]>
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; | |
205 | size_t io_pages; | |
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 JA |
233 | wait_queue_head_t sqo_wait; |
234 | unsigned sqo_stop; | |
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; | |
244 | } ____cacheline_aligned_in_smp; | |
245 | ||
6b06314c JA |
246 | /* |
247 | * If used, fixed file set. Writers must ensure that ->refs is dead, | |
248 | * readers must ensure that ->refs is alive as long as the file* is | |
249 | * used. Only updated through io_uring_register(2). | |
250 | */ | |
251 | struct file **user_files; | |
252 | unsigned nr_user_files; | |
253 | ||
edafccee JA |
254 | /* if used, fixed mapped user buffers */ |
255 | unsigned nr_user_bufs; | |
256 | struct io_mapped_ubuf *user_bufs; | |
257 | ||
2b188cc1 JA |
258 | struct user_struct *user; |
259 | ||
260 | struct completion ctx_done; | |
261 | ||
262 | struct { | |
263 | struct mutex uring_lock; | |
264 | wait_queue_head_t wait; | |
265 | } ____cacheline_aligned_in_smp; | |
266 | ||
267 | struct { | |
268 | spinlock_t completion_lock; | |
def596e9 JA |
269 | bool poll_multi_file; |
270 | /* | |
271 | * ->poll_list is protected by the ctx->uring_lock for | |
272 | * io_uring instances that don't use IORING_SETUP_SQPOLL. | |
273 | * For SQPOLL, only the single threaded io_sq_thread() will | |
274 | * manipulate the list, hence no extra locking is needed there. | |
275 | */ | |
276 | struct list_head poll_list; | |
221c5eb2 | 277 | struct list_head cancel_list; |
2b188cc1 JA |
278 | } ____cacheline_aligned_in_smp; |
279 | ||
31b51510 JA |
280 | struct async_list pending_async[2]; |
281 | ||
2b188cc1 JA |
282 | #if defined(CONFIG_UNIX) |
283 | struct socket *ring_sock; | |
284 | #endif | |
285 | }; | |
286 | ||
287 | struct sqe_submit { | |
288 | const struct io_uring_sqe *sqe; | |
289 | unsigned short index; | |
290 | bool has_user; | |
def596e9 | 291 | bool needs_lock; |
6c271ce2 | 292 | bool needs_fixed_file; |
2b188cc1 JA |
293 | }; |
294 | ||
09bb8394 JA |
295 | /* |
296 | * First field must be the file pointer in all the | |
297 | * iocb unions! See also 'struct kiocb' in <linux/fs.h> | |
298 | */ | |
221c5eb2 JA |
299 | struct io_poll_iocb { |
300 | struct file *file; | |
301 | struct wait_queue_head *head; | |
302 | __poll_t events; | |
8c838788 | 303 | bool done; |
221c5eb2 JA |
304 | bool canceled; |
305 | struct wait_queue_entry wait; | |
306 | }; | |
307 | ||
09bb8394 JA |
308 | /* |
309 | * NOTE! Each of the iocb union members has the file pointer | |
310 | * as the first entry in their struct definition. So you can | |
311 | * access the file pointer through any of the sub-structs, | |
312 | * or directly as just 'ki_filp' in this struct. | |
313 | */ | |
2b188cc1 | 314 | struct io_kiocb { |
221c5eb2 | 315 | union { |
09bb8394 | 316 | struct file *file; |
221c5eb2 JA |
317 | struct kiocb rw; |
318 | struct io_poll_iocb poll; | |
319 | }; | |
2b188cc1 JA |
320 | |
321 | struct sqe_submit submit; | |
322 | ||
323 | struct io_ring_ctx *ctx; | |
324 | struct list_head list; | |
325 | unsigned int flags; | |
c16361c1 | 326 | refcount_t refs; |
8449eeda | 327 | #define REQ_F_NOWAIT 1 /* must not punt to workers */ |
def596e9 | 328 | #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */ |
6b06314c | 329 | #define REQ_F_FIXED_FILE 4 /* ctx owns file */ |
31b51510 | 330 | #define REQ_F_SEQ_PREV 8 /* sequential with previous */ |
d530a402 | 331 | #define REQ_F_PREPPED 16 /* prep already done */ |
de0617e4 JA |
332 | #define REQ_F_IO_DRAIN 32 /* drain existing IO first */ |
333 | #define REQ_F_IO_DRAINED 64 /* drain done */ | |
2b188cc1 | 334 | u64 user_data; |
de0617e4 JA |
335 | u32 error; |
336 | u32 sequence; | |
2b188cc1 JA |
337 | |
338 | struct work_struct work; | |
339 | }; | |
340 | ||
341 | #define IO_PLUG_THRESHOLD 2 | |
def596e9 | 342 | #define IO_IOPOLL_BATCH 8 |
2b188cc1 | 343 | |
9a56a232 JA |
344 | struct io_submit_state { |
345 | struct blk_plug plug; | |
346 | ||
2579f913 JA |
347 | /* |
348 | * io_kiocb alloc cache | |
349 | */ | |
350 | void *reqs[IO_IOPOLL_BATCH]; | |
351 | unsigned int free_reqs; | |
352 | unsigned int cur_req; | |
353 | ||
9a56a232 JA |
354 | /* |
355 | * File reference cache | |
356 | */ | |
357 | struct file *file; | |
358 | unsigned int fd; | |
359 | unsigned int has_refs; | |
360 | unsigned int used_refs; | |
361 | unsigned int ios_left; | |
362 | }; | |
363 | ||
de0617e4 JA |
364 | static void io_sq_wq_submit_work(struct work_struct *work); |
365 | ||
2b188cc1 JA |
366 | static struct kmem_cache *req_cachep; |
367 | ||
368 | static const struct file_operations io_uring_fops; | |
369 | ||
370 | struct sock *io_uring_get_socket(struct file *file) | |
371 | { | |
372 | #if defined(CONFIG_UNIX) | |
373 | if (file->f_op == &io_uring_fops) { | |
374 | struct io_ring_ctx *ctx = file->private_data; | |
375 | ||
376 | return ctx->ring_sock->sk; | |
377 | } | |
378 | #endif | |
379 | return NULL; | |
380 | } | |
381 | EXPORT_SYMBOL(io_uring_get_socket); | |
382 | ||
383 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) | |
384 | { | |
385 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
386 | ||
387 | complete(&ctx->ctx_done); | |
388 | } | |
389 | ||
390 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) | |
391 | { | |
392 | struct io_ring_ctx *ctx; | |
31b51510 | 393 | int i; |
2b188cc1 JA |
394 | |
395 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
396 | if (!ctx) | |
397 | return NULL; | |
398 | ||
399 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, 0, GFP_KERNEL)) { | |
400 | kfree(ctx); | |
401 | return NULL; | |
402 | } | |
403 | ||
404 | ctx->flags = p->flags; | |
405 | init_waitqueue_head(&ctx->cq_wait); | |
406 | init_completion(&ctx->ctx_done); | |
407 | mutex_init(&ctx->uring_lock); | |
408 | init_waitqueue_head(&ctx->wait); | |
31b51510 JA |
409 | for (i = 0; i < ARRAY_SIZE(ctx->pending_async); i++) { |
410 | spin_lock_init(&ctx->pending_async[i].lock); | |
411 | INIT_LIST_HEAD(&ctx->pending_async[i].list); | |
412 | atomic_set(&ctx->pending_async[i].cnt, 0); | |
413 | } | |
2b188cc1 | 414 | spin_lock_init(&ctx->completion_lock); |
def596e9 | 415 | INIT_LIST_HEAD(&ctx->poll_list); |
221c5eb2 | 416 | INIT_LIST_HEAD(&ctx->cancel_list); |
de0617e4 | 417 | INIT_LIST_HEAD(&ctx->defer_list); |
2b188cc1 JA |
418 | return ctx; |
419 | } | |
420 | ||
de0617e4 JA |
421 | static inline bool io_sequence_defer(struct io_ring_ctx *ctx, |
422 | struct io_kiocb *req) | |
423 | { | |
424 | if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) != REQ_F_IO_DRAIN) | |
425 | return false; | |
426 | ||
427 | return req->sequence > ctx->cached_cq_tail + ctx->sq_ring->dropped; | |
428 | } | |
429 | ||
430 | static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx) | |
431 | { | |
432 | struct io_kiocb *req; | |
433 | ||
434 | if (list_empty(&ctx->defer_list)) | |
435 | return NULL; | |
436 | ||
437 | req = list_first_entry(&ctx->defer_list, struct io_kiocb, list); | |
438 | if (!io_sequence_defer(ctx, req)) { | |
439 | list_del_init(&req->list); | |
440 | return req; | |
441 | } | |
442 | ||
443 | return NULL; | |
444 | } | |
445 | ||
446 | static void __io_commit_cqring(struct io_ring_ctx *ctx) | |
2b188cc1 JA |
447 | { |
448 | struct io_cq_ring *ring = ctx->cq_ring; | |
449 | ||
450 | if (ctx->cached_cq_tail != READ_ONCE(ring->r.tail)) { | |
451 | /* order cqe stores with ring update */ | |
452 | smp_store_release(&ring->r.tail, ctx->cached_cq_tail); | |
453 | ||
2b188cc1 JA |
454 | if (wq_has_sleeper(&ctx->cq_wait)) { |
455 | wake_up_interruptible(&ctx->cq_wait); | |
456 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
457 | } | |
458 | } | |
459 | } | |
460 | ||
de0617e4 JA |
461 | static void io_commit_cqring(struct io_ring_ctx *ctx) |
462 | { | |
463 | struct io_kiocb *req; | |
464 | ||
465 | __io_commit_cqring(ctx); | |
466 | ||
467 | while ((req = io_get_deferred_req(ctx)) != NULL) { | |
468 | req->flags |= REQ_F_IO_DRAINED; | |
469 | queue_work(ctx->sqo_wq, &req->work); | |
470 | } | |
471 | } | |
472 | ||
2b188cc1 JA |
473 | static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx) |
474 | { | |
475 | struct io_cq_ring *ring = ctx->cq_ring; | |
476 | unsigned tail; | |
477 | ||
478 | tail = ctx->cached_cq_tail; | |
115e12e5 SB |
479 | /* |
480 | * writes to the cq entry need to come after reading head; the | |
481 | * control dependency is enough as we're using WRITE_ONCE to | |
482 | * fill the cq entry | |
483 | */ | |
74f464e9 | 484 | if (tail - READ_ONCE(ring->r.head) == ring->ring_entries) |
2b188cc1 JA |
485 | return NULL; |
486 | ||
487 | ctx->cached_cq_tail++; | |
488 | return &ring->cqes[tail & ctx->cq_mask]; | |
489 | } | |
490 | ||
491 | static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data, | |
492 | long res, unsigned ev_flags) | |
493 | { | |
494 | struct io_uring_cqe *cqe; | |
495 | ||
496 | /* | |
497 | * If we can't get a cq entry, userspace overflowed the | |
498 | * submission (by quite a lot). Increment the overflow count in | |
499 | * the ring. | |
500 | */ | |
501 | cqe = io_get_cqring(ctx); | |
502 | if (cqe) { | |
503 | WRITE_ONCE(cqe->user_data, ki_user_data); | |
504 | WRITE_ONCE(cqe->res, res); | |
505 | WRITE_ONCE(cqe->flags, ev_flags); | |
506 | } else { | |
507 | unsigned overflow = READ_ONCE(ctx->cq_ring->overflow); | |
508 | ||
509 | WRITE_ONCE(ctx->cq_ring->overflow, overflow + 1); | |
510 | } | |
511 | } | |
512 | ||
8c838788 JA |
513 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
514 | { | |
515 | if (waitqueue_active(&ctx->wait)) | |
516 | wake_up(&ctx->wait); | |
517 | if (waitqueue_active(&ctx->sqo_wait)) | |
518 | wake_up(&ctx->sqo_wait); | |
519 | } | |
520 | ||
521 | static void io_cqring_add_event(struct io_ring_ctx *ctx, u64 user_data, | |
2b188cc1 JA |
522 | long res, unsigned ev_flags) |
523 | { | |
524 | unsigned long flags; | |
525 | ||
526 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
8c838788 | 527 | io_cqring_fill_event(ctx, user_data, res, ev_flags); |
2b188cc1 JA |
528 | io_commit_cqring(ctx); |
529 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
530 | ||
8c838788 | 531 | io_cqring_ev_posted(ctx); |
2b188cc1 JA |
532 | } |
533 | ||
534 | static void io_ring_drop_ctx_refs(struct io_ring_ctx *ctx, unsigned refs) | |
535 | { | |
536 | percpu_ref_put_many(&ctx->refs, refs); | |
537 | ||
538 | if (waitqueue_active(&ctx->wait)) | |
539 | wake_up(&ctx->wait); | |
540 | } | |
541 | ||
2579f913 JA |
542 | static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx, |
543 | struct io_submit_state *state) | |
2b188cc1 | 544 | { |
fd6fab2c | 545 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
2b188cc1 JA |
546 | struct io_kiocb *req; |
547 | ||
548 | if (!percpu_ref_tryget(&ctx->refs)) | |
549 | return NULL; | |
550 | ||
2579f913 | 551 | if (!state) { |
fd6fab2c | 552 | req = kmem_cache_alloc(req_cachep, gfp); |
2579f913 JA |
553 | if (unlikely(!req)) |
554 | goto out; | |
555 | } else if (!state->free_reqs) { | |
556 | size_t sz; | |
557 | int ret; | |
558 | ||
559 | sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs)); | |
fd6fab2c JA |
560 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs); |
561 | ||
562 | /* | |
563 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
564 | * retry single alloc to be on the safe side. | |
565 | */ | |
566 | if (unlikely(ret <= 0)) { | |
567 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
568 | if (!state->reqs[0]) | |
569 | goto out; | |
570 | ret = 1; | |
571 | } | |
2579f913 JA |
572 | state->free_reqs = ret - 1; |
573 | state->cur_req = 1; | |
574 | req = state->reqs[0]; | |
575 | } else { | |
576 | req = state->reqs[state->cur_req]; | |
577 | state->free_reqs--; | |
578 | state->cur_req++; | |
2b188cc1 JA |
579 | } |
580 | ||
2579f913 JA |
581 | req->ctx = ctx; |
582 | req->flags = 0; | |
e65ef56d JA |
583 | /* one is dropped after submission, the other at completion */ |
584 | refcount_set(&req->refs, 2); | |
2579f913 JA |
585 | return req; |
586 | out: | |
2b188cc1 JA |
587 | io_ring_drop_ctx_refs(ctx, 1); |
588 | return NULL; | |
589 | } | |
590 | ||
def596e9 JA |
591 | static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr) |
592 | { | |
593 | if (*nr) { | |
594 | kmem_cache_free_bulk(req_cachep, *nr, reqs); | |
595 | io_ring_drop_ctx_refs(ctx, *nr); | |
596 | *nr = 0; | |
597 | } | |
598 | } | |
599 | ||
2b188cc1 JA |
600 | static void io_free_req(struct io_kiocb *req) |
601 | { | |
09bb8394 JA |
602 | if (req->file && !(req->flags & REQ_F_FIXED_FILE)) |
603 | fput(req->file); | |
e65ef56d JA |
604 | io_ring_drop_ctx_refs(req->ctx, 1); |
605 | kmem_cache_free(req_cachep, req); | |
606 | } | |
607 | ||
608 | static void io_put_req(struct io_kiocb *req) | |
609 | { | |
610 | if (refcount_dec_and_test(&req->refs)) | |
611 | io_free_req(req); | |
2b188cc1 JA |
612 | } |
613 | ||
def596e9 JA |
614 | /* |
615 | * Find and free completed poll iocbs | |
616 | */ | |
617 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
618 | struct list_head *done) | |
619 | { | |
620 | void *reqs[IO_IOPOLL_BATCH]; | |
621 | struct io_kiocb *req; | |
09bb8394 | 622 | int to_free; |
def596e9 | 623 | |
09bb8394 | 624 | to_free = 0; |
def596e9 JA |
625 | while (!list_empty(done)) { |
626 | req = list_first_entry(done, struct io_kiocb, list); | |
627 | list_del(&req->list); | |
628 | ||
629 | io_cqring_fill_event(ctx, req->user_data, req->error, 0); | |
def596e9 JA |
630 | (*nr_events)++; |
631 | ||
09bb8394 JA |
632 | if (refcount_dec_and_test(&req->refs)) { |
633 | /* If we're not using fixed files, we have to pair the | |
634 | * completion part with the file put. Use regular | |
635 | * completions for those, only batch free for fixed | |
636 | * file. | |
637 | */ | |
638 | if (req->flags & REQ_F_FIXED_FILE) { | |
639 | reqs[to_free++] = req; | |
640 | if (to_free == ARRAY_SIZE(reqs)) | |
641 | io_free_req_many(ctx, reqs, &to_free); | |
6b06314c | 642 | } else { |
09bb8394 | 643 | io_free_req(req); |
6b06314c | 644 | } |
9a56a232 | 645 | } |
def596e9 | 646 | } |
def596e9 | 647 | |
09bb8394 | 648 | io_commit_cqring(ctx); |
def596e9 JA |
649 | io_free_req_many(ctx, reqs, &to_free); |
650 | } | |
651 | ||
652 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
653 | long min) | |
654 | { | |
655 | struct io_kiocb *req, *tmp; | |
656 | LIST_HEAD(done); | |
657 | bool spin; | |
658 | int ret; | |
659 | ||
660 | /* | |
661 | * Only spin for completions if we don't have multiple devices hanging | |
662 | * off our complete list, and we're under the requested amount. | |
663 | */ | |
664 | spin = !ctx->poll_multi_file && *nr_events < min; | |
665 | ||
666 | ret = 0; | |
667 | list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) { | |
668 | struct kiocb *kiocb = &req->rw; | |
669 | ||
670 | /* | |
671 | * Move completed entries to our local list. If we find a | |
672 | * request that requires polling, break out and complete | |
673 | * the done list first, if we have entries there. | |
674 | */ | |
675 | if (req->flags & REQ_F_IOPOLL_COMPLETED) { | |
676 | list_move_tail(&req->list, &done); | |
677 | continue; | |
678 | } | |
679 | if (!list_empty(&done)) | |
680 | break; | |
681 | ||
682 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
683 | if (ret < 0) | |
684 | break; | |
685 | ||
686 | if (ret && spin) | |
687 | spin = false; | |
688 | ret = 0; | |
689 | } | |
690 | ||
691 | if (!list_empty(&done)) | |
692 | io_iopoll_complete(ctx, nr_events, &done); | |
693 | ||
694 | return ret; | |
695 | } | |
696 | ||
697 | /* | |
698 | * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a | |
699 | * non-spinning poll check - we'll still enter the driver poll loop, but only | |
700 | * as a non-spinning completion check. | |
701 | */ | |
702 | static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
703 | long min) | |
704 | { | |
705 | while (!list_empty(&ctx->poll_list)) { | |
706 | int ret; | |
707 | ||
708 | ret = io_do_iopoll(ctx, nr_events, min); | |
709 | if (ret < 0) | |
710 | return ret; | |
711 | if (!min || *nr_events >= min) | |
712 | return 0; | |
713 | } | |
714 | ||
715 | return 1; | |
716 | } | |
717 | ||
718 | /* | |
719 | * We can't just wait for polled events to come to us, we have to actively | |
720 | * find and complete them. | |
721 | */ | |
722 | static void io_iopoll_reap_events(struct io_ring_ctx *ctx) | |
723 | { | |
724 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
725 | return; | |
726 | ||
727 | mutex_lock(&ctx->uring_lock); | |
728 | while (!list_empty(&ctx->poll_list)) { | |
729 | unsigned int nr_events = 0; | |
730 | ||
731 | io_iopoll_getevents(ctx, &nr_events, 1); | |
732 | } | |
733 | mutex_unlock(&ctx->uring_lock); | |
734 | } | |
735 | ||
736 | static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, | |
737 | long min) | |
738 | { | |
739 | int ret = 0; | |
740 | ||
741 | do { | |
742 | int tmin = 0; | |
743 | ||
744 | if (*nr_events < min) | |
745 | tmin = min - *nr_events; | |
746 | ||
747 | ret = io_iopoll_getevents(ctx, nr_events, tmin); | |
748 | if (ret <= 0) | |
749 | break; | |
750 | ret = 0; | |
751 | } while (min && !*nr_events && !need_resched()); | |
752 | ||
753 | return ret; | |
754 | } | |
755 | ||
2b188cc1 JA |
756 | static void kiocb_end_write(struct kiocb *kiocb) |
757 | { | |
758 | if (kiocb->ki_flags & IOCB_WRITE) { | |
759 | struct inode *inode = file_inode(kiocb->ki_filp); | |
760 | ||
761 | /* | |
762 | * Tell lockdep we inherited freeze protection from submission | |
763 | * thread. | |
764 | */ | |
765 | if (S_ISREG(inode->i_mode)) | |
766 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); | |
767 | file_end_write(kiocb->ki_filp); | |
768 | } | |
769 | } | |
770 | ||
771 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
772 | { | |
773 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
774 | ||
775 | kiocb_end_write(kiocb); | |
776 | ||
2b188cc1 | 777 | io_cqring_add_event(req->ctx, req->user_data, res, 0); |
e65ef56d | 778 | io_put_req(req); |
2b188cc1 JA |
779 | } |
780 | ||
def596e9 JA |
781 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
782 | { | |
783 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
784 | ||
785 | kiocb_end_write(kiocb); | |
786 | ||
787 | req->error = res; | |
788 | if (res != -EAGAIN) | |
789 | req->flags |= REQ_F_IOPOLL_COMPLETED; | |
790 | } | |
791 | ||
792 | /* | |
793 | * After the iocb has been issued, it's safe to be found on the poll list. | |
794 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
795 | * find it from a io_iopoll_getevents() thread before the issuer is done | |
796 | * accessing the kiocb cookie. | |
797 | */ | |
798 | static void io_iopoll_req_issued(struct io_kiocb *req) | |
799 | { | |
800 | struct io_ring_ctx *ctx = req->ctx; | |
801 | ||
802 | /* | |
803 | * Track whether we have multiple files in our lists. This will impact | |
804 | * how we do polling eventually, not spinning if we're on potentially | |
805 | * different devices. | |
806 | */ | |
807 | if (list_empty(&ctx->poll_list)) { | |
808 | ctx->poll_multi_file = false; | |
809 | } else if (!ctx->poll_multi_file) { | |
810 | struct io_kiocb *list_req; | |
811 | ||
812 | list_req = list_first_entry(&ctx->poll_list, struct io_kiocb, | |
813 | list); | |
814 | if (list_req->rw.ki_filp != req->rw.ki_filp) | |
815 | ctx->poll_multi_file = true; | |
816 | } | |
817 | ||
818 | /* | |
819 | * For fast devices, IO may have already completed. If it has, add | |
820 | * it to the front so we find it first. | |
821 | */ | |
822 | if (req->flags & REQ_F_IOPOLL_COMPLETED) | |
823 | list_add(&req->list, &ctx->poll_list); | |
824 | else | |
825 | list_add_tail(&req->list, &ctx->poll_list); | |
826 | } | |
827 | ||
3d6770fb | 828 | static void io_file_put(struct io_submit_state *state) |
9a56a232 | 829 | { |
3d6770fb | 830 | if (state->file) { |
9a56a232 JA |
831 | int diff = state->has_refs - state->used_refs; |
832 | ||
833 | if (diff) | |
834 | fput_many(state->file, diff); | |
835 | state->file = NULL; | |
836 | } | |
837 | } | |
838 | ||
839 | /* | |
840 | * Get as many references to a file as we have IOs left in this submission, | |
841 | * assuming most submissions are for one file, or at least that each file | |
842 | * has more than one submission. | |
843 | */ | |
844 | static struct file *io_file_get(struct io_submit_state *state, int fd) | |
845 | { | |
846 | if (!state) | |
847 | return fget(fd); | |
848 | ||
849 | if (state->file) { | |
850 | if (state->fd == fd) { | |
851 | state->used_refs++; | |
852 | state->ios_left--; | |
853 | return state->file; | |
854 | } | |
3d6770fb | 855 | io_file_put(state); |
9a56a232 JA |
856 | } |
857 | state->file = fget_many(fd, state->ios_left); | |
858 | if (!state->file) | |
859 | return NULL; | |
860 | ||
861 | state->fd = fd; | |
862 | state->has_refs = state->ios_left; | |
863 | state->used_refs = 1; | |
864 | state->ios_left--; | |
865 | return state->file; | |
866 | } | |
867 | ||
2b188cc1 JA |
868 | /* |
869 | * If we tracked the file through the SCM inflight mechanism, we could support | |
870 | * any file. For now, just ensure that anything potentially problematic is done | |
871 | * inline. | |
872 | */ | |
873 | static bool io_file_supports_async(struct file *file) | |
874 | { | |
875 | umode_t mode = file_inode(file)->i_mode; | |
876 | ||
877 | if (S_ISBLK(mode) || S_ISCHR(mode)) | |
878 | return true; | |
879 | if (S_ISREG(mode) && file->f_op != &io_uring_fops) | |
880 | return true; | |
881 | ||
882 | return false; | |
883 | } | |
884 | ||
6c271ce2 | 885 | static int io_prep_rw(struct io_kiocb *req, const struct sqe_submit *s, |
8358e3a8 | 886 | bool force_nonblock) |
2b188cc1 | 887 | { |
6c271ce2 | 888 | const struct io_uring_sqe *sqe = s->sqe; |
def596e9 | 889 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 | 890 | struct kiocb *kiocb = &req->rw; |
09bb8394 JA |
891 | unsigned ioprio; |
892 | int ret; | |
2b188cc1 | 893 | |
09bb8394 JA |
894 | if (!req->file) |
895 | return -EBADF; | |
2b188cc1 | 896 | /* For -EAGAIN retry, everything is already prepped */ |
d530a402 | 897 | if (req->flags & REQ_F_PREPPED) |
2b188cc1 JA |
898 | return 0; |
899 | ||
09bb8394 JA |
900 | if (force_nonblock && !io_file_supports_async(req->file)) |
901 | force_nonblock = false; | |
6b06314c | 902 | |
2b188cc1 JA |
903 | kiocb->ki_pos = READ_ONCE(sqe->off); |
904 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); | |
905 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); | |
906 | ||
907 | ioprio = READ_ONCE(sqe->ioprio); | |
908 | if (ioprio) { | |
909 | ret = ioprio_check_cap(ioprio); | |
910 | if (ret) | |
09bb8394 | 911 | return ret; |
2b188cc1 JA |
912 | |
913 | kiocb->ki_ioprio = ioprio; | |
914 | } else | |
915 | kiocb->ki_ioprio = get_current_ioprio(); | |
916 | ||
917 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
918 | if (unlikely(ret)) | |
09bb8394 | 919 | return ret; |
8449eeda SB |
920 | |
921 | /* don't allow async punt if RWF_NOWAIT was requested */ | |
922 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
923 | req->flags |= REQ_F_NOWAIT; | |
924 | ||
925 | if (force_nonblock) | |
2b188cc1 | 926 | kiocb->ki_flags |= IOCB_NOWAIT; |
8449eeda | 927 | |
def596e9 | 928 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
929 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
930 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 931 | return -EOPNOTSUPP; |
2b188cc1 | 932 | |
def596e9 JA |
933 | req->error = 0; |
934 | kiocb->ki_flags |= IOCB_HIPRI; | |
935 | kiocb->ki_complete = io_complete_rw_iopoll; | |
936 | } else { | |
09bb8394 JA |
937 | if (kiocb->ki_flags & IOCB_HIPRI) |
938 | return -EINVAL; | |
def596e9 JA |
939 | kiocb->ki_complete = io_complete_rw; |
940 | } | |
d530a402 | 941 | req->flags |= REQ_F_PREPPED; |
2b188cc1 | 942 | return 0; |
2b188cc1 JA |
943 | } |
944 | ||
945 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
946 | { | |
947 | switch (ret) { | |
948 | case -EIOCBQUEUED: | |
949 | break; | |
950 | case -ERESTARTSYS: | |
951 | case -ERESTARTNOINTR: | |
952 | case -ERESTARTNOHAND: | |
953 | case -ERESTART_RESTARTBLOCK: | |
954 | /* | |
955 | * We can't just restart the syscall, since previously | |
956 | * submitted sqes may already be in progress. Just fail this | |
957 | * IO with EINTR. | |
958 | */ | |
959 | ret = -EINTR; | |
960 | /* fall through */ | |
961 | default: | |
962 | kiocb->ki_complete(kiocb, ret, 0); | |
963 | } | |
964 | } | |
965 | ||
edafccee JA |
966 | static int io_import_fixed(struct io_ring_ctx *ctx, int rw, |
967 | const struct io_uring_sqe *sqe, | |
968 | struct iov_iter *iter) | |
969 | { | |
970 | size_t len = READ_ONCE(sqe->len); | |
971 | struct io_mapped_ubuf *imu; | |
972 | unsigned index, buf_index; | |
973 | size_t offset; | |
974 | u64 buf_addr; | |
975 | ||
976 | /* attempt to use fixed buffers without having provided iovecs */ | |
977 | if (unlikely(!ctx->user_bufs)) | |
978 | return -EFAULT; | |
979 | ||
980 | buf_index = READ_ONCE(sqe->buf_index); | |
981 | if (unlikely(buf_index >= ctx->nr_user_bufs)) | |
982 | return -EFAULT; | |
983 | ||
984 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); | |
985 | imu = &ctx->user_bufs[index]; | |
986 | buf_addr = READ_ONCE(sqe->addr); | |
987 | ||
988 | /* overflow */ | |
989 | if (buf_addr + len < buf_addr) | |
990 | return -EFAULT; | |
991 | /* not inside the mapped region */ | |
992 | if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len) | |
993 | return -EFAULT; | |
994 | ||
995 | /* | |
996 | * May not be a start of buffer, set size appropriately | |
997 | * and advance us to the beginning. | |
998 | */ | |
999 | offset = buf_addr - imu->ubuf; | |
1000 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
1001 | if (offset) | |
1002 | iov_iter_advance(iter, offset); | |
875f1d07 JA |
1003 | |
1004 | /* don't drop a reference to these pages */ | |
1005 | iter->type |= ITER_BVEC_FLAG_NO_REF; | |
edafccee JA |
1006 | return 0; |
1007 | } | |
1008 | ||
2b188cc1 JA |
1009 | static int io_import_iovec(struct io_ring_ctx *ctx, int rw, |
1010 | const struct sqe_submit *s, struct iovec **iovec, | |
1011 | struct iov_iter *iter) | |
1012 | { | |
1013 | const struct io_uring_sqe *sqe = s->sqe; | |
1014 | void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
1015 | size_t sqe_len = READ_ONCE(sqe->len); | |
edafccee JA |
1016 | u8 opcode; |
1017 | ||
1018 | /* | |
1019 | * We're reading ->opcode for the second time, but the first read | |
1020 | * doesn't care whether it's _FIXED or not, so it doesn't matter | |
1021 | * whether ->opcode changes concurrently. The first read does care | |
1022 | * about whether it is a READ or a WRITE, so we don't trust this read | |
1023 | * for that purpose and instead let the caller pass in the read/write | |
1024 | * flag. | |
1025 | */ | |
1026 | opcode = READ_ONCE(sqe->opcode); | |
1027 | if (opcode == IORING_OP_READ_FIXED || | |
1028 | opcode == IORING_OP_WRITE_FIXED) { | |
e0c5c576 | 1029 | int ret = io_import_fixed(ctx, rw, sqe, iter); |
edafccee JA |
1030 | *iovec = NULL; |
1031 | return ret; | |
1032 | } | |
2b188cc1 JA |
1033 | |
1034 | if (!s->has_user) | |
1035 | return -EFAULT; | |
1036 | ||
1037 | #ifdef CONFIG_COMPAT | |
1038 | if (ctx->compat) | |
1039 | return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV, | |
1040 | iovec, iter); | |
1041 | #endif | |
1042 | ||
1043 | return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter); | |
1044 | } | |
1045 | ||
31b51510 JA |
1046 | /* |
1047 | * Make a note of the last file/offset/direction we punted to async | |
1048 | * context. We'll use this information to see if we can piggy back a | |
1049 | * sequential request onto the previous one, if it's still hasn't been | |
1050 | * completed by the async worker. | |
1051 | */ | |
1052 | static void io_async_list_note(int rw, struct io_kiocb *req, size_t len) | |
1053 | { | |
1054 | struct async_list *async_list = &req->ctx->pending_async[rw]; | |
1055 | struct kiocb *kiocb = &req->rw; | |
1056 | struct file *filp = kiocb->ki_filp; | |
1057 | off_t io_end = kiocb->ki_pos + len; | |
1058 | ||
1059 | if (filp == async_list->file && kiocb->ki_pos == async_list->io_end) { | |
1060 | unsigned long max_pages; | |
1061 | ||
1062 | /* Use 8x RA size as a decent limiter for both reads/writes */ | |
1063 | max_pages = filp->f_ra.ra_pages; | |
1064 | if (!max_pages) | |
b5420237 | 1065 | max_pages = VM_READAHEAD_PAGES; |
31b51510 JA |
1066 | max_pages *= 8; |
1067 | ||
1068 | /* If max pages are exceeded, reset the state */ | |
1069 | len >>= PAGE_SHIFT; | |
1070 | if (async_list->io_pages + len <= max_pages) { | |
1071 | req->flags |= REQ_F_SEQ_PREV; | |
1072 | async_list->io_pages += len; | |
1073 | } else { | |
1074 | io_end = 0; | |
1075 | async_list->io_pages = 0; | |
1076 | } | |
1077 | } | |
1078 | ||
1079 | /* New file? Reset state. */ | |
1080 | if (async_list->file != filp) { | |
1081 | async_list->io_pages = 0; | |
1082 | async_list->file = filp; | |
1083 | } | |
1084 | async_list->io_end = io_end; | |
1085 | } | |
1086 | ||
e0c5c576 | 1087 | static int io_read(struct io_kiocb *req, const struct sqe_submit *s, |
8358e3a8 | 1088 | bool force_nonblock) |
2b188cc1 JA |
1089 | { |
1090 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1091 | struct kiocb *kiocb = &req->rw; | |
1092 | struct iov_iter iter; | |
1093 | struct file *file; | |
31b51510 | 1094 | size_t iov_count; |
e0c5c576 | 1095 | int ret; |
2b188cc1 | 1096 | |
8358e3a8 | 1097 | ret = io_prep_rw(req, s, force_nonblock); |
2b188cc1 JA |
1098 | if (ret) |
1099 | return ret; | |
1100 | file = kiocb->ki_filp; | |
1101 | ||
2b188cc1 | 1102 | if (unlikely(!(file->f_mode & FMODE_READ))) |
09bb8394 | 1103 | return -EBADF; |
2b188cc1 | 1104 | if (unlikely(!file->f_op->read_iter)) |
09bb8394 | 1105 | return -EINVAL; |
2b188cc1 JA |
1106 | |
1107 | ret = io_import_iovec(req->ctx, READ, s, &iovec, &iter); | |
1108 | if (ret) | |
09bb8394 | 1109 | return ret; |
2b188cc1 | 1110 | |
31b51510 JA |
1111 | iov_count = iov_iter_count(&iter); |
1112 | ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 JA |
1113 | if (!ret) { |
1114 | ssize_t ret2; | |
1115 | ||
1116 | /* Catch -EAGAIN return for forced non-blocking submission */ | |
1117 | ret2 = call_read_iter(file, kiocb, &iter); | |
31b51510 | 1118 | if (!force_nonblock || ret2 != -EAGAIN) { |
2b188cc1 | 1119 | io_rw_done(kiocb, ret2); |
31b51510 JA |
1120 | } else { |
1121 | /* | |
1122 | * If ->needs_lock is true, we're already in async | |
1123 | * context. | |
1124 | */ | |
1125 | if (!s->needs_lock) | |
1126 | io_async_list_note(READ, req, iov_count); | |
2b188cc1 | 1127 | ret = -EAGAIN; |
31b51510 | 1128 | } |
2b188cc1 JA |
1129 | } |
1130 | kfree(iovec); | |
2b188cc1 JA |
1131 | return ret; |
1132 | } | |
1133 | ||
e0c5c576 | 1134 | static int io_write(struct io_kiocb *req, const struct sqe_submit *s, |
8358e3a8 | 1135 | bool force_nonblock) |
2b188cc1 JA |
1136 | { |
1137 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1138 | struct kiocb *kiocb = &req->rw; | |
1139 | struct iov_iter iter; | |
1140 | struct file *file; | |
31b51510 | 1141 | size_t iov_count; |
e0c5c576 | 1142 | int ret; |
2b188cc1 | 1143 | |
8358e3a8 | 1144 | ret = io_prep_rw(req, s, force_nonblock); |
2b188cc1 JA |
1145 | if (ret) |
1146 | return ret; | |
2b188cc1 | 1147 | |
2b188cc1 JA |
1148 | file = kiocb->ki_filp; |
1149 | if (unlikely(!(file->f_mode & FMODE_WRITE))) | |
09bb8394 | 1150 | return -EBADF; |
2b188cc1 | 1151 | if (unlikely(!file->f_op->write_iter)) |
09bb8394 | 1152 | return -EINVAL; |
2b188cc1 JA |
1153 | |
1154 | ret = io_import_iovec(req->ctx, WRITE, s, &iovec, &iter); | |
1155 | if (ret) | |
09bb8394 | 1156 | return ret; |
2b188cc1 | 1157 | |
31b51510 JA |
1158 | iov_count = iov_iter_count(&iter); |
1159 | ||
1160 | ret = -EAGAIN; | |
1161 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT)) { | |
1162 | /* If ->needs_lock is true, we're already in async context. */ | |
1163 | if (!s->needs_lock) | |
1164 | io_async_list_note(WRITE, req, iov_count); | |
1165 | goto out_free; | |
1166 | } | |
1167 | ||
1168 | ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 | 1169 | if (!ret) { |
9bf7933f RP |
1170 | ssize_t ret2; |
1171 | ||
2b188cc1 JA |
1172 | /* |
1173 | * Open-code file_start_write here to grab freeze protection, | |
1174 | * which will be released by another thread in | |
1175 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
1176 | * released so that it doesn't complain about the held lock when | |
1177 | * we return to userspace. | |
1178 | */ | |
1179 | if (S_ISREG(file_inode(file)->i_mode)) { | |
1180 | __sb_start_write(file_inode(file)->i_sb, | |
1181 | SB_FREEZE_WRITE, true); | |
1182 | __sb_writers_release(file_inode(file)->i_sb, | |
1183 | SB_FREEZE_WRITE); | |
1184 | } | |
1185 | kiocb->ki_flags |= IOCB_WRITE; | |
9bf7933f RP |
1186 | |
1187 | ret2 = call_write_iter(file, kiocb, &iter); | |
1188 | if (!force_nonblock || ret2 != -EAGAIN) { | |
1189 | io_rw_done(kiocb, ret2); | |
1190 | } else { | |
1191 | /* | |
1192 | * If ->needs_lock is true, we're already in async | |
1193 | * context. | |
1194 | */ | |
1195 | if (!s->needs_lock) | |
1196 | io_async_list_note(WRITE, req, iov_count); | |
1197 | ret = -EAGAIN; | |
1198 | } | |
2b188cc1 | 1199 | } |
31b51510 | 1200 | out_free: |
2b188cc1 | 1201 | kfree(iovec); |
2b188cc1 JA |
1202 | return ret; |
1203 | } | |
1204 | ||
1205 | /* | |
1206 | * IORING_OP_NOP just posts a completion event, nothing else. | |
1207 | */ | |
1208 | static int io_nop(struct io_kiocb *req, u64 user_data) | |
1209 | { | |
1210 | struct io_ring_ctx *ctx = req->ctx; | |
1211 | long err = 0; | |
1212 | ||
def596e9 JA |
1213 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
1214 | return -EINVAL; | |
1215 | ||
2b188cc1 | 1216 | io_cqring_add_event(ctx, user_data, err, 0); |
e65ef56d | 1217 | io_put_req(req); |
2b188cc1 JA |
1218 | return 0; |
1219 | } | |
1220 | ||
c992fe29 CH |
1221 | static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1222 | { | |
6b06314c | 1223 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 1224 | |
09bb8394 JA |
1225 | if (!req->file) |
1226 | return -EBADF; | |
d530a402 JA |
1227 | /* Prep already done (EAGAIN retry) */ |
1228 | if (req->flags & REQ_F_PREPPED) | |
c992fe29 CH |
1229 | return 0; |
1230 | ||
6b06314c | 1231 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 1232 | return -EINVAL; |
edafccee | 1233 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
1234 | return -EINVAL; |
1235 | ||
d530a402 | 1236 | req->flags |= REQ_F_PREPPED; |
c992fe29 CH |
1237 | return 0; |
1238 | } | |
1239 | ||
1240 | static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
1241 | bool force_nonblock) | |
1242 | { | |
1243 | loff_t sqe_off = READ_ONCE(sqe->off); | |
1244 | loff_t sqe_len = READ_ONCE(sqe->len); | |
1245 | loff_t end = sqe_off + sqe_len; | |
1246 | unsigned fsync_flags; | |
1247 | int ret; | |
1248 | ||
1249 | fsync_flags = READ_ONCE(sqe->fsync_flags); | |
1250 | if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC)) | |
1251 | return -EINVAL; | |
1252 | ||
1253 | ret = io_prep_fsync(req, sqe); | |
1254 | if (ret) | |
1255 | return ret; | |
1256 | ||
1257 | /* fsync always requires a blocking context */ | |
1258 | if (force_nonblock) | |
1259 | return -EAGAIN; | |
1260 | ||
1261 | ret = vfs_fsync_range(req->rw.ki_filp, sqe_off, | |
1262 | end > 0 ? end : LLONG_MAX, | |
1263 | fsync_flags & IORING_FSYNC_DATASYNC); | |
1264 | ||
c992fe29 | 1265 | io_cqring_add_event(req->ctx, sqe->user_data, ret, 0); |
e65ef56d | 1266 | io_put_req(req); |
c992fe29 CH |
1267 | return 0; |
1268 | } | |
1269 | ||
221c5eb2 JA |
1270 | static void io_poll_remove_one(struct io_kiocb *req) |
1271 | { | |
1272 | struct io_poll_iocb *poll = &req->poll; | |
1273 | ||
1274 | spin_lock(&poll->head->lock); | |
1275 | WRITE_ONCE(poll->canceled, true); | |
1276 | if (!list_empty(&poll->wait.entry)) { | |
1277 | list_del_init(&poll->wait.entry); | |
1278 | queue_work(req->ctx->sqo_wq, &req->work); | |
1279 | } | |
1280 | spin_unlock(&poll->head->lock); | |
1281 | ||
1282 | list_del_init(&req->list); | |
1283 | } | |
1284 | ||
1285 | static void io_poll_remove_all(struct io_ring_ctx *ctx) | |
1286 | { | |
1287 | struct io_kiocb *req; | |
1288 | ||
1289 | spin_lock_irq(&ctx->completion_lock); | |
1290 | while (!list_empty(&ctx->cancel_list)) { | |
1291 | req = list_first_entry(&ctx->cancel_list, struct io_kiocb,list); | |
1292 | io_poll_remove_one(req); | |
1293 | } | |
1294 | spin_unlock_irq(&ctx->completion_lock); | |
1295 | } | |
1296 | ||
1297 | /* | |
1298 | * Find a running poll command that matches one specified in sqe->addr, | |
1299 | * and remove it if found. | |
1300 | */ | |
1301 | static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
1302 | { | |
1303 | struct io_ring_ctx *ctx = req->ctx; | |
1304 | struct io_kiocb *poll_req, *next; | |
1305 | int ret = -ENOENT; | |
1306 | ||
1307 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1308 | return -EINVAL; | |
1309 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index || | |
1310 | sqe->poll_events) | |
1311 | return -EINVAL; | |
1312 | ||
1313 | spin_lock_irq(&ctx->completion_lock); | |
1314 | list_for_each_entry_safe(poll_req, next, &ctx->cancel_list, list) { | |
1315 | if (READ_ONCE(sqe->addr) == poll_req->user_data) { | |
1316 | io_poll_remove_one(poll_req); | |
1317 | ret = 0; | |
1318 | break; | |
1319 | } | |
1320 | } | |
1321 | spin_unlock_irq(&ctx->completion_lock); | |
1322 | ||
1323 | io_cqring_add_event(req->ctx, sqe->user_data, ret, 0); | |
e65ef56d | 1324 | io_put_req(req); |
221c5eb2 JA |
1325 | return 0; |
1326 | } | |
1327 | ||
8c838788 JA |
1328 | static void io_poll_complete(struct io_ring_ctx *ctx, struct io_kiocb *req, |
1329 | __poll_t mask) | |
221c5eb2 | 1330 | { |
8c838788 JA |
1331 | req->poll.done = true; |
1332 | io_cqring_fill_event(ctx, req->user_data, mangle_poll(mask), 0); | |
1333 | io_commit_cqring(ctx); | |
221c5eb2 JA |
1334 | } |
1335 | ||
1336 | static void io_poll_complete_work(struct work_struct *work) | |
1337 | { | |
1338 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
1339 | struct io_poll_iocb *poll = &req->poll; | |
1340 | struct poll_table_struct pt = { ._key = poll->events }; | |
1341 | struct io_ring_ctx *ctx = req->ctx; | |
1342 | __poll_t mask = 0; | |
1343 | ||
1344 | if (!READ_ONCE(poll->canceled)) | |
1345 | mask = vfs_poll(poll->file, &pt) & poll->events; | |
1346 | ||
1347 | /* | |
1348 | * Note that ->ki_cancel callers also delete iocb from active_reqs after | |
1349 | * calling ->ki_cancel. We need the ctx_lock roundtrip here to | |
1350 | * synchronize with them. In the cancellation case the list_del_init | |
1351 | * itself is not actually needed, but harmless so we keep it in to | |
1352 | * avoid further branches in the fast path. | |
1353 | */ | |
1354 | spin_lock_irq(&ctx->completion_lock); | |
1355 | if (!mask && !READ_ONCE(poll->canceled)) { | |
1356 | add_wait_queue(poll->head, &poll->wait); | |
1357 | spin_unlock_irq(&ctx->completion_lock); | |
1358 | return; | |
1359 | } | |
1360 | list_del_init(&req->list); | |
8c838788 | 1361 | io_poll_complete(ctx, req, mask); |
221c5eb2 JA |
1362 | spin_unlock_irq(&ctx->completion_lock); |
1363 | ||
8c838788 JA |
1364 | io_cqring_ev_posted(ctx); |
1365 | io_put_req(req); | |
221c5eb2 JA |
1366 | } |
1367 | ||
1368 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
1369 | void *key) | |
1370 | { | |
1371 | struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb, | |
1372 | wait); | |
1373 | struct io_kiocb *req = container_of(poll, struct io_kiocb, poll); | |
1374 | struct io_ring_ctx *ctx = req->ctx; | |
1375 | __poll_t mask = key_to_poll(key); | |
8c838788 | 1376 | unsigned long flags; |
221c5eb2 JA |
1377 | |
1378 | /* for instances that support it check for an event match first: */ | |
8c838788 JA |
1379 | if (mask && !(mask & poll->events)) |
1380 | return 0; | |
221c5eb2 | 1381 | |
8c838788 | 1382 | list_del_init(&poll->wait.entry); |
221c5eb2 | 1383 | |
8c838788 JA |
1384 | if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) { |
1385 | list_del(&req->list); | |
1386 | io_poll_complete(ctx, req, mask); | |
1387 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
221c5eb2 | 1388 | |
8c838788 JA |
1389 | io_cqring_ev_posted(ctx); |
1390 | io_put_req(req); | |
1391 | } else { | |
1392 | queue_work(ctx->sqo_wq, &req->work); | |
221c5eb2 JA |
1393 | } |
1394 | ||
221c5eb2 JA |
1395 | return 1; |
1396 | } | |
1397 | ||
1398 | struct io_poll_table { | |
1399 | struct poll_table_struct pt; | |
1400 | struct io_kiocb *req; | |
1401 | int error; | |
1402 | }; | |
1403 | ||
1404 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
1405 | struct poll_table_struct *p) | |
1406 | { | |
1407 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
1408 | ||
1409 | if (unlikely(pt->req->poll.head)) { | |
1410 | pt->error = -EINVAL; | |
1411 | return; | |
1412 | } | |
1413 | ||
1414 | pt->error = 0; | |
1415 | pt->req->poll.head = head; | |
1416 | add_wait_queue(head, &pt->req->poll.wait); | |
1417 | } | |
1418 | ||
1419 | static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
1420 | { | |
1421 | struct io_poll_iocb *poll = &req->poll; | |
1422 | struct io_ring_ctx *ctx = req->ctx; | |
1423 | struct io_poll_table ipt; | |
8c838788 | 1424 | bool cancel = false; |
221c5eb2 JA |
1425 | __poll_t mask; |
1426 | u16 events; | |
221c5eb2 JA |
1427 | |
1428 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1429 | return -EINVAL; | |
1430 | if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
1431 | return -EINVAL; | |
09bb8394 JA |
1432 | if (!poll->file) |
1433 | return -EBADF; | |
221c5eb2 JA |
1434 | |
1435 | INIT_WORK(&req->work, io_poll_complete_work); | |
1436 | events = READ_ONCE(sqe->poll_events); | |
1437 | poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP; | |
1438 | ||
221c5eb2 | 1439 | poll->head = NULL; |
8c838788 | 1440 | poll->done = false; |
221c5eb2 JA |
1441 | poll->canceled = false; |
1442 | ||
1443 | ipt.pt._qproc = io_poll_queue_proc; | |
1444 | ipt.pt._key = poll->events; | |
1445 | ipt.req = req; | |
1446 | ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */ | |
1447 | ||
1448 | /* initialized the list so that we can do list_empty checks */ | |
1449 | INIT_LIST_HEAD(&poll->wait.entry); | |
1450 | init_waitqueue_func_entry(&poll->wait, io_poll_wake); | |
1451 | ||
221c5eb2 | 1452 | mask = vfs_poll(poll->file, &ipt.pt) & poll->events; |
221c5eb2 JA |
1453 | |
1454 | spin_lock_irq(&ctx->completion_lock); | |
8c838788 JA |
1455 | if (likely(poll->head)) { |
1456 | spin_lock(&poll->head->lock); | |
1457 | if (unlikely(list_empty(&poll->wait.entry))) { | |
1458 | if (ipt.error) | |
1459 | cancel = true; | |
1460 | ipt.error = 0; | |
1461 | mask = 0; | |
1462 | } | |
1463 | if (mask || ipt.error) | |
1464 | list_del_init(&poll->wait.entry); | |
1465 | else if (cancel) | |
1466 | WRITE_ONCE(poll->canceled, true); | |
1467 | else if (!poll->done) /* actually waiting for an event */ | |
1468 | list_add_tail(&req->list, &ctx->cancel_list); | |
1469 | spin_unlock(&poll->head->lock); | |
1470 | } | |
1471 | if (mask) { /* no async, we'd stolen it */ | |
1472 | req->error = mangle_poll(mask); | |
221c5eb2 | 1473 | ipt.error = 0; |
8c838788 | 1474 | io_poll_complete(ctx, req, mask); |
221c5eb2 | 1475 | } |
221c5eb2 JA |
1476 | spin_unlock_irq(&ctx->completion_lock); |
1477 | ||
8c838788 JA |
1478 | if (mask) { |
1479 | io_cqring_ev_posted(ctx); | |
e65ef56d | 1480 | io_put_req(req); |
221c5eb2 | 1481 | } |
8c838788 | 1482 | return ipt.error; |
221c5eb2 JA |
1483 | } |
1484 | ||
de0617e4 JA |
1485 | static int io_req_defer(struct io_ring_ctx *ctx, struct io_kiocb *req, |
1486 | const struct io_uring_sqe *sqe) | |
1487 | { | |
1488 | struct io_uring_sqe *sqe_copy; | |
1489 | ||
1490 | if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) | |
1491 | return 0; | |
1492 | ||
1493 | sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL); | |
1494 | if (!sqe_copy) | |
1495 | return -EAGAIN; | |
1496 | ||
1497 | spin_lock_irq(&ctx->completion_lock); | |
1498 | if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) { | |
1499 | spin_unlock_irq(&ctx->completion_lock); | |
1500 | kfree(sqe_copy); | |
1501 | return 0; | |
1502 | } | |
1503 | ||
1504 | memcpy(sqe_copy, sqe, sizeof(*sqe_copy)); | |
1505 | req->submit.sqe = sqe_copy; | |
1506 | ||
1507 | INIT_WORK(&req->work, io_sq_wq_submit_work); | |
1508 | list_add_tail(&req->list, &ctx->defer_list); | |
1509 | spin_unlock_irq(&ctx->completion_lock); | |
1510 | return -EIOCBQUEUED; | |
1511 | } | |
1512 | ||
2b188cc1 | 1513 | static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
8358e3a8 | 1514 | const struct sqe_submit *s, bool force_nonblock) |
2b188cc1 | 1515 | { |
e0c5c576 | 1516 | int ret, opcode; |
2b188cc1 JA |
1517 | |
1518 | if (unlikely(s->index >= ctx->sq_entries)) | |
1519 | return -EINVAL; | |
1520 | req->user_data = READ_ONCE(s->sqe->user_data); | |
1521 | ||
1522 | opcode = READ_ONCE(s->sqe->opcode); | |
1523 | switch (opcode) { | |
1524 | case IORING_OP_NOP: | |
1525 | ret = io_nop(req, req->user_data); | |
1526 | break; | |
1527 | case IORING_OP_READV: | |
edafccee JA |
1528 | if (unlikely(s->sqe->buf_index)) |
1529 | return -EINVAL; | |
8358e3a8 | 1530 | ret = io_read(req, s, force_nonblock); |
2b188cc1 JA |
1531 | break; |
1532 | case IORING_OP_WRITEV: | |
edafccee JA |
1533 | if (unlikely(s->sqe->buf_index)) |
1534 | return -EINVAL; | |
8358e3a8 | 1535 | ret = io_write(req, s, force_nonblock); |
edafccee JA |
1536 | break; |
1537 | case IORING_OP_READ_FIXED: | |
8358e3a8 | 1538 | ret = io_read(req, s, force_nonblock); |
edafccee JA |
1539 | break; |
1540 | case IORING_OP_WRITE_FIXED: | |
8358e3a8 | 1541 | ret = io_write(req, s, force_nonblock); |
2b188cc1 | 1542 | break; |
c992fe29 CH |
1543 | case IORING_OP_FSYNC: |
1544 | ret = io_fsync(req, s->sqe, force_nonblock); | |
1545 | break; | |
221c5eb2 JA |
1546 | case IORING_OP_POLL_ADD: |
1547 | ret = io_poll_add(req, s->sqe); | |
1548 | break; | |
1549 | case IORING_OP_POLL_REMOVE: | |
1550 | ret = io_poll_remove(req, s->sqe); | |
1551 | break; | |
2b188cc1 JA |
1552 | default: |
1553 | ret = -EINVAL; | |
1554 | break; | |
1555 | } | |
1556 | ||
def596e9 JA |
1557 | if (ret) |
1558 | return ret; | |
1559 | ||
1560 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
1561 | if (req->error == -EAGAIN) | |
1562 | return -EAGAIN; | |
1563 | ||
1564 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
1565 | if (s->needs_lock) | |
1566 | mutex_lock(&ctx->uring_lock); | |
1567 | io_iopoll_req_issued(req); | |
1568 | if (s->needs_lock) | |
1569 | mutex_unlock(&ctx->uring_lock); | |
1570 | } | |
1571 | ||
1572 | return 0; | |
2b188cc1 JA |
1573 | } |
1574 | ||
31b51510 JA |
1575 | static struct async_list *io_async_list_from_sqe(struct io_ring_ctx *ctx, |
1576 | const struct io_uring_sqe *sqe) | |
1577 | { | |
1578 | switch (sqe->opcode) { | |
1579 | case IORING_OP_READV: | |
1580 | case IORING_OP_READ_FIXED: | |
1581 | return &ctx->pending_async[READ]; | |
1582 | case IORING_OP_WRITEV: | |
1583 | case IORING_OP_WRITE_FIXED: | |
1584 | return &ctx->pending_async[WRITE]; | |
1585 | default: | |
1586 | return NULL; | |
1587 | } | |
1588 | } | |
1589 | ||
edafccee JA |
1590 | static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe) |
1591 | { | |
1592 | u8 opcode = READ_ONCE(sqe->opcode); | |
1593 | ||
1594 | return !(opcode == IORING_OP_READ_FIXED || | |
1595 | opcode == IORING_OP_WRITE_FIXED); | |
1596 | } | |
1597 | ||
2b188cc1 JA |
1598 | static void io_sq_wq_submit_work(struct work_struct *work) |
1599 | { | |
1600 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
2b188cc1 | 1601 | struct io_ring_ctx *ctx = req->ctx; |
31b51510 JA |
1602 | struct mm_struct *cur_mm = NULL; |
1603 | struct async_list *async_list; | |
1604 | LIST_HEAD(req_list); | |
edafccee | 1605 | mm_segment_t old_fs; |
2b188cc1 JA |
1606 | int ret; |
1607 | ||
31b51510 JA |
1608 | async_list = io_async_list_from_sqe(ctx, req->submit.sqe); |
1609 | restart: | |
1610 | do { | |
1611 | struct sqe_submit *s = &req->submit; | |
1612 | const struct io_uring_sqe *sqe = s->sqe; | |
2b188cc1 | 1613 | |
8449eeda | 1614 | /* Ensure we clear previously set non-block flag */ |
31b51510 JA |
1615 | req->rw.ki_flags &= ~IOCB_NOWAIT; |
1616 | ||
1617 | ret = 0; | |
1618 | if (io_sqe_needs_user(sqe) && !cur_mm) { | |
1619 | if (!mmget_not_zero(ctx->sqo_mm)) { | |
1620 | ret = -EFAULT; | |
1621 | } else { | |
1622 | cur_mm = ctx->sqo_mm; | |
1623 | use_mm(cur_mm); | |
1624 | old_fs = get_fs(); | |
1625 | set_fs(USER_DS); | |
1626 | } | |
1627 | } | |
1628 | ||
1629 | if (!ret) { | |
1630 | s->has_user = cur_mm != NULL; | |
1631 | s->needs_lock = true; | |
1632 | do { | |
8358e3a8 | 1633 | ret = __io_submit_sqe(ctx, req, s, false); |
31b51510 JA |
1634 | /* |
1635 | * We can get EAGAIN for polled IO even though | |
1636 | * we're forcing a sync submission from here, | |
1637 | * since we can't wait for request slots on the | |
1638 | * block side. | |
1639 | */ | |
1640 | if (ret != -EAGAIN) | |
1641 | break; | |
1642 | cond_resched(); | |
1643 | } while (1); | |
1644 | } | |
817869d2 JA |
1645 | |
1646 | /* drop submission reference */ | |
1647 | io_put_req(req); | |
1648 | ||
31b51510 JA |
1649 | if (ret) { |
1650 | io_cqring_add_event(ctx, sqe->user_data, ret, 0); | |
e65ef56d | 1651 | io_put_req(req); |
31b51510 JA |
1652 | } |
1653 | ||
1654 | /* async context always use a copy of the sqe */ | |
1655 | kfree(sqe); | |
1656 | ||
1657 | if (!async_list) | |
1658 | break; | |
1659 | if (!list_empty(&req_list)) { | |
1660 | req = list_first_entry(&req_list, struct io_kiocb, | |
1661 | list); | |
1662 | list_del(&req->list); | |
1663 | continue; | |
1664 | } | |
1665 | if (list_empty(&async_list->list)) | |
1666 | break; | |
1667 | ||
1668 | req = NULL; | |
1669 | spin_lock(&async_list->lock); | |
1670 | if (list_empty(&async_list->list)) { | |
1671 | spin_unlock(&async_list->lock); | |
1672 | break; | |
1673 | } | |
1674 | list_splice_init(&async_list->list, &req_list); | |
1675 | spin_unlock(&async_list->lock); | |
1676 | ||
1677 | req = list_first_entry(&req_list, struct io_kiocb, list); | |
1678 | list_del(&req->list); | |
1679 | } while (req); | |
edafccee JA |
1680 | |
1681 | /* | |
31b51510 JA |
1682 | * Rare case of racing with a submitter. If we find the count has |
1683 | * dropped to zero AND we have pending work items, then restart | |
1684 | * the processing. This is a tiny race window. | |
edafccee | 1685 | */ |
31b51510 JA |
1686 | if (async_list) { |
1687 | ret = atomic_dec_return(&async_list->cnt); | |
1688 | while (!ret && !list_empty(&async_list->list)) { | |
1689 | spin_lock(&async_list->lock); | |
1690 | atomic_inc(&async_list->cnt); | |
1691 | list_splice_init(&async_list->list, &req_list); | |
1692 | spin_unlock(&async_list->lock); | |
1693 | ||
1694 | if (!list_empty(&req_list)) { | |
1695 | req = list_first_entry(&req_list, | |
1696 | struct io_kiocb, list); | |
1697 | list_del(&req->list); | |
1698 | goto restart; | |
1699 | } | |
1700 | ret = atomic_dec_return(&async_list->cnt); | |
edafccee | 1701 | } |
edafccee | 1702 | } |
2b188cc1 | 1703 | |
31b51510 | 1704 | if (cur_mm) { |
edafccee | 1705 | set_fs(old_fs); |
31b51510 JA |
1706 | unuse_mm(cur_mm); |
1707 | mmput(cur_mm); | |
2b188cc1 | 1708 | } |
31b51510 | 1709 | } |
2b188cc1 | 1710 | |
31b51510 JA |
1711 | /* |
1712 | * See if we can piggy back onto previously submitted work, that is still | |
1713 | * running. We currently only allow this if the new request is sequential | |
1714 | * to the previous one we punted. | |
1715 | */ | |
1716 | static bool io_add_to_prev_work(struct async_list *list, struct io_kiocb *req) | |
1717 | { | |
1718 | bool ret = false; | |
1719 | ||
1720 | if (!list) | |
1721 | return false; | |
1722 | if (!(req->flags & REQ_F_SEQ_PREV)) | |
1723 | return false; | |
1724 | if (!atomic_read(&list->cnt)) | |
1725 | return false; | |
1726 | ||
1727 | ret = true; | |
1728 | spin_lock(&list->lock); | |
1729 | list_add_tail(&req->list, &list->list); | |
1730 | if (!atomic_read(&list->cnt)) { | |
1731 | list_del_init(&req->list); | |
1732 | ret = false; | |
1733 | } | |
1734 | spin_unlock(&list->lock); | |
1735 | return ret; | |
2b188cc1 JA |
1736 | } |
1737 | ||
09bb8394 JA |
1738 | static bool io_op_needs_file(const struct io_uring_sqe *sqe) |
1739 | { | |
1740 | int op = READ_ONCE(sqe->opcode); | |
1741 | ||
1742 | switch (op) { | |
1743 | case IORING_OP_NOP: | |
1744 | case IORING_OP_POLL_REMOVE: | |
1745 | return false; | |
1746 | default: | |
1747 | return true; | |
1748 | } | |
1749 | } | |
1750 | ||
1751 | static int io_req_set_file(struct io_ring_ctx *ctx, const struct sqe_submit *s, | |
1752 | struct io_submit_state *state, struct io_kiocb *req) | |
1753 | { | |
1754 | unsigned flags; | |
1755 | int fd; | |
1756 | ||
1757 | flags = READ_ONCE(s->sqe->flags); | |
1758 | fd = READ_ONCE(s->sqe->fd); | |
1759 | ||
de0617e4 JA |
1760 | if (flags & IOSQE_IO_DRAIN) { |
1761 | req->flags |= REQ_F_IO_DRAIN; | |
1762 | req->sequence = ctx->cached_sq_head - 1; | |
1763 | } | |
1764 | ||
09bb8394 JA |
1765 | if (!io_op_needs_file(s->sqe)) { |
1766 | req->file = NULL; | |
1767 | return 0; | |
1768 | } | |
1769 | ||
1770 | if (flags & IOSQE_FIXED_FILE) { | |
1771 | if (unlikely(!ctx->user_files || | |
1772 | (unsigned) fd >= ctx->nr_user_files)) | |
1773 | return -EBADF; | |
1774 | req->file = ctx->user_files[fd]; | |
1775 | req->flags |= REQ_F_FIXED_FILE; | |
1776 | } else { | |
1777 | if (s->needs_fixed_file) | |
1778 | return -EBADF; | |
1779 | req->file = io_file_get(state, fd); | |
1780 | if (unlikely(!req->file)) | |
1781 | return -EBADF; | |
1782 | } | |
1783 | ||
1784 | return 0; | |
1785 | } | |
1786 | ||
9a56a232 JA |
1787 | static int io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s, |
1788 | struct io_submit_state *state) | |
2b188cc1 JA |
1789 | { |
1790 | struct io_kiocb *req; | |
e0c5c576 | 1791 | int ret; |
2b188cc1 JA |
1792 | |
1793 | /* enforce forwards compatibility on users */ | |
de0617e4 | 1794 | if (unlikely(s->sqe->flags & ~(IOSQE_FIXED_FILE | IOSQE_IO_DRAIN))) |
2b188cc1 JA |
1795 | return -EINVAL; |
1796 | ||
2579f913 | 1797 | req = io_get_req(ctx, state); |
2b188cc1 JA |
1798 | if (unlikely(!req)) |
1799 | return -EAGAIN; | |
1800 | ||
09bb8394 JA |
1801 | ret = io_req_set_file(ctx, s, state, req); |
1802 | if (unlikely(ret)) | |
1803 | goto out; | |
2b188cc1 | 1804 | |
de0617e4 JA |
1805 | ret = io_req_defer(ctx, req, s->sqe); |
1806 | if (ret) { | |
1807 | if (ret == -EIOCBQUEUED) | |
1808 | ret = 0; | |
1809 | return ret; | |
1810 | } | |
1811 | ||
8358e3a8 | 1812 | ret = __io_submit_sqe(ctx, req, s, true); |
8449eeda | 1813 | if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) { |
2b188cc1 JA |
1814 | struct io_uring_sqe *sqe_copy; |
1815 | ||
1816 | sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL); | |
1817 | if (sqe_copy) { | |
31b51510 JA |
1818 | struct async_list *list; |
1819 | ||
2b188cc1 JA |
1820 | memcpy(sqe_copy, s->sqe, sizeof(*sqe_copy)); |
1821 | s->sqe = sqe_copy; | |
1822 | ||
1823 | memcpy(&req->submit, s, sizeof(*s)); | |
31b51510 JA |
1824 | list = io_async_list_from_sqe(ctx, s->sqe); |
1825 | if (!io_add_to_prev_work(list, req)) { | |
1826 | if (list) | |
1827 | atomic_inc(&list->cnt); | |
1828 | INIT_WORK(&req->work, io_sq_wq_submit_work); | |
1829 | queue_work(ctx->sqo_wq, &req->work); | |
1830 | } | |
e65ef56d JA |
1831 | |
1832 | /* | |
1833 | * Queued up for async execution, worker will release | |
1834 | * submit reference when the iocb is actually | |
1835 | * submitted. | |
1836 | */ | |
1837 | return 0; | |
2b188cc1 JA |
1838 | } |
1839 | } | |
e65ef56d | 1840 | |
09bb8394 | 1841 | out: |
e65ef56d JA |
1842 | /* drop submission reference */ |
1843 | io_put_req(req); | |
1844 | ||
1845 | /* and drop final reference, if we failed */ | |
2b188cc1 | 1846 | if (ret) |
e65ef56d | 1847 | io_put_req(req); |
2b188cc1 JA |
1848 | |
1849 | return ret; | |
1850 | } | |
1851 | ||
9a56a232 JA |
1852 | /* |
1853 | * Batched submission is done, ensure local IO is flushed out. | |
1854 | */ | |
1855 | static void io_submit_state_end(struct io_submit_state *state) | |
1856 | { | |
1857 | blk_finish_plug(&state->plug); | |
3d6770fb | 1858 | io_file_put(state); |
2579f913 JA |
1859 | if (state->free_reqs) |
1860 | kmem_cache_free_bulk(req_cachep, state->free_reqs, | |
1861 | &state->reqs[state->cur_req]); | |
9a56a232 JA |
1862 | } |
1863 | ||
1864 | /* | |
1865 | * Start submission side cache. | |
1866 | */ | |
1867 | static void io_submit_state_start(struct io_submit_state *state, | |
1868 | struct io_ring_ctx *ctx, unsigned max_ios) | |
1869 | { | |
1870 | blk_start_plug(&state->plug); | |
2579f913 | 1871 | state->free_reqs = 0; |
9a56a232 JA |
1872 | state->file = NULL; |
1873 | state->ios_left = max_ios; | |
1874 | } | |
1875 | ||
2b188cc1 JA |
1876 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
1877 | { | |
1878 | struct io_sq_ring *ring = ctx->sq_ring; | |
1879 | ||
1880 | if (ctx->cached_sq_head != READ_ONCE(ring->r.head)) { | |
1881 | /* | |
1882 | * Ensure any loads from the SQEs are done at this point, | |
1883 | * since once we write the new head, the application could | |
1884 | * write new data to them. | |
1885 | */ | |
1886 | smp_store_release(&ring->r.head, ctx->cached_sq_head); | |
2b188cc1 JA |
1887 | } |
1888 | } | |
1889 | ||
2b188cc1 JA |
1890 | /* |
1891 | * Fetch an sqe, if one is available. Note that s->sqe will point to memory | |
1892 | * that is mapped by userspace. This means that care needs to be taken to | |
1893 | * ensure that reads are stable, as we cannot rely on userspace always | |
1894 | * being a good citizen. If members of the sqe are validated and then later | |
1895 | * used, it's important that those reads are done through READ_ONCE() to | |
1896 | * prevent a re-load down the line. | |
1897 | */ | |
1898 | static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s) | |
1899 | { | |
1900 | struct io_sq_ring *ring = ctx->sq_ring; | |
1901 | unsigned head; | |
1902 | ||
1903 | /* | |
1904 | * The cached sq head (or cq tail) serves two purposes: | |
1905 | * | |
1906 | * 1) allows us to batch the cost of updating the user visible | |
1907 | * head updates. | |
1908 | * 2) allows the kernel side to track the head on its own, even | |
1909 | * though the application is the one updating it. | |
1910 | */ | |
1911 | head = ctx->cached_sq_head; | |
e523a29c SB |
1912 | /* make sure SQ entry isn't read before tail */ |
1913 | if (head == smp_load_acquire(&ring->r.tail)) | |
2b188cc1 JA |
1914 | return false; |
1915 | ||
1916 | head = READ_ONCE(ring->array[head & ctx->sq_mask]); | |
1917 | if (head < ctx->sq_entries) { | |
1918 | s->index = head; | |
1919 | s->sqe = &ctx->sq_sqes[head]; | |
1920 | ctx->cached_sq_head++; | |
1921 | return true; | |
1922 | } | |
1923 | ||
1924 | /* drop invalid entries */ | |
1925 | ctx->cached_sq_head++; | |
1926 | ring->dropped++; | |
2b188cc1 JA |
1927 | return false; |
1928 | } | |
1929 | ||
6c271ce2 JA |
1930 | static int io_submit_sqes(struct io_ring_ctx *ctx, struct sqe_submit *sqes, |
1931 | unsigned int nr, bool has_user, bool mm_fault) | |
1932 | { | |
1933 | struct io_submit_state state, *statep = NULL; | |
1934 | int ret, i, submitted = 0; | |
1935 | ||
1936 | if (nr > IO_PLUG_THRESHOLD) { | |
1937 | io_submit_state_start(&state, ctx, nr); | |
1938 | statep = &state; | |
1939 | } | |
1940 | ||
1941 | for (i = 0; i < nr; i++) { | |
1942 | if (unlikely(mm_fault)) { | |
1943 | ret = -EFAULT; | |
1944 | } else { | |
1945 | sqes[i].has_user = has_user; | |
1946 | sqes[i].needs_lock = true; | |
1947 | sqes[i].needs_fixed_file = true; | |
1948 | ret = io_submit_sqe(ctx, &sqes[i], statep); | |
1949 | } | |
1950 | if (!ret) { | |
1951 | submitted++; | |
1952 | continue; | |
1953 | } | |
1954 | ||
1955 | io_cqring_add_event(ctx, sqes[i].sqe->user_data, ret, 0); | |
1956 | } | |
1957 | ||
1958 | if (statep) | |
1959 | io_submit_state_end(&state); | |
1960 | ||
1961 | return submitted; | |
1962 | } | |
1963 | ||
1964 | static int io_sq_thread(void *data) | |
1965 | { | |
1966 | struct sqe_submit sqes[IO_IOPOLL_BATCH]; | |
1967 | struct io_ring_ctx *ctx = data; | |
1968 | struct mm_struct *cur_mm = NULL; | |
1969 | mm_segment_t old_fs; | |
1970 | DEFINE_WAIT(wait); | |
1971 | unsigned inflight; | |
1972 | unsigned long timeout; | |
1973 | ||
1974 | old_fs = get_fs(); | |
1975 | set_fs(USER_DS); | |
1976 | ||
1977 | timeout = inflight = 0; | |
1978 | while (!kthread_should_stop() && !ctx->sqo_stop) { | |
1979 | bool all_fixed, mm_fault = false; | |
1980 | int i; | |
1981 | ||
1982 | if (inflight) { | |
1983 | unsigned nr_events = 0; | |
1984 | ||
1985 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
1986 | /* | |
1987 | * We disallow the app entering submit/complete | |
1988 | * with polling, but we still need to lock the | |
1989 | * ring to prevent racing with polled issue | |
1990 | * that got punted to a workqueue. | |
1991 | */ | |
1992 | mutex_lock(&ctx->uring_lock); | |
1993 | io_iopoll_check(ctx, &nr_events, 0); | |
1994 | mutex_unlock(&ctx->uring_lock); | |
1995 | } else { | |
1996 | /* | |
1997 | * Normal IO, just pretend everything completed. | |
1998 | * We don't have to poll completions for that. | |
1999 | */ | |
2000 | nr_events = inflight; | |
2001 | } | |
2002 | ||
2003 | inflight -= nr_events; | |
2004 | if (!inflight) | |
2005 | timeout = jiffies + ctx->sq_thread_idle; | |
2006 | } | |
2007 | ||
2008 | if (!io_get_sqring(ctx, &sqes[0])) { | |
2009 | /* | |
2010 | * We're polling. If we're within the defined idle | |
2011 | * period, then let us spin without work before going | |
2012 | * to sleep. | |
2013 | */ | |
2014 | if (inflight || !time_after(jiffies, timeout)) { | |
2015 | cpu_relax(); | |
2016 | continue; | |
2017 | } | |
2018 | ||
2019 | /* | |
2020 | * Drop cur_mm before scheduling, we can't hold it for | |
2021 | * long periods (or over schedule()). Do this before | |
2022 | * adding ourselves to the waitqueue, as the unuse/drop | |
2023 | * may sleep. | |
2024 | */ | |
2025 | if (cur_mm) { | |
2026 | unuse_mm(cur_mm); | |
2027 | mmput(cur_mm); | |
2028 | cur_mm = NULL; | |
2029 | } | |
2030 | ||
2031 | prepare_to_wait(&ctx->sqo_wait, &wait, | |
2032 | TASK_INTERRUPTIBLE); | |
2033 | ||
2034 | /* Tell userspace we may need a wakeup call */ | |
2035 | ctx->sq_ring->flags |= IORING_SQ_NEED_WAKEUP; | |
0d7bae69 SB |
2036 | /* make sure to read SQ tail after writing flags */ |
2037 | smp_mb(); | |
6c271ce2 JA |
2038 | |
2039 | if (!io_get_sqring(ctx, &sqes[0])) { | |
2040 | if (kthread_should_stop()) { | |
2041 | finish_wait(&ctx->sqo_wait, &wait); | |
2042 | break; | |
2043 | } | |
2044 | if (signal_pending(current)) | |
2045 | flush_signals(current); | |
2046 | schedule(); | |
2047 | finish_wait(&ctx->sqo_wait, &wait); | |
2048 | ||
2049 | ctx->sq_ring->flags &= ~IORING_SQ_NEED_WAKEUP; | |
6c271ce2 JA |
2050 | continue; |
2051 | } | |
2052 | finish_wait(&ctx->sqo_wait, &wait); | |
2053 | ||
2054 | ctx->sq_ring->flags &= ~IORING_SQ_NEED_WAKEUP; | |
6c271ce2 JA |
2055 | } |
2056 | ||
2057 | i = 0; | |
2058 | all_fixed = true; | |
2059 | do { | |
2060 | if (all_fixed && io_sqe_needs_user(sqes[i].sqe)) | |
2061 | all_fixed = false; | |
2062 | ||
2063 | i++; | |
2064 | if (i == ARRAY_SIZE(sqes)) | |
2065 | break; | |
2066 | } while (io_get_sqring(ctx, &sqes[i])); | |
2067 | ||
2068 | /* Unless all new commands are FIXED regions, grab mm */ | |
2069 | if (!all_fixed && !cur_mm) { | |
2070 | mm_fault = !mmget_not_zero(ctx->sqo_mm); | |
2071 | if (!mm_fault) { | |
2072 | use_mm(ctx->sqo_mm); | |
2073 | cur_mm = ctx->sqo_mm; | |
2074 | } | |
2075 | } | |
2076 | ||
2077 | inflight += io_submit_sqes(ctx, sqes, i, cur_mm != NULL, | |
2078 | mm_fault); | |
2079 | ||
2080 | /* Commit SQ ring head once we've consumed all SQEs */ | |
2081 | io_commit_sqring(ctx); | |
2082 | } | |
2083 | ||
2084 | set_fs(old_fs); | |
2085 | if (cur_mm) { | |
2086 | unuse_mm(cur_mm); | |
2087 | mmput(cur_mm); | |
2088 | } | |
06058632 JA |
2089 | |
2090 | if (kthread_should_park()) | |
2091 | kthread_parkme(); | |
2092 | ||
6c271ce2 JA |
2093 | return 0; |
2094 | } | |
2095 | ||
2b188cc1 JA |
2096 | static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit) |
2097 | { | |
9a56a232 | 2098 | struct io_submit_state state, *statep = NULL; |
5c8b0b54 | 2099 | int i, submit = 0; |
2b188cc1 | 2100 | |
9a56a232 JA |
2101 | if (to_submit > IO_PLUG_THRESHOLD) { |
2102 | io_submit_state_start(&state, ctx, to_submit); | |
2103 | statep = &state; | |
2104 | } | |
2b188cc1 JA |
2105 | |
2106 | for (i = 0; i < to_submit; i++) { | |
2107 | struct sqe_submit s; | |
5c8b0b54 | 2108 | int ret; |
2b188cc1 JA |
2109 | |
2110 | if (!io_get_sqring(ctx, &s)) | |
2111 | break; | |
2112 | ||
2113 | s.has_user = true; | |
def596e9 | 2114 | s.needs_lock = false; |
6c271ce2 | 2115 | s.needs_fixed_file = false; |
5c8b0b54 | 2116 | submit++; |
def596e9 | 2117 | |
9a56a232 | 2118 | ret = io_submit_sqe(ctx, &s, statep); |
5c8b0b54 JA |
2119 | if (ret) |
2120 | io_cqring_add_event(ctx, s.sqe->user_data, ret, 0); | |
2b188cc1 JA |
2121 | } |
2122 | io_commit_sqring(ctx); | |
2123 | ||
9a56a232 JA |
2124 | if (statep) |
2125 | io_submit_state_end(statep); | |
2b188cc1 | 2126 | |
5c8b0b54 | 2127 | return submit; |
2b188cc1 JA |
2128 | } |
2129 | ||
2130 | static unsigned io_cqring_events(struct io_cq_ring *ring) | |
2131 | { | |
2132 | return READ_ONCE(ring->r.tail) - READ_ONCE(ring->r.head); | |
2133 | } | |
2134 | ||
2135 | /* | |
2136 | * Wait until events become available, if we don't already have some. The | |
2137 | * application must reap them itself, as they reside on the shared cq ring. | |
2138 | */ | |
2139 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
2140 | const sigset_t __user *sig, size_t sigsz) | |
2141 | { | |
2142 | struct io_cq_ring *ring = ctx->cq_ring; | |
2143 | sigset_t ksigmask, sigsaved; | |
2144 | DEFINE_WAIT(wait); | |
2145 | int ret; | |
2146 | ||
2147 | /* See comment at the top of this file */ | |
2148 | smp_rmb(); | |
2149 | if (io_cqring_events(ring) >= min_events) | |
2150 | return 0; | |
2151 | ||
2152 | if (sig) { | |
9e75ad5d AB |
2153 | #ifdef CONFIG_COMPAT |
2154 | if (in_compat_syscall()) | |
2155 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
2156 | &ksigmask, &sigsaved, sigsz); | |
2157 | else | |
2158 | #endif | |
2159 | ret = set_user_sigmask(sig, &ksigmask, | |
2160 | &sigsaved, sigsz); | |
2161 | ||
2b188cc1 JA |
2162 | if (ret) |
2163 | return ret; | |
2164 | } | |
2165 | ||
2166 | do { | |
2167 | prepare_to_wait(&ctx->wait, &wait, TASK_INTERRUPTIBLE); | |
2168 | ||
2169 | ret = 0; | |
2170 | /* See comment at the top of this file */ | |
2171 | smp_rmb(); | |
2172 | if (io_cqring_events(ring) >= min_events) | |
2173 | break; | |
2174 | ||
2175 | schedule(); | |
2176 | ||
2177 | ret = -EINTR; | |
2178 | if (signal_pending(current)) | |
2179 | break; | |
2180 | } while (1); | |
2181 | ||
2182 | finish_wait(&ctx->wait, &wait); | |
2183 | ||
2184 | if (sig) | |
2185 | restore_user_sigmask(sig, &sigsaved); | |
2186 | ||
2187 | return READ_ONCE(ring->r.head) == READ_ONCE(ring->r.tail) ? ret : 0; | |
2188 | } | |
2189 | ||
6b06314c JA |
2190 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
2191 | { | |
2192 | #if defined(CONFIG_UNIX) | |
2193 | if (ctx->ring_sock) { | |
2194 | struct sock *sock = ctx->ring_sock->sk; | |
2195 | struct sk_buff *skb; | |
2196 | ||
2197 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
2198 | kfree_skb(skb); | |
2199 | } | |
2200 | #else | |
2201 | int i; | |
2202 | ||
2203 | for (i = 0; i < ctx->nr_user_files; i++) | |
2204 | fput(ctx->user_files[i]); | |
2205 | #endif | |
2206 | } | |
2207 | ||
2208 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) | |
2209 | { | |
2210 | if (!ctx->user_files) | |
2211 | return -ENXIO; | |
2212 | ||
2213 | __io_sqe_files_unregister(ctx); | |
2214 | kfree(ctx->user_files); | |
2215 | ctx->user_files = NULL; | |
2216 | ctx->nr_user_files = 0; | |
2217 | return 0; | |
2218 | } | |
2219 | ||
6c271ce2 JA |
2220 | static void io_sq_thread_stop(struct io_ring_ctx *ctx) |
2221 | { | |
2222 | if (ctx->sqo_thread) { | |
2223 | ctx->sqo_stop = 1; | |
2224 | mb(); | |
06058632 | 2225 | kthread_park(ctx->sqo_thread); |
6c271ce2 JA |
2226 | kthread_stop(ctx->sqo_thread); |
2227 | ctx->sqo_thread = NULL; | |
2228 | } | |
2229 | } | |
2230 | ||
6b06314c JA |
2231 | static void io_finish_async(struct io_ring_ctx *ctx) |
2232 | { | |
6c271ce2 JA |
2233 | io_sq_thread_stop(ctx); |
2234 | ||
6b06314c JA |
2235 | if (ctx->sqo_wq) { |
2236 | destroy_workqueue(ctx->sqo_wq); | |
2237 | ctx->sqo_wq = NULL; | |
2238 | } | |
2239 | } | |
2240 | ||
2241 | #if defined(CONFIG_UNIX) | |
2242 | static void io_destruct_skb(struct sk_buff *skb) | |
2243 | { | |
2244 | struct io_ring_ctx *ctx = skb->sk->sk_user_data; | |
2245 | ||
2246 | io_finish_async(ctx); | |
2247 | unix_destruct_scm(skb); | |
2248 | } | |
2249 | ||
2250 | /* | |
2251 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
2252 | * the io_uring can be safely unregistered on process exit, even if we have | |
2253 | * loops in the file referencing. | |
2254 | */ | |
2255 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
2256 | { | |
2257 | struct sock *sk = ctx->ring_sock->sk; | |
2258 | struct scm_fp_list *fpl; | |
2259 | struct sk_buff *skb; | |
2260 | int i; | |
2261 | ||
2262 | if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) { | |
2263 | unsigned long inflight = ctx->user->unix_inflight + nr; | |
2264 | ||
2265 | if (inflight > task_rlimit(current, RLIMIT_NOFILE)) | |
2266 | return -EMFILE; | |
2267 | } | |
2268 | ||
2269 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); | |
2270 | if (!fpl) | |
2271 | return -ENOMEM; | |
2272 | ||
2273 | skb = alloc_skb(0, GFP_KERNEL); | |
2274 | if (!skb) { | |
2275 | kfree(fpl); | |
2276 | return -ENOMEM; | |
2277 | } | |
2278 | ||
2279 | skb->sk = sk; | |
2280 | skb->destructor = io_destruct_skb; | |
2281 | ||
2282 | fpl->user = get_uid(ctx->user); | |
2283 | for (i = 0; i < nr; i++) { | |
2284 | fpl->fp[i] = get_file(ctx->user_files[i + offset]); | |
2285 | unix_inflight(fpl->user, fpl->fp[i]); | |
2286 | } | |
2287 | ||
2288 | fpl->max = fpl->count = nr; | |
2289 | UNIXCB(skb).fp = fpl; | |
2290 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); | |
2291 | skb_queue_head(&sk->sk_receive_queue, skb); | |
2292 | ||
2293 | for (i = 0; i < nr; i++) | |
2294 | fput(fpl->fp[i]); | |
2295 | ||
2296 | return 0; | |
2297 | } | |
2298 | ||
2299 | /* | |
2300 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
2301 | * causes regular reference counting to break down. We rely on the UNIX | |
2302 | * garbage collection to take care of this problem for us. | |
2303 | */ | |
2304 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
2305 | { | |
2306 | unsigned left, total; | |
2307 | int ret = 0; | |
2308 | ||
2309 | total = 0; | |
2310 | left = ctx->nr_user_files; | |
2311 | while (left) { | |
2312 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
2313 | int ret; | |
2314 | ||
2315 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
2316 | if (ret) | |
2317 | break; | |
2318 | left -= this_files; | |
2319 | total += this_files; | |
2320 | } | |
2321 | ||
2322 | if (!ret) | |
2323 | return 0; | |
2324 | ||
2325 | while (total < ctx->nr_user_files) { | |
2326 | fput(ctx->user_files[total]); | |
2327 | total++; | |
2328 | } | |
2329 | ||
2330 | return ret; | |
2331 | } | |
2332 | #else | |
2333 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
2334 | { | |
2335 | return 0; | |
2336 | } | |
2337 | #endif | |
2338 | ||
2339 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, | |
2340 | unsigned nr_args) | |
2341 | { | |
2342 | __s32 __user *fds = (__s32 __user *) arg; | |
2343 | int fd, ret = 0; | |
2344 | unsigned i; | |
2345 | ||
2346 | if (ctx->user_files) | |
2347 | return -EBUSY; | |
2348 | if (!nr_args) | |
2349 | return -EINVAL; | |
2350 | if (nr_args > IORING_MAX_FIXED_FILES) | |
2351 | return -EMFILE; | |
2352 | ||
2353 | ctx->user_files = kcalloc(nr_args, sizeof(struct file *), GFP_KERNEL); | |
2354 | if (!ctx->user_files) | |
2355 | return -ENOMEM; | |
2356 | ||
2357 | for (i = 0; i < nr_args; i++) { | |
2358 | ret = -EFAULT; | |
2359 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) | |
2360 | break; | |
2361 | ||
2362 | ctx->user_files[i] = fget(fd); | |
2363 | ||
2364 | ret = -EBADF; | |
2365 | if (!ctx->user_files[i]) | |
2366 | break; | |
2367 | /* | |
2368 | * Don't allow io_uring instances to be registered. If UNIX | |
2369 | * isn't enabled, then this causes a reference cycle and this | |
2370 | * instance can never get freed. If UNIX is enabled we'll | |
2371 | * handle it just fine, but there's still no point in allowing | |
2372 | * a ring fd as it doesn't support regular read/write anyway. | |
2373 | */ | |
2374 | if (ctx->user_files[i]->f_op == &io_uring_fops) { | |
2375 | fput(ctx->user_files[i]); | |
2376 | break; | |
2377 | } | |
2378 | ctx->nr_user_files++; | |
2379 | ret = 0; | |
2380 | } | |
2381 | ||
2382 | if (ret) { | |
2383 | for (i = 0; i < ctx->nr_user_files; i++) | |
2384 | fput(ctx->user_files[i]); | |
2385 | ||
2386 | kfree(ctx->user_files); | |
25adf50f | 2387 | ctx->user_files = NULL; |
6b06314c JA |
2388 | ctx->nr_user_files = 0; |
2389 | return ret; | |
2390 | } | |
2391 | ||
2392 | ret = io_sqe_files_scm(ctx); | |
2393 | if (ret) | |
2394 | io_sqe_files_unregister(ctx); | |
2395 | ||
2396 | return ret; | |
2397 | } | |
2398 | ||
6c271ce2 JA |
2399 | static int io_sq_offload_start(struct io_ring_ctx *ctx, |
2400 | struct io_uring_params *p) | |
2b188cc1 JA |
2401 | { |
2402 | int ret; | |
2403 | ||
6c271ce2 | 2404 | init_waitqueue_head(&ctx->sqo_wait); |
2b188cc1 JA |
2405 | mmgrab(current->mm); |
2406 | ctx->sqo_mm = current->mm; | |
2407 | ||
6c271ce2 | 2408 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3ec482d1 JA |
2409 | ret = -EPERM; |
2410 | if (!capable(CAP_SYS_ADMIN)) | |
2411 | goto err; | |
2412 | ||
917257da JA |
2413 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
2414 | if (!ctx->sq_thread_idle) | |
2415 | ctx->sq_thread_idle = HZ; | |
2416 | ||
6c271ce2 | 2417 | if (p->flags & IORING_SETUP_SQ_AFF) { |
975554b0 MR |
2418 | int cpu = array_index_nospec(p->sq_thread_cpu, |
2419 | nr_cpu_ids); | |
6c271ce2 | 2420 | |
917257da | 2421 | ret = -EINVAL; |
975554b0 | 2422 | if (!cpu_possible(cpu)) |
917257da JA |
2423 | goto err; |
2424 | ||
6c271ce2 JA |
2425 | ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread, |
2426 | ctx, cpu, | |
2427 | "io_uring-sq"); | |
2428 | } else { | |
2429 | ctx->sqo_thread = kthread_create(io_sq_thread, ctx, | |
2430 | "io_uring-sq"); | |
2431 | } | |
2432 | if (IS_ERR(ctx->sqo_thread)) { | |
2433 | ret = PTR_ERR(ctx->sqo_thread); | |
2434 | ctx->sqo_thread = NULL; | |
2435 | goto err; | |
2436 | } | |
2437 | wake_up_process(ctx->sqo_thread); | |
2438 | } else if (p->flags & IORING_SETUP_SQ_AFF) { | |
2439 | /* Can't have SQ_AFF without SQPOLL */ | |
2440 | ret = -EINVAL; | |
2441 | goto err; | |
2442 | } | |
2443 | ||
2b188cc1 JA |
2444 | /* Do QD, or 2 * CPUS, whatever is smallest */ |
2445 | ctx->sqo_wq = alloc_workqueue("io_ring-wq", WQ_UNBOUND | WQ_FREEZABLE, | |
2446 | min(ctx->sq_entries - 1, 2 * num_online_cpus())); | |
2447 | if (!ctx->sqo_wq) { | |
2448 | ret = -ENOMEM; | |
2449 | goto err; | |
2450 | } | |
2451 | ||
2452 | return 0; | |
2453 | err: | |
6c271ce2 | 2454 | io_sq_thread_stop(ctx); |
2b188cc1 JA |
2455 | mmdrop(ctx->sqo_mm); |
2456 | ctx->sqo_mm = NULL; | |
2457 | return ret; | |
2458 | } | |
2459 | ||
2460 | static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages) | |
2461 | { | |
2462 | atomic_long_sub(nr_pages, &user->locked_vm); | |
2463 | } | |
2464 | ||
2465 | static int io_account_mem(struct user_struct *user, unsigned long nr_pages) | |
2466 | { | |
2467 | unsigned long page_limit, cur_pages, new_pages; | |
2468 | ||
2469 | /* Don't allow more pages than we can safely lock */ | |
2470 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
2471 | ||
2472 | do { | |
2473 | cur_pages = atomic_long_read(&user->locked_vm); | |
2474 | new_pages = cur_pages + nr_pages; | |
2475 | if (new_pages > page_limit) | |
2476 | return -ENOMEM; | |
2477 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
2478 | new_pages) != cur_pages); | |
2479 | ||
2480 | return 0; | |
2481 | } | |
2482 | ||
2483 | static void io_mem_free(void *ptr) | |
2484 | { | |
52e04ef4 MR |
2485 | struct page *page; |
2486 | ||
2487 | if (!ptr) | |
2488 | return; | |
2b188cc1 | 2489 | |
52e04ef4 | 2490 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
2491 | if (put_page_testzero(page)) |
2492 | free_compound_page(page); | |
2493 | } | |
2494 | ||
2495 | static void *io_mem_alloc(size_t size) | |
2496 | { | |
2497 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
2498 | __GFP_NORETRY; | |
2499 | ||
2500 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
2501 | } | |
2502 | ||
2503 | static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries) | |
2504 | { | |
2505 | struct io_sq_ring *sq_ring; | |
2506 | struct io_cq_ring *cq_ring; | |
2507 | size_t bytes; | |
2508 | ||
2509 | bytes = struct_size(sq_ring, array, sq_entries); | |
2510 | bytes += array_size(sizeof(struct io_uring_sqe), sq_entries); | |
2511 | bytes += struct_size(cq_ring, cqes, cq_entries); | |
2512 | ||
2513 | return (bytes + PAGE_SIZE - 1) / PAGE_SIZE; | |
2514 | } | |
2515 | ||
edafccee JA |
2516 | static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx) |
2517 | { | |
2518 | int i, j; | |
2519 | ||
2520 | if (!ctx->user_bufs) | |
2521 | return -ENXIO; | |
2522 | ||
2523 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
2524 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
2525 | ||
2526 | for (j = 0; j < imu->nr_bvecs; j++) | |
2527 | put_page(imu->bvec[j].bv_page); | |
2528 | ||
2529 | if (ctx->account_mem) | |
2530 | io_unaccount_mem(ctx->user, imu->nr_bvecs); | |
d4ef6475 | 2531 | kvfree(imu->bvec); |
edafccee JA |
2532 | imu->nr_bvecs = 0; |
2533 | } | |
2534 | ||
2535 | kfree(ctx->user_bufs); | |
2536 | ctx->user_bufs = NULL; | |
2537 | ctx->nr_user_bufs = 0; | |
2538 | return 0; | |
2539 | } | |
2540 | ||
2541 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
2542 | void __user *arg, unsigned index) | |
2543 | { | |
2544 | struct iovec __user *src; | |
2545 | ||
2546 | #ifdef CONFIG_COMPAT | |
2547 | if (ctx->compat) { | |
2548 | struct compat_iovec __user *ciovs; | |
2549 | struct compat_iovec ciov; | |
2550 | ||
2551 | ciovs = (struct compat_iovec __user *) arg; | |
2552 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
2553 | return -EFAULT; | |
2554 | ||
2555 | dst->iov_base = (void __user *) (unsigned long) ciov.iov_base; | |
2556 | dst->iov_len = ciov.iov_len; | |
2557 | return 0; | |
2558 | } | |
2559 | #endif | |
2560 | src = (struct iovec __user *) arg; | |
2561 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
2562 | return -EFAULT; | |
2563 | return 0; | |
2564 | } | |
2565 | ||
2566 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, | |
2567 | unsigned nr_args) | |
2568 | { | |
2569 | struct vm_area_struct **vmas = NULL; | |
2570 | struct page **pages = NULL; | |
2571 | int i, j, got_pages = 0; | |
2572 | int ret = -EINVAL; | |
2573 | ||
2574 | if (ctx->user_bufs) | |
2575 | return -EBUSY; | |
2576 | if (!nr_args || nr_args > UIO_MAXIOV) | |
2577 | return -EINVAL; | |
2578 | ||
2579 | ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf), | |
2580 | GFP_KERNEL); | |
2581 | if (!ctx->user_bufs) | |
2582 | return -ENOMEM; | |
2583 | ||
2584 | for (i = 0; i < nr_args; i++) { | |
2585 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
2586 | unsigned long off, start, end, ubuf; | |
2587 | int pret, nr_pages; | |
2588 | struct iovec iov; | |
2589 | size_t size; | |
2590 | ||
2591 | ret = io_copy_iov(ctx, &iov, arg, i); | |
2592 | if (ret) | |
2593 | break; | |
2594 | ||
2595 | /* | |
2596 | * Don't impose further limits on the size and buffer | |
2597 | * constraints here, we'll -EINVAL later when IO is | |
2598 | * submitted if they are wrong. | |
2599 | */ | |
2600 | ret = -EFAULT; | |
2601 | if (!iov.iov_base || !iov.iov_len) | |
2602 | goto err; | |
2603 | ||
2604 | /* arbitrary limit, but we need something */ | |
2605 | if (iov.iov_len > SZ_1G) | |
2606 | goto err; | |
2607 | ||
2608 | ubuf = (unsigned long) iov.iov_base; | |
2609 | end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
2610 | start = ubuf >> PAGE_SHIFT; | |
2611 | nr_pages = end - start; | |
2612 | ||
2613 | if (ctx->account_mem) { | |
2614 | ret = io_account_mem(ctx->user, nr_pages); | |
2615 | if (ret) | |
2616 | goto err; | |
2617 | } | |
2618 | ||
2619 | ret = 0; | |
2620 | if (!pages || nr_pages > got_pages) { | |
2621 | kfree(vmas); | |
2622 | kfree(pages); | |
d4ef6475 | 2623 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), |
edafccee | 2624 | GFP_KERNEL); |
d4ef6475 | 2625 | vmas = kvmalloc_array(nr_pages, |
edafccee JA |
2626 | sizeof(struct vm_area_struct *), |
2627 | GFP_KERNEL); | |
2628 | if (!pages || !vmas) { | |
2629 | ret = -ENOMEM; | |
2630 | if (ctx->account_mem) | |
2631 | io_unaccount_mem(ctx->user, nr_pages); | |
2632 | goto err; | |
2633 | } | |
2634 | got_pages = nr_pages; | |
2635 | } | |
2636 | ||
d4ef6475 | 2637 | imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec), |
edafccee JA |
2638 | GFP_KERNEL); |
2639 | ret = -ENOMEM; | |
2640 | if (!imu->bvec) { | |
2641 | if (ctx->account_mem) | |
2642 | io_unaccount_mem(ctx->user, nr_pages); | |
2643 | goto err; | |
2644 | } | |
2645 | ||
2646 | ret = 0; | |
2647 | down_read(¤t->mm->mmap_sem); | |
2648 | pret = get_user_pages_longterm(ubuf, nr_pages, FOLL_WRITE, | |
2649 | pages, vmas); | |
2650 | if (pret == nr_pages) { | |
2651 | /* don't support file backed memory */ | |
2652 | for (j = 0; j < nr_pages; j++) { | |
2653 | struct vm_area_struct *vma = vmas[j]; | |
2654 | ||
2655 | if (vma->vm_file && | |
2656 | !is_file_hugepages(vma->vm_file)) { | |
2657 | ret = -EOPNOTSUPP; | |
2658 | break; | |
2659 | } | |
2660 | } | |
2661 | } else { | |
2662 | ret = pret < 0 ? pret : -EFAULT; | |
2663 | } | |
2664 | up_read(¤t->mm->mmap_sem); | |
2665 | if (ret) { | |
2666 | /* | |
2667 | * if we did partial map, or found file backed vmas, | |
2668 | * release any pages we did get | |
2669 | */ | |
2670 | if (pret > 0) { | |
2671 | for (j = 0; j < pret; j++) | |
2672 | put_page(pages[j]); | |
2673 | } | |
2674 | if (ctx->account_mem) | |
2675 | io_unaccount_mem(ctx->user, nr_pages); | |
d4ef6475 | 2676 | kvfree(imu->bvec); |
edafccee JA |
2677 | goto err; |
2678 | } | |
2679 | ||
2680 | off = ubuf & ~PAGE_MASK; | |
2681 | size = iov.iov_len; | |
2682 | for (j = 0; j < nr_pages; j++) { | |
2683 | size_t vec_len; | |
2684 | ||
2685 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
2686 | imu->bvec[j].bv_page = pages[j]; | |
2687 | imu->bvec[j].bv_len = vec_len; | |
2688 | imu->bvec[j].bv_offset = off; | |
2689 | off = 0; | |
2690 | size -= vec_len; | |
2691 | } | |
2692 | /* store original address for later verification */ | |
2693 | imu->ubuf = ubuf; | |
2694 | imu->len = iov.iov_len; | |
2695 | imu->nr_bvecs = nr_pages; | |
2696 | ||
2697 | ctx->nr_user_bufs++; | |
2698 | } | |
d4ef6475 MR |
2699 | kvfree(pages); |
2700 | kvfree(vmas); | |
edafccee JA |
2701 | return 0; |
2702 | err: | |
d4ef6475 MR |
2703 | kvfree(pages); |
2704 | kvfree(vmas); | |
edafccee JA |
2705 | io_sqe_buffer_unregister(ctx); |
2706 | return ret; | |
2707 | } | |
2708 | ||
2b188cc1 JA |
2709 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
2710 | { | |
6b06314c | 2711 | io_finish_async(ctx); |
2b188cc1 JA |
2712 | if (ctx->sqo_mm) |
2713 | mmdrop(ctx->sqo_mm); | |
def596e9 JA |
2714 | |
2715 | io_iopoll_reap_events(ctx); | |
edafccee | 2716 | io_sqe_buffer_unregister(ctx); |
6b06314c | 2717 | io_sqe_files_unregister(ctx); |
def596e9 | 2718 | |
2b188cc1 JA |
2719 | #if defined(CONFIG_UNIX) |
2720 | if (ctx->ring_sock) | |
2721 | sock_release(ctx->ring_sock); | |
2722 | #endif | |
2723 | ||
2724 | io_mem_free(ctx->sq_ring); | |
2725 | io_mem_free(ctx->sq_sqes); | |
2726 | io_mem_free(ctx->cq_ring); | |
2727 | ||
2728 | percpu_ref_exit(&ctx->refs); | |
2729 | if (ctx->account_mem) | |
2730 | io_unaccount_mem(ctx->user, | |
2731 | ring_pages(ctx->sq_entries, ctx->cq_entries)); | |
2732 | free_uid(ctx->user); | |
2733 | kfree(ctx); | |
2734 | } | |
2735 | ||
2736 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
2737 | { | |
2738 | struct io_ring_ctx *ctx = file->private_data; | |
2739 | __poll_t mask = 0; | |
2740 | ||
2741 | poll_wait(file, &ctx->cq_wait, wait); | |
4f7067c3 SB |
2742 | /* |
2743 | * synchronizes with barrier from wq_has_sleeper call in | |
2744 | * io_commit_cqring | |
2745 | */ | |
2b188cc1 | 2746 | smp_rmb(); |
fb775faa SB |
2747 | if (READ_ONCE(ctx->sq_ring->r.tail) - ctx->cached_sq_head != |
2748 | ctx->sq_ring->ring_entries) | |
2b188cc1 JA |
2749 | mask |= EPOLLOUT | EPOLLWRNORM; |
2750 | if (READ_ONCE(ctx->cq_ring->r.head) != ctx->cached_cq_tail) | |
2751 | mask |= EPOLLIN | EPOLLRDNORM; | |
2752 | ||
2753 | return mask; | |
2754 | } | |
2755 | ||
2756 | static int io_uring_fasync(int fd, struct file *file, int on) | |
2757 | { | |
2758 | struct io_ring_ctx *ctx = file->private_data; | |
2759 | ||
2760 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
2761 | } | |
2762 | ||
2763 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) | |
2764 | { | |
2765 | mutex_lock(&ctx->uring_lock); | |
2766 | percpu_ref_kill(&ctx->refs); | |
2767 | mutex_unlock(&ctx->uring_lock); | |
2768 | ||
221c5eb2 | 2769 | io_poll_remove_all(ctx); |
def596e9 | 2770 | io_iopoll_reap_events(ctx); |
2b188cc1 JA |
2771 | wait_for_completion(&ctx->ctx_done); |
2772 | io_ring_ctx_free(ctx); | |
2773 | } | |
2774 | ||
2775 | static int io_uring_release(struct inode *inode, struct file *file) | |
2776 | { | |
2777 | struct io_ring_ctx *ctx = file->private_data; | |
2778 | ||
2779 | file->private_data = NULL; | |
2780 | io_ring_ctx_wait_and_kill(ctx); | |
2781 | return 0; | |
2782 | } | |
2783 | ||
2784 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
2785 | { | |
2786 | loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT; | |
2787 | unsigned long sz = vma->vm_end - vma->vm_start; | |
2788 | struct io_ring_ctx *ctx = file->private_data; | |
2789 | unsigned long pfn; | |
2790 | struct page *page; | |
2791 | void *ptr; | |
2792 | ||
2793 | switch (offset) { | |
2794 | case IORING_OFF_SQ_RING: | |
2795 | ptr = ctx->sq_ring; | |
2796 | break; | |
2797 | case IORING_OFF_SQES: | |
2798 | ptr = ctx->sq_sqes; | |
2799 | break; | |
2800 | case IORING_OFF_CQ_RING: | |
2801 | ptr = ctx->cq_ring; | |
2802 | break; | |
2803 | default: | |
2804 | return -EINVAL; | |
2805 | } | |
2806 | ||
2807 | page = virt_to_head_page(ptr); | |
2808 | if (sz > (PAGE_SIZE << compound_order(page))) | |
2809 | return -EINVAL; | |
2810 | ||
2811 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
2812 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
2813 | } | |
2814 | ||
2815 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, | |
2816 | u32, min_complete, u32, flags, const sigset_t __user *, sig, | |
2817 | size_t, sigsz) | |
2818 | { | |
2819 | struct io_ring_ctx *ctx; | |
2820 | long ret = -EBADF; | |
2821 | int submitted = 0; | |
2822 | struct fd f; | |
2823 | ||
6c271ce2 | 2824 | if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP)) |
2b188cc1 JA |
2825 | return -EINVAL; |
2826 | ||
2827 | f = fdget(fd); | |
2828 | if (!f.file) | |
2829 | return -EBADF; | |
2830 | ||
2831 | ret = -EOPNOTSUPP; | |
2832 | if (f.file->f_op != &io_uring_fops) | |
2833 | goto out_fput; | |
2834 | ||
2835 | ret = -ENXIO; | |
2836 | ctx = f.file->private_data; | |
2837 | if (!percpu_ref_tryget(&ctx->refs)) | |
2838 | goto out_fput; | |
2839 | ||
6c271ce2 JA |
2840 | /* |
2841 | * For SQ polling, the thread will do all submissions and completions. | |
2842 | * Just return the requested submit count, and wake the thread if | |
2843 | * we were asked to. | |
2844 | */ | |
2845 | if (ctx->flags & IORING_SETUP_SQPOLL) { | |
2846 | if (flags & IORING_ENTER_SQ_WAKEUP) | |
2847 | wake_up(&ctx->sqo_wait); | |
2848 | submitted = to_submit; | |
2849 | goto out_ctx; | |
2850 | } | |
2851 | ||
2b188cc1 JA |
2852 | ret = 0; |
2853 | if (to_submit) { | |
2854 | to_submit = min(to_submit, ctx->sq_entries); | |
2855 | ||
2856 | mutex_lock(&ctx->uring_lock); | |
2857 | submitted = io_ring_submit(ctx, to_submit); | |
2858 | mutex_unlock(&ctx->uring_lock); | |
2b188cc1 JA |
2859 | } |
2860 | if (flags & IORING_ENTER_GETEVENTS) { | |
def596e9 JA |
2861 | unsigned nr_events = 0; |
2862 | ||
2b188cc1 JA |
2863 | min_complete = min(min_complete, ctx->cq_entries); |
2864 | ||
def596e9 JA |
2865 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
2866 | mutex_lock(&ctx->uring_lock); | |
2867 | ret = io_iopoll_check(ctx, &nr_events, min_complete); | |
2868 | mutex_unlock(&ctx->uring_lock); | |
2869 | } else { | |
2870 | ret = io_cqring_wait(ctx, min_complete, sig, sigsz); | |
2871 | } | |
2b188cc1 JA |
2872 | } |
2873 | ||
2874 | out_ctx: | |
2875 | io_ring_drop_ctx_refs(ctx, 1); | |
2876 | out_fput: | |
2877 | fdput(f); | |
2878 | return submitted ? submitted : ret; | |
2879 | } | |
2880 | ||
2881 | static const struct file_operations io_uring_fops = { | |
2882 | .release = io_uring_release, | |
2883 | .mmap = io_uring_mmap, | |
2884 | .poll = io_uring_poll, | |
2885 | .fasync = io_uring_fasync, | |
2886 | }; | |
2887 | ||
2888 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
2889 | struct io_uring_params *p) | |
2890 | { | |
2891 | struct io_sq_ring *sq_ring; | |
2892 | struct io_cq_ring *cq_ring; | |
2893 | size_t size; | |
2894 | ||
2895 | sq_ring = io_mem_alloc(struct_size(sq_ring, array, p->sq_entries)); | |
2896 | if (!sq_ring) | |
2897 | return -ENOMEM; | |
2898 | ||
2899 | ctx->sq_ring = sq_ring; | |
2900 | sq_ring->ring_mask = p->sq_entries - 1; | |
2901 | sq_ring->ring_entries = p->sq_entries; | |
2902 | ctx->sq_mask = sq_ring->ring_mask; | |
2903 | ctx->sq_entries = sq_ring->ring_entries; | |
2904 | ||
2905 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
2906 | if (size == SIZE_MAX) | |
2907 | return -EOVERFLOW; | |
2908 | ||
2909 | ctx->sq_sqes = io_mem_alloc(size); | |
52e04ef4 | 2910 | if (!ctx->sq_sqes) |
2b188cc1 | 2911 | return -ENOMEM; |
2b188cc1 JA |
2912 | |
2913 | cq_ring = io_mem_alloc(struct_size(cq_ring, cqes, p->cq_entries)); | |
52e04ef4 | 2914 | if (!cq_ring) |
2b188cc1 | 2915 | return -ENOMEM; |
2b188cc1 JA |
2916 | |
2917 | ctx->cq_ring = cq_ring; | |
2918 | cq_ring->ring_mask = p->cq_entries - 1; | |
2919 | cq_ring->ring_entries = p->cq_entries; | |
2920 | ctx->cq_mask = cq_ring->ring_mask; | |
2921 | ctx->cq_entries = cq_ring->ring_entries; | |
2922 | return 0; | |
2923 | } | |
2924 | ||
2925 | /* | |
2926 | * Allocate an anonymous fd, this is what constitutes the application | |
2927 | * visible backing of an io_uring instance. The application mmaps this | |
2928 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
2929 | * we have to tie this fd to a socket for file garbage collection purposes. | |
2930 | */ | |
2931 | static int io_uring_get_fd(struct io_ring_ctx *ctx) | |
2932 | { | |
2933 | struct file *file; | |
2934 | int ret; | |
2935 | ||
2936 | #if defined(CONFIG_UNIX) | |
2937 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, | |
2938 | &ctx->ring_sock); | |
2939 | if (ret) | |
2940 | return ret; | |
2941 | #endif | |
2942 | ||
2943 | ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
2944 | if (ret < 0) | |
2945 | goto err; | |
2946 | ||
2947 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, | |
2948 | O_RDWR | O_CLOEXEC); | |
2949 | if (IS_ERR(file)) { | |
2950 | put_unused_fd(ret); | |
2951 | ret = PTR_ERR(file); | |
2952 | goto err; | |
2953 | } | |
2954 | ||
2955 | #if defined(CONFIG_UNIX) | |
2956 | ctx->ring_sock->file = file; | |
6b06314c | 2957 | ctx->ring_sock->sk->sk_user_data = ctx; |
2b188cc1 JA |
2958 | #endif |
2959 | fd_install(ret, file); | |
2960 | return ret; | |
2961 | err: | |
2962 | #if defined(CONFIG_UNIX) | |
2963 | sock_release(ctx->ring_sock); | |
2964 | ctx->ring_sock = NULL; | |
2965 | #endif | |
2966 | return ret; | |
2967 | } | |
2968 | ||
2969 | static int io_uring_create(unsigned entries, struct io_uring_params *p) | |
2970 | { | |
2971 | struct user_struct *user = NULL; | |
2972 | struct io_ring_ctx *ctx; | |
2973 | bool account_mem; | |
2974 | int ret; | |
2975 | ||
2976 | if (!entries || entries > IORING_MAX_ENTRIES) | |
2977 | return -EINVAL; | |
2978 | ||
2979 | /* | |
2980 | * Use twice as many entries for the CQ ring. It's possible for the | |
2981 | * application to drive a higher depth than the size of the SQ ring, | |
2982 | * since the sqes are only used at submission time. This allows for | |
2983 | * some flexibility in overcommitting a bit. | |
2984 | */ | |
2985 | p->sq_entries = roundup_pow_of_two(entries); | |
2986 | p->cq_entries = 2 * p->sq_entries; | |
2987 | ||
2988 | user = get_uid(current_user()); | |
2989 | account_mem = !capable(CAP_IPC_LOCK); | |
2990 | ||
2991 | if (account_mem) { | |
2992 | ret = io_account_mem(user, | |
2993 | ring_pages(p->sq_entries, p->cq_entries)); | |
2994 | if (ret) { | |
2995 | free_uid(user); | |
2996 | return ret; | |
2997 | } | |
2998 | } | |
2999 | ||
3000 | ctx = io_ring_ctx_alloc(p); | |
3001 | if (!ctx) { | |
3002 | if (account_mem) | |
3003 | io_unaccount_mem(user, ring_pages(p->sq_entries, | |
3004 | p->cq_entries)); | |
3005 | free_uid(user); | |
3006 | return -ENOMEM; | |
3007 | } | |
3008 | ctx->compat = in_compat_syscall(); | |
3009 | ctx->account_mem = account_mem; | |
3010 | ctx->user = user; | |
3011 | ||
3012 | ret = io_allocate_scq_urings(ctx, p); | |
3013 | if (ret) | |
3014 | goto err; | |
3015 | ||
6c271ce2 | 3016 | ret = io_sq_offload_start(ctx, p); |
2b188cc1 JA |
3017 | if (ret) |
3018 | goto err; | |
3019 | ||
3020 | ret = io_uring_get_fd(ctx); | |
3021 | if (ret < 0) | |
3022 | goto err; | |
3023 | ||
3024 | memset(&p->sq_off, 0, sizeof(p->sq_off)); | |
3025 | p->sq_off.head = offsetof(struct io_sq_ring, r.head); | |
3026 | p->sq_off.tail = offsetof(struct io_sq_ring, r.tail); | |
3027 | p->sq_off.ring_mask = offsetof(struct io_sq_ring, ring_mask); | |
3028 | p->sq_off.ring_entries = offsetof(struct io_sq_ring, ring_entries); | |
3029 | p->sq_off.flags = offsetof(struct io_sq_ring, flags); | |
3030 | p->sq_off.dropped = offsetof(struct io_sq_ring, dropped); | |
3031 | p->sq_off.array = offsetof(struct io_sq_ring, array); | |
3032 | ||
3033 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
3034 | p->cq_off.head = offsetof(struct io_cq_ring, r.head); | |
3035 | p->cq_off.tail = offsetof(struct io_cq_ring, r.tail); | |
3036 | p->cq_off.ring_mask = offsetof(struct io_cq_ring, ring_mask); | |
3037 | p->cq_off.ring_entries = offsetof(struct io_cq_ring, ring_entries); | |
3038 | p->cq_off.overflow = offsetof(struct io_cq_ring, overflow); | |
3039 | p->cq_off.cqes = offsetof(struct io_cq_ring, cqes); | |
3040 | return ret; | |
3041 | err: | |
3042 | io_ring_ctx_wait_and_kill(ctx); | |
3043 | return ret; | |
3044 | } | |
3045 | ||
3046 | /* | |
3047 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
3048 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
3049 | * params structure passed in. | |
3050 | */ | |
3051 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
3052 | { | |
3053 | struct io_uring_params p; | |
3054 | long ret; | |
3055 | int i; | |
3056 | ||
3057 | if (copy_from_user(&p, params, sizeof(p))) | |
3058 | return -EFAULT; | |
3059 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
3060 | if (p.resv[i]) | |
3061 | return -EINVAL; | |
3062 | } | |
3063 | ||
6c271ce2 JA |
3064 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
3065 | IORING_SETUP_SQ_AFF)) | |
2b188cc1 JA |
3066 | return -EINVAL; |
3067 | ||
3068 | ret = io_uring_create(entries, &p); | |
3069 | if (ret < 0) | |
3070 | return ret; | |
3071 | ||
3072 | if (copy_to_user(params, &p, sizeof(p))) | |
3073 | return -EFAULT; | |
3074 | ||
3075 | return ret; | |
3076 | } | |
3077 | ||
3078 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
3079 | struct io_uring_params __user *, params) | |
3080 | { | |
3081 | return io_uring_setup(entries, params); | |
3082 | } | |
3083 | ||
edafccee JA |
3084 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
3085 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
3086 | __releases(ctx->uring_lock) |
3087 | __acquires(ctx->uring_lock) | |
edafccee JA |
3088 | { |
3089 | int ret; | |
3090 | ||
35fa71a0 JA |
3091 | /* |
3092 | * We're inside the ring mutex, if the ref is already dying, then | |
3093 | * someone else killed the ctx or is already going through | |
3094 | * io_uring_register(). | |
3095 | */ | |
3096 | if (percpu_ref_is_dying(&ctx->refs)) | |
3097 | return -ENXIO; | |
3098 | ||
edafccee | 3099 | percpu_ref_kill(&ctx->refs); |
b19062a5 JA |
3100 | |
3101 | /* | |
3102 | * Drop uring mutex before waiting for references to exit. If another | |
3103 | * thread is currently inside io_uring_enter() it might need to grab | |
3104 | * the uring_lock to make progress. If we hold it here across the drain | |
3105 | * wait, then we can deadlock. It's safe to drop the mutex here, since | |
3106 | * no new references will come in after we've killed the percpu ref. | |
3107 | */ | |
3108 | mutex_unlock(&ctx->uring_lock); | |
edafccee | 3109 | wait_for_completion(&ctx->ctx_done); |
b19062a5 | 3110 | mutex_lock(&ctx->uring_lock); |
edafccee JA |
3111 | |
3112 | switch (opcode) { | |
3113 | case IORING_REGISTER_BUFFERS: | |
3114 | ret = io_sqe_buffer_register(ctx, arg, nr_args); | |
3115 | break; | |
3116 | case IORING_UNREGISTER_BUFFERS: | |
3117 | ret = -EINVAL; | |
3118 | if (arg || nr_args) | |
3119 | break; | |
3120 | ret = io_sqe_buffer_unregister(ctx); | |
3121 | break; | |
6b06314c JA |
3122 | case IORING_REGISTER_FILES: |
3123 | ret = io_sqe_files_register(ctx, arg, nr_args); | |
3124 | break; | |
3125 | case IORING_UNREGISTER_FILES: | |
3126 | ret = -EINVAL; | |
3127 | if (arg || nr_args) | |
3128 | break; | |
3129 | ret = io_sqe_files_unregister(ctx); | |
3130 | break; | |
edafccee JA |
3131 | default: |
3132 | ret = -EINVAL; | |
3133 | break; | |
3134 | } | |
3135 | ||
3136 | /* bring the ctx back to life */ | |
3137 | reinit_completion(&ctx->ctx_done); | |
3138 | percpu_ref_reinit(&ctx->refs); | |
3139 | return ret; | |
3140 | } | |
3141 | ||
3142 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
3143 | void __user *, arg, unsigned int, nr_args) | |
3144 | { | |
3145 | struct io_ring_ctx *ctx; | |
3146 | long ret = -EBADF; | |
3147 | struct fd f; | |
3148 | ||
3149 | f = fdget(fd); | |
3150 | if (!f.file) | |
3151 | return -EBADF; | |
3152 | ||
3153 | ret = -EOPNOTSUPP; | |
3154 | if (f.file->f_op != &io_uring_fops) | |
3155 | goto out_fput; | |
3156 | ||
3157 | ctx = f.file->private_data; | |
3158 | ||
3159 | mutex_lock(&ctx->uring_lock); | |
3160 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
3161 | mutex_unlock(&ctx->uring_lock); | |
3162 | out_fput: | |
3163 | fdput(f); | |
3164 | return ret; | |
3165 | } | |
3166 | ||
2b188cc1 JA |
3167 | static int __init io_uring_init(void) |
3168 | { | |
3169 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC); | |
3170 | return 0; | |
3171 | }; | |
3172 | __initcall(io_uring_init); |