962 | if (state->fd == fd) { | |
963 | state->used_refs++; | |
964 | state->ios_left--; | |
965 | return state->file; | |
966 | } | |
3d6770fb | 967 | io_file_put(state); |
9a56a232 JA |
968 | } |
969 | state->file = fget_many(fd, state->ios_left); | |
970 | if (!state->file) | |
971 | return NULL; | |
972 | ||
973 | state->fd = fd; | |
974 | state->has_refs = state->ios_left; | |
975 | state->used_refs = 1; | |
976 | state->ios_left--; | |
977 | return state->file; | |
978 | } | |
979 | ||
2b188cc1 JA |
980 | /* |
981 | * If we tracked the file through the SCM inflight mechanism, we could support | |
982 | * any file. For now, just ensure that anything potentially problematic is done | |
983 | * inline. | |
984 | */ | |
985 | static bool io_file_supports_async(struct file *file) | |
986 | { | |
987 | umode_t mode = file_inode(file)->i_mode; | |
988 | ||
989 | if (S_ISBLK(mode) || S_ISCHR(mode)) | |
990 | return true; | |
991 | if (S_ISREG(mode) && file->f_op != &io_uring_fops) | |
992 | return true; | |
993 | ||
994 | return false; | |
995 | } | |
996 | ||
6c271ce2 | 997 | static int io_prep_rw(struct io_kiocb *req, const struct sqe_submit *s, |
8358e3a8 | 998 | bool force_nonblock) |
2b188cc1 | 999 | { |
6c271ce2 | 1000 | const struct io_uring_sqe *sqe = s->sqe; |
def596e9 | 1001 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 | 1002 | struct kiocb *kiocb = &req->rw; |
09bb8394 JA |
1003 | unsigned ioprio; |
1004 | int ret; | |
2b188cc1 | 1005 | |
09bb8394 JA |
1006 | if (!req->file) |
1007 | return -EBADF; | |
2b188cc1 | 1008 | |
09bb8394 JA |
1009 | if (force_nonblock && !io_file_supports_async(req->file)) |
1010 | force_nonblock = false; | |
6b06314c | 1011 | |
2b188cc1 JA |
1012 | kiocb->ki_pos = READ_ONCE(sqe->off); |
1013 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); | |
1014 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); | |
1015 | ||
1016 | ioprio = READ_ONCE(sqe->ioprio); | |
1017 | if (ioprio) { | |
1018 | ret = ioprio_check_cap(ioprio); | |
1019 | if (ret) | |
09bb8394 | 1020 | return ret; |
2b188cc1 JA |
1021 | |
1022 | kiocb->ki_ioprio = ioprio; | |
1023 | } else | |
1024 | kiocb->ki_ioprio = get_current_ioprio(); | |
1025 | ||
1026 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
1027 | if (unlikely(ret)) | |
09bb8394 | 1028 | return ret; |
8449eeda SB |
1029 | |
1030 | /* don't allow async punt if RWF_NOWAIT was requested */ | |
1031 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
1032 | req->flags |= REQ_F_NOWAIT; | |
1033 | ||
1034 | if (force_nonblock) | |
2b188cc1 | 1035 | kiocb->ki_flags |= IOCB_NOWAIT; |
8449eeda | 1036 | |
def596e9 | 1037 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
1038 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
1039 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 1040 | return -EOPNOTSUPP; |
2b188cc1 | 1041 | |
def596e9 JA |
1042 | kiocb->ki_flags |= IOCB_HIPRI; |
1043 | kiocb->ki_complete = io_complete_rw_iopoll; | |
1044 | } else { | |
09bb8394 JA |
1045 | if (kiocb->ki_flags & IOCB_HIPRI) |
1046 | return -EINVAL; | |
def596e9 JA |
1047 | kiocb->ki_complete = io_complete_rw; |
1048 | } | |
2b188cc1 | 1049 | return 0; |
2b188cc1 JA |
1050 | } |
1051 | ||
1052 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
1053 | { | |
1054 | switch (ret) { | |
1055 | case -EIOCBQUEUED: | |
1056 | break; | |
1057 | case -ERESTARTSYS: | |
1058 | case -ERESTARTNOINTR: | |
1059 | case -ERESTARTNOHAND: | |
1060 | case -ERESTART_RESTARTBLOCK: | |
1061 | /* | |
1062 | * We can't just restart the syscall, since previously | |
1063 | * submitted sqes may already be in progress. Just fail this | |
1064 | * IO with EINTR. | |
1065 | */ | |
1066 | ret = -EINTR; | |
1067 | /* fall through */ | |
1068 | default: | |
1069 | kiocb->ki_complete(kiocb, ret, 0); | |
1070 | } | |
1071 | } | |
1072 | ||
edafccee JA |
1073 | static int io_import_fixed(struct io_ring_ctx *ctx, int rw, |
1074 | const struct io_uring_sqe *sqe, | |
1075 | struct iov_iter *iter) | |
1076 | { | |
1077 | size_t len = READ_ONCE(sqe->len); | |
1078 | struct io_mapped_ubuf *imu; | |
1079 | unsigned index, buf_index; | |
1080 | size_t offset; | |
1081 | u64 buf_addr; | |
1082 | ||
1083 | /* attempt to use fixed buffers without having provided iovecs */ | |
1084 | if (unlikely(!ctx->user_bufs)) | |
1085 | return -EFAULT; | |
1086 | ||
1087 | buf_index = READ_ONCE(sqe->buf_index); | |
1088 | if (unlikely(buf_index >= ctx->nr_user_bufs)) | |
1089 | return -EFAULT; | |
1090 | ||
1091 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); | |
1092 | imu = &ctx->user_bufs[index]; | |
1093 | buf_addr = READ_ONCE(sqe->addr); | |
1094 | ||
1095 | /* overflow */ | |
1096 | if (buf_addr + len < buf_addr) | |
1097 | return -EFAULT; | |
1098 | /* not inside the mapped region */ | |
1099 | if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len) | |
1100 | return -EFAULT; | |
1101 | ||
1102 | /* | |
1103 | * May not be a start of buffer, set size appropriately | |
1104 | * and advance us to the beginning. | |
1105 | */ | |
1106 | offset = buf_addr - imu->ubuf; | |
1107 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
bd11b3a3 JA |
1108 | |
1109 | if (offset) { | |
1110 | /* | |
1111 | * Don't use iov_iter_advance() here, as it's really slow for | |
1112 | * using the latter parts of a big fixed buffer - it iterates | |
1113 | * over each segment manually. We can cheat a bit here, because | |
1114 | * we know that: | |
1115 | * | |
1116 | * 1) it's a BVEC iter, we set it up | |
1117 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
1118 | * first and last bvec | |
1119 | * | |
1120 | * So just find our index, and adjust the iterator afterwards. | |
1121 | * If the offset is within the first bvec (or the whole first | |
1122 | * bvec, just use iov_iter_advance(). This makes it easier | |
1123 | * since we can just skip the first segment, which may not | |
1124 | * be PAGE_SIZE aligned. | |
1125 | */ | |
1126 | const struct bio_vec *bvec = imu->bvec; | |
1127 | ||
1128 | if (offset <= bvec->bv_len) { | |
1129 | iov_iter_advance(iter, offset); | |
1130 | } else { | |
1131 | unsigned long seg_skip; | |
1132 | ||
1133 | /* skip first vec */ | |
1134 | offset -= bvec->bv_len; | |
1135 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
1136 | ||
1137 | iter->bvec = bvec + seg_skip; | |
1138 | iter->nr_segs -= seg_skip; | |
99c79f66 | 1139 | iter->count -= bvec->bv_len + offset; |
bd11b3a3 | 1140 | iter->iov_offset = offset & ~PAGE_MASK; |
bd11b3a3 JA |
1141 | } |
1142 | } | |
1143 | ||
edafccee JA |
1144 | return 0; |
1145 | } | |
1146 | ||
87e5e6da JA |
1147 | static ssize_t io_import_iovec(struct io_ring_ctx *ctx, int rw, |
1148 | const struct sqe_submit *s, struct iovec **iovec, | |
1149 | struct iov_iter *iter) | |
2b188cc1 JA |
1150 | { |
1151 | const struct io_uring_sqe *sqe = s->sqe; | |
1152 | void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
1153 | size_t sqe_len = READ_ONCE(sqe->len); | |
edafccee JA |
1154 | u8 opcode; |
1155 | ||
1156 | /* | |
1157 | * We're reading ->opcode for the second time, but the first read | |
1158 | * doesn't care whether it's _FIXED or not, so it doesn't matter | |
1159 | * whether ->opcode changes concurrently. The first read does care | |
1160 | * about whether it is a READ or a WRITE, so we don't trust this read | |
1161 | * for that purpose and instead let the caller pass in the read/write | |
1162 | * flag. | |
1163 | */ | |
1164 | opcode = READ_ONCE(sqe->opcode); | |
1165 | if (opcode == IORING_OP_READ_FIXED || | |
1166 | opcode == IORING_OP_WRITE_FIXED) { | |
87e5e6da | 1167 | ssize_t ret = io_import_fixed(ctx, rw, sqe, iter); |
edafccee JA |
1168 | *iovec = NULL; |
1169 | return ret; | |
1170 | } | |
2b188cc1 JA |
1171 | |
1172 | if (!s->has_user) | |
1173 | return -EFAULT; | |
1174 | ||
1175 | #ifdef CONFIG_COMPAT | |
1176 | if (ctx->compat) | |
1177 | return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV, | |
1178 | iovec, iter); | |
1179 | #endif | |
1180 | ||
1181 | return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter); | |
1182 | } | |
1183 | ||
31b51510 JA |
1184 | /* |
1185 | * Make a note of the last file/offset/direction we punted to async | |
1186 | * context. We'll use this information to see if we can piggy back a | |
1187 | * sequential request onto the previous one, if it's still hasn't been | |
1188 | * completed by the async worker. | |
1189 | */ | |
1190 | static void io_async_list_note(int rw, struct io_kiocb *req, size_t len) | |
1191 | { | |
1192 | struct async_list *async_list = &req->ctx->pending_async[rw]; | |
1193 | struct kiocb *kiocb = &req->rw; | |
1194 | struct file *filp = kiocb->ki_filp; | |
1195 | off_t io_end = kiocb->ki_pos + len; | |
1196 | ||
1197 | if (filp == async_list->file && kiocb->ki_pos == async_list->io_end) { | |
9310a7ba | 1198 | unsigned long max_bytes; |
31b51510 JA |
1199 | |
1200 | /* Use 8x RA size as a decent limiter for both reads/writes */ | |
9310a7ba ZL |
1201 | max_bytes = filp->f_ra.ra_pages << (PAGE_SHIFT + 3); |
1202 | if (!max_bytes) | |
1203 | max_bytes = VM_READAHEAD_PAGES << (PAGE_SHIFT + 3); | |
1204 | ||
1205 | /* If max len are exceeded, reset the state */ | |
1206 | if (async_list->io_len + len <= max_bytes) { | |
31b51510 | 1207 | req->flags |= REQ_F_SEQ_PREV; |
9310a7ba | 1208 | async_list->io_len += len; |
31b51510 JA |
1209 | } else { |
1210 | io_end = 0; | |
9310a7ba | 1211 | async_list->io_len = 0; |
31b51510 JA |
1212 | } |
1213 | } | |
1214 | ||
1215 | /* New file? Reset state. */ | |
1216 | if (async_list->file != filp) { | |
9310a7ba | 1217 | async_list->io_len = 0; |
31b51510 JA |
1218 | async_list->file = filp; |
1219 | } | |
1220 | async_list->io_end = io_end; | |
1221 | } | |
1222 | ||
e0c5c576 | 1223 | static int io_read(struct io_kiocb *req, const struct sqe_submit *s, |
8358e3a8 | 1224 | bool force_nonblock) |
2b188cc1 JA |
1225 | { |
1226 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1227 | struct kiocb *kiocb = &req->rw; | |
1228 | struct iov_iter iter; | |
1229 | struct file *file; | |
31b51510 | 1230 | size_t iov_count; |
9d93a3f5 | 1231 | ssize_t read_size, ret; |
2b188cc1 | 1232 | |
8358e3a8 | 1233 | ret = io_prep_rw(req, s, force_nonblock); |
2b188cc1 JA |
1234 | if (ret) |
1235 | return ret; | |
1236 | file = kiocb->ki_filp; | |
1237 | ||
2b188cc1 | 1238 | if (unlikely(!(file->f_mode & FMODE_READ))) |
09bb8394 | 1239 | return -EBADF; |
2b188cc1 | 1240 | if (unlikely(!file->f_op->read_iter)) |
09bb8394 | 1241 | return -EINVAL; |
2b188cc1 JA |
1242 | |
1243 | ret = io_import_iovec(req->ctx, READ, s, &iovec, &iter); | |
87e5e6da | 1244 | if (ret < 0) |
09bb8394 | 1245 | return ret; |
2b188cc1 | 1246 | |
9d93a3f5 | 1247 | read_size = ret; |
9e645e11 JA |
1248 | if (req->flags & REQ_F_LINK) |
1249 | req->result = read_size; | |
1250 | ||
31b51510 JA |
1251 | iov_count = iov_iter_count(&iter); |
1252 | ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 JA |
1253 | if (!ret) { |
1254 | ssize_t ret2; | |
1255 | ||
2b188cc1 | 1256 | ret2 = call_read_iter(file, kiocb, &iter); |
9d93a3f5 JA |
1257 | /* |
