]>
Commit | Line | Data |
---|---|---|
2b188cc1 JA |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* | |
3 | * Shared application/kernel submission and completion ring pairs, for | |
4 | * supporting fast/efficient IO. | |
5 | * | |
6 | * A note on the read/write ordering memory barriers that are matched between | |
1e84b97b SB |
7 | * the application and kernel side. |
8 | * | |
9 | * After the application reads the CQ ring tail, it must use an | |
10 | * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses | |
11 | * before writing the tail (using smp_load_acquire to read the tail will | |
12 | * do). It also needs a smp_mb() before updating CQ head (ordering the | |
13 | * entry load(s) with the head store), pairing with an implicit barrier | |
14 | * through a control-dependency in io_get_cqring (smp_store_release to | |
15 | * store head will do). Failure to do so could lead to reading invalid | |
16 | * CQ entries. | |
17 | * | |
18 | * Likewise, the application must use an appropriate smp_wmb() before | |
19 | * writing the SQ tail (ordering SQ entry stores with the tail store), | |
20 | * which pairs with smp_load_acquire in io_get_sqring (smp_store_release | |
21 | * to store the tail will do). And it needs a barrier ordering the SQ | |
22 | * head load before writing new SQ entries (smp_load_acquire to read | |
23 | * head will do). | |
24 | * | |
25 | * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application | |
26 | * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after* | |
27 | * updating the SQ tail; a full memory barrier smp_mb() is needed | |
28 | * between. | |
2b188cc1 JA |
29 | * |
30 | * Also see the examples in the liburing library: | |
31 | * | |
32 | * git://git.kernel.dk/liburing | |
33 | * | |
34 | * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens | |
35 | * from data shared between the kernel and application. This is done both | |
36 | * for ordering purposes, but also to ensure that once a value is loaded from | |
37 | * data that the application could potentially modify, it remains stable. | |
38 | * | |
39 | * Copyright (C) 2018-2019 Jens Axboe | |
c992fe29 | 40 | * Copyright (c) 2018-2019 Christoph Hellwig |
2b188cc1 JA |
41 | */ |
42 | #include <linux/kernel.h> | |
43 | #include <linux/init.h> | |
44 | #include <linux/errno.h> | |
45 | #include <linux/syscalls.h> | |
46 | #include <linux/compat.h> | |
47 | #include <linux/refcount.h> | |
48 | #include <linux/uio.h> | |
49 | ||
50 | #include <linux/sched/signal.h> | |
51 | #include <linux/fs.h> | |
52 | #include <linux/file.h> | |
53 | #include <linux/fdtable.h> | |
54 | #include <linux/mm.h> | |
55 | #include <linux/mman.h> | |
56 | #include <linux/mmu_context.h> | |
57 | #include <linux/percpu.h> | |
58 | #include <linux/slab.h> | |
59 | #include <linux/workqueue.h> | |
6c271ce2 | 60 | #include <linux/kthread.h> |
2b188cc1 | 61 | #include <linux/blkdev.h> |
edafccee | 62 | #include <linux/bvec.h> |
2b188cc1 JA |
63 | #include <linux/net.h> |
64 | #include <net/sock.h> | |
65 | #include <net/af_unix.h> | |
6b06314c | 66 | #include <net/scm.h> |
2b188cc1 JA |
67 | #include <linux/anon_inodes.h> |
68 | #include <linux/sched/mm.h> | |
69 | #include <linux/uaccess.h> | |
70 | #include <linux/nospec.h> | |
edafccee JA |
71 | #include <linux/sizes.h> |
72 | #include <linux/hugetlb.h> | |
2b188cc1 JA |
73 | |
74 | #include <uapi/linux/io_uring.h> | |
75 | ||
76 | #include "internal.h" | |
77 | ||
78 | #define IORING_MAX_ENTRIES 4096 | |
6b06314c | 79 | #define IORING_MAX_FIXED_FILES 1024 |
2b188cc1 JA |
80 | |
81 | struct io_uring { | |
82 | u32 head ____cacheline_aligned_in_smp; | |
83 | u32 tail ____cacheline_aligned_in_smp; | |
84 | }; | |
85 | ||
1e84b97b SB |
86 | /* |
87 | * This data is shared with the application through the mmap at offset | |
88 | * IORING_OFF_SQ_RING. | |
89 | * | |
90 | * The offsets to the member fields are published through struct | |
91 | * io_sqring_offsets when calling io_uring_setup. | |
92 | */ | |
2b188cc1 | 93 | struct io_sq_ring { |
1e84b97b SB |
94 | /* |
95 | * Head and tail offsets into the ring; the offsets need to be | |
96 | * masked to get valid indices. | |
97 | * | |
98 | * The kernel controls head and the application controls tail. | |
99 | */ | |
2b188cc1 | 100 | struct io_uring r; |
1e84b97b SB |
101 | /* |
102 | * Bitmask to apply to head and tail offsets (constant, equals | |
103 | * ring_entries - 1) | |
104 | */ | |
2b188cc1 | 105 | u32 ring_mask; |
1e84b97b | 106 | /* Ring size (constant, power of 2) */ |
2b188cc1 | 107 | u32 ring_entries; |
1e84b97b SB |
108 | /* |
109 | * Number of invalid entries dropped by the kernel due to | |
110 | * invalid index stored in array | |
111 | * | |
112 | * Written by the kernel, shouldn't be modified by the | |
113 | * application (i.e. get number of "new events" by comparing to | |
114 | * cached value). | |
115 | * | |
116 | * After a new SQ head value was read by the application this | |
117 | * counter includes all submissions that were dropped reaching | |
118 | * the new SQ head (and possibly more). | |
119 | */ | |
2b188cc1 | 120 | u32 dropped; |
1e84b97b SB |
121 | /* |
122 | * Runtime flags | |
123 | * | |
124 | * Written by the kernel, shouldn't be modified by the | |
125 | * application. | |
126 | * | |
127 | * The application needs a full memory barrier before checking | |
128 | * for IORING_SQ_NEED_WAKEUP after updating the sq tail. | |
129 | */ | |
2b188cc1 | 130 | u32 flags; |
1e84b97b SB |
131 | /* |
132 | * Ring buffer of indices into array of io_uring_sqe, which is | |
133 | * mmapped by the application using the IORING_OFF_SQES offset. | |
134 | * | |
135 | * This indirection could e.g. be used to assign fixed | |
136 | * io_uring_sqe entries to operations and only submit them to | |
137 | * the queue when needed. | |
138 | * | |
139 | * The kernel modifies neither the indices array nor the entries | |
140 | * array. | |
141 | */ | |
2b188cc1 JA |
142 | u32 array[]; |
143 | }; | |
144 | ||
1e84b97b SB |
145 | /* |
146 | * This data is shared with the application through the mmap at offset | |
147 | * IORING_OFF_CQ_RING. | |
148 | * | |
149 | * The offsets to the member fields are published through struct | |
150 | * io_cqring_offsets when calling io_uring_setup. | |
151 | */ | |
2b188cc1 | 152 | struct io_cq_ring { |
1e84b97b SB |
153 | /* |
154 | * Head and tail offsets into the ring; the offsets need to be | |
155 | * masked to get valid indices. | |
156 | * | |
157 | * The application controls head and the kernel tail. | |
158 | */ | |
2b188cc1 | 159 | struct io_uring r; |
1e84b97b SB |
160 | /* |
161 | * Bitmask to apply to head and tail offsets (constant, equals | |
162 | * ring_entries - 1) | |
163 | */ | |
2b188cc1 | 164 | u32 ring_mask; |
1e84b97b | 165 | /* Ring size (constant, power of 2) */ |
2b188cc1 | 166 | u32 ring_entries; |
1e84b97b SB |
167 | /* |
168 | * Number of completion events lost because the queue was full; | |
169 | * this should be avoided by the application by making sure | |
170 | * there are not more requests pending thatn there is space in | |
171 | * the completion queue. | |
172 | * | |
173 | * Written by the kernel, shouldn't be modified by the | |
174 | * application (i.e. get number of "new events" by comparing to | |
175 | * cached value). | |
176 | * | |
177 | * As completion events come in out of order this counter is not | |
178 | * ordered with any other data. | |
179 | */ | |
2b188cc1 | 180 | u32 overflow; |
1e84b97b SB |
181 | /* |
182 | * Ring buffer of completion events. | |
183 | * | |
184 | * The kernel writes completion events fresh every time they are | |
185 | * produced, so the application is allowed to modify pending | |
186 | * entries. | |
187 | */ | |
2b188cc1 JA |
188 | struct io_uring_cqe cqes[]; |
189 | }; | |
190 | ||
edafccee JA |
191 | struct io_mapped_ubuf { |
192 | u64 ubuf; | |
193 | size_t len; | |
194 | struct bio_vec *bvec; | |
195 | unsigned int nr_bvecs; | |
196 | }; | |
197 | ||
31b51510 JA |
198 | struct async_list { |
199 | spinlock_t lock; | |
200 | atomic_t cnt; | |
201 | struct list_head list; | |
202 | ||
203 | struct file *file; | |
204 | off_t io_end; | |
9310a7ba | 205 | size_t io_len; |
31b51510 JA |
206 | }; |
207 | ||
2b188cc1 JA |
208 | struct io_ring_ctx { |
209 | struct { | |
210 | struct percpu_ref refs; | |
211 | } ____cacheline_aligned_in_smp; | |
212 | ||
213 | struct { | |
214 | unsigned int flags; | |
215 | bool compat; | |
216 | bool account_mem; | |
217 | ||
218 | /* SQ ring */ | |
219 | struct io_sq_ring *sq_ring; | |
220 | unsigned cached_sq_head; | |
221 | unsigned sq_entries; | |
222 | unsigned sq_mask; | |
6c271ce2 | 223 | unsigned sq_thread_idle; |
2b188cc1 | 224 | struct io_uring_sqe *sq_sqes; |
de0617e4 JA |
225 | |
226 | struct list_head defer_list; | |
2b188cc1 JA |
227 | } ____cacheline_aligned_in_smp; |
228 | ||
229 | /* IO offload */ | |
230 | struct workqueue_struct *sqo_wq; | |