1258 | * In case of a short read, punt to async. This can happen | |
1259 | * if we have data partially cached. Alternatively we can | |
1260 | * return the short read, in which case the application will | |
1261 | * need to issue another SQE and wait for it. That SQE will | |
1262 | * need async punt anyway, so it's more efficient to do it | |
1263 | * here. | |
1264 | */ | |
1265 | if (force_nonblock && ret2 > 0 && ret2 < read_size) | |
1266 | ret2 = -EAGAIN; | |
1267 | /* Catch -EAGAIN return for forced non-blocking submission */ | |
31b51510 | 1268 | if (!force_nonblock || ret2 != -EAGAIN) { |
2b188cc1 | 1269 | io_rw_done(kiocb, ret2); |
31b51510 JA |
1270 | } else { |
1271 | /* | |
1272 | * If ->needs_lock is true, we're already in async | |
1273 | * context. | |
1274 | */ | |
1275 | if (!s->needs_lock) | |
1276 | io_async_list_note(READ, req, iov_count); | |
2b188cc1 | 1277 | ret = -EAGAIN; |
31b51510 | 1278 | } |
2b188cc1 JA |
1279 | } |
1280 | kfree(iovec); | |
2b188cc1 JA |
1281 | return ret; |
1282 | } | |
1283 | ||
e0c5c576 | 1284 | static int io_write(struct io_kiocb *req, const struct sqe_submit *s, |
8358e3a8 | 1285 | bool force_nonblock) |
2b188cc1 JA |
1286 | { |
1287 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1288 | struct kiocb *kiocb = &req->rw; | |
1289 | struct iov_iter iter; | |
1290 | struct file *file; | |
31b51510 | 1291 | size_t iov_count; |
87e5e6da | 1292 | ssize_t ret; |
2b188cc1 | 1293 | |
8358e3a8 | 1294 | ret = io_prep_rw(req, s, force_nonblock); |
2b188cc1 JA |
1295 | if (ret) |
1296 | return ret; | |
2b188cc1 | 1297 | |
2b188cc1 JA |
1298 | file = kiocb->ki_filp; |
1299 | if (unlikely(!(file->f_mode & FMODE_WRITE))) | |
09bb8394 | 1300 | return -EBADF; |
2b188cc1 | 1301 | if (unlikely(!file->f_op->write_iter)) |
09bb8394 | 1302 | return -EINVAL; |
2b188cc1 JA |
1303 | |
1304 | ret = io_import_iovec(req->ctx, WRITE, s, &iovec, &iter); | |
87e5e6da | 1305 | if (ret < 0) |
09bb8394 | 1306 | return ret; |
2b188cc1 | 1307 | |
9e645e11 JA |
1308 | if (req->flags & REQ_F_LINK) |
1309 | req->result = ret; | |
1310 | ||
31b51510 JA |
1311 | iov_count = iov_iter_count(&iter); |
1312 | ||
1313 | ret = -EAGAIN; | |
1314 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT)) { | |
1315 | /* If ->needs_lock is true, we're already in async context. */ | |
1316 | if (!s->needs_lock) | |
1317 | io_async_list_note(WRITE, req, iov_count); | |
1318 | goto out_free; | |
1319 | } | |
1320 | ||
1321 | ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 | 1322 | if (!ret) { |
9bf7933f RP |
1323 | ssize_t ret2; |
1324 | ||
2b188cc1 JA |
1325 | /* |
1326 | * Open-code file_start_write here to grab freeze protection, | |
1327 | * which will be released by another thread in | |
1328 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
1329 | * released so that it doesn't complain about the held lock when | |
1330 | * we return to userspace. | |
1331 | */ | |
1332 | if (S_ISREG(file_inode(file)->i_mode)) { | |
1333 | __sb_start_write(file_inode(file)->i_sb, | |
1334 | SB_FREEZE_WRITE, true); | |
1335 | __sb_writers_release(file_inode(file)->i_sb, | |
1336 | SB_FREEZE_WRITE); | |
1337 | } | |
1338 | kiocb->ki_flags |= IOCB_WRITE; | |
9bf7933f RP |
1339 | |
1340 | ret2 = call_write_iter(file, kiocb, &iter); | |
1341 | if (!force_nonblock || ret2 != -EAGAIN) { | |
1342 | io_rw_done(kiocb, ret2); | |
1343 | } else { | |
1344 | /* | |
1345 | * If ->needs_lock is true, we're already in async | |
1346 | * context. | |
1347 | */ | |
1348 | if (!s->needs_lock) | |
1349 | io_async_list_note(WRITE, req, iov_count); | |
1350 | ret = -EAGAIN; | |
1351 | } | |
2b188cc1 | 1352 | } |
31b51510 | 1353 | out_free: |
2b188cc1 | 1354 | kfree(iovec); |
2b188cc1 JA |
1355 | return ret; |
1356 | } | |
1357 | ||
1358 | /* | |
1359 | * IORING_OP_NOP just posts a completion event, nothing else. | |
1360 | */ | |
1361 | static int io_nop(struct io_kiocb *req, u64 user_data) | |
1362 | { | |
1363 | struct io_ring_ctx *ctx = req->ctx; | |
1364 | long err = 0; | |
1365 | ||
def596e9 JA |
1366 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
1367 | return -EINVAL; | |
1368 | ||
c71ffb67 | 1369 | io_cqring_add_event(ctx, user_data, err); |
e65ef56d | 1370 | io_put_req(req); |
2b188cc1 JA |
1371 | return 0; |
1372 | } | |
1373 | ||
c992fe29 CH |
1374 | static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1375 | { | |
6b06314c | 1376 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 1377 | |
09bb8394 JA |
1378 | if (!req->file) |
1379 | return -EBADF; | |
c992fe29 | 1380 | |
6b06314c | 1381 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 1382 | return -EINVAL; |
edafccee | 1383 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
1384 | return -EINVAL; |
1385 | ||
c992fe29 CH |
1386 | return 0; |
1387 | } | |
1388 | ||
1389 | static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
1390 | bool force_nonblock) | |
1391 | { | |
1392 | loff_t sqe_off = READ_ONCE(sqe->off); | |
1393 | loff_t sqe_len = READ_ONCE(sqe->len); | |
1394 | loff_t end = sqe_off + sqe_len; | |
1395 | unsigned fsync_flags; | |
1396 | int ret; | |
1397 | ||
1398 | fsync_flags = READ_ONCE(sqe->fsync_flags); | |
1399 | if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC)) | |
1400 | return -EINVAL; | |
1401 | ||
1402 | ret = io_prep_fsync(req, sqe); | |
1403 | if (ret) | |
1404 | return ret; | |
1405 | ||
1406 | /* fsync always requires a blocking context */ | |
1407 | if (force_nonblock) | |
1408 | return -EAGAIN; | |
1409 | ||
1410 | ret = vfs_fsync_range(req->rw.ki_filp, sqe_off, | |
1411 | end > 0 ? end : LLONG_MAX, | |
1412 | fsync_flags & IORING_FSYNC_DATASYNC); | |
1413 | ||
9e645e11 JA |
1414 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1415 | req->flags |= REQ_F_FAIL_LINK; | |
c71ffb67 | 1416 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
e65ef56d | 1417 | io_put_req(req); |
c992fe29 CH |
1418 | return 0; |
1419 | } | |
1420 | ||
5d17b4a4 JA |
1421 | static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1422 | { | |
1423 | struct io_ring_ctx *ctx = req->ctx; | |
1424 | int ret = 0; | |
1425 | ||
1426 | if (!req->file) | |
1427 | return -EBADF; | |
5d17b4a4 JA |
1428 | |
1429 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
1430 | return -EINVAL; | |
1431 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) | |
1432 | return -EINVAL; | |
1433 | ||
5d17b4a4 JA |
1434 | return ret; |
1435 | } | |
1436 | ||
1437 | static int io_sync_file_range(struct io_kiocb *req, | |
1438 | const struct io_uring_sqe *sqe, | |
1439 | bool force_nonblock) | |
1440 | { | |
1441 | loff_t sqe_off; | |
1442 | loff_t sqe_len; | |
1443 | unsigned flags; | |
1444 | int ret; | |
1445 | ||
1446 | ret = io_prep_sfr(req, sqe); | |
1447 | if (ret) | |
1448 | return ret; | |
1449 | ||
1450 | /* sync_file_range always requires a blocking context */ | |
1451 | if (force_nonblock) | |
1452 | return -EAGAIN; | |
1453 | ||
1454 | sqe_off = READ_ONCE(sqe->off); | |
1455 | sqe_len = READ_ONCE(sqe->len); | |
1456 | flags = READ_ONCE(sqe->sync_range_flags); | |
1457 | ||
1458 | ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags); | |
1459 | ||
9e645e11 JA |
1460 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1461 | req->flags |= REQ_F_FAIL_LINK; | |
c71ffb67 | 1462 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
5d17b4a4 JA |
1463 | io_put_req(req); |
1464 | return 0; | |
1465 | } | |
1466 | ||
0fa03c62 | 1467 | #if defined(CONFIG_NET) |
aa1fa28f JA |
1468 | static int io_send_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
1469 | bool force_nonblock, | |
1470 | long (*fn)(struct socket *, struct user_msghdr __user *, | |
1471 | unsigned int)) | |
1472 | { | |
0fa03c62 JA |
1473 | struct socket *sock; |
1474 | int ret; | |
1475 | ||
1476 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1477 | return -EINVAL; | |
1478 | ||
1479 | sock = sock_from_file(req->file, &ret); | |
1480 | if (sock) { | |
1481 | struct user_msghdr __user *msg; | |
1482 | unsigned flags; | |
1483 | ||
1484 | flags = READ_ONCE(sqe->msg_flags); | |
1485 | if (flags & MSG_DONTWAIT) | |
1486 | req->flags |= REQ_F_NOWAIT; | |
1487 | else if (force_nonblock) | |
1488 | flags |= MSG_DONTWAIT; | |
1489 | ||
1490 | msg = (struct user_msghdr __user *) (unsigned long) | |
1491 | READ_ONCE(sqe->addr); | |
1492 | ||
aa1fa28f | 1493 | ret = fn(sock, msg, flags); |
0fa03c62 JA |
1494 | if (force_nonblock && ret == -EAGAIN) |
1495 | return ret; | |
1496 | } | |
1497 | ||
c71ffb67 | 1498 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
5d17b4a4 JA |
1499 | io_put_req(req); |
1500 | return 0; | |
1501 | } | |
aa1fa28f JA |
1502 | #endif |
1503 | ||
1504 | static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
1505 | bool force_nonblock) | |
1506 | { | |
1507 | #if defined(CONFIG_NET) | |
1508 | return io_send_recvmsg(req, sqe, force_nonblock, __sys_sendmsg_sock); | |
1509 | #else | |
1510 | return -EOPNOTSUPP; | |
1511 | #endif | |
1512 | } | |
1513 | ||
1514 | static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
1515 | bool force_nonblock) | |
1516 | { | |
1517 | #if defined(CONFIG_NET) | |
1518 | return io_send_recvmsg(req, sqe, force_nonblock, __sys_recvmsg_sock); | |
0fa03c62 JA |
1519 | #else |
1520 | return -EOPNOTSUPP; | |
1521 | #endif | |
1522 | } | |
5d17b4a4 | 1523 | |
221c5eb2 JA |
1524 | static void io_poll_remove_one(struct io_kiocb *req) |
1525 | { | |
1526 | struct io_poll_iocb *poll = &req->poll; | |
1527 | ||
1528 | spin_lock(&poll->head->lock); | |
1529 | WRITE_ONCE(poll->canceled, true); | |
1530 | if (!list_empty(&poll->wait.entry)) { | |
1531 | list_del_init(&poll->wait.entry); | |
1532 | queue_work(req->ctx->sqo_wq, &req->work); | |
1533 | } | |
1534 | spin_unlock(&poll->head->lock); | |
1535 | ||
1536 | list_del_init(&req->list); | |
1537 | } | |
1538 | ||
1539 | static void io_poll_remove_all(struct io_ring_ctx *ctx) | |
1540 | { | |
1541 | struct io_kiocb *req; | |
1542 | ||
1543 | spin_lock_irq(&ctx->completion_lock); | |
1544 | while (!list_empty(&ctx->cancel_list)) { | |
1545 | req = list_first_entry(&ctx->cancel_list, struct io_kiocb,list); | |
1546 | io_poll_remove_one(req); | |
1547 | } | |
1548 | spin_unlock_irq(&ctx->completion_lock); | |
1549 | } | |
1550 | ||
1551 | /* | |
1552 | * Find a running poll command that matches one specified in sqe->addr, | |
1553 | * and remove it if found. | |
1554 | */ | |
1555 | static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
1556 | { | |
1557 | struct io_ring_ctx *ctx = req->ctx; | |
1558 | struct io_kiocb *poll_req, *next; | |
1559 | int ret = -ENOENT; | |
1560 | ||
1561 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1562 | return -EINVAL; | |
1563 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index || | |
1564 | sqe->poll_events) | |
1565 | return -EINVAL; | |
1566 | ||
1567 | spin_lock_irq(&ctx->completion_lock); | |
1568 | list_for_each_entry_safe(poll_req, next, &ctx->cancel_list, list) { | |
1569 | if (READ_ONCE(sqe->addr) == poll_req->user_data) { | |
1570 | io_poll_remove_one(poll_req); | |
1571 | ret = 0; | |
1572 | break; | |
1573 | } | |
1574 | } | |
1575 | spin_unlock_irq(&ctx->completion_lock); | |
1576 | ||
c71ffb67 | 1577 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
e65ef56d | 1578 | io_put_req(req); |
221c5eb2 JA |
1579 | return 0; |
1580 | } | |
1581 | ||
8c838788 JA |
1582 | static void io_poll_complete(struct io_ring_ctx *ctx, struct io_kiocb *req, |
1583 | __poll_t mask) | |
221c5eb2 | 1584 | { |
8c838788 | 1585 | req->poll.done = true; |
c71ffb67 | 1586 | io_cqring_fill_event(ctx, req->user_data, mangle_poll(mask)); |
8c838788 | 1587 | io_commit_cqring(ctx); |
221c5eb2 JA |
1588 | } |
1589 | ||
1590 | static void io_poll_complete_work(struct work_struct *work) | |
1591 | { | |
1592 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
1593 | struct io_poll_iocb *poll = &req->poll; | |
1594 | struct poll_table_struct pt = { ._key = poll->events }; | |
1595 | struct io_ring_ctx *ctx = req->ctx; | |
1596 | __poll_t mask = 0; | |
1597 | ||
1598 | if (!READ_ONCE(poll->canceled)) | |
1599 | mask = vfs_poll(poll->file, &pt) & poll->events; | |
1600 | ||
1601 | /* | |
1602 | * Note that ->ki_cancel callers also delete iocb from active_reqs after | |
1603 | * calling ->ki_cancel. We need the ctx_lock roundtrip here to | |
1604 | * synchronize with them. In the cancellation case the list_del_init | |
1605 | * itself is not actually needed, but harmless so we keep it in to | |
1606 | * avoid further branches in the fast path. | |
1607 | */ | |
1608 | spin_lock_irq(&ctx->completion_lock); | |
1609 | if (!mask && !READ_ONCE(poll->canceled)) { | |
1610 | add_wait_queue(poll->head, &poll->wait); | |
1611 | spin_unlock_irq(&ctx->completion_lock); | |
1612 | return; | |
1613 | } | |
1614 | list_del_init(&req->list); | |
8c838788 | 1615 | io_poll_complete(ctx, req, mask); |
221c5eb2 JA |
1616 | spin_unlock_irq(&ctx->completion_lock); |
1617 | ||
8c838788 JA |
1618 | io_cqring_ev_posted(ctx); |
1619 | io_put_req(req); | |
221c5eb2 JA |
1620 | } |
1621 | ||
1622 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
1623 | void *key) | |
1624 | { | |
1625 | struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb, | |
1626 | wait); | |
1627 | struct io_kiocb *req = container_of(poll, struct io_kiocb, poll); | |
1628 | struct io_ring_ctx *ctx = req->ctx; | |
1629 | __poll_t mask = key_to_poll(key); | |
8c838788 | 1630 | unsigned long flags; |
221c5eb2 JA |
1631 | |
1632 | /* for instances that support it check for an event match first: */ | |
8c838788 JA |
1633 | if (mask && !(mask & poll->events)) |
1634 | return 0; | |
221c5eb2 | 1635 | |
8c838788 | 1636 | list_del_init(&poll->wait.entry); |
221c5eb2 | 1637 | |
8c838788 JA |
1638 | if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) { |
1639 | list_del(&req->list); | |
1640 | io_poll_complete(ctx, req, mask); | |
1641 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
221c5eb2 | 1642 | |
8c838788 JA |
1643 | io_cqring_ev_posted(ctx); |
1644 | io_put_req(req); | |
1645 | } else { | |
1646 | queue_work(ctx->sqo_wq, &req->work); | |
221c5eb2 JA |
1647 | } |
1648 | ||
221c5eb2 JA |
1649 | return 1; |
1650 | } | |
1651 | ||
1652 | struct io_poll_table { | |
1653 | struct poll_table_struct pt; | |
1654 | struct io_kiocb *req; | |
1655 | int error; | |
1656 | }; | |
1657 | ||
1658 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
1659 | struct poll_table_struct *p) | |
1660 | { | |
1661 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
1662 | ||
1663 | if (unlikely(pt->req->poll.head)) { | |
1664 | pt->error = -EINVAL; | |
1665 | return; | |
1666 | } | |
1667 | ||
1668 | pt->error = 0; | |
1669 | pt->req->poll.head = head; | |
1670 | add_wait_queue(head, &pt->req->poll.wait); | |
1671 | } | |
1672 | ||
1673 | static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
1674 | { | |
1675 | struct io_poll_iocb *poll = &req->poll; | |
1676 | struct io_ring_ctx *ctx = req->ctx; | |
1677 | struct io_poll_table ipt; | |
8c838788 | 1678 | bool cancel = false; |
221c5eb2 JA |
1679 | __poll_t mask; |
1680 | u16 events; | |
221c5eb2 JA |
1681 | |
1682 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1683 | return -EINVAL; | |
1684 | if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
1685 | return -EINVAL; | |
09bb8394 JA |
1686 | if (!poll->file) |
1687 | return -EBADF; | |
221c5eb2 JA |
1688 | |
1689 | INIT_WORK(&req->work, io_poll_complete_work); | |
1690 | events = READ_ONCE(sqe->poll_events); | |
1691 | poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP; | |
1692 | ||
221c5eb2 | 1693 | poll->head = NULL; |
8c838788 | 1694 | poll->done = false; |
221c5eb2 JA |
1695 | poll->canceled = false; |
1696 | ||
1697 | ipt.pt._qproc = io_poll_queue_proc; | |
1698 | ipt.pt._key = poll->events; | |
1699 | ipt.req = req; | |
1700 | ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */ | |
1701 | ||
1702 | /* initialized the list so that we can do list_empty checks */ | |
1703 | INIT_LIST_HEAD(&poll->wait.entry); | |
1704 | init_waitqueue_func_entry(&poll->wait, io_poll_wake); | |
1705 | ||
36703247 JA |
1706 | INIT_LIST_HEAD(&req->list); |
1707 | ||
221c5eb2 | 1708 | mask = vfs_poll(poll->file, &ipt.pt) & poll->events; |
221c5eb2 JA |
1709 | |
1710 | spin_lock_irq(&ctx->completion_lock); | |
8c838788 JA |
1711 | if (likely(poll->head)) { |
1712 | spin_lock(&poll->head->lock); | |
1713 | if (unlikely(list_empty(&poll->wait.entry))) { | |
1714 | if (ipt.error) | |
1715 | cancel = true; | |
1716 | ipt.error = 0; | |
1717 | mask = 0; | |
1718 | } | |
1719 | if (mask || ipt.error) | |
1720 | list_del_init(&poll->wait.entry); | |
1721 | else if (cancel) | |
1722 | WRITE_ONCE(poll->canceled, true); | |
1723 | else if (!poll->done) /* actually waiting for an event */ | |
1724 | list_add_tail(&req->list, &ctx->cancel_list); | |
1725 | spin_unlock(&poll->head->lock); | |
1726 | } | |
1727 | if (mask) { /* no async, we'd stolen it */ | |
221c5eb2 | 1728 | ipt.error = 0; |
8c838788 | 1729 | io_poll_complete(ctx, req, mask); |
221c5eb2 | 1730 | } |
221c5eb2 JA |
1731 | spin_unlock_irq(&ctx->completion_lock); |
1732 | ||
8c838788 JA |
1733 | if (mask) { |
1734 | io_cqring_ev_posted(ctx); | |
e65ef56d | 1735 | io_put_req(req); |
221c5eb2 | 1736 | } |
8c838788 | 1737 | return ipt.error; |
221c5eb2 JA |
1738 | } |
1739 | ||
de0617e4 JA |
1740 | static int io_req_defer(struct io_ring_ctx *ctx, struct io_kiocb *req, |
1741 | const struct io_uring_sqe *sqe) | |
1742 | { | |
1743 | struct io_uring_sqe *sqe_copy; | |
1744 | ||
1745 | if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) | |
1746 | return 0; | |
1747 | ||
1748 | sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL); | |
1749 | if (!sqe_copy) | |
1750 | return -EAGAIN; | |
1751 | ||
1752 | spin_lock_irq(&ctx->completion_lock); | |
1753 | if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) { | |
1754 | spin_unlock_irq(&ctx->completion_lock); | |
1755 | kfree(sqe_copy); | |
1756 | return 0; | |
1757 | } | |
1758 | ||
1759 | memcpy(sqe_copy, sqe, sizeof(*sqe_copy)); | |
1760 | req->submit.sqe = sqe_copy; | |
1761 | ||
1762 | INIT_WORK(&req->work, io_sq_wq_submit_work); | |
1763 | list_add_tail(&req->list, &ctx->defer_list); | |
1764 | spin_unlock_irq(&ctx->completion_lock); | |
1765 | return -EIOCBQUEUED; | |
1766 | } | |
1767 | ||
2b188cc1 | 1768 | static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
8358e3a8 | 1769 | const struct sqe_submit *s, bool force_nonblock) |
2b188cc1 | 1770 | { |
e0c5c576 | 1771 | int ret, opcode; |
2b188cc1 | 1772 | |
9e645e11 JA |
1773 | req->user_data = READ_ONCE(s->sqe->user_data); |
1774 | ||
2b188cc1 JA |
1775 | if (unlikely(s->index >= ctx->sq_entries)) |
1776 | return -EINVAL; | |
2b188cc1 JA |
1777 | |
1778 | opcode = READ_ONCE(s->sqe->opcode); | |
1779 | switch (opcode) { | |
1780 | case IORING_OP_NOP: | |
1781 | ret = io_nop(req, req->user_data); | |
1782 | break; | |
1783 | case IORING_OP_READV: | |
edafccee JA |
1784 | if (unlikely(s->sqe->buf_index)) |
1785 | return -EINVAL; | |
8358e3a8 | 1786 | ret = io_read(req, s, force_nonblock); |
2b188cc1 JA |
1787 | break; |
1788 | case IORING_OP_WRITEV: | |
edafccee JA |
1789 | if (unlikely(s->sqe->buf_index)) |
1790 | return -EINVAL; | |
8358e3a8 | 1791 | ret = io_write(req, s, force_nonblock); |
edafccee JA |
1792 | break; |
1793 | case IORING_OP_READ_FIXED: | |
8358e3a8 | 1794 | ret = io_read(req, s, force_nonblock); |
edafccee JA |
1795 | break; |
1796 | case IORING_OP_WRITE_FIXED: | |
8358e3a8 | 1797 | ret = io_write(req, s, force_nonblock); |
2b188cc1 | 1798 | break; |
c992fe29 CH |
1799 | case IORING_OP_FSYNC: |
1800 | ret = io_fsync(req, s->sqe, force_nonblock); | |
1801 | break; | |
221c5eb2 JA |
1802 | case IORING_OP_POLL_ADD: |
1803 | ret = io_poll_add(req, s->sqe); | |
1804 | break; | |
1805 | case IORING_OP_POLL_REMOVE: | |
1806 | ret = io_poll_remove(req, s->sqe); | |
1807 | break; | |
5d17b4a4 JA |
1808 | case IORING_OP_SYNC_FILE_RANGE: |
1809 | ret = io_sync_file_range(req, s->sqe, force_nonblock); | |
1810 | break; | |
0fa03c62 JA |
1811 | case IORING_OP_SENDMSG: |
1812 | ret = io_sendmsg(req, s->sqe, force_nonblock); | |
1813 | break; | |
aa1fa28f JA |
1814 | case IORING_OP_RECVMSG: |
1815 | ret = io_recvmsg(req, s->sqe, force_nonblock); | |
1816 | break; | |
2b188cc1 JA |
1817 | default: |
1818 | ret = -EINVAL; | |
1819 | break; | |
1820 | } | |
1821 | ||
def596e9 JA |
1822 | if (ret) |
1823 | return ret; | |
1824 | ||
1825 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
9e645e11 | 1826 | if (req->result == -EAGAIN) |
def596e9 JA |
1827 | return -EAGAIN; |
1828 | ||
1829 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
1830 | if (s->needs_lock) | |
1831 | mutex_lock(&ctx->uring_lock); | |
1832 | io_iopoll_req_issued(req); | |
1833 | if (s->needs_lock) | |
1834 | mutex_unlock(&ctx->uring_lock); | |
1835 | } | |
1836 | ||
1837 | return 0; | |
2b188cc1 JA |
1838 | } |
1839 | ||
31b51510 JA |
1840 | static struct async_list *io_async_list_from_sqe(struct io_ring_ctx *ctx, |
1841 | const struct io_uring_sqe *sqe) | |
1842 | { | |
1843 | switch (sqe->opcode) { | |
1844 | case IORING_OP_READV: | |
1845 | case IORING_OP_READ_FIXED: | |
1846 | return &ctx->pending_async[READ]; | |
1847 | case IORING_OP_WRITEV: | |
1848 | case IORING_OP_WRITE_FIXED: | |
1849 | return &ctx->pending_async[WRITE]; | |
1850 | default: | |
1851 | return NULL; | |
1852 | } | |
1853 | } | |
1854 | ||
edafccee JA |
1855 | static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe) |
1856 | { | |
1857 | u8 opcode = READ_ONCE(sqe->opcode); | |
1858 | ||
1859 | return !(opcode == IORING_OP_READ_FIXED || | |
1860 | opcode == IORING_OP_WRITE_FIXED); | |
1861 | } | |
1862 | ||
2b188cc1 JA |
1863 | static void io_sq_wq_submit_work(struct work_struct *work) |
1864 | { | |
1865 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
2b188cc1 | 1866 | struct io_ring_ctx *ctx = req->ctx; |
31b51510 JA |
1867 | struct mm_struct *cur_mm = NULL; |
1868 | struct async_list *async_list; | |
1869 | LIST_HEAD(req_list); | |
edafccee | 1870 | mm_segment_t old_fs; |
2b188cc1 JA |
1871 | int ret; |
1872 | ||
31b51510 JA |
1873 | async_list = io_async_list_from_sqe(ctx, req->submit.sqe); |
1874 | restart: | |
1875 | do { | |
1876 | struct sqe_submit *s = &req->submit; | |
1877 | const struct io_uring_sqe *sqe = s->sqe; | |
d0ee8791 | 1878 | unsigned int flags = req->flags; |
2b188cc1 | 1879 | |
8449eeda | 1880 | /* Ensure we clear previously set non-block flag */ |
31b51510 JA |
1881 | req->rw.ki_flags &= ~IOCB_NOWAIT; |
1882 | ||
1883 | ret = 0; | |
1884 | if (io_sqe_needs_user(sqe) && !cur_mm) { | |
1885 | if (!mmget_not_zero(ctx->sqo_mm)) { | |
1886 | ret = -EFAULT; | |
1887 | } else { | |
1888 | cur_mm = ctx->sqo_mm; | |
1889 | use_mm(cur_mm); | |
1890 | old_fs = get_fs(); | |
1891 | set_fs(USER_DS); | |
1892 | } | |
1893 | } | |
1894 | ||
1895 | if (!ret) { | |
1896 | s->has_user = cur_mm != NULL; | |
1897 | s->needs_lock = true; | |
1898 | do { | |
8358e3a8 | 1899 | ret = __io_submit_sqe(ctx, req, s, false); |
31b51510 JA |
1900 | /* |
1901 | * We can get EAGAIN for polled IO even though | |
1902 | * we're forcing a sync submission from here, | |
1903 | * since we can't wait for request slots on the | |
1904 | * block side. | |
1905 | */ | |
1906 | if (ret != -EAGAIN) | |
1907 | break; | |
1908 | cond_resched(); | |
1909 | } while (1); | |
1910 | } | |
817869d2 JA |
1911 | |
1912 | /* drop submission reference */ | |
1913 | io_put_req(req); | |
1914 | ||
31b51510 | 1915 | if (ret) { |
c71ffb67 | 1916 | io_cqring_add_event(ctx, sqe->user_data, ret); |
e65ef56d | 1917 | io_put_req(req); |
31b51510 JA |
1918 | } |
1919 | ||
1920 | /* async context always use a copy of the sqe */ | |
1921 | kfree(sqe); | |
1922 | ||
f7b76ac9 | 1923 | /* req from defer and link list needn't decrease async cnt */ |
d0ee8791 | 1924 | if (flags & (REQ_F_IO_DRAINED | REQ_F_LINK_DONE)) |
f7b76ac9 ZL |
1925 | goto out; |
1926 | ||
31b51510 JA |
1927 | if (!async_list) |
1928 | break; | |
1929 | if (!list_empty(&req_list)) { | |
1930 | req = list_first_entry(&req_list, struct io_kiocb, | |
1931 | list); | |
1932 | list_del(&req->list); | |
1933 | continue; | |
1934 | } | |
1935 | if (list_empty(&async_list->list)) | |
1936 | break; | |
1937 | ||
1938 | req = NULL; | |
1939 | spin_lock(&async_list->lock); | |
1940 | if (list_empty(&async_list->list)) { | |
1941 | spin_unlock(&async_list->lock); | |
1942 | break; | |
1943 | } | |
1944 | list_splice_init(&async_list->list, &req_list); | |
1945 | spin_unlock(&async_list->lock); | |
1946 | ||
1947 | req = list_first_entry(&req_list, struct io_kiocb, list); | |
1948 | list_del(&req->list); | |
1949 | } while (req); | |
edafccee JA |
1950 | |
1951 | /* | |
31b51510 JA |
1952 | * Rare case of racing with a submitter. If we find the count has |
1953 | * dropped to zero AND we have pending work items, then restart | |
1954 | * the processing. This is a tiny race window. | |
edafccee | 1955 | */ |
31b51510 JA |
1956 | if (async_list) { |
1957 | ret = atomic_dec_return(&async_list->cnt); | |
1958 | while (!ret && !list_empty(&async_list->list)) { | |
1959 | spin_lock(&async_list->lock); | |
1960 | atomic_inc(&async_list->cnt); | |
1961 | list_splice_init(&async_list->list, &req_list); | |
1962 | spin_unlock(&async_list->lock); | |
1963 | ||
1964 | if (!list_empty(&req_list)) { | |
1965 | req = list_first_entry(&req_list, | |
1966 | struct io_kiocb, list); | |
1967 | list_del(&req->list); | |
1968 | goto restart; | |
1969 | } | |
1970 | ret = atomic_dec_return(&async_list->cnt); | |
edafccee | 1971 | } |
edafccee | 1972 | } |
2b188cc1 | 1973 | |
f7b76ac9 | 1974 | out: |
31b51510 | 1975 | if (cur_mm) { |
edafccee | 1976 | set_fs(old_fs); |
31b51510 JA |
1977 | unuse_mm(cur_mm); |
1978 | mmput(cur_mm); | |
2b188cc1 | 1979 | } |
31b51510 | 1980 | } |
2b188cc1 | 1981 | |
31b51510 JA |
1982 | /* |
1983 | * See if we can piggy back onto previously submitted work, that is still | |
1984 | * running. We currently only allow this if the new request is sequential | |
1985 | * to the previous one we punted. | |
1986 | */ | |
1987 | static bool io_add_to_prev_work(struct async_list *list, struct io_kiocb *req) | |
1988 | { | |
1989 | bool ret = false; | |
1990 | ||
1991 | if (!list) | |
1992 | return false; | |
1993 | if (!(req->flags & REQ_F_SEQ_PREV)) | |
1994 | return false; | |
1995 | if (!atomic_read(&list->cnt)) | |
1996 | return false; | |
1997 | ||
1998 | ret = true; | |
1999 | spin_lock(&list->lock); | |
2000 | list_add_tail(&req->list, &list->list); | |
c0e48f9d ZL |
2001 | /* |
2002 | * Ensure we see a simultaneous modification from io_sq_wq_submit_work() | |
2003 | */ | |
2004 | smp_mb(); | |
31b51510 JA |
2005 | if (!atomic_read(&list->cnt)) { |
2006 | list_del_init(&req->list); | |
2007 | ret = false; | |
2008 | } | |
2009 | spin_unlock(&list->lock); | |
2010 | return ret; | |
2b188cc1 JA |
2011 | } |
2012 | ||
09bb8394 JA |
2013 | static bool io_op_needs_file(const struct io_uring_sqe *sqe) |
2014 | { | |
2015 | int op = READ_ONCE(sqe->opcode); | |
2016 | ||
2017 | switch (op) { | |
2018 | case IORING_OP_NOP: | |
2019 | case IORING_OP_POLL_REMOVE: | |
2020 | return false; | |
2021 | default: | |
2022 | return true; | |
2023 | } | |
2024 | } | |
2025 | ||
2026 | static int io_req_set_file(struct io_ring_ctx *ctx, const struct sqe_submit *s, | |
2027 | struct io_submit_state *state, struct io_kiocb *req) | |
2028 | { | |
2029 | unsigned flags; | |
2030 | int fd; | |
2031 | ||
2032 | flags = READ_ONCE(s->sqe->flags); | |
2033 | fd = READ_ONCE(s->sqe->fd); | |
2034 | ||
de0617e4 JA |
2035 | if (flags & IOSQE_IO_DRAIN) { |
2036 | req->flags |= REQ_F_IO_DRAIN; | |
2037 | req->sequence = ctx->cached_sq_head - 1; | |
2038 | } | |
2039 | ||
60c112b0 | 2040 | if (!io_op_needs_file(s->sqe)) |
09bb8394 | 2041 | return 0; |
09bb8394 JA |
2042 | |
2043 | if (flags & IOSQE_FIXED_FILE) { | |
2044 | if (unlikely(!ctx->user_files || | |
2045 | (unsigned) fd >= ctx->nr_user_files)) | |
2046 | return -EBADF; | |
2047 | req->file = ctx->user_files[fd]; | |
2048 | req->flags |= REQ_F_FIXED_FILE; | |
2049 | } else { | |
2050 | if (s->needs_fixed_file) | |
2051 | return -EBADF; | |
2052 | req->file = io_file_get(state, fd); | |
2053 | if (unlikely(!req->file)) | |
2054 | return -EBADF; | |
2055 | } | |
2056 | ||
2057 | return 0; | |
2058 | } | |
2059 | ||
9e645e11 JA |
2060 | static int io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
2061 | struct sqe_submit *s) | |
2b188cc1 | 2062 | { |
e0c5c576 | 2063 | int ret; |
2b188cc1 | 2064 | |
a982eeb0 JL |
2065 | ret = io_req_defer(ctx, req, s->sqe); |
2066 | if (ret) { | |
2067 | if (ret != -EIOCBQUEUED) { | |
2068 | io_free_req(req); | |
2069 | io_cqring_add_event(ctx, s->sqe->user_data, ret); | |
2070 | } | |
2071 | return 0; | |
2072 | } | |
2073 | ||
8358e3a8 | 2074 | ret = __io_submit_sqe(ctx, req, s, true); |
8449eeda | 2075 | if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) { |
2b188cc1 JA |
2076 | struct io_uring_sqe *sqe_copy; |
2077 | ||
2078 | sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL); | |
2079 | if (sqe_copy) { | |
31b51510 JA |
2080 | struct async_list *list; |
2081 | ||
2b188cc1 JA |
2082 | memcpy(sqe_copy, s->sqe, sizeof(*sqe_copy)); |
2083 | s->sqe = sqe_copy; | |
2084 | ||
2085 | memcpy(&req->submit, s, sizeof(*s)); | |
31b51510 JA |
2086 | list = io_async_list_from_sqe(ctx, s->sqe); |
2087 | if (!io_add_to_prev_work(list, req)) { | |
2088 | if (list) | |
2089 | atomic_inc(&list->cnt); | |
2090 | INIT_WORK(&req->work, io_sq_wq_submit_work); | |
2091 | queue_work(ctx->sqo_wq, &req->work); | |
2092 | } | |
e65ef56d JA |
2093 | |
2094 | /* | |
2095 | * Queued up for async execution, worker will release | |
9e645e11 | 2096 | * submit reference when the iocb is actually submitted. |
e65ef56d JA |
2097 | */ |
2098 | return 0; | |
2b188cc1 JA |
2099 | } |
2100 | } | |
e65ef56d JA |
2101 | |
2102 | /* drop submission reference */ | |
2103 | io_put_req(req); | |
2104 | ||
2105 | /* and drop final reference, if we failed */ | |
9e645e11 JA |
2106 | if (ret) { |
2107 | io_cqring_add_event(ctx, req->user_data, ret); | |
2108 | if (req->flags & REQ_F_LINK) | |
2109 | req->flags |= REQ_F_FAIL_LINK; | |
e65ef56d | 2110 | io_put_req(req); |
9e645e11 | 2111 | } |
2b188cc1 JA |
2112 | |
2113 | return ret; | |
2114 | } | |
2115 | ||
9e645e11 JA |
2116 | #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK) |
2117 | ||
2118 | static void io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s, | |
2119 | struct io_submit_state *state, struct io_kiocb **link) | |
2120 | { | |
2121 | struct io_uring_sqe *sqe_copy; | |
2122 | struct io_kiocb *req; | |
2123 | int ret; | |
2124 | ||
2125 | /* enforce forwards compatibility on users */ | |
2126 | if (unlikely(s->sqe->flags & ~SQE_VALID_FLAGS)) { | |
2127 | ret = -EINVAL; | |
2128 | goto err; | |
2129 | } | |
2130 | ||
2131 | req = io_get_req(ctx, state); | |
2132 | if (unlikely(!req)) { | |
2133 | ret = -EAGAIN; | |
2134 | goto err; | |
2135 | } | |
2136 | ||
2137 | ret = io_req_set_file(ctx, s, state, req); | |
2138 | if (unlikely(ret)) { | |
2139 | err_req: | |
2140 | io_free_req(req); | |
2141 | err: | |
2142 | io_cqring_add_event(ctx, s->sqe->user_data, ret); | |
2143 | return; | |
2144 | } | |
2145 | ||
9e645e11 JA |
2146 | /* |
2147 | * If we already have a head request, queue this one for async | |
2148 | * submittal once the head completes. If we don't have a head but | |
2149 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
2150 | * submitted sync once the chain is complete. If none of those | |
2151 | * conditions are true (normal request), then just queue it. | |
2152 | */ | |
2153 | if (*link) { | |
2154 | struct io_kiocb *prev = *link; | |
2155 | ||
2156 | sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL); | |
2157 | if (!sqe_copy) { | |
2158 | ret = -EAGAIN; | |
2159 | goto err_req; | |
2160 | } | |
2161 | ||
2162 | s->sqe = sqe_copy; | |
2163 | memcpy(&req->submit, s, sizeof(*s)); | |
2164 | list_add_tail(&req->list, &prev->link_list); | |
2165 | } else if (s->sqe->flags & IOSQE_IO_LINK) { | |
2166 | req->flags |= REQ_F_LINK; | |
2167 | ||
2168 | memcpy(&req->submit, s, sizeof(*s)); | |
2169 | INIT_LIST_HEAD(&req->link_list); | |
2170 | *link = req; | |
2171 | } else { | |
2172 | io_queue_sqe(ctx, req, s); | |
2173 | } | |
2174 | } | |
2175 | ||
9a56a232 JA |
2176 | /* |
2177 | * Batched submission is done, ensure local IO is flushed out. | |
2178 | */ | |
2179 | static void io_submit_state_end(struct io_submit_state *state) | |
2180 | { | |
2181 | blk_finish_plug(&state->plug); | |
3d6770fb | 2182 | io_file_put(state); |
2579f913 JA |
2183 | if (state->free_reqs) |
2184 | kmem_cache_free_bulk(req_cachep, state->free_reqs, | |
2185 | &state->reqs[state->cur_req]); | |
9a56a232 JA |
2186 | } |
2187 | ||
2188 | /* | |
2189 | * Start submission side cache. | |
2190 | */ | |
2191 | static void io_submit_state_start(struct io_submit_state *state, | |
2192 | struct io_ring_ctx *ctx, unsigned max_ios) | |
2193 | { | |
2194 | blk_start_plug(&state->plug); | |
2579f913 | 2195 | state->free_reqs = 0; |
9a56a232 JA |
2196 | state->file = NULL; |
2197 | state->ios_left = max_ios; | |
2198 | } | |
2199 | ||
2b188cc1 JA |
2200 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
2201 | { | |
2202 | struct io_sq_ring *ring = ctx->sq_ring; | |
2203 | ||
2204 | if (ctx->cached_sq_head != READ_ONCE(ring->r.head)) { | |
2205 | /* | |
2206 | * Ensure any loads from the SQEs are done at this point, | |
2207 | * since once we write the new head, the application could | |
2208 | * write new data to them. | |
2209 | */ | |
2210 | smp_store_release(&ring->r.head, ctx->cached_sq_head); | |
2b188cc1 JA |
2211 | } |
2212 | } | |
2213 | ||
2b188cc1 JA |
2214 | /* |
2215 | * Fetch an sqe, if one is available. Note that s->sqe will point to memory | |
2216 | * that is mapped by userspace. This means that care needs to be taken to | |
2217 | * ensure that reads are stable, as we cannot rely on userspace always | |
2218 | * being a good citizen. If members of the sqe are validated and then later | |
2219 | * used, it's important that those reads are done through READ_ONCE() to | |
2220 | * prevent a re-load down the line. | |
2221 | */ | |
2222 | static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s) | |
2223 | { | |
2224 | struct io_sq_ring *ring = ctx->sq_ring; | |
2225 | unsigned head; | |
2226 | ||
2227 | /* | |
2228 | * The cached sq head (or cq tail) serves two purposes: | |
2229 | * | |
2230 | * 1) allows us to batch the cost of updating the user visible | |
2231 | * head updates. | |
2232 | * 2) allows the kernel side to track the head on its own, even | |
2233 | * though the application is the one updating it. | |
2234 | */ | |
2235 | head = ctx->cached_sq_head; | |
e523a29c SB |
2236 | /* make sure SQ entry isn't read before tail */ |
2237 | if (head == smp_load_acquire(&ring->r.tail)) | |
2b188cc1 JA |
2238 | return false; |
2239 | ||
2240 | head = READ_ONCE(ring->array[head & ctx->sq_mask]); | |
2241 | if (head < ctx->sq_entries) { | |
2242 | s->index = head; | |
2243 | s->sqe = &ctx->sq_sqes[head]; | |
2244 | ctx->cached_sq_head++; | |
2245 | return true; | |
2246 | } | |
2247 | ||
2248 | /* drop invalid entries */ | |
2249 | ctx->cached_sq_head++; | |
2250 | ring->dropped++; | |
2b188cc1 JA |
2251 | return false; |
2252 | } | |
2253 | ||
6c271ce2 JA |
2254 | static int io_submit_sqes(struct io_ring_ctx *ctx, struct sqe_submit *sqes, |
2255 | unsigned int nr, bool has_user, bool mm_fault) | |
2256 | { | |
2257 | struct io_submit_state state, *statep = NULL; | |
9e645e11 JA |
2258 | struct io_kiocb *link = NULL; |
2259 | bool prev_was_link = false; | |
2260 | int i, submitted = 0; | |
6c271ce2 JA |
2261 | |
2262 | if (nr > IO_PLUG_THRESHOLD) { | |
2263 | io_submit_state_start(&state, ctx, nr); | |
2264 | statep = &state; | |
2265 | } | |
2266 | ||
2267 | for (i = 0; i < nr; i++) { | |
9e645e11 JA |
2268 | /* |
2269 | * If previous wasn't linked and we have a linked command, | |
2270 | * that's the end of the chain. Submit the previous link. | |
2271 | */ | |
2272 | if (!prev_was_link && link) { | |
2273 | io_queue_sqe(ctx, link, &link->submit); | |
2274 | link = NULL; | |
2275 | } | |
2276 | prev_was_link = (sqes[i].sqe->flags & IOSQE_IO_LINK) != 0; | |
2277 | ||
6c271ce2 | 2278 | if (unlikely(mm_fault)) { |
9e645e11 JA |
2279 | io_cqring_add_event(ctx, sqes[i].sqe->user_data, |
2280 | -EFAULT); | |
6c271ce2 JA |
2281 | } else { |
2282 | sqes[i].has_user = has_user; | |
2283 | sqes[i].needs_lock = true; | |
2284 | sqes[i].needs_fixed_file = true; | |
9e645e11 | 2285 | io_submit_sqe(ctx, &sqes[i], statep, &link); |
6c271ce2 | 2286 | submitted++; |
6c271ce2 | 2287 | } |
6c271ce2 JA |
2288 | } |
2289 | ||
9e645e11 JA |
2290 | if (link) |
2291 | io_queue_sqe(ctx, link, &link->submit); | |
6c271ce2 JA |
2292 | if (statep) |
2293 | io_submit_state_end(&state); | |
2294 | ||
2295 | return submitted; | |
2296 | } | |
2297 | ||
2298 | static int io_sq_thread(void *data) | |
2299 | { | |
2300 | struct sqe_submit sqes[IO_IOPOLL_BATCH]; | |
2301 | struct io_ring_ctx *ctx = data; | |
2302 | struct mm_struct *cur_mm = NULL; | |
2303 | mm_segment_t old_fs; | |
2304 | DEFINE_WAIT(wait); | |
2305 | unsigned inflight; | |
2306 | unsigned long timeout; | |
2307 | ||
a4c0b3de JL |
2308 | complete(&ctx->sqo_thread_started); |
2309 | ||
6c271ce2 JA |
2310 | old_fs = get_fs(); |
2311 | set_fs(USER_DS); | |
2312 | ||
2313 | timeout = inflight = 0; | |
2bbcd6d3 | 2314 | while (!kthread_should_park()) { |
6c271ce2 JA |
2315 | bool all_fixed, mm_fault = false; |
2316 | int i; | |
2317 | ||
2318 | if (inflight) { | |
2319 | unsigned nr_events = 0; | |
2320 | ||
2321 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
6c271ce2 | 2322 | io_iopoll_check(ctx, &nr_events, 0); |
6c271ce2 JA |
2323 | } else { |
2324 | /* | |
2325 | * Normal IO, just pretend everything completed. | |
2326 | * We don't have to poll completions for that. | |
2327 | */ | |
2328 | nr_events = inflight; | |
2329 | } | |
2330 | ||
2331 | inflight -= nr_events; | |
2332 | if (!inflight) | |
2333 | timeout = jiffies + ctx->sq_thread_idle; | |
2334 | } | |
2335 | ||
2336 | if (!io_get_sqring(ctx, &sqes[0])) { | |
2337 | /* | |
2338 | * We're polling. If we're within the defined idle | |
2339 | * period, then let us spin without work before going | |
2340 | * to sleep. | |
2341 | */ | |
2342 | if (inflight || !time_after(jiffies, timeout)) { | |
2343 | cpu_relax(); | |
2344 | continue; | |
2345 | } | |
2346 | ||
2347 | /* | |
2348 | * Drop cur_mm before scheduling, we can't hold it for | |
2349 | * long periods (or over schedule()). Do this before | |
2350 | * adding ourselves to the waitqueue, as the unuse/drop | |
2351 | * may sleep. | |
2352 | */ | |
2353 | if (cur_mm) { | |
2354 | unuse_mm(cur_mm); | |
2355 | mmput(cur_mm); | |
2356 | cur_mm = NULL; | |
2357 | } | |
2358 | ||
2359 | prepare_to_wait(&ctx->sqo_wait, &wait, | |
2360 | TASK_INTERRUPTIBLE); | |
2361 | ||
2362 | /* Tell userspace we may need a wakeup call */ | |
2363 | ctx->sq_ring->flags |= IORING_SQ_NEED_WAKEUP; | |
0d7bae69 SB |
2364 | /* make sure to read SQ tail after writing flags */ |
2365 | smp_mb(); | |
6c271ce2 JA |
2366 | |
2367 | if (!io_get_sqring(ctx, &sqes[0])) { | |
2bbcd6d3 | 2368 | if (kthread_should_park()) { |
6c271ce2 JA |
2369 | finish_wait(&ctx->sqo_wait, &wait); |
2370 | break; | |
2371 | } | |
2372 | if (signal_pending(current)) | |
2373 | flush_signals(current); | |
2374 | schedule(); | |
2375 | finish_wait(&ctx->sqo_wait, &wait); | |
2376 | ||
2377 | ctx->sq_ring->flags &= ~IORING_SQ_NEED_WAKEUP; | |
6c271ce2 JA |
2378 | continue; |
2379 | } | |
2380 | finish_wait(&ctx->sqo_wait, &wait); | |
2381 | ||
2382 | ctx->sq_ring->flags &= ~IORING_SQ_NEED_WAKEUP; | |
6c271ce2 JA |
2383 | } |
2384 | ||
2385 | i = 0; | |
2386 | all_fixed = true; | |
2387 | do { | |
2388 | if (all_fixed && io_sqe_needs_user(sqes[i].sqe)) | |
2389 | all_fixed = false; | |
2390 | ||
2391 | i++; | |
2392 | if (i == ARRAY_SIZE(sqes)) | |
2393 | break; | |
2394 | } while (io_get_sqring(ctx, &sqes[i])); | |
2395 | ||
2396 | /* Unless all new commands are FIXED regions, grab mm */ | |
2397 | if (!all_fixed && !cur_mm) { | |
2398 | mm_fault = !mmget_not_zero(ctx->sqo_mm); | |
2399 | if (!mm_fault) { | |
2400 | use_mm(ctx->sqo_mm); | |
2401 | cur_mm = ctx->sqo_mm; | |
2402 | } | |
2403 | } | |
2404 | ||
2405 | inflight += io_submit_sqes(ctx, sqes, i, cur_mm != NULL, | |
2406 | mm_fault); | |
2407 | ||
2408 | /* Commit SQ ring head once we've consumed all SQEs */ | |
2409 | io_commit_sqring(ctx); | |
2410 | } | |
2411 | ||
2412 | set_fs(old_fs); | |
2413 | if (cur_mm) { | |
2414 | unuse_mm(cur_mm); | |
2415 | mmput(cur_mm); | |
2416 | } | |
06058632 | 2417 | |
2bbcd6d3 | 2418 | kthread_parkme(); |
06058632 | 2419 | |
6c271ce2 JA |
2420 | return 0; |
2421 | } | |
2422 | ||
2b188cc1 JA |
2423 | static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit) |
2424 | { | |
9a56a232 | 2425 | struct io_submit_state state, *statep = NULL; |
9e645e11 JA |
2426 | struct io_kiocb *link = NULL; |
2427 | bool prev_was_link = false; | |
5c8b0b54 | 2428 | int i, submit = 0; |
2b188cc1 | 2429 | |
9a56a232 JA |
2430 | if (to_submit > IO_PLUG_THRESHOLD) { |
2431 | io_submit_state_start(&state, ctx, to_submit); | |
2432 | statep = &state; | |
2433 | } | |
2b188cc1 JA |
2434 | |
2435 | for (i = 0; i < to_submit; i++) { | |
2436 | struct sqe_submit s; | |
2437 | ||
2438 | if (!io_get_sqring(ctx, &s)) | |
2439 | break; | |
2440 | ||
9e645e11 JA |
2441 | /* |
2442 | * If previous wasn't linked and we have a linked command, | |
2443 | * that's the end of the chain. Submit the previous link. | |
2444 | */ | |
2445 | if (!prev_was_link && link) { | |
2446 | io_queue_sqe(ctx, link, &link->submit); | |
2447 | link = NULL; | |
2448 | } | |
2449 | prev_was_link = (s.sqe->flags & IOSQE_IO_LINK) != 0; | |
2450 | ||
2b188cc1 | 2451 | s.has_user = true; |
def596e9 | 2452 | s.needs_lock = false; |
6c271ce2 | 2453 | s.needs_fixed_file = false; |
5c8b0b54 | 2454 | submit++; |
9e645e11 | 2455 | io_submit_sqe(ctx, &s, statep, &link); |
2b188cc1 JA |
2456 | } |
2457 | io_commit_sqring(ctx); | |
2458 | ||
9e645e11 JA |
2459 | if (link) |
2460 | io_queue_sqe(ctx, link, &link->submit); | |
9a56a232 JA |
2461 | if (statep) |
2462 | io_submit_state_end(statep); | |
2b188cc1 | 2463 | |
5c8b0b54 | 2464 | return submit; |
2b188cc1 JA |
2465 | } |
2466 | ||
2b188cc1 JA |
2467 | /* |
2468 | * Wait until events become available, if we don't already have some. The | |
2469 | * application must reap them itself, as they reside on the shared cq ring. | |
2470 | */ | |
2471 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
2472 | const sigset_t __user *sig, size_t sigsz) | |
2473 | { | |
2474 | struct io_cq_ring *ring = ctx->cq_ring; | |
2b188cc1 JA |
2475 | int ret; |
2476 | ||
2b188cc1 JA |
2477 | if (io_cqring_events(ring) >= min_events) |
2478 | return 0; | |
2479 | ||
2480 | if (sig) { | |
9e75ad5d AB |
2481 | #ifdef CONFIG_COMPAT |
2482 | if (in_compat_syscall()) | |
2483 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 2484 | sigsz); |
9e75ad5d AB |
2485 | else |
2486 | #endif | |
b772434b | 2487 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 2488 | |
2b188cc1 JA |
2489 | if (ret) |
2490 | return ret; | |
2491 | } | |
2492 | ||
fdb288a6 | 2493 | ret = wait_event_interruptible(ctx->wait, io_cqring_events(ring) >= min_events); |
b772434b | 2494 | restore_saved_sigmask_unless(ret == -ERESTARTSYS); |
97abc889 ON |
2495 | if (ret == -ERESTARTSYS) |
2496 | ret = -EINTR; | |
2b188cc1 JA |
2497 | |
2498 | return READ_ONCE(ring->r.head) == READ_ONCE(ring->r.tail) ? ret : 0; | |
2499 | } | |
2500 | ||
6b06314c JA |
2501 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
2502 | { | |
2503 | #if defined(CONFIG_UNIX) | |
2504 | if (ctx->ring_sock) { | |
2505 | struct sock *sock = ctx->ring_sock->sk; | |
2506 | struct sk_buff *skb; | |
2507 | ||
2508 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
2509 | kfree_skb(skb); | |
2510 | } | |
2511 | #else | |
2512 | int i; | |
2513 | ||
2514 | for (i = 0; i < ctx->nr_user_files; i++) | |
2515 | fput(ctx->user_files[i]); | |
2516 | #endif | |
2517 | } | |
2518 | ||
2519 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) | |
2520 | { | |
2521 | if (!ctx->user_files) | |
2522 | return -ENXIO; | |
2523 | ||
2524 | __io_sqe_files_unregister(ctx); | |
2525 | kfree(ctx->user_files); | |
2526 | ctx->user_files = NULL; | |
2527 | ctx->nr_user_files = 0; | |
2528 | return 0; | |
2529 | } | |
2530 | ||
6c271ce2 JA |
2531 | static void io_sq_thread_stop(struct io_ring_ctx *ctx) |
2532 | { | |
2533 | if (ctx->sqo_thread) { | |
a4c0b3de | 2534 | wait_for_completion(&ctx->sqo_thread_started); |
2bbcd6d3 RP |
2535 | /* |
2536 | * The park is a bit of a work-around, without it we get | |
2537 | * warning spews on shutdown with SQPOLL set and affinity | |
2538 | * set to a single CPU. | |
2539 | */ | |
06058632 | 2540 | kthread_park(ctx->sqo_thread); |
6c271ce2 JA |
2541 | kthread_stop(ctx->sqo_thread); |
2542 | ctx->sqo_thread = NULL; | |
2543 | } | |
2544 | } | |
2545 | ||
6b06314c JA |
2546 | static void io_finish_async(struct io_ring_ctx *ctx) |
2547 | { | |
6c271ce2 JA |
2548 | io_sq_thread_stop(ctx); |
2549 | ||
6b06314c JA |
2550 | if (ctx->sqo_wq) { |
2551 | destroy_workqueue(ctx->sqo_wq); | |
2552 | ctx->sqo_wq = NULL; | |
2553 | } | |
2554 | } | |
2555 | ||
2556 | #if defined(CONFIG_UNIX) | |
2557 | static void io_destruct_skb(struct sk_buff *skb) | |
2558 | { | |
2559 | struct io_ring_ctx *ctx = skb->sk->sk_user_data; | |
2560 | ||
2561 | io_finish_async(ctx); | |
2562 | unix_destruct_scm(skb); | |
2563 | } | |
2564 | ||
2565 | /* | |
2566 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
2567 | * the io_uring can be safely unregistered on process exit, even if we have | |
2568 | * loops in the file referencing. | |
2569 | */ | |
2570 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
2571 | { | |
2572 | struct sock *sk = ctx->ring_sock->sk; | |
2573 | struct scm_fp_list *fpl; | |
2574 | struct sk_buff *skb; | |
2575 | int i; | |
2576 | ||
2577 | if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) { | |
2578 | unsigned long inflight = ctx->user->unix_inflight + nr; | |
2579 | ||
2580 | if (inflight > task_rlimit(current, RLIMIT_NOFILE)) | |
2581 | return -EMFILE; | |
2582 | } | |
2583 | ||
2584 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); | |
2585 | if (!fpl) | |
2586 | return -ENOMEM; | |
2587 | ||
2588 | skb = alloc_skb(0, GFP_KERNEL); | |
2589 | if (!skb) { | |
2590 | kfree(fpl); | |
2591 | return -ENOMEM; | |
2592 | } | |
2593 | ||
2594 | skb->sk = sk; | |
2595 | skb->destructor = io_destruct_skb; | |
2596 | ||
2597 | fpl->user = get_uid(ctx->user); | |
2598 | for (i = 0; i < nr; i++) { | |
2599 | fpl->fp[i] = get_file(ctx->user_files[i + offset]); | |
2600 | unix_inflight(fpl->user, fpl->fp[i]); | |
2601 | } | |
2602 | ||
2603 | fpl->max = fpl->count = nr; | |
2604 | UNIXCB(skb).fp = fpl; | |
2605 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); | |
2606 | skb_queue_head(&sk->sk_receive_queue, skb); | |
2607 | ||
2608 | for (i = 0; i < nr; i++) | |
2609 | fput(fpl->fp[i]); | |
2610 | ||
2611 | return 0; | |
2612 | } | |
2613 | ||
2614 | /* | |
2615 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
2616 | * causes regular reference counting to break down. We rely on the UNIX | |
2617 | * garbage collection to take care of this problem for us. | |
2618 | */ | |
2619 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
2620 | { | |
2621 | unsigned left, total; | |
2622 | int ret = 0; | |
2623 | ||
2624 | total = 0; | |
2625 | left = ctx->nr_user_files; | |
2626 | while (left) { | |
2627 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
2628 | |
2629 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
2630 | if (ret) | |
2631 | break; | |
2632 | left -= this_files; | |
2633 | total += this_files; | |
2634 | } | |
2635 | ||
2636 | if (!ret) | |
2637 | return 0; | |
2638 | ||
2639 | while (total < ctx->nr_user_files) { | |
2640 | fput(ctx->user_files[total]); | |
2641 | total++; | |
2642 | } | |
2643 | ||
2644 | return ret; | |
2645 | } | |
2646 | #else | |
2647 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
2648 | { | |
2649 | return 0; | |
2650 | } | |
2651 | #endif | |
2652 | ||
2653 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, | |
2654 | unsigned nr_args) | |
2655 | { | |
2656 | __s32 __user *fds = (__s32 __user *) arg; | |
2657 | int fd, ret = 0; | |
2658 | unsigned i; | |
2659 | ||
2660 | if (ctx->user_files) | |
2661 | return -EBUSY; | |
2662 | if (!nr_args) | |
2663 | return -EINVAL; | |
2664 | if (nr_args > IORING_MAX_FIXED_FILES) | |
2665 | return -EMFILE; | |
2666 | ||
2667 | ctx->user_files = kcalloc(nr_args, sizeof(struct file *), GFP_KERNEL); | |
2668 | if (!ctx->user_files) | |
2669 | return -ENOMEM; | |
2670 | ||
2671 | for (i = 0; i < nr_args; i++) { | |
2672 | ret = -EFAULT; | |
2673 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) | |
2674 | break; | |
2675 | ||
2676 | ctx->user_files[i] = fget(fd); | |
2677 | ||
2678 | ret = -EBADF; | |
2679 | if (!ctx->user_files[i]) | |
2680 | break; | |
2681 | /* | |
2682 | * Don't allow io_uring instances to be registered. If UNIX | |
2683 | * isn't enabled, then this causes a reference cycle and this | |
2684 | * instance can never get freed. If UNIX is enabled we'll | |
2685 | * handle it just fine, but there's still no point in allowing | |
2686 | * a ring fd as it doesn't support regular read/write anyway. | |
2687 | */ | |
2688 | if (ctx->user_files[i]->f_op == &io_uring_fops) { | |
2689 | fput(ctx->user_files[i]); | |
2690 | break; | |
2691 | } | |
2692 | ctx->nr_user_files++; | |
2693 | ret = 0; | |
2694 | } | |
2695 | ||
2696 | if (ret) { | |
2697 | for (i = 0; i < ctx->nr_user_files; i++) | |
2698 | fput(ctx->user_files[i]); | |
2699 | ||
2700 | kfree(ctx->user_files); | |
25adf50f | 2701 | ctx->user_files = NULL; |
6b06314c JA |
2702 | ctx->nr_user_files = 0; |
2703 | return ret; | |
2704 | } | |
2705 | ||
2706 | ret = io_sqe_files_scm(ctx); | |
2707 | if (ret) | |
2708 | io_sqe_files_unregister(ctx); | |
2709 | ||
2710 | return ret; | |
2711 | } | |
2712 | ||
6c271ce2 JA |
2713 | static int io_sq_offload_start(struct io_ring_ctx *ctx, |
2714 | struct io_uring_params *p) | |
2b188cc1 JA |
2715 | { |
2716 | int ret; | |
2717 | ||
6c271ce2 | 2718 | init_waitqueue_head(&ctx->sqo_wait); |
2b188cc1 JA |
2719 | mmgrab(current->mm); |
2720 | ctx->sqo_mm = current->mm; | |
2721 | ||
6c271ce2 | 2722 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3ec482d1 JA |
2723 | ret = -EPERM; |
2724 | if (!capable(CAP_SYS_ADMIN)) | |
2725 | goto err; | |
2726 | ||
917257da JA |
2727 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
2728 | if (!ctx->sq_thread_idle) | |
2729 | ctx->sq_thread_idle = HZ; | |
2730 | ||
6c271ce2 | 2731 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 2732 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 2733 | |
917257da | 2734 | ret = -EINVAL; |
44a9bd18 JA |
2735 | if (cpu >= nr_cpu_ids) |
2736 | goto err; | |
7889f44d | 2737 | if (!cpu_online(cpu)) |
917257da JA |
2738 | goto err; |
2739 | ||
6c271ce2 JA |
2740 | ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread, |
2741 | ctx, cpu, | |
2742 | "io_uring-sq"); | |
2743 | } else { | |
2744 | ctx->sqo_thread = kthread_create(io_sq_thread, ctx, | |
2745 | "io_uring-sq"); | |
2746 | } | |
2747 | if (IS_ERR(ctx->sqo_thread)) { | |
2748 | ret = PTR_ERR(ctx->sqo_thread); | |
2749 | ctx->sqo_thread = NULL; | |
2750 | goto err; | |
2751 | } | |
2752 | wake_up_process(ctx->sqo_thread); | |
2753 | } else if (p->flags & IORING_SETUP_SQ_AFF) { | |
2754 | /* Can't have SQ_AFF without SQPOLL */ | |
2755 | ret = -EINVAL; | |
2756 | goto err; | |
2757 | } | |
2758 | ||
2b188cc1 JA |
2759 | /* Do QD, or 2 * CPUS, whatever is smallest */ |
2760 | ctx->sqo_wq = alloc_workqueue("io_ring-wq", WQ_UNBOUND | WQ_FREEZABLE, | |
2761 | min(ctx->sq_entries - 1, 2 * num_online_cpus())); | |
2762 | if (!ctx->sqo_wq) { | |
2763 | ret = -ENOMEM; | |
2764 | goto err; | |
2765 | } | |
2766 | ||
2767 | return 0; | |
2768 | err: | |
6c271ce2 | 2769 | io_sq_thread_stop(ctx); |
2b188cc1 JA |
2770 | mmdrop(ctx->sqo_mm); |
2771 | ctx->sqo_mm = NULL; | |
2772 | return ret; | |
2773 | } | |
2774 | ||
2775 | static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages) | |
2776 | { | |
2777 | atomic_long_sub(nr_pages, &user->locked_vm); | |
2778 | } | |
2779 | ||
2780 | static int io_account_mem(struct user_struct *user, unsigned long nr_pages) | |
2781 | { | |
2782 | unsigned long page_limit, cur_pages, new_pages; | |
2783 | ||
2784 | /* Don't allow more pages than we can safely lock */ | |
2785 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
2786 | ||
2787 | do { | |
2788 | cur_pages = atomic_long_read(&user->locked_vm); | |
2789 | new_pages = cur_pages + nr_pages; | |
2790 | if (new_pages > page_limit) | |
2791 | return -ENOMEM; | |
2792 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
2793 | new_pages) != cur_pages); | |
2794 | ||
2795 | return 0; | |
2796 | } | |
2797 | ||
2798 | static void io_mem_free(void *ptr) | |
2799 | { | |
52e04ef4 MR |
2800 | struct page *page; |
2801 | ||
2802 | if (!ptr) | |
2803 | return; | |
2b188cc1 | 2804 | |
52e04ef4 | 2805 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
2806 | if (put_page_testzero(page)) |
2807 | free_compound_page(page); | |
2808 | } | |
2809 | ||
2810 | static void *io_mem_alloc(size_t size) | |
2811 | { | |
2812 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
2813 | __GFP_NORETRY; | |
2814 | ||
2815 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
2816 | } | |
2817 | ||
2818 | static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries) | |
2819 | { | |
2820 | struct io_sq_ring *sq_ring; | |
2821 | struct io_cq_ring *cq_ring; | |
2822 | size_t bytes; | |
2823 | ||
2824 | bytes = struct_size(sq_ring, array, sq_entries); | |
2825 | bytes += array_size(sizeof(struct io_uring_sqe), sq_entries); | |
2826 | bytes += struct_size(cq_ring, cqes, cq_entries); | |
2827 | ||
2828 | return (bytes + PAGE_SIZE - 1) / PAGE_SIZE; | |
2829 | } | |
2830 | ||
edafccee JA |
2831 | static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx) |
2832 | { | |
2833 | int i, j; | |
2834 | ||
2835 | if (!ctx->user_bufs) | |
2836 | return -ENXIO; | |
2837 | ||
2838 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
2839 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
2840 | ||
2841 | for (j = 0; j < imu->nr_bvecs; j++) | |
2842 | put_page(imu->bvec[j].bv_page); | |
2843 | ||
2844 | if (ctx->account_mem) | |
2845 | io_unaccount_mem(ctx->user, imu->nr_bvecs); | |
d4ef6475 | 2846 | kvfree(imu->bvec); |
edafccee JA |
2847 | imu->nr_bvecs = 0; |
2848 | } | |
2849 | ||
2850 | kfree(ctx->user_bufs); | |
2851 | ctx->user_bufs = NULL; | |
2852 | ctx->nr_user_bufs = 0; | |
2853 | return 0; | |
2854 | } | |
2855 | ||
2856 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
2857 | void __user *arg, unsigned index) | |
2858 | { | |
2859 | struct iovec __user *src; | |
2860 | ||
2861 | #ifdef CONFIG_COMPAT | |
2862 | if (ctx->compat) { | |
2863 | struct compat_iovec __user *ciovs; | |
2864 | struct compat_iovec ciov; | |
2865 | ||
2866 | ciovs = (struct compat_iovec __user *) arg; | |
2867 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
2868 | return -EFAULT; | |
2869 | ||
2870 | dst->iov_base = (void __user *) (unsigned long) ciov.iov_base; | |
2871 | dst->iov_len = ciov.iov_len; | |
2872 | return 0; | |
2873 | } | |
2874 | #endif | |
2875 | src = (struct iovec __user *) arg; | |
2876 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
2877 | return -EFAULT; | |
2878 | return 0; | |
2879 | } | |
2880 | ||
2881 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, | |
2882 | unsigned nr_args) | |
2883 | { | |
2884 | struct vm_area_struct **vmas = NULL; | |
2885 | struct page **pages = NULL; | |
2886 | int i, j, got_pages = 0; | |
2887 | int ret = -EINVAL; | |
2888 | ||
2889 | if (ctx->user_bufs) | |
2890 | return -EBUSY; | |
2891 | if (!nr_args || nr_args > UIO_MAXIOV) | |
2892 | return -EINVAL; | |
2893 | ||
2894 | ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf), | |
2895 | GFP_KERNEL); | |
2896 | if (!ctx->user_bufs) | |
2897 | return -ENOMEM; | |
2898 | ||
2899 | for (i = 0; i < nr_args; i++) { | |
2900 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
2901 | unsigned long off, start, end, ubuf; | |
2902 | int pret, nr_pages; | |
2903 | struct iovec iov; | |
2904 | size_t size; | |
2905 | ||
2906 | ret = io_copy_iov(ctx, &iov, arg, i); | |
2907 | if (ret) | |
a278682d | 2908 | goto err; |
edafccee JA |
2909 | |
2910 | /* | |
2911 | * Don't impose further limits on the size and buffer | |
2912 | * constraints here, we'll -EINVAL later when IO is | |
2913 | * submitted if they are wrong. | |
2914 | */ | |
2915 | ret = -EFAULT; | |
2916 | if (!iov.iov_base || !iov.iov_len) | |
2917 | goto err; | |
2918 | ||
2919 | /* arbitrary limit, but we need something */ | |
2920 | if (iov.iov_len > SZ_1G) | |
2921 | goto err; | |
2922 | ||
2923 | ubuf = (unsigned long) iov.iov_base; | |
2924 | end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
2925 | start = ubuf >> PAGE_SHIFT; | |
2926 | nr_pages = end - start; | |
2927 | ||
2928 | if (ctx->account_mem) { | |
2929 | ret = io_account_mem(ctx->user, nr_pages); | |
2930 | if (ret) | |
2931 | goto err; | |
2932 | } | |
2933 | ||
2934 | ret = 0; | |
2935 | if (!pages || nr_pages > got_pages) { | |
2936 | kfree(vmas); | |
2937 | kfree(pages); | |
d4ef6475 | 2938 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), |
edafccee | 2939 | GFP_KERNEL); |
d4ef6475 | 2940 | vmas = kvmalloc_array(nr_pages, |
edafccee JA |
2941 | sizeof(struct vm_area_struct *), |
2942 | GFP_KERNEL); | |
2943 | if (!pages || !vmas) { | |
2944 | ret = -ENOMEM; | |
2945 | if (ctx->account_mem) | |
2946 | io_unaccount_mem(ctx->user, nr_pages); | |
2947 | goto err; | |
2948 | } | |
2949 | got_pages = nr_pages; | |
2950 | } | |
2951 | ||
d4ef6475 | 2952 | imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec), |
edafccee JA |
2953 | GFP_KERNEL); |
2954 | ret = -ENOMEM; | |
2955 | if (!imu->bvec) { | |
2956 | if (ctx->account_mem) | |
2957 | io_unaccount_mem(ctx->user, nr_pages); | |
2958 | goto err; | |
2959 | } | |
2960 | ||
2961 | ret = 0; | |
2962 | down_read(¤t->mm->mmap_sem); | |
932f4a63 IW |
2963 | pret = get_user_pages(ubuf, nr_pages, |
2964 | FOLL_WRITE | FOLL_LONGTERM, | |
2965 | pages, vmas); | |
edafccee JA |
2966 | if (pret == nr_pages) { |
2967 | /* don't support file backed memory */ | |
2968 | for (j = 0; j < nr_pages; j++) { | |
2969 | struct vm_area_struct *vma = vmas[j]; | |
2970 | ||
2971 | if (vma->vm_file && | |
2972 | !is_file_hugepages(vma->vm_file)) { | |
2973 | ret = -EOPNOTSUPP; | |
2974 | break; | |
2975 | } | |
2976 | } | |
2977 | } else { | |
2978 | ret = pret < 0 ? pret : -EFAULT; | |
2979 | } | |
2980 | up_read(¤t->mm->mmap_sem); | |
2981 | if (ret) { | |
2982 | /* | |
2983 | * if we did partial map, or found file backed vmas, | |
2984 | * release any pages we did get | |
2985 | */ | |
2986 | if (pret > 0) { | |
2987 | for (j = 0; j < pret; j++) | |
2988 | put_page(pages[j]); | |
2989 | } | |
2990 | if (ctx->account_mem) | |
2991 | io_unaccount_mem(ctx->user, nr_pages); | |
d4ef6475 | 2992 | kvfree(imu->bvec); |
edafccee JA |
2993 | goto err; |
2994 | } | |
2995 | ||
2996 | off = ubuf & ~PAGE_MASK; | |
2997 | size = iov.iov_len; | |
2998 | for (j = 0; j < nr_pages; j++) { | |
2999 | size_t vec_len; | |
3000 | ||
3001 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
3002 | imu->bvec[j].bv_page = pages[j]; | |
3003 | imu->bvec[j].bv_len = vec_len; | |
3004 | imu->bvec[j].bv_offset = off; | |
3005 | off = 0; | |
3006 | size -= vec_len; | |
3007 | } | |
3008 | /* store original address for later verification */ | |
3009 | imu->ubuf = ubuf; | |
3010 | imu->len = iov.iov_len; | |
3011 | imu->nr_bvecs = nr_pages; | |
3012 | ||
3013 | ctx->nr_user_bufs++; | |
3014 | } | |
d4ef6475 MR |
3015 | kvfree(pages); |
3016 | kvfree(vmas); | |
edafccee JA |
3017 | return 0; |
3018 | err: | |
d4ef6475 MR |
3019 | kvfree(pages); |
3020 | kvfree(vmas); | |
edafccee JA |
3021 | io_sqe_buffer_unregister(ctx); |
3022 | return ret; | |
3023 | } | |
3024 | ||
9b402849 JA |
3025 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
3026 | { | |
3027 | __s32 __user *fds = arg; | |
3028 | int fd; | |
3029 | ||
3030 | if (ctx->cq_ev_fd) | |
3031 | return -EBUSY; | |
3032 | ||
3033 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
3034 | return -EFAULT; | |
3035 | ||
3036 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
3037 | if (IS_ERR(ctx->cq_ev_fd)) { | |
3038 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
3039 | ctx->cq_ev_fd = NULL; | |
3040 | return ret; | |
3041 | } | |
3042 | ||
3043 | return 0; | |
3044 | } | |
3045 | ||
3046 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
3047 | { | |
3048 | if (ctx->cq_ev_fd) { | |
3049 | eventfd_ctx_put(ctx->cq_ev_fd); | |
3050 | ctx->cq_ev_fd = NULL; | |
3051 | return 0; | |
3052 | } | |
3053 | ||
3054 | return -ENXIO; | |
3055 | } | |
3056 | ||
2b188cc1 JA |
3057 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
3058 | { | |
6b06314c | 3059 | io_finish_async(ctx); |
2b188cc1 JA |
3060 | if (ctx->sqo_mm) |
3061 | mmdrop(ctx->sqo_mm); | |
def596e9 JA |
3062 | |
3063 | io_iopoll_reap_events(ctx); | |
edafccee | 3064 | io_sqe_buffer_unregister(ctx); |
6b06314c | 3065 | io_sqe_files_unregister(ctx); |
9b402849 | 3066 | io_eventfd_unregister(ctx); |
def596e9 | 3067 | |
2b188cc1 | 3068 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
3069 | if (ctx->ring_sock) { |
3070 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 3071 | sock_release(ctx->ring_sock); |
355e8d26 | 3072 | } |
2b188cc1 JA |
3073 | #endif |
3074 | ||
3075 | io_mem_free(ctx->sq_ring); | |
3076 | io_mem_free(ctx->sq_sqes); | |
3077 | io_mem_free(ctx->cq_ring); | |
3078 | ||
3079 | percpu_ref_exit(&ctx->refs); | |
3080 | if (ctx->account_mem) | |
3081 | io_unaccount_mem(ctx->user, | |
3082 | ring_pages(ctx->sq_entries, ctx->cq_entries)); | |
3083 | free_uid(ctx->user); | |
3084 | kfree(ctx); | |
3085 | } | |
3086 | ||
3087 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
3088 | { | |
3089 | struct io_ring_ctx *ctx = file->private_data; | |
3090 | __poll_t mask = 0; | |
3091 | ||
3092 | poll_wait(file, &ctx->cq_wait, wait); | |
4f7067c3 SB |
3093 | /* |
3094 | * synchronizes with barrier from wq_has_sleeper call in | |
3095 | * io_commit_cqring | |
3096 | */ | |
2b188cc1 | 3097 | smp_rmb(); |
fb775faa SB |
3098 | if (READ_ONCE(ctx->sq_ring->r.tail) - ctx->cached_sq_head != |
3099 | ctx->sq_ring->ring_entries) | |
2b188cc1 JA |
3100 | mask |= EPOLLOUT | EPOLLWRNORM; |
3101 | if (READ_ONCE(ctx->cq_ring->r.head) != ctx->cached_cq_tail) | |
3102 | mask |= EPOLLIN | EPOLLRDNORM; | |
3103 | ||
3104 | return mask; | |
3105 | } | |
3106 | ||
3107 | static int io_uring_fasync(int fd, struct file *file, int on) | |
3108 | { | |
3109 | struct io_ring_ctx *ctx = file->private_data; | |
3110 | ||
3111 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
3112 | } | |
3113 | ||
3114 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) | |
3115 | { | |
3116 | mutex_lock(&ctx->uring_lock); | |
3117 | percpu_ref_kill(&ctx->refs); | |
3118 | mutex_unlock(&ctx->uring_lock); | |
3119 | ||
221c5eb2 | 3120 | io_poll_remove_all(ctx); |
def596e9 | 3121 | io_iopoll_reap_events(ctx); |
2b188cc1 JA |
3122 | wait_for_completion(&ctx->ctx_done); |
3123 | io_ring_ctx_free(ctx); | |
3124 | } | |
3125 | ||
3126 | static int io_uring_release(struct inode *inode, struct file *file) | |
3127 | { | |
3128 | struct io_ring_ctx *ctx = file->private_data; | |
3129 | ||
3130 | file->private_data = NULL; | |
3131 | io_ring_ctx_wait_and_kill(ctx); | |
3132 | return 0; | |
3133 | } | |
3134 | ||
3135 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
3136 | { | |
3137 | loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT; | |
3138 | unsigned long sz = vma->vm_end - vma->vm_start; | |
3139 | struct io_ring_ctx *ctx = file->private_data; | |
3140 | unsigned long pfn; | |
3141 | struct page *page; | |
3142 | void *ptr; | |
3143 | ||
3144 | switch (offset) { | |
3145 | case IORING_OFF_SQ_RING: | |
3146 | ptr = ctx->sq_ring; | |
3147 | break; | |
3148 | case IORING_OFF_SQES: | |
3149 | ptr = ctx->sq_sqes; | |
3150 | break; | |
3151 | case IORING_OFF_CQ_RING: | |
3152 | ptr = ctx->cq_ring; | |
3153 | break; | |
3154 | default: | |
3155 | return -EINVAL; | |
3156 | } | |
3157 | ||
3158 | page = virt_to_head_page(ptr); | |
3159 | if (sz > (PAGE_SIZE << compound_order(page))) | |
3160 | return -EINVAL; | |
3161 | ||
3162 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
3163 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
3164 | } | |
3165 | ||
3166 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, | |
3167 | u32, min_complete, u32, flags, const sigset_t __user *, sig, | |
3168 | size_t, sigsz) | |
3169 | { | |
3170 | struct io_ring_ctx *ctx; | |
3171 | long ret = -EBADF; | |
3172 | int submitted = 0; | |
3173 | struct fd f; | |
3174 | ||
6c271ce2 | 3175 | if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP)) |
2b188cc1 JA |
3176 | return -EINVAL; |
3177 | ||
3178 | f = fdget(fd); | |
3179 | if (!f.file) | |
3180 | return -EBADF; | |
3181 | ||
3182 | ret = -EOPNOTSUPP; | |
3183 | if (f.file->f_op != &io_uring_fops) | |
3184 | goto out_fput; | |
3185 | ||
3186 | ret = -ENXIO; | |
3187 | ctx = f.file->private_data; | |
3188 | if (!percpu_ref_tryget(&ctx->refs)) | |
3189 | goto out_fput; | |
3190 | ||
6c271ce2 JA |
3191 | /* |
3192 | * For SQ polling, the thread will do all submissions and completions. | |
3193 | * Just return the requested submit count, and wake the thread if | |
3194 | * we were asked to. | |
3195 | */ | |
3196 | if (ctx->flags & IORING_SETUP_SQPOLL) { | |
3197 | if (flags & IORING_ENTER_SQ_WAKEUP) | |
3198 | wake_up(&ctx->sqo_wait); | |
3199 | submitted = to_submit; | |
3200 | goto out_ctx; | |
3201 | } | |
3202 | ||
2b188cc1 JA |
3203 | ret = 0; |
3204 | if (to_submit) { | |
3205 | to_submit = min(to_submit, ctx->sq_entries); | |
3206 | ||
3207 | mutex_lock(&ctx->uring_lock); | |
3208 | submitted = io_ring_submit(ctx, to_submit); | |
3209 | mutex_unlock(&ctx->uring_lock); | |
2b188cc1 JA |
3210 | } |
3211 | if (flags & IORING_ENTER_GETEVENTS) { | |
def596e9 JA |
3212 | unsigned nr_events = 0; |
3213 | ||
2b188cc1 JA |
3214 | min_complete = min(min_complete, ctx->cq_entries); |
3215 | ||
def596e9 | 3216 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 | 3217 | ret = io_iopoll_check(ctx, &nr_events, min_complete); |
def596e9 JA |
3218 | } else { |
3219 | ret = io_cqring_wait(ctx, min_complete, sig, sigsz); | |
3220 | } | |
2b188cc1 JA |
3221 | } |
3222 | ||
3223 | out_ctx: | |
3224 | io_ring_drop_ctx_refs(ctx, 1); | |
3225 | out_fput: | |
3226 | fdput(f); | |
3227 | return submitted ? submitted : ret; | |
3228 | } | |
3229 | ||
3230 | static const struct file_operations io_uring_fops = { | |
3231 | .release = io_uring_release, | |
3232 | .mmap = io_uring_mmap, | |
3233 | .poll = io_uring_poll, | |
3234 | .fasync = io_uring_fasync, | |
3235 | }; | |
3236 | ||
3237 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
3238 | struct io_uring_params *p) | |
3239 | { | |
3240 | struct io_sq_ring *sq_ring; | |
3241 | struct io_cq_ring *cq_ring; | |
3242 | size_t size; | |
3243 | ||
3244 | sq_ring = io_mem_alloc(struct_size(sq_ring, array, p->sq_entries)); | |
3245 | if (!sq_ring) | |
3246 | return -ENOMEM; | |
3247 | ||
3248 | ctx->sq_ring = sq_ring; | |
3249 | sq_ring->ring_mask = p->sq_entries - 1; | |
3250 | sq_ring->ring_entries = p->sq_entries; | |
3251 | ctx->sq_mask = sq_ring->ring_mask; | |
3252 | ctx->sq_entries = sq_ring->ring_entries; | |
3253 | ||
3254 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
3255 | if (size == SIZE_MAX) | |
3256 | return -EOVERFLOW; | |
3257 | ||
3258 | ctx->sq_sqes = io_mem_alloc(size); | |
52e04ef4 | 3259 | if (!ctx->sq_sqes) |
2b188cc1 | 3260 | return -ENOMEM; |
2b188cc1 JA |
3261 | |
3262 | cq_ring = io_mem_alloc(struct_size(cq_ring, cqes, p->cq_entries)); | |
52e04ef4 | 3263 | if (!cq_ring) |
2b188cc1 | 3264 | return -ENOMEM; |
2b188cc1 JA |
3265 | |
3266 | ctx->cq_ring = cq_ring; | |
3267 | cq_ring->ring_mask = p->cq_entries - 1; | |
3268 | cq_ring->ring_entries = p->cq_entries; | |
3269 | ctx->cq_mask = cq_ring->ring_mask; | |
3270 | ctx->cq_entries = cq_ring->ring_entries; | |
3271 | return 0; | |
3272 | } | |
3273 | ||
3274 | /* | |
3275 | * Allocate an anonymous fd, this is what constitutes the application | |
3276 | * visible backing of an io_uring instance. The application mmaps this | |
3277 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
3278 | * we have to tie this fd to a socket for file garbage collection purposes. | |
3279 | */ | |
3280 | static int io_uring_get_fd(struct io_ring_ctx *ctx) | |
3281 | { | |
3282 | struct file *file; | |
3283 | int ret; | |
3284 | ||
3285 | #if defined(CONFIG_UNIX) | |
3286 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, | |
3287 | &ctx->ring_sock); | |
3288 | if (ret) | |
3289 | return ret; | |
3290 | #endif | |
3291 | ||
3292 | ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
3293 | if (ret < 0) | |
3294 | goto err; | |
3295 | ||
3296 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, | |
3297 | O_RDWR | O_CLOEXEC); | |
3298 | if (IS_ERR(file)) { | |
3299 | put_unused_fd(ret); | |
3300 | ret = PTR_ERR(file); | |
3301 | goto err; | |
3302 | } | |
3303 | ||
3304 | #if defined(CONFIG_UNIX) | |
3305 | ctx->ring_sock->file = file; | |
6b06314c | 3306 | ctx->ring_sock->sk->sk_user_data = ctx; |
2b188cc1 JA |
3307 | #endif |
3308 | fd_install(ret, file); | |
3309 | return ret; | |
3310 | err: | |
3311 | #if defined(CONFIG_UNIX) | |
3312 | sock_release(ctx->ring_sock); | |
3313 | ctx->ring_sock = NULL; | |
3314 | #endif | |
3315 | return ret; | |
3316 | } | |
3317 | ||
3318 | static int io_uring_create(unsigned entries, struct io_uring_params *p) | |
3319 | { | |
3320 | struct user_struct *user = NULL; | |
3321 | struct io_ring_ctx *ctx; | |
3322 | bool account_mem; | |
3323 | int ret; | |
3324 | ||
3325 | if (!entries || entries > IORING_MAX_ENTRIES) | |
3326 | return -EINVAL; | |
3327 | ||
3328 | /* | |
3329 | * Use twice as many entries for the CQ ring. It's possible for the | |
3330 | * application to drive a higher depth than the size of the SQ ring, | |
3331 | * since the sqes are only used at submission time. This allows for | |
3332 | * some flexibility in overcommitting a bit. | |
3333 | */ | |
3334 | p->sq_entries = roundup_pow_of_two(entries); | |
3335 | p->cq_entries = 2 * p->sq_entries; | |
3336 | ||
3337 | user = get_uid(current_user()); | |
3338 | account_mem = !capable(CAP_IPC_LOCK); | |
3339 | ||
3340 | if (account_mem) { | |
3341 | ret = io_account_mem(user, | |
3342 | ring_pages(p->sq_entries, p->cq_entries)); | |
3343 | if (ret) { | |
3344 | free_uid(user); | |
3345 | return ret; | |
3346 | } | |
3347 | } | |
3348 | ||
3349 | ctx = io_ring_ctx_alloc(p); | |
3350 | if (!ctx) { | |
3351 | if (account_mem) | |
3352 | io_unaccount_mem(user, ring_pages(p->sq_entries, | |
3353 | p->cq_entries)); | |
3354 | free_uid(user); | |
3355 | return -ENOMEM; | |
3356 | } | |
3357 | ctx->compat = in_compat_syscall(); | |
3358 | ctx->account_mem = account_mem; | |
3359 | ctx->user = user; | |
3360 | ||
3361 | ret = io_allocate_scq_urings(ctx, p); | |
3362 | if (ret) | |
3363 | goto err; | |
3364 | ||
6c271ce2 | 3365 | ret = io_sq_offload_start(ctx, p); |
2b188cc1 JA |
3366 | if (ret) |
3367 | goto err; | |
3368 | ||
3369 | ret = io_uring_get_fd(ctx); | |
3370 | if (ret < 0) | |
3371 | goto err; | |
3372 | ||
3373 | memset(&p->sq_off, 0, sizeof(p->sq_off)); | |
3374 | p->sq_off.head = offsetof(struct io_sq_ring, r.head); | |
3375 | p->sq_off.tail = offsetof(struct io_sq_ring, r.tail); | |
3376 | p->sq_off.ring_mask = offsetof(struct io_sq_ring, ring_mask); | |
3377 | p->sq_off.ring_entries = offsetof(struct io_sq_ring, ring_entries); | |
3378 | p->sq_off.flags = offsetof(struct io_sq_ring, flags); | |
3379 | p->sq_off.dropped = offsetof(struct io_sq_ring, dropped); | |
3380 | p->sq_off.array = offsetof(struct io_sq_ring, array); | |
3381 | ||
3382 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
3383 | p->cq_off.head = offsetof(struct io_cq_ring, r.head); | |
3384 | p->cq_off.tail = offsetof(struct io_cq_ring, r.tail); | |
3385 | p->cq_off.ring_mask = offsetof(struct io_cq_ring, ring_mask); | |
3386 | p->cq_off.ring_entries = offsetof(struct io_cq_ring, ring_entries); | |
3387 | p->cq_off.overflow = offsetof(struct io_cq_ring, overflow); | |
3388 | p->cq_off.cqes = offsetof(struct io_cq_ring, cqes); | |
3389 | return ret; | |
3390 | err: | |
3391 | io_ring_ctx_wait_and_kill(ctx); | |
3392 | return ret; | |
3393 | } | |
3394 | ||
3395 | /* | |
3396 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
3397 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
3398 | * params structure passed in. | |
3399 | */ | |
3400 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
3401 | { | |
3402 | struct io_uring_params p; | |
3403 | long ret; | |
3404 | int i; | |
3405 | ||
3406 | if (copy_from_user(&p, params, sizeof(p))) | |
3407 | return -EFAULT; | |
3408 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
3409 | if (p.resv[i]) | |
3410 | return -EINVAL; | |
3411 | } | |
3412 | ||
6c271ce2 JA |
3413 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
3414 | IORING_SETUP_SQ_AFF)) | |
2b188cc1 JA |
3415 | return -EINVAL; |
3416 | ||
3417 | ret = io_uring_create(entries, &p); | |
3418 | if (ret < 0) | |
3419 | return ret; | |
3420 | ||
3421 | if (copy_to_user(params, &p, sizeof(p))) | |
3422 | return -EFAULT; | |
3423 | ||
3424 | return ret; | |
3425 | } | |
3426 | ||
3427 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
3428 | struct io_uring_params __user *, params) | |
3429 | { | |
3430 | return io_uring_setup(entries, params); | |
3431 | } | |
3432 | ||
edafccee JA |
3433 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
3434 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
3435 | __releases(ctx->uring_lock) |
3436 | __acquires(ctx->uring_lock) | |
edafccee JA |
3437 | { |
3438 | int ret; | |
3439 | ||
35fa71a0 JA |
3440 | /* |
3441 | * We're inside the ring mutex, if the ref is already dying, then | |
3442 | * someone else killed the ctx or is already going through | |
3443 | * io_uring_register(). | |
3444 | */ | |
3445 | if (percpu_ref_is_dying(&ctx->refs)) | |
3446 | return -ENXIO; | |
3447 | ||
edafccee | 3448 | percpu_ref_kill(&ctx->refs); |
b19062a5 JA |
3449 | |
3450 | /* | |
3451 | * Drop uring mutex before waiting for references to exit. If another | |
3452 | * thread is currently inside io_uring_enter() it might need to grab | |
3453 | * the uring_lock to make progress. If we hold it here across the drain | |
3454 | * wait, then we can deadlock. It's safe to drop the mutex here, since | |
3455 | * no new references will come in after we've killed the percpu ref. | |
3456 | */ | |
3457 | mutex_unlock(&ctx->uring_lock); | |
edafccee | 3458 | wait_for_completion(&ctx->ctx_done); |
b19062a5 | 3459 | mutex_lock(&ctx->uring_lock); |
edafccee JA |
3460 | |
3461 | switch (opcode) { | |
3462 | case IORING_REGISTER_BUFFERS: | |
3463 | ret = io_sqe_buffer_register(ctx, arg, nr_args); | |
3464 | break; | |
3465 | case IORING_UNREGISTER_BUFFERS: | |
3466 | ret = -EINVAL; | |
3467 | if (arg || nr_args) | |
3468 | break; | |
3469 | ret = io_sqe_buffer_unregister(ctx); | |
3470 | break; | |
6b06314c JA |
3471 | case IORING_REGISTER_FILES: |
3472 | ret = io_sqe_files_register(ctx, arg, nr_args); | |
3473 | break; | |
3474 | case IORING_UNREGISTER_FILES: | |
3475 | ret = -EINVAL; | |
3476 | if (arg || nr_args) | |
3477 | break; | |
3478 | ret = io_sqe_files_unregister(ctx); | |
3479 | break; | |
9b402849 JA |
3480 | case IORING_REGISTER_EVENTFD: |
3481 | ret = -EINVAL; | |
3482 | if (nr_args != 1) | |
3483 | break; | |
3484 | ret = io_eventfd_register(ctx, arg); | |
3485 | break; | |
3486 | case IORING_UNREGISTER_EVENTFD: | |
3487 | ret = -EINVAL; | |
3488 | if (arg || nr_args) | |
3489 | break; | |
3490 | ret = io_eventfd_unregister(ctx); | |
3491 | break; | |
edafccee JA |
3492 | default: |
3493 | ret = -EINVAL; | |
3494 | break; | |
3495 | } | |
3496 | ||
3497 | /* bring the ctx back to life */ | |
3498 | reinit_completion(&ctx->ctx_done); | |
3499 | percpu_ref_reinit(&ctx->refs); | |
3500 | return ret; | |
3501 | } | |
3502 | ||
3503 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
3504 | void __user *, arg, unsigned int, nr_args) | |
3505 | { | |
3506 | struct io_ring_ctx *ctx; | |
3507 | long ret = -EBADF; | |
3508 | struct fd f; | |
3509 | ||
3510 | f = fdget(fd); | |
3511 | if (!f.file) | |
3512 | return -EBADF; | |
3513 | ||
3514 | ret = -EOPNOTSUPP; | |
3515 | if (f.file->f_op != &io_uring_fops) | |
3516 | goto out_fput; | |
3517 | ||
3518 | ctx = f.file->private_data; | |
3519 | ||
3520 | mutex_lock(&ctx->uring_lock); | |
3521 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
3522 | mutex_unlock(&ctx->uring_lock); | |
3523 | out_fput: | |
3524 | fdput(f); | |
3525 | return ret; | |
3526 | } | |
3527 | ||
2b188cc1 JA |
3528 | static int __init io_uring_init(void) |
3529 | { | |
3530 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC); | |
3531 | return 0; | |
3532 | }; | |
3533 | __initcall(io_uring_init); |