6c271ce2 | 231 | struct task_struct *sqo_thread; /* if using sq thread polling */ |
2b188cc1 | 232 | struct mm_struct *sqo_mm; |
6c271ce2 | 233 | wait_queue_head_t sqo_wait; |
a4c0b3de | 234 | struct completion sqo_thread_started; |
2b188cc1 JA |
235 | |
236 | struct { | |
237 | /* CQ ring */ | |
238 | struct io_cq_ring *cq_ring; | |
239 | unsigned cached_cq_tail; | |
240 | unsigned cq_entries; | |
241 | unsigned cq_mask; | |
242 | struct wait_queue_head cq_wait; | |
243 | struct fasync_struct *cq_fasync; | |
9b402849 | 244 | struct eventfd_ctx *cq_ev_fd; |
2b188cc1 JA |
245 | } ____cacheline_aligned_in_smp; |
246 | ||
6b06314c JA |
247 | /* |
248 | * If used, fixed file set. Writers must ensure that ->refs is dead, | |
249 | * readers must ensure that ->refs is alive as long as the file* is | |
250 | * used. Only updated through io_uring_register(2). | |
251 | */ | |
252 | struct file **user_files; | |
253 | unsigned nr_user_files; | |
254 | ||
edafccee JA |
255 | /* if used, fixed mapped user buffers */ |
256 | unsigned nr_user_bufs; | |
257 | struct io_mapped_ubuf *user_bufs; | |
258 | ||
2b188cc1 JA |
259 | struct user_struct *user; |
260 | ||
261 | struct completion ctx_done; | |
262 | ||
263 | struct { | |
264 | struct mutex uring_lock; | |
265 | wait_queue_head_t wait; | |
266 | } ____cacheline_aligned_in_smp; | |
267 | ||
268 | struct { | |
269 | spinlock_t completion_lock; | |
def596e9 JA |
270 | bool poll_multi_file; |
271 | /* | |
272 | * ->poll_list is protected by the ctx->uring_lock for | |
273 | * io_uring instances that don't use IORING_SETUP_SQPOLL. | |
274 | * For SQPOLL, only the single threaded io_sq_thread() will | |
275 | * manipulate the list, hence no extra locking is needed there. | |
276 | */ | |
277 | struct list_head poll_list; | |
221c5eb2 | 278 | struct list_head cancel_list; |
2b188cc1 JA |
279 | } ____cacheline_aligned_in_smp; |
280 | ||
31b51510 JA |
281 | struct async_list pending_async[2]; |
282 | ||
2b188cc1 JA |
283 | #if defined(CONFIG_UNIX) |
284 | struct socket *ring_sock; | |
285 | #endif | |
286 | }; | |
287 | ||
288 | struct sqe_submit { | |
289 | const struct io_uring_sqe *sqe; | |
290 | unsigned short index; | |
291 | bool has_user; | |
def596e9 | 292 | bool needs_lock; |
6c271ce2 | 293 | bool needs_fixed_file; |
2b188cc1 JA |
294 | }; |
295 | ||
09bb8394 JA |
296 | /* |
297 | * First field must be the file pointer in all the | |
298 | * iocb unions! See also 'struct kiocb' in <linux/fs.h> | |
299 | */ | |
221c5eb2 JA |
300 | struct io_poll_iocb { |
301 | struct file *file; | |
302 | struct wait_queue_head *head; | |
303 | __poll_t events; | |
8c838788 | 304 | bool done; |
221c5eb2 JA |
305 | bool canceled; |
306 | struct wait_queue_entry wait; | |
307 | }; | |
308 | ||
09bb8394 JA |
309 | /* |
310 | * NOTE! Each of the iocb union members has the file pointer | |
311 | * as the first entry in their struct definition. So you can | |
312 | * access the file pointer through any of the sub-structs, | |
313 | * or directly as just 'ki_filp' in this struct. | |
314 | */ | |
2b188cc1 | 315 | struct io_kiocb { |
221c5eb2 | 316 | union { |
09bb8394 | 317 | struct file *file; |
221c5eb2 JA |
318 | struct kiocb rw; |
319 | struct io_poll_iocb poll; | |
320 | }; | |
2b188cc1 JA |
321 | |
322 | struct sqe_submit submit; | |
323 | ||
324 | struct io_ring_ctx *ctx; | |
325 | struct list_head list; | |
9e645e11 | 326 | struct list_head link_list; |
2b188cc1 | 327 | unsigned int flags; |
c16361c1 | 328 | refcount_t refs; |
8449eeda | 329 | #define REQ_F_NOWAIT 1 /* must not punt to workers */ |
def596e9 | 330 | #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */ |
6b06314c | 331 | #define REQ_F_FIXED_FILE 4 /* ctx owns file */ |
31b51510 | 332 | #define REQ_F_SEQ_PREV 8 /* sequential with previous */ |
e2033e33 SB |
333 | #define REQ_F_IO_DRAIN 16 /* drain existing IO first */ |
334 | #define REQ_F_IO_DRAINED 32 /* drain done */ | |
9e645e11 | 335 | #define REQ_F_LINK 64 /* linked sqes */ |
f7b76ac9 ZL |
336 | #define REQ_F_LINK_DONE 128 /* linked sqes done */ |
337 | #define REQ_F_FAIL_LINK 256 /* fail rest of links */ | |
2b188cc1 | 338 | u64 user_data; |
9e645e11 | 339 | u32 result; |
de0617e4 | 340 | u32 sequence; |
2b188cc1 JA |
341 | |
342 | struct work_struct work; | |
343 | }; | |
344 | ||
345 | #define IO_PLUG_THRESHOLD 2 | |
def596e9 | 346 | #define IO_IOPOLL_BATCH 8 |
2b188cc1 | 347 | |
9a56a232 JA |
348 | struct io_submit_state { |
349 | struct blk_plug plug; | |
350 | ||
2579f913 JA |
351 | /* |
352 | * io_kiocb alloc cache | |
353 | */ | |
354 | void *reqs[IO_IOPOLL_BATCH]; | |
355 | unsigned int free_reqs; | |
356 | unsigned int cur_req; | |
357 | ||
9a56a232 JA |
358 | /* |
359 | * File reference cache | |
360 | */ | |
361 | struct file *file; | |
362 | unsigned int fd; | |
363 | unsigned int has_refs; | |
364 | unsigned int used_refs; | |
365 | unsigned int ios_left; | |
366 | }; | |
367 | ||
de0617e4 JA |
368 | static void io_sq_wq_submit_work(struct work_struct *work); |
369 | ||
2b188cc1 JA |
370 | static struct kmem_cache *req_cachep; |
371 | ||
372 | static const struct file_operations io_uring_fops; | |
373 | ||
374 | struct sock *io_uring_get_socket(struct file *file) | |
375 | { | |
376 | #if defined(CONFIG_UNIX) | |
377 | if (file->f_op == &io_uring_fops) { | |
378 | struct io_ring_ctx *ctx = file->private_data; | |
379 | ||
380 | return ctx->ring_sock->sk; | |
381 | } | |
382 | #endif | |
383 | return NULL; | |
384 | } | |
385 | EXPORT_SYMBOL(io_uring_get_socket); | |
386 | ||
387 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) | |
388 | { | |
389 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
390 | ||
391 | complete(&ctx->ctx_done); | |
392 | } | |
393 | ||
394 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) | |
395 | { | |
396 | struct io_ring_ctx *ctx; | |
31b51510 | 397 | int i; |
2b188cc1 JA |
398 | |
399 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
400 | if (!ctx) | |
401 | return NULL; | |
402 | ||
21482896 RG |
403 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
404 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) { | |
2b188cc1 JA |
405 | kfree(ctx); |
406 | return NULL; | |
407 | } | |
408 | ||
409 | ctx->flags = p->flags; | |
410 | init_waitqueue_head(&ctx->cq_wait); | |
411 | init_completion(&ctx->ctx_done); | |
a4c0b3de | 412 | init_completion(&ctx->sqo_thread_started); |
2b188cc1 JA |
413 | mutex_init(&ctx->uring_lock); |
414 | init_waitqueue_head(&ctx->wait); | |
31b51510 JA |
415 | for (i = 0; i < ARRAY_SIZE(ctx->pending_async); i++) { |
416 | spin_lock_init(&ctx->pending_async[i].lock); | |
417 | INIT_LIST_HEAD(&ctx->pending_async[i].list); | |
418 | atomic_set(&ctx->pending_async[i].cnt, 0); | |
419 | } | |
2b188cc1 | 420 | spin_lock_init(&ctx->completion_lock); |
def596e9 | 421 | INIT_LIST_HEAD(&ctx->poll_list); |
221c5eb2 | 422 | INIT_LIST_HEAD(&ctx->cancel_list); |
de0617e4 | 423 | INIT_LIST_HEAD(&ctx->defer_list); |
2b188cc1 JA |
424 | return ctx; |
425 | } | |
426 | ||
de0617e4 JA |
427 | static inline bool io_sequence_defer(struct io_ring_ctx *ctx, |
428 | struct io_kiocb *req) | |
429 | { | |
430 | if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) != REQ_F_IO_DRAIN) | |
431 | return false; | |
432 | ||
dbd0f6d6 | 433 | return req->sequence != ctx->cached_cq_tail + ctx->sq_ring->dropped; |
de0617e4 JA |
434 | } |
435 | ||
436 | static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx) | |
437 | { | |
438 | struct io_kiocb *req; | |
439 | ||
440 | if (list_empty(&ctx->defer_list)) | |
441 | return NULL; | |
442 | ||
443 | req = list_first_entry(&ctx->defer_list, struct io_kiocb, list); | |
444 | if (!io_sequence_defer(ctx, req)) { | |
445 | list_del_init(&req->list); | |
446 | return req; | |
447 | } | |
448 | ||
449 | return NULL; | |
450 | } | |
451 | ||
452 | static void __io_commit_cqring(struct io_ring_ctx *ctx) | |
2b188cc1 JA |
453 | { |
454 | struct io_cq_ring *ring = ctx->cq_ring; | |
455 | ||
456 | if (ctx->cached_cq_tail != READ_ONCE(ring->r.tail)) { | |
457 | /* order cqe stores with ring update */ | |
458 | smp_store_release(&ring->r.tail, ctx->cached_cq_tail); | |
459 | ||
2b188cc1 JA |
460 | if (wq_has_sleeper(&ctx->cq_wait)) { |
461 | wake_up_interruptible(&ctx->cq_wait); | |
462 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
463 | } | |
464 | } | |
465 | } | |
466 | ||
de0617e4 JA |
467 | static void io_commit_cqring(struct io_ring_ctx *ctx) |
468 | { | |
469 | struct io_kiocb *req; | |
470 | ||
471 | __io_commit_cqring(ctx); | |
472 | ||
473 | while ((req = io_get_deferred_req(ctx)) != NULL) { | |
474 | req->flags |= REQ_F_IO_DRAINED; | |
475 | queue_work(ctx->sqo_wq, &req->work); | |
476 | } | |
477 | } | |
478 | ||
2b188cc1 JA |
479 | static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx) |
480 | { | |
481 | struct io_cq_ring *ring = ctx->cq_ring; | |
482 | unsigned tail; | |
483 | ||
484 | tail = ctx->cached_cq_tail; | |
115e12e5 SB |
485 | /* |
486 | * writes to the cq entry need to come after reading head; the | |
487 | * control dependency is enough as we're using WRITE_ONCE to | |
488 | * fill the cq entry | |
489 | */ | |
74f464e9 | 490 | if (tail - READ_ONCE(ring->r.head) == ring->ring_entries) |
2b188cc1 JA |
491 | return NULL; |
492 | ||
493 | ctx->cached_cq_tail++; | |
494 | return &ring->cqes[tail & ctx->cq_mask]; | |
495 | } | |
496 | ||
497 | static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data, | |
c71ffb67 | 498 | long res) |
2b188cc1 JA |
499 | { |
500 | struct io_uring_cqe *cqe; | |
501 | ||
502 | /* | |
503 | * If we can't get a cq entry, userspace overflowed the | |
504 | * submission (by quite a lot). Increment the overflow count in | |
505 | * the ring. | |
506 | */ | |
507 | cqe = io_get_cqring(ctx); | |
508 | if (cqe) { | |
509 | WRITE_ONCE(cqe->user_data, ki_user_data); | |
510 | WRITE_ONCE(cqe->res, res); | |
c71ffb67 | 511 | WRITE_ONCE(cqe->flags, 0); |
2b188cc1 JA |
512 | } else { |
513 | unsigned overflow = READ_ONCE(ctx->cq_ring->overflow); | |
514 | ||
515 | WRITE_ONCE(ctx->cq_ring->overflow, overflow + 1); | |
516 | } | |
517 | } | |
518 | ||
8c838788 JA |
519 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
520 | { | |
521 | if (waitqueue_active(&ctx->wait)) | |
522 | wake_up(&ctx->wait); | |
523 | if (waitqueue_active(&ctx->sqo_wait)) | |
524 | wake_up(&ctx->sqo_wait); | |
9b402849 JA |
525 | if (ctx->cq_ev_fd) |
526 | eventfd_signal(ctx->cq_ev_fd, 1); | |
8c838788 JA |
527 | } |
528 | ||
529 | static void io_cqring_add_event(struct io_ring_ctx *ctx, u64 user_data, | |
c71ffb67 | 530 | long res) |
2b188cc1 JA |
531 | { |
532 | unsigned long flags; | |
533 | ||
534 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
c71ffb67 | 535 | io_cqring_fill_event(ctx, user_data, res); |
2b188cc1 JA |
536 | io_commit_cqring(ctx); |
537 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
538 | ||
8c838788 | 539 | io_cqring_ev_posted(ctx); |
2b188cc1 JA |
540 | } |
541 | ||
542 | static void io_ring_drop_ctx_refs(struct io_ring_ctx *ctx, unsigned refs) | |
543 | { | |
544 | percpu_ref_put_many(&ctx->refs, refs); | |
545 | ||
546 | if (waitqueue_active(&ctx->wait)) | |
547 | wake_up(&ctx->wait); | |
548 | } | |
549 | ||
2579f913 JA |
550 | static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx, |
551 | struct io_submit_state *state) | |
2b188cc1 | 552 | { |
fd6fab2c | 553 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
2b188cc1 JA |
554 | struct io_kiocb *req; |
555 | ||
556 | if (!percpu_ref_tryget(&ctx->refs)) | |
557 | return NULL; | |
558 | ||
2579f913 | 559 | if (!state) { |
fd6fab2c | 560 | req = kmem_cache_alloc(req_cachep, gfp); |
2579f913 JA |
561 | if (unlikely(!req)) |
562 | goto out; | |
563 | } else if (!state->free_reqs) { | |
564 | size_t sz; | |
565 | int ret; | |
566 | ||
567 | sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs)); | |
fd6fab2c JA |
568 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs); |
569 | ||
570 | /* | |
571 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
572 | * retry single alloc to be on the safe side. | |
573 | */ | |
574 | if (unlikely(ret <= 0)) { | |
575 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
576 | if (!state->reqs[0]) | |
577 | goto out; | |
578 | ret = 1; | |
579 | } | |
2579f913 JA |
580 | state->free_reqs = ret - 1; |
581 | state->cur_req = 1; | |
582 | req = state->reqs[0]; | |
583 | } else { | |
584 | req = state->reqs[state->cur_req]; | |
585 | state->free_reqs--; | |
586 | state->cur_req++; | |
2b188cc1 JA |
587 | } |
588 | ||
60c112b0 | 589 | req->file = NULL; |
2579f913 JA |
590 | req->ctx = ctx; |
591 | req->flags = 0; | |
e65ef56d JA |
592 | /* one is dropped after submission, the other at completion */ |
593 | refcount_set(&req->refs, 2); | |
9e645e11 | 594 | req->result = 0; |
2579f913 JA |
595 | return req; |
596 | out: | |
2b188cc1 JA |
597 | io_ring_drop_ctx_refs(ctx, 1); |
598 | return NULL; | |
599 | } | |
600 | ||
def596e9 JA |
601 | static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr) |
602 | { | |
603 | if (*nr) { | |
604 | kmem_cache_free_bulk(req_cachep, *nr, reqs); | |
605 | io_ring_drop_ctx_refs(ctx, *nr); | |
606 | *nr = 0; | |
607 | } | |
608 | } | |
609 | ||
9e645e11 | 610 | static void __io_free_req(struct io_kiocb *req) |
2b188cc1 | 611 | { |
09bb8394 JA |
612 | if (req->file && !(req->flags & REQ_F_FIXED_FILE)) |
613 | fput(req->file); | |
e65ef56d JA |
614 | io_ring_drop_ctx_refs(req->ctx, 1); |
615 | kmem_cache_free(req_cachep, req); | |
616 | } | |
617 | ||
9e645e11 JA |
618 | static void io_req_link_next(struct io_kiocb *req) |
619 | { | |
620 | struct io_kiocb *nxt; | |
621 | ||
622 | /* | |
623 | * The list should never be empty when we are called here. But could | |
624 | * potentially happen if the chain is messed up, check to be on the | |
625 | * safe side. | |
626 | */ | |
627 | nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list); | |
628 | if (nxt) { | |
629 | list_del(&nxt->list); | |
630 | if (!list_empty(&req->link_list)) { | |
631 | INIT_LIST_HEAD(&nxt->link_list); | |
632 | list_splice(&req->link_list, &nxt->link_list); | |
633 | nxt->flags |= REQ_F_LINK; | |
634 | } | |
635 | ||
f7b76ac9 | 636 | nxt->flags |= REQ_F_LINK_DONE; |
9e645e11 JA |
637 | INIT_WORK(&nxt->work, io_sq_wq_submit_work); |
638 | queue_work(req->ctx->sqo_wq, &nxt->work); | |
639 | } | |
640 | } | |
641 | ||
642 | /* | |
643 | * Called if REQ_F_LINK is set, and we fail the head request | |
644 | */ | |
645 | static void io_fail_links(struct io_kiocb *req) | |
646 | { | |
647 | struct io_kiocb *link; | |
648 | ||
649 | while (!list_empty(&req->link_list)) { | |
650 | link = list_first_entry(&req->link_list, struct io_kiocb, list); | |
651 | list_del(&link->list); | |
652 | ||
653 | io_cqring_add_event(req->ctx, link->user_data, -ECANCELED); | |
654 | __io_free_req(link); | |
655 | } | |
656 | } | |
657 | ||
658 | static void io_free_req(struct io_kiocb *req) | |
659 | { | |
660 | /* | |
661 | * If LINK is set, we have dependent requests in this chain. If we | |
662 | * didn't fail this request, queue the first one up, moving any other | |
663 | * dependencies to the next request. In case of failure, fail the rest | |
664 | * of the chain. | |
665 | */ | |
666 | if (req->flags & REQ_F_LINK) { | |
667 | if (req->flags & REQ_F_FAIL_LINK) | |
668 | io_fail_links(req); | |
669 | else | |
670 | io_req_link_next(req); | |
671 | } | |
672 | ||
673 | __io_free_req(req); | |
674 | } | |
675 | ||
e65ef56d JA |
676 | static void io_put_req(struct io_kiocb *req) |
677 | { | |
678 | if (refcount_dec_and_test(&req->refs)) | |
679 | io_free_req(req); | |
2b188cc1 JA |
680 | } |
681 | ||
def596e9 JA |
682 | /* |
683 | * Find and free completed poll iocbs | |
684 | */ | |
685 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
686 | struct list_head *done) | |
687 | { | |
688 | void *reqs[IO_IOPOLL_BATCH]; | |
689 | struct io_kiocb *req; | |
09bb8394 | 690 | int to_free; |
def596e9 | 691 | |
09bb8394 | 692 | to_free = 0; |
def596e9 JA |
693 | while (!list_empty(done)) { |
694 | req = list_first_entry(done, struct io_kiocb, list); | |
695 | list_del(&req->list); | |
696 | ||
9e645e11 | 697 | io_cqring_fill_event(ctx, req->user_data, req->result); |
def596e9 JA |
698 | (*nr_events)++; |
699 | ||
09bb8394 JA |
700 | if (refcount_dec_and_test(&req->refs)) { |
701 | /* If we're not using fixed files, we have to pair the | |
702 | * completion part with the file put. Use regular | |
703 | * completions for those, only batch free for fixed | |
9e645e11 | 704 | * file and non-linked commands. |
09bb8394 | 705 | */ |
9e645e11 JA |
706 | if ((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) == |
707 | REQ_F_FIXED_FILE) { | |
09bb8394 JA |
708 | reqs[to_free++] = req; |
709 | if (to_free == ARRAY_SIZE(reqs)) | |
710 | io_free_req_many(ctx, reqs, &to_free); | |
6b06314c | 711 | } else { |
09bb8394 | 712 | io_free_req(req); |
6b06314c | 713 | } |
9a56a232 | 714 | } |
def596e9 | 715 | } |
def596e9 | 716 | |
09bb8394 | 717 | io_commit_cqring(ctx); |
def596e9 JA |
718 | io_free_req_many(ctx, reqs, &to_free); |
719 | } | |
720 | ||
721 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
722 | long min) | |
723 | { | |
724 | struct io_kiocb *req, *tmp; | |
725 | LIST_HEAD(done); | |
726 | bool spin; | |
727 | int ret; | |
728 | ||
729 | /* | |
730 | * Only spin for completions if we don't have multiple devices hanging | |
731 | * off our complete list, and we're under the requested amount. | |
732 | */ | |
733 | spin = !ctx->poll_multi_file && *nr_events < min; | |
734 | ||
735 | ret = 0; | |
736 | list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) { | |
737 | struct kiocb *kiocb = &req->rw; | |
738 | ||
739 | /* | |
740 | * Move completed entries to our local list. If we find a | |
741 | * request that requires polling, break out and complete | |
742 | * the done list first, if we have entries there. | |
743 | */ | |
744 | if (req->flags & REQ_F_IOPOLL_COMPLETED) { | |
745 | list_move_tail(&req->list, &done); | |
746 | continue; | |
747 | } | |
748 | if (!list_empty(&done)) | |
749 | break; | |
750 | ||
751 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
752 | if (ret < 0) | |
753 | break; | |
754 | ||
755 | if (ret && spin) | |
756 | spin = false; | |
757 | ret = 0; | |
758 | } | |
759 | ||
760 | if (!list_empty(&done)) | |
761 | io_iopoll_complete(ctx, nr_events, &done); | |
762 | ||
763 | return ret; | |
764 | } | |
765 | ||
766 | /* | |
767 | * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a | |
768 | * non-spinning poll check - we'll still enter the driver poll loop, but only | |
769 | * as a non-spinning completion check. | |
770 | */ | |
771 | static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
772 | long min) | |
773 | { | |
774 | while (!list_empty(&ctx->poll_list)) { | |
775 | int ret; | |
776 | ||
777 | ret = io_do_iopoll(ctx, nr_events, min); | |
778 | if (ret < 0) | |
779 | return ret; | |
780 | if (!min || *nr_events >= min) | |
781 | return 0; | |
782 | } | |
783 | ||
784 | return 1; | |
785 | } | |
786 | ||
787 | /* | |
788 | * We can't just wait for polled events to come to us, we have to actively | |
789 | * find and complete them. | |
790 | */ | |
791 | static void io_iopoll_reap_events(struct io_ring_ctx *ctx) | |
792 | { | |
793 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
794 | return; | |
795 | ||
796 | mutex_lock(&ctx->uring_lock); | |
797 | while (!list_empty(&ctx->poll_list)) { | |
798 | unsigned int nr_events = 0; | |
799 | ||
800 | io_iopoll_getevents(ctx, &nr_events, 1); | |
801 | } | |
802 | mutex_unlock(&ctx->uring_lock); | |
803 | } | |
804 | ||
805 | static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, | |
806 | long min) | |
807 | { | |
808 | int ret = 0; | |
809 | ||
810 | do { | |
811 | int tmin = 0; | |
812 | ||
813 | if (*nr_events < min) | |
814 | tmin = min - *nr_events; | |
815 | ||
816 | ret = io_iopoll_getevents(ctx, nr_events, tmin); | |
817 | if (ret <= 0) | |
818 | break; | |
819 | ret = 0; | |
820 | } while (min && !*nr_events && !need_resched()); | |
821 | ||
822 | return ret; | |
823 | } | |
824 | ||
2b188cc1 JA |
825 | static void kiocb_end_write(struct kiocb *kiocb) |
826 | { | |
827 | if (kiocb->ki_flags & IOCB_WRITE) { | |
828 | struct inode *inode = file_inode(kiocb->ki_filp); | |
829 | ||
830 | /* | |
831 | * Tell lockdep we inherited freeze protection from submission | |
832 | * thread. | |
833 | */ | |
834 | if (S_ISREG(inode->i_mode)) | |
835 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); | |
836 | file_end_write(kiocb->ki_filp); | |
837 | } | |
838 | } | |
839 | ||
840 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
841 | { | |
842 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
843 | ||
844 | kiocb_end_write(kiocb); | |
845 | ||
9e645e11 JA |
846 | if ((req->flags & REQ_F_LINK) && res != req->result) |
847 | req->flags |= REQ_F_FAIL_LINK; | |
c71ffb67 | 848 | io_cqring_add_event(req->ctx, req->user_data, res); |
e65ef56d | 849 | io_put_req(req); |
2b188cc1 JA |
850 | } |
851 | ||
def596e9 JA |
852 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
853 | { | |
854 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
855 | ||
856 | kiocb_end_write(kiocb); | |
857 | ||
9e645e11 JA |
858 | if ((req->flags & REQ_F_LINK) && res != req->result) |
859 | req->flags |= REQ_F_FAIL_LINK; | |
860 | req->result = res; | |
def596e9 JA |
861 | if (res != -EAGAIN) |
862 | req->flags |= REQ_F_IOPOLL_COMPLETED; | |
863 | } | |
864 | ||
865 | /* | |
866 | * After the iocb has been issued, it's safe to be found on the poll list. | |
867 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
868 | * find it from a io_iopoll_getevents() thread before the issuer is done | |
869 | * accessing the kiocb cookie. | |
870 | */ | |
871 | static void io_iopoll_req_issued(struct io_kiocb *req) | |
872 | { | |
873 | struct io_ring_ctx *ctx = req->ctx; | |
874 | ||
875 | /* | |
876 | * Track whether we have multiple files in our lists. This will impact | |
877 | * how we do polling eventually, not spinning if we're on potentially | |
878 | * different devices. | |
879 | */ | |
880 | if (list_empty(&ctx->poll_list)) { | |
881 | ctx->poll_multi_file = false; | |
882 | } else if (!ctx->poll_multi_file) { | |
883 | struct io_kiocb *list_req; | |
884 | ||
885 | list_req = list_first_entry(&ctx->poll_list, struct io_kiocb, | |
886 | list); | |
887 | if (list_req->rw.ki_filp != req->rw.ki_filp) | |
888 | ctx->poll_multi_file = true; | |
889 | } | |
890 | ||
891 | /* | |
892 | * For fast devices, IO may have already completed. If it has, add | |
893 | * it to the front so we find it first. | |
894 | */ | |
895 | if (req->flags & REQ_F_IOPOLL_COMPLETED) | |
896 | list_add(&req->list, &ctx->poll_list); | |
897 | else | |
898 | list_add_tail(&req->list, &ctx->poll_list); | |
899 | } | |
900 | ||
3d6770fb | 901 | static void io_file_put(struct io_submit_state *state) |
9a56a232 | 902 | { |
3d6770fb | 903 | if (state->file) { |
9a56a232 JA |
904 | int diff = state->has_refs - state->used_refs; |
905 | ||
906 | if (diff) | |
907 | fput_many(state->file, diff); | |
908 | state->file = NULL; | |
909 | } | |
910 | } | |
911 | ||
912 | /* | |
913 | * Get as many references to a file as we have IOs left in this submission, | |
914 | * assuming most submissions are for one file, or at least that each file | |
915 | * has more than one submission. | |
916 | */ | |
917 | static struct file *io_file_get(struct io_submit_state *state, int fd) | |
918 | { | |
919 | if (!state) | |
920 | return fget(fd); | |
921 | ||
922 | if (state->file) { | |
923 | if (state->fd == fd) { | |
924 | state->used_refs++; | |
925 | state->ios_left--; | |
926 | return state->file; | |
927 | } | |
3d6770fb | 928 | io_file_put(state); |
9a56a232 JA |
929 | } |
930 | state->file = fget_many(fd, state->ios_left); | |
931 | if (!state->file) | |
932 | return NULL; | |
933 | ||
934 | state->fd = fd; | |
935 | state->has_refs = state->ios_left; | |
936 | state->used_refs = 1; | |
937 | state->ios_left--; | |
938 | return state->file; | |
939 | } | |
940 | ||
2b188cc1 JA |
941 | /* |
942 | * If we tracked the file through the SCM inflight mechanism, we could support | |
943 | * any file. For now, just ensure that anything potentially problematic is done | |
944 | * inline. | |
945 | */ | |
946 | static bool io_file_supports_async(struct file *file) | |
947 | { | |
948 | umode_t mode = file_inode(file)->i_mode; | |
949 | ||
950 | if (S_ISBLK(mode) || S_ISCHR(mode)) | |
951 | return true; | |
952 | if (S_ISREG(mode) && file->f_op != &io_uring_fops) | |
953 | return true; | |
954 | ||
955 | return false; | |
956 | } | |
957 | ||
6c271ce2 | 958 | static int io_prep_rw(struct io_kiocb *req, const struct sqe_submit *s, |
8358e3a8 | 959 | bool force_nonblock) |
2b188cc1 | 960 | { |
6c271ce2 | 961 | const struct io_uring_sqe *sqe = s->sqe; |
def596e9 | 962 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 | 963 | struct kiocb *kiocb = &req->rw; |
09bb8394 JA |
964 | unsigned ioprio; |
965 | int ret; | |
2b188cc1 | 966 | |
09bb8394 JA |
967 | if (!req->file) |
968 | return -EBADF; | |
2b188cc1 | 969 | |
09bb8394 JA |
970 | if (force_nonblock && !io_file_supports_async(req->file)) |
971 | force_nonblock = false; | |
6b06314c | 972 | |
2b188cc1 JA |
973 | kiocb->ki_pos = READ_ONCE(sqe->off); |
974 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); | |
975 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); | |
976 | ||
977 | ioprio = READ_ONCE(sqe->ioprio); | |
978 | if (ioprio) { | |
979 | ret = ioprio_check_cap(ioprio); | |
980 | if (ret) | |
09bb8394 | 981 | return ret; |
2b188cc1 JA |
982 | |
983 | kiocb->ki_ioprio = ioprio; | |
984 | } else | |
985 | kiocb->ki_ioprio = get_current_ioprio(); | |
986 | ||
987 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
988 | if (unlikely(ret)) | |
09bb8394 | 989 | return ret; |
8449eeda SB |
990 | |
991 | /* don't allow async punt if RWF_NOWAIT was requested */ | |
992 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
993 | req->flags |= REQ_F_NOWAIT; | |
994 | ||
995 | if (force_nonblock) | |
2b188cc1 | 996 | kiocb->ki_flags |= IOCB_NOWAIT; |
8449eeda | 997 | |
def596e9 | 998 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
999 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
1000 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 1001 | return -EOPNOTSUPP; |
2b188cc1 | 1002 | |
def596e9 JA |
1003 | kiocb->ki_flags |= IOCB_HIPRI; |
1004 | kiocb->ki_complete = io_complete_rw_iopoll; | |
1005 | } else { | |
09bb8394 JA |
1006 | if (kiocb->ki_flags & IOCB_HIPRI) |
1007 | return -EINVAL; | |
def596e9 JA |
1008 | kiocb->ki_complete = io_complete_rw; |
1009 | } | |
2b188cc1 | 1010 | return 0; |
2b188cc1 JA |
1011 | } |
1012 | ||
1013 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
1014 | { | |
1015 | switch (ret) { | |
1016 | case -EIOCBQUEUED: | |
1017 | break; | |
1018 | case -ERESTARTSYS: | |
1019 | case -ERESTARTNOINTR: | |
1020 | case -ERESTARTNOHAND: | |
1021 | case -ERESTART_RESTARTBLOCK: | |
1022 | /* | |
1023 | * We can't just restart the syscall, since previously | |
1024 | * submitted sqes may already be in progress. Just fail this | |
1025 | * IO with EINTR. | |
1026 | */ | |
1027 | ret = -EINTR; | |
1028 | /* fall through */ | |
1029 | default: | |
1030 | kiocb->ki_complete(kiocb, ret, 0); | |
1031 | } | |
1032 | } | |
1033 | ||
edafccee JA |
1034 | static int io_import_fixed(struct io_ring_ctx *ctx, int rw, |
1035 | const struct io_uring_sqe *sqe, | |
1036 | struct iov_iter *iter) | |
1037 | { | |
1038 | size_t len = READ_ONCE(sqe->len); | |
1039 | struct io_mapped_ubuf *imu; | |
1040 | unsigned index, buf_index; | |
1041 | size_t offset; | |
1042 | u64 buf_addr; | |
1043 | ||
1044 | /* attempt to use fixed buffers without having provided iovecs */ | |
1045 | if (unlikely(!ctx->user_bufs)) | |
1046 | return -EFAULT; | |
1047 | ||
1048 | buf_index = READ_ONCE(sqe->buf_index); | |
1049 | if (unlikely(buf_index >= ctx->nr_user_bufs)) | |
1050 | return -EFAULT; | |
1051 | ||
1052 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); | |
1053 | imu = &ctx->user_bufs[index]; | |
1054 | buf_addr = READ_ONCE(sqe->addr); | |
1055 | ||
1056 | /* overflow */ | |
1057 | if (buf_addr + len < buf_addr) | |
1058 | return -EFAULT; | |
1059 | /* not inside the mapped region */ | |
1060 | if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len) | |
1061 | return -EFAULT; | |
1062 | ||
1063 | /* | |
1064 | * May not be a start of buffer, set size appropriately | |
1065 | * and advance us to the beginning. | |
1066 | */ | |
1067 | offset = buf_addr - imu->ubuf; | |
1068 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
bd11b3a3 JA |
1069 | |
1070 | if (offset) { | |
1071 | /* | |
1072 | * Don't use iov_iter_advance() here, as it's really slow for | |
1073 | * using the latter parts of a big fixed buffer - it iterates | |
1074 | * over each segment manually. We can cheat a bit here, because | |
1075 | * we know that: | |
1076 | * | |
1077 | * 1) it's a BVEC iter, we set it up | |
1078 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
1079 | * first and last bvec | |
1080 | * | |
1081 | * So just find our index, and adjust the iterator afterwards. | |
1082 | * If the offset is within the first bvec (or the whole first | |
1083 | * bvec, just use iov_iter_advance(). This makes it easier | |
1084 | * since we can just skip the first segment, which may not | |
1085 | * be PAGE_SIZE aligned. | |
1086 | */ | |
1087 | const struct bio_vec *bvec = imu->bvec; | |
1088 | ||
1089 | if (offset <= bvec->bv_len) { | |
1090 | iov_iter_advance(iter, offset); | |
1091 | } else { | |
1092 | unsigned long seg_skip; | |
1093 | ||
1094 | /* skip first vec */ | |
1095 | offset -= bvec->bv_len; | |
1096 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
1097 | ||
1098 | iter->bvec = bvec + seg_skip; | |
1099 | iter->nr_segs -= seg_skip; | |
1100 | iter->count -= (seg_skip << PAGE_SHIFT); | |
1101 | iter->iov_offset = offset & ~PAGE_MASK; | |
1102 | if (iter->iov_offset) | |
1103 | iter->count -= iter->iov_offset; | |
1104 | } | |
1105 | } | |
1106 | ||
edafccee JA |
1107 | return 0; |
1108 | } | |
1109 | ||
87e5e6da JA |
1110 | static ssize_t io_import_iovec(struct io_ring_ctx *ctx, int rw, |
1111 | const struct sqe_submit *s, struct iovec **iovec, | |
1112 | struct iov_iter *iter) | |
2b188cc1 JA |
1113 | { |
1114 | const struct io_uring_sqe *sqe = s->sqe; | |
1115 | void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
1116 | size_t sqe_len = READ_ONCE(sqe->len); | |
edafccee JA |
1117 | u8 opcode; |
1118 | ||
1119 | /* | |
1120 | * We're reading ->opcode for the second time, but the first read | |
1121 | * doesn't care whether it's _FIXED or not, so it doesn't matter | |
1122 | * whether ->opcode changes concurrently. The first read does care | |
1123 | * about whether it is a READ or a WRITE, so we don't trust this read | |
1124 | * for that purpose and instead let the caller pass in the read/write | |
1125 | * flag. | |
1126 | */ | |
1127 | opcode = READ_ONCE(sqe->opcode); | |
1128 | if (opcode == IORING_OP_READ_FIXED || | |
1129 | opcode == IORING_OP_WRITE_FIXED) { | |
87e5e6da | 1130 | ssize_t ret = io_import_fixed(ctx, rw, sqe, iter); |
edafccee JA |
1131 | *iovec = NULL; |
1132 | return ret; | |
1133 | } | |
2b188cc1 JA |
1134 | |
1135 | if (!s->has_user) | |
1136 | return -EFAULT; | |
1137 | ||
1138 | #ifdef CONFIG_COMPAT | |
1139 | if (ctx->compat) | |
1140 | return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV, | |
1141 | iovec, iter); | |
1142 | #endif | |
1143 | ||
1144 | return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter); | |
1145 | } | |
1146 | ||
31b51510 JA |
1147 | /* |
1148 | * Make a note of the last file/offset/direction we punted to async | |
1149 | * context. We'll use this information to see if we can piggy back a | |
1150 | * sequential request onto the previous one, if it's still hasn't been | |
1151 | * completed by the async worker. | |
1152 | */ | |
1153 | static void io_async_list_note(int rw, struct io_kiocb *req, size_t len) | |
1154 | { | |
1155 | struct async_list *async_list = &req->ctx->pending_async[rw]; | |
1156 | struct kiocb *kiocb = &req->rw; | |
1157 | struct file *filp = kiocb->ki_filp; | |
1158 | off_t io_end = kiocb->ki_pos + len; | |
1159 | ||
1160 | if (filp == async_list->file && kiocb->ki_pos == async_list->io_end) { | |
9310a7ba | 1161 | unsigned long max_bytes; |
31b51510 JA |
1162 | |
1163 | /* Use 8x RA size as a decent limiter for both reads/writes */ | |
9310a7ba ZL |
1164 | max_bytes = filp->f_ra.ra_pages << (PAGE_SHIFT + 3); |
1165 | if (!max_bytes) | |
1166 | max_bytes = VM_READAHEAD_PAGES << (PAGE_SHIFT + 3); | |
1167 | ||
1168 | /* If max len are exceeded, reset the state */ | |
1169 | if (async_list->io_len + len <= max_bytes) { | |
31b51510 | 1170 | req->flags |= REQ_F_SEQ_PREV; |
9310a7ba | 1171 | async_list->io_len += len; |
31b51510 JA |
1172 | } else { |
1173 | io_end = 0; | |
9310a7ba | 1174 | async_list->io_len = 0; |
31b51510 JA |
1175 | } |
1176 | } | |
1177 | ||
1178 | /* New file? Reset state. */ | |
1179 | if (async_list->file != filp) { | |
9310a7ba | 1180 | async_list->io_len = 0; |
31b51510 JA |
1181 | async_list->file = filp; |
1182 | } | |
1183 | async_list->io_end = io_end; | |
1184 | } | |
1185 | ||
e0c5c576 | 1186 | static int io_read(struct io_kiocb *req, const struct sqe_submit *s, |
8358e3a8 | 1187 | bool force_nonblock) |
2b188cc1 JA |
1188 | { |
1189 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1190 | struct kiocb *kiocb = &req->rw; | |
1191 | struct iov_iter iter; | |
1192 | struct file *file; | |
31b51510 | 1193 | size_t iov_count; |
9d93a3f5 | 1194 | ssize_t read_size, ret; |
2b188cc1 | 1195 | |
8358e3a8 | 1196 | ret = io_prep_rw(req, s, force_nonblock); |
2b188cc1 JA |
1197 | if (ret) |
1198 | return ret; | |
1199 | file = kiocb->ki_filp; | |
1200 | ||
2b188cc1 | 1201 | if (unlikely(!(file->f_mode & FMODE_READ))) |
09bb8394 | 1202 | return -EBADF; |
2b188cc1 | 1203 | if (unlikely(!file->f_op->read_iter)) |
09bb8394 | 1204 | return -EINVAL; |
2b188cc1 JA |
1205 | |
1206 | ret = io_import_iovec(req->ctx, READ, s, &iovec, &iter); | |
87e5e6da | 1207 | if (ret < 0) |
09bb8394 | 1208 | return ret; |
2b188cc1 | 1209 | |
9d93a3f5 | 1210 | read_size = ret; |
9e645e11 JA |
1211 | if (req->flags & REQ_F_LINK) |
1212 | req->result = read_size; | |
1213 | ||
31b51510 JA |
1214 | iov_count = iov_iter_count(&iter); |
1215 | ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 JA |
1216 | if (!ret) { |
1217 | ssize_t ret2; | |
1218 | ||
2b188cc1 | 1219 | ret2 = call_read_iter(file, kiocb, &iter); |
9d93a3f5 JA |
1220 | /* |
1221 | * In case of a short read, punt to async. This can happen | |
1222 | * if we have data partially cached. Alternatively we can | |
1223 | * return the short read, in which case the application will | |
1224 | * need to issue another SQE and wait for it. That SQE will | |
1225 | * need async punt anyway, so it's more efficient to do it | |
1226 | * here. | |
1227 | */ | |
1228 | if (force_nonblock && ret2 > 0 && ret2 < read_size) | |
1229 | ret2 = -EAGAIN; | |
1230 | /* Catch -EAGAIN return for forced non-blocking submission */ | |
31b51510 | 1231 | if (!force_nonblock || ret2 != -EAGAIN) { |
2b188cc1 | 1232 | io_rw_done(kiocb, ret2); |
31b51510 JA |
1233 | } else { |
1234 | /* | |
1235 | * If ->needs_lock is true, we're already in async | |
1236 | * context. | |
1237 | */ | |
1238 | if (!s->needs_lock) | |
1239 | io_async_list_note(READ, req, iov_count); | |
2b188cc1 | 1240 | ret = -EAGAIN; |
31b51510 | 1241 | } |
2b188cc1 JA |
1242 | } |
1243 | kfree(iovec); | |
2b188cc1 JA |
1244 | return ret; |
1245 | } | |
1246 | ||
e0c5c576 | 1247 | static int io_write(struct io_kiocb *req, const struct sqe_submit *s, |
8358e3a8 | 1248 | bool force_nonblock) |
2b188cc1 JA |
1249 | { |
1250 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1251 | struct kiocb *kiocb = &req->rw; | |
1252 | struct iov_iter iter; | |
1253 | struct file *file; | |
31b51510 | 1254 | size_t iov_count; |
87e5e6da | 1255 | ssize_t ret; |
2b188cc1 | 1256 | |
8358e3a8 | 1257 | ret = io_prep_rw(req, s, force_nonblock); |
2b188cc1 JA |
1258 | if (ret) |
1259 | return ret; | |
2b188cc1 | 1260 | |
2b188cc1 JA |
1261 | file = kiocb->ki_filp; |
1262 | if (unlikely(!(file->f_mode & FMODE_WRITE))) | |
09bb8394 | 1263 | return -EBADF; |
2b188cc1 | 1264 | if (unlikely(!file->f_op->write_iter)) |
09bb8394 | 1265 | return -EINVAL; |
2b188cc1 JA |
1266 | |
1267 | ret = io_import_iovec(req->ctx, WRITE, s, &iovec, &iter); | |
87e5e6da | 1268 | if (ret < 0) |
09bb8394 | 1269 | return ret; |
2b188cc1 | 1270 | |
9e645e11 JA |
1271 | if (req->flags & REQ_F_LINK) |
1272 | req->result = ret; | |
1273 | ||
31b51510 JA |
1274 | iov_count = iov_iter_count(&iter); |
1275 | ||
1276 | ret = -EAGAIN; | |
1277 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT)) { | |
1278 | /* If ->needs_lock is true, we're already in async context. */ | |
1279 | if (!s->needs_lock) | |
1280 | io_async_list_note(WRITE, req, iov_count); | |
1281 | goto out_free; | |
1282 | } | |
1283 | ||
1284 | ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 | 1285 | if (!ret) { |
9bf7933f RP |
1286 | ssize_t ret2; |
1287 | ||
2b188cc1 JA |
1288 | /* |
1289 | * Open-code file_start_write here to grab freeze protection, | |
1290 | * which will be released by another thread in | |
1291 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
1292 | * released so that it doesn't complain about the held lock when | |
1293 | * we return to userspace. | |
1294 | */ | |
1295 | if (S_ISREG(file_inode(file)->i_mode)) { | |
1296 | __sb_start_write(file_inode(file)->i_sb, | |
1297 | SB_FREEZE_WRITE, true); | |
1298 | __sb_writers_release(file_inode(file)->i_sb, | |
1299 | SB_FREEZE_WRITE); | |
1300 | } | |
1301 | kiocb->ki_flags |= IOCB_WRITE; | |
9bf7933f RP |
1302 | |
1303 | ret2 = call_write_iter(file, kiocb, &iter); | |
1304 | if (!force_nonblock || ret2 != -EAGAIN) { | |
1305 | io_rw_done(kiocb, ret2); | |
1306 | } else { | |
1307 | /* | |
1308 | * If ->needs_lock is true, we're already in async | |
1309 | * context. | |
1310 | */ | |
1311 | if (!s->needs_lock) | |
1312 | io_async_list_note(WRITE, req, iov_count); | |
1313 | ret = -EAGAIN; | |
1314 | } | |
2b188cc1 | 1315 | } |
31b51510 | 1316 | out_free: |
2b188cc1 | 1317 | kfree(iovec); |
2b188cc1 JA |
1318 | return ret; |
1319 | } | |
1320 | ||
1321 | /* | |
1322 | * IORING_OP_NOP just posts a completion event, nothing else. | |
1323 | */ | |
1324 | static int io_nop(struct io_kiocb *req, u64 user_data) | |
1325 | { | |
1326 | struct io_ring_ctx *ctx = req->ctx; | |
1327 | long err = 0; | |
1328 | ||
def596e9 JA |
1329 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
1330 | return -EINVAL; | |
1331 | ||
c71ffb67 | 1332 | io_cqring_add_event(ctx, user_data, err); |
e65ef56d | 1333 | io_put_req(req); |
2b188cc1 JA |
1334 | return 0; |
1335 | } | |
1336 | ||
c992fe29 CH |
1337 | static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1338 | { | |
6b06314c | 1339 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 1340 | |
09bb8394 JA |
1341 | if (!req->file) |
1342 | return -EBADF; | |
c992fe29 | 1343 | |
6b06314c | 1344 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 1345 | return -EINVAL; |
edafccee | 1346 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
1347 | return -EINVAL; |
1348 | ||
c992fe29 CH |
1349 | return 0; |
1350 | } | |
1351 | ||
1352 | static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
1353 | bool force_nonblock) | |
1354 | { | |
1355 | loff_t sqe_off = READ_ONCE(sqe->off); | |
1356 | loff_t sqe_len = READ_ONCE(sqe->len); | |
1357 | loff_t end = sqe_off + sqe_len; | |
1358 | unsigned fsync_flags; | |
1359 | int ret; | |
1360 | ||
1361 | fsync_flags = READ_ONCE(sqe->fsync_flags); | |
1362 | if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC)) | |
1363 | return -EINVAL; | |
1364 | ||
1365 | ret = io_prep_fsync(req, sqe); | |
1366 | if (ret) | |
1367 | return ret; | |
1368 | ||
1369 | /* fsync always requires a blocking context */ | |
1370 | if (force_nonblock) | |
1371 | return -EAGAIN; | |
1372 | ||
1373 | ret = vfs_fsync_range(req->rw.ki_filp, sqe_off, | |
1374 | end > 0 ? end : LLONG_MAX, | |
1375 | fsync_flags & IORING_FSYNC_DATASYNC); | |
1376 | ||
9e645e11 JA |
1377 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1378 | req->flags |= REQ_F_FAIL_LINK; | |
c71ffb67 | 1379 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
e65ef56d | 1380 | io_put_req(req); |
c992fe29 CH |
1381 | return 0; |
1382 | } | |
1383 | ||
5d17b4a4 JA |
1384 | static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1385 | { | |
1386 | struct io_ring_ctx *ctx = req->ctx; | |
1387 | int ret = 0; | |
1388 | ||
1389 | if (!req->file) | |
1390 | return -EBADF; | |
5d17b4a4 JA |
1391 | |
1392 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
1393 | return -EINVAL; | |
1394 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) | |
1395 | return -EINVAL; | |
1396 | ||
5d17b4a4 JA |
1397 | return ret; |
1398 | } | |
1399 | ||
1400 | static int io_sync_file_range(struct io_kiocb *req, | |
1401 | const struct io_uring_sqe *sqe, | |
1402 | bool force_nonblock) | |
1403 | { | |
1404 | loff_t sqe_off; | |
1405 | loff_t sqe_len; | |
1406 | unsigned flags; | |
1407 | int ret; | |
1408 | ||
1409 | ret = io_prep_sfr(req, sqe); | |
1410 | if (ret) | |
1411 | return ret; | |
1412 | ||
1413 | /* sync_file_range always requires a blocking context */ | |
1414 | if (force_nonblock) | |
1415 | return -EAGAIN; | |
1416 | ||
1417 | sqe_off = READ_ONCE(sqe->off); | |
1418 | sqe_len = READ_ONCE(sqe->len); | |
1419 | flags = READ_ONCE(sqe->sync_range_flags); | |
1420 | ||
1421 | ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags); | |
1422 | ||
9e645e11 JA |
1423 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1424 | req->flags |= REQ_F_FAIL_LINK; | |
c71ffb67 | 1425 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
5d17b4a4 JA |
1426 | io_put_req(req); |
1427 | return 0; | |
1428 | } | |
1429 | ||
0fa03c62 | 1430 | #if defined(CONFIG_NET) |
aa1fa28f JA |
1431 | static int io_send_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
1432 | bool force_nonblock, | |
1433 | long (*fn)(struct socket *, struct user_msghdr __user *, | |
1434 | unsigned int)) | |
1435 | { | |
0fa03c62 JA |
1436 | struct socket *sock; |
1437 | int ret; | |
1438 | ||
1439 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1440 | return -EINVAL; | |
1441 | ||
1442 | sock = sock_from_file(req->file, &ret); | |
1443 | if (sock) { | |
1444 | struct user_msghdr __user *msg; | |
1445 | unsigned flags; | |
1446 | ||
1447 | flags = READ_ONCE(sqe->msg_flags); | |
1448 | if (flags & MSG_DONTWAIT) | |
1449 | req->flags |= REQ_F_NOWAIT; | |
1450 | else if (force_nonblock) | |
1451 | flags |= MSG_DONTWAIT; | |
1452 | ||
1453 | msg = (struct user_msghdr __user *) (unsigned long) | |
1454 | READ_ONCE(sqe->addr); | |
1455 | ||
aa1fa28f | 1456 | ret = fn(sock, msg, flags); |
0fa03c62 JA |
1457 | if (force_nonblock && ret == -EAGAIN) |
1458 | return ret; | |
1459 | } | |
1460 | ||
c71ffb67 | 1461 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
5d17b4a4 JA |
1462 | io_put_req(req); |
1463 | return 0; | |
1464 | } | |
aa1fa28f JA |
1465 | #endif |
1466 | ||
1467 | static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
1468 | bool force_nonblock) | |
1469 | { | |
1470 | #if defined(CONFIG_NET) | |
1471 | return io_send_recvmsg(req, sqe, force_nonblock, __sys_sendmsg_sock); | |
1472 | #else | |
1473 | return -EOPNOTSUPP; | |
1474 | #endif | |
1475 | } | |
1476 | ||
1477 | static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
1478 | bool force_nonblock) | |
1479 | { | |
1480 | #if defined(CONFIG_NET) | |
1481 | return io_send_recvmsg(req, sqe, force_nonblock, __sys_recvmsg_sock); | |
0fa03c62 JA |
1482 | #else |
1483 | return -EOPNOTSUPP; | |
1484 | #endif | |
1485 | } | |
5d17b4a4 | 1486 | |
221c5eb2 JA |
1487 | static void io_poll_remove_one(struct io_kiocb *req) |
1488 | { | |
1489 | struct io_poll_iocb *poll = &req->poll; | |
1490 | ||
1491 | spin_lock(&poll->head->lock); | |
1492 | WRITE_ONCE(poll->canceled, true); | |
1493 | if (!list_empty(&poll->wait.entry)) { | |
1494 | list_del_init(&poll->wait.entry); | |
1495 | queue_work(req->ctx->sqo_wq, &req->work); | |
1496 | } | |
1497 | spin_unlock(&poll->head->lock); | |
1498 | ||
1499 | list_del_init(&req->list); | |
1500 | } | |
1501 | ||
1502 | static void io_poll_remove_all(struct io_ring_ctx *ctx) | |
1503 | { | |
1504 | struct io_kiocb *req; | |
1505 | ||
1506 | spin_lock_irq(&ctx->completion_lock); | |
1507 | while (!list_empty(&ctx->cancel_list)) { | |
1508 | req = list_first_entry(&ctx->cancel_list, struct io_kiocb,list); | |
1509 | io_poll_remove_one(req); | |
1510 | } | |
1511 | spin_unlock_irq(&ctx->completion_lock); | |
1512 | } | |
1513 | ||
1514 | /* | |
1515 | * Find a running poll command that matches one specified in sqe->addr, | |
1516 | * and remove it if found. | |
1517 | */ | |
1518 | static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
1519 | { | |
1520 | struct io_ring_ctx *ctx = req->ctx; | |
1521 | struct io_kiocb *poll_req, *next; | |
1522 | int ret = -ENOENT; | |
1523 | ||
1524 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1525 | return -EINVAL; | |
1526 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index || | |
1527 | sqe->poll_events) | |
1528 | return -EINVAL; | |
1529 | ||
1530 | spin_lock_irq(&ctx->completion_lock); | |
1531 | list_for_each_entry_safe(poll_req, next, &ctx->cancel_list, list) { | |
1532 | if (READ_ONCE(sqe->addr) == poll_req->user_data) { | |
1533 | io_poll_remove_one(poll_req); | |
1534 | ret = 0; | |
1535 | break; | |
1536 | } | |
1537 | } | |
1538 | spin_unlock_irq(&ctx->completion_lock); | |
1539 | ||
c71ffb67 | 1540 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
e65ef56d | 1541 | io_put_req(req); |
221c5eb2 JA |
1542 | return 0; |
1543 | } | |
1544 | ||
8c838788 JA |
1545 | static void io_poll_complete(struct io_ring_ctx *ctx, struct io_kiocb *req, |
1546 | __poll_t mask) | |
221c5eb2 | 1547 | { |
8c838788 | 1548 | req->poll.done = true; |
c71ffb67 | 1549 | io_cqring_fill_event(ctx, req->user_data, mangle_poll(mask)); |
8c838788 | 1550 | io_commit_cqring(ctx); |
221c5eb2 JA |
1551 | } |
1552 | ||
1553 | static void io_poll_complete_work(struct work_struct *work) | |
1554 | { | |
1555 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
1556 | struct io_poll_iocb *poll = &req->poll; | |
1557 | struct poll_table_struct pt = { ._key = poll->events }; | |
1558 | struct io_ring_ctx *ctx = req->ctx; | |
1559 | __poll_t mask = 0; | |
1560 | ||
1561 | if (!READ_ONCE(poll->canceled)) | |
1562 | mask = vfs_poll(poll->file, &pt) & poll->events; | |
1563 | ||
1564 | /* | |
1565 | * Note that ->ki_cancel callers also delete iocb from active_reqs after | |
1566 | * calling ->ki_cancel. We need the ctx_lock roundtrip here to | |
1567 | * synchronize with them. In the cancellation case the list_del_init | |
1568 | * itself is not actually needed, but harmless so we keep it in to | |
1569 | * avoid further branches in the fast path. | |
1570 | */ | |
1571 | spin_lock_irq(&ctx->completion_lock); | |
1572 | if (!mask && !READ_ONCE(poll->canceled)) { | |
1573 | add_wait_queue(poll->head, &poll->wait); | |
1574 | spin_unlock_irq(&ctx->completion_lock); | |
1575 | return; | |
1576 | } | |
1577 | list_del_init(&req->list); | |
8c838788 | 1578 | io_poll_complete(ctx, req, mask); |
221c5eb2 JA |
1579 | spin_unlock_irq(&ctx->completion_lock); |
1580 | ||
8c838788 JA |
1581 | io_cqring_ev_posted(ctx); |
1582 | io_put_req(req); | |
221c5eb2 JA |
1583 | } |
1584 | ||
1585 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
1586 | void *key) | |
1587 | { | |
1588 | struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb, | |
1589 | wait); | |
1590 | struct io_kiocb *req = container_of(poll, struct io_kiocb, poll); | |
1591 | struct io_ring_ctx *ctx = req->ctx; | |
1592 | __poll_t mask = key_to_poll(key); | |
8c838788 | 1593 | unsigned long flags; |
221c5eb2 JA |
1594 | |
1595 | /* for instances that support it check for an event match first: */ | |
8c838788 JA |
1596 | if (mask && !(mask & poll->events)) |
1597 | return 0; | |
221c5eb2 | 1598 | |
8c838788 | 1599 | list_del_init(&poll->wait.entry); |
221c5eb2 | 1600 | |
8c838788 JA |
1601 | if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) { |
1602 | list_del(&req->list); | |
1603 | io_poll_complete(ctx, req, mask); | |
1604 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
221c5eb2 | 1605 | |
8c838788 JA |
1606 | io_cqring_ev_posted(ctx); |
1607 | io_put_req(req); | |
1608 | } else { | |
1609 | queue_work(ctx->sqo_wq, &req->work); | |
221c5eb2 JA |
1610 | } |
1611 | ||
221c5eb2 JA |
1612 | return 1; |
1613 | } | |
1614 | ||
1615 | struct io_poll_table { | |
1616 | struct poll_table_struct pt; | |
1617 | struct io_kiocb *req; | |
1618 | int error; | |
1619 | }; | |
1620 | ||
1621 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
1622 | struct poll_table_struct *p) | |
1623 | { | |
1624 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
1625 | ||
1626 | if (unlikely(pt->req->poll.head)) { | |
1627 | pt->error = -EINVAL; | |
1628 | return; | |
1629 | } | |
1630 | ||
1631 | pt->error = 0; | |
1632 | pt->req->poll.head = head; | |
1633 | add_wait_queue(head, &pt->req->poll.wait); | |
1634 | } | |
1635 | ||
1636 | static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
1637 | { | |
1638 | struct io_poll_iocb *poll = &req->poll; | |
1639 | struct io_ring_ctx *ctx = req->ctx; | |
1640 | struct io_poll_table ipt; | |
8c838788 | 1641 | bool cancel = false; |
221c5eb2 JA |
1642 | __poll_t mask; |
1643 | u16 events; | |
221c5eb2 JA |
1644 | |
1645 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1646 | return -EINVAL; | |
1647 | if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
1648 | return -EINVAL; | |
09bb8394 JA |
1649 | if (!poll->file) |
1650 | return -EBADF; | |
221c5eb2 JA |
1651 | |
1652 | INIT_WORK(&req->work, io_poll_complete_work); | |
1653 | events = READ_ONCE(sqe->poll_events); | |
1654 | poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP; | |
1655 | ||
221c5eb2 | 1656 | poll->head = NULL; |
8c838788 | 1657 | poll->done = false; |
221c5eb2 JA |
1658 | poll->canceled = false; |
1659 | ||
1660 | ipt.pt._qproc = io_poll_queue_proc; | |
1661 | ipt.pt._key = poll->events; | |
1662 | ipt.req = req; | |
1663 | ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */ | |
1664 | ||
1665 | /* initialized the list so that we can do list_empty checks */ | |
1666 | INIT_LIST_HEAD(&poll->wait.entry); | |
1667 | init_waitqueue_func_entry(&poll->wait, io_poll_wake); | |
1668 | ||
36703247 JA |
1669 | INIT_LIST_HEAD(&req->list); |
1670 | ||
221c5eb2 | 1671 | mask = vfs_poll(poll->file, &ipt.pt) & poll->events; |
221c5eb2 JA |
1672 | |
1673 | spin_lock_irq(&ctx->completion_lock); | |
8c838788 JA |
1674 | if (likely(poll->head)) { |
1675 | spin_lock(&poll->head->lock); | |
1676 | if (unlikely(list_empty(&poll->wait.entry))) { | |
1677 | if (ipt.error) | |
1678 | cancel = true; | |
1679 | ipt.error = 0; | |
1680 | mask = 0; | |
1681 | } | |
1682 | if (mask || ipt.error) | |
1683 | list_del_init(&poll->wait.entry); | |
1684 | else if (cancel) | |
1685 | WRITE_ONCE(poll->canceled, true); | |
1686 | else if (!poll->done) /* actually waiting for an event */ | |
1687 | list_add_tail(&req->list, &ctx->cancel_list); | |
1688 | spin_unlock(&poll->head->lock); | |
1689 | } | |
1690 | if (mask) { /* no async, we'd stolen it */ | |
221c5eb2 | 1691 | ipt.error = 0; |
8c838788 | 1692 | io_poll_complete(ctx, req, mask); |
221c5eb2 | 1693 | } |
221c5eb2 JA |
1694 | spin_unlock_irq(&ctx->completion_lock); |
1695 | ||
8c838788 JA |
1696 | if (mask) { |
1697 | io_cqring_ev_posted(ctx); | |
e65ef56d | 1698 | io_put_req(req); |
221c5eb2 | 1699 | } |
8c838788 | 1700 | return ipt.error; |
221c5eb2 JA |
1701 | } |
1702 | ||
de0617e4 JA |
1703 | static int io_req_defer(struct io_ring_ctx *ctx, struct io_kiocb *req, |
1704 | const struct io_uring_sqe *sqe) | |
1705 | { | |
1706 | struct io_uring_sqe *sqe_copy; | |
1707 | ||
1708 | if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) | |
1709 | return 0; | |
1710 | ||
1711 | sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL); | |
1712 | if (!sqe_copy) | |
1713 | return -EAGAIN; | |
1714 | ||
1715 | spin_lock_irq(&ctx->completion_lock); | |
1716 | if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) { | |
1717 | spin_unlock_irq(&ctx->completion_lock); | |
1718 | kfree(sqe_copy); | |
1719 | return 0; | |
1720 | } | |
1721 | ||
1722 | memcpy(sqe_copy, sqe, sizeof(*sqe_copy)); | |
1723 | req->submit.sqe = sqe_copy; | |
1724 | ||
1725 | INIT_WORK(&req->work, io_sq_wq_submit_work); | |
1726 | list_add_tail(&req->list, &ctx->defer_list); | |
1727 | spin_unlock_irq(&ctx->completion_lock); | |
1728 | return -EIOCBQUEUED; | |
1729 | } | |
1730 | ||
2b188cc1 | 1731 | static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
8358e3a8 | 1732 | const struct sqe_submit *s, bool force_nonblock) |
2b188cc1 | 1733 | { |
e0c5c576 | 1734 | int ret, opcode; |
2b188cc1 | 1735 | |
9e645e11 JA |
1736 | req->user_data = READ_ONCE(s->sqe->user_data); |
1737 | ||
2b188cc1 JA |
1738 | if (unlikely(s->index >= ctx->sq_entries)) |
1739 | return -EINVAL; | |
2b188cc1 JA |
1740 | |
1741 | opcode = READ_ONCE(s->sqe->opcode); | |
1742 | switch (opcode) { | |
1743 | case IORING_OP_NOP: | |
1744 | ret = io_nop(req, req->user_data); | |
1745 | break; | |
1746 | case IORING_OP_READV: | |
edafccee JA |
1747 | if (unlikely(s->sqe->buf_index)) |
1748 | return -EINVAL; | |
8358e3a8 | 1749 | ret = io_read(req, s, force_nonblock); |
2b188cc1 JA |
1750 | break; |
1751 | case IORING_OP_WRITEV: | |
edafccee JA |
1752 | if (unlikely(s->sqe->buf_index)) |
1753 | return -EINVAL; | |
8358e3a8 | 1754 | ret = io_write(req, s, force_nonblock); |
edafccee JA |
1755 | break; |
1756 | case IORING_OP_READ_FIXED: | |
8358e3a8 | 1757 | ret = io_read(req, s, force_nonblock); |
edafccee JA |
1758 | break; |
1759 | case IORING_OP_WRITE_FIXED: | |
8358e3a8 | 1760 | ret = io_write(req, s, force_nonblock); |
2b188cc1 | 1761 | break; |
c992fe29 CH |
1762 | case IORING_OP_FSYNC: |
1763 | ret = io_fsync(req, s->sqe, force_nonblock); | |
1764 | break; | |
221c5eb2 JA |
1765 | case IORING_OP_POLL_ADD: |
1766 | ret = io_poll_add(req, s->sqe); | |
1767 | break; | |
1768 | case IORING_OP_POLL_REMOVE: | |
1769 | ret = io_poll_remove(req, s->sqe); | |
1770 | break; | |
5d17b4a4 JA |
1771 | case IORING_OP_SYNC_FILE_RANGE: |
1772 | ret = io_sync_file_range(req, s->sqe, force_nonblock); | |
1773 | break; | |
0fa03c62 JA |
1774 | case IORING_OP_SENDMSG: |
1775 | ret = io_sendmsg(req, s->sqe, force_nonblock); | |
1776 | break; | |
aa1fa28f JA |
1777 | case IORING_OP_RECVMSG: |
1778 | ret = io_recvmsg(req, s->sqe, force_nonblock); | |
1779 | break; | |
2b188cc1 JA |
1780 | default: |
1781 | ret = -EINVAL; | |
1782 | break; | |
1783 | } | |
1784 | ||
def596e9 JA |
1785 | if (ret) |
1786 | return ret; | |
1787 | ||
1788 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
9e645e11 | 1789 | if (req->result == -EAGAIN) |
def596e9 JA |
1790 | return -EAGAIN; |
1791 | ||
1792 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
1793 | if (s->needs_lock) | |
1794 | mutex_lock(&ctx->uring_lock); | |
1795 | io_iopoll_req_issued(req); | |
1796 | if (s->needs_lock) | |
1797 | mutex_unlock(&ctx->uring_lock); | |
1798 | } | |
1799 | ||
1800 | return 0; | |
2b188cc1 JA |
1801 | } |
1802 | ||
31b51510 JA |
1803 | static struct async_list *io_async_list_from_sqe(struct io_ring_ctx *ctx, |
1804 | const struct io_uring_sqe *sqe) | |
1805 | { | |
1806 | switch (sqe->opcode) { | |
1807 | case IORING_OP_READV: | |
1808 | case IORING_OP_READ_FIXED: | |
1809 | return &ctx->pending_async[READ]; | |
1810 | case IORING_OP_WRITEV: | |
1811 | case IORING_OP_WRITE_FIXED: | |
1812 | return &ctx->pending_async[WRITE]; | |
1813 | default: | |
1814 | return NULL; | |
1815 | } | |
1816 | } | |
1817 | ||
edafccee JA |
1818 | static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe) |
1819 | { | |
1820 | u8 opcode = READ_ONCE(sqe->opcode); | |
1821 | ||
1822 | return !(opcode == IORING_OP_READ_FIXED || | |
1823 | opcode == IORING_OP_WRITE_FIXED); | |
1824 | } | |
1825 | ||
2b188cc1 JA |
1826 | static void io_sq_wq_submit_work(struct work_struct *work) |
1827 | { | |
1828 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
2b188cc1 | 1829 | struct io_ring_ctx *ctx = req->ctx; |
31b51510 JA |
1830 | struct mm_struct *cur_mm = NULL; |
1831 | struct async_list *async_list; | |
1832 | LIST_HEAD(req_list); | |
edafccee | 1833 | mm_segment_t old_fs; |
2b188cc1 JA |
1834 | int ret; |
1835 | ||
31b51510 JA |
1836 | async_list = io_async_list_from_sqe(ctx, req->submit.sqe); |
1837 | restart: | |
1838 | do { | |
1839 | struct sqe_submit *s = &req->submit; | |
1840 | const struct io_uring_sqe *sqe = s->sqe; | |
d0ee8791 | 1841 | unsigned int flags = req->flags; |
2b188cc1 | 1842 | |
8449eeda | 1843 | /* Ensure we clear previously set non-block flag */ |
31b51510 JA |
1844 | req->rw.ki_flags &= ~IOCB_NOWAIT; |
1845 | ||
1846 | ret = 0; | |
1847 | if (io_sqe_needs_user(sqe) && !cur_mm) { | |
1848 | if (!mmget_not_zero(ctx->sqo_mm)) { | |
1849 | ret = -EFAULT; | |
1850 | } else { | |
1851 | cur_mm = ctx->sqo_mm; | |
1852 | use_mm(cur_mm); | |
1853 | old_fs = get_fs(); | |
1854 | set_fs(USER_DS); | |
1855 | } | |
1856 | } | |
1857 | ||
1858 | if (!ret) { | |
1859 | s->has_user = cur_mm != NULL; | |
1860 | s->needs_lock = true; | |
1861 | do { | |
8358e3a8 | 1862 | ret = __io_submit_sqe(ctx, req, s, false); |
31b51510 JA |
1863 | /* |
1864 | * We can get EAGAIN for polled IO even though | |
1865 | * we're forcing a sync submission from here, | |
1866 | * since we can't wait for request slots on the | |
1867 | * block side. | |
1868 | */ | |
1869 | if (ret != -EAGAIN) | |
1870 | break; | |
1871 | cond_resched(); | |
1872 | } while (1); | |
1873 | } | |
817869d2 JA |
1874 | |
1875 | /* drop submission reference */ | |
1876 | io_put_req(req); | |
1877 | ||
31b51510 | 1878 | if (ret) { |
c71ffb67 | 1879 | io_cqring_add_event(ctx, sqe->user_data, ret); |
e65ef56d | 1880 | io_put_req(req); |
31b51510 JA |
1881 | } |
1882 | ||
1883 | /* async context always use a copy of the sqe */ | |
1884 | kfree(sqe); | |
1885 | ||
f7b76ac9 | 1886 | /* req from defer and link list needn't decrease async cnt */ |
d0ee8791 | 1887 | if (flags & (REQ_F_IO_DRAINED | REQ_F_LINK_DONE)) |
f7b76ac9 ZL |
1888 | goto out; |
1889 | ||
31b51510 JA |
1890 | if (!async_list) |
1891 | break; | |
1892 | if (!list_empty(&req_list)) { | |
1893 | req = list_first_entry(&req_list, struct io_kiocb, | |
1894 | list); | |
1895 | list_del(&req->list); | |
1896 | continue; | |
1897 | } | |
1898 | if (list_empty(&async_list->list)) | |
1899 | break; | |
1900 | ||
1901 | req = NULL; | |
1902 | spin_lock(&async_list->lock); | |
1903 | if (list_empty(&async_list->list)) { | |
1904 | spin_unlock(&async_list->lock); | |
1905 | break; | |
1906 | } | |
1907 | list_splice_init(&async_list->list, &req_list); | |
1908 | spin_unlock(&async_list->lock); | |
1909 | ||
1910 | req = list_first_entry(&req_list, struct io_kiocb, list); | |
1911 | list_del(&req->list); | |
1912 | } while (req); | |
edafccee JA |
1913 | |
1914 | /* | |
31b51510 JA |
1915 | * Rare case of racing with a submitter. If we find the count has |
1916 | * dropped to zero AND we have pending work items, then restart | |
1917 | * the processing. This is a tiny race window. | |
edafccee | 1918 | */ |
31b51510 JA |
1919 | if (async_list) { |
1920 | ret = atomic_dec_return(&async_list->cnt); | |
1921 | while (!ret && !list_empty(&async_list->list)) { | |
1922 | spin_lock(&async_list->lock); | |
1923 | atomic_inc(&async_list->cnt); | |
1924 | list_splice_init(&async_list->list, &req_list); | |
1925 | spin_unlock(&async_list->lock); | |
1926 | ||
1927 | if (!list_empty(&req_list)) { | |
1928 | req = list_first_entry(&req_list, | |
1929 | struct io_kiocb, list); | |
1930 | list_del(&req->list); | |
1931 | goto restart; | |
1932 | } | |
1933 | ret = atomic_dec_return(&async_list->cnt); | |
edafccee | 1934 | } |
edafccee | 1935 | } |
2b188cc1 | 1936 | |
f7b76ac9 | 1937 | out: |
31b51510 | 1938 | if (cur_mm) { |
edafccee | 1939 | set_fs(old_fs); |
31b51510 JA |
1940 | unuse_mm(cur_mm); |
1941 | mmput(cur_mm); | |
2b188cc1 | 1942 | } |
31b51510 | 1943 | } |
2b188cc1 | 1944 | |
31b51510 JA |
1945 | /* |
1946 | * See if we can piggy back onto previously submitted work, that is still | |
1947 | * running. We currently only allow this if the new request is sequential | |
1948 | * to the previous one we punted. | |
1949 | */ | |
1950 | static bool io_add_to_prev_work(struct async_list *list, struct io_kiocb *req) | |
1951 | { | |
1952 | bool ret = false; | |
1953 | ||
1954 | if (!list) | |
1955 | return false; | |
1956 | if (!(req->flags & REQ_F_SEQ_PREV)) | |
1957 | return false; | |
1958 | if (!atomic_read(&list->cnt)) | |
1959 | return false; | |
1960 | ||
1961 | ret = true; | |
1962 | spin_lock(&list->lock); | |
1963 | list_add_tail(&req->list, &list->list); | |
c0e48f9d ZL |
1964 | /* |
1965 | * Ensure we see a simultaneous modification from io_sq_wq_submit_work() | |
1966 | */ | |
1967 | smp_mb(); | |
31b51510 JA |
1968 | if (!atomic_read(&list->cnt)) { |
1969 | list_del_init(&req->list); | |
1970 | ret = false; | |
1971 | } | |
1972 | spin_unlock(&list->lock); | |
1973 | return ret; | |
2b188cc1 JA |
1974 | } |
1975 | ||
09bb8394 JA |
1976 | static bool io_op_needs_file(const struct io_uring_sqe *sqe) |
1977 | { | |
1978 | int op = READ_ONCE(sqe->opcode); | |
1979 | ||
1980 | switch (op) { | |
1981 | case IORING_OP_NOP: | |
1982 | case IORING_OP_POLL_REMOVE: | |
1983 | return false; | |
1984 | default: | |
1985 | return true; | |
1986 | } | |
1987 | } | |
1988 | ||
1989 | static int io_req_set_file(struct io_ring_ctx *ctx, const struct sqe_submit *s, | |
1990 | struct io_submit_state *state, struct io_kiocb *req) | |
1991 | { | |
1992 | unsigned flags; | |
1993 | int fd; | |
1994 | ||
1995 | flags = READ_ONCE(s->sqe->flags); | |
1996 | fd = READ_ONCE(s->sqe->fd); | |
1997 | ||
de0617e4 JA |
1998 | if (flags & IOSQE_IO_DRAIN) { |
1999 | req->flags |= REQ_F_IO_DRAIN; | |
2000 | req->sequence = ctx->cached_sq_head - 1; | |
2001 | } | |
2002 | ||
60c112b0 | 2003 | if (!io_op_needs_file(s->sqe)) |
09bb8394 | 2004 | return 0; |
09bb8394 JA |
2005 | |
2006 | if (flags & IOSQE_FIXED_FILE) { | |
2007 | if (unlikely(!ctx->user_files || | |
2008 | (unsigned) fd >= ctx->nr_user_files)) | |
2009 | return -EBADF; | |
2010 | req->file = ctx->user_files[fd]; | |
2011 | req->flags |= REQ_F_FIXED_FILE; | |
2012 | } else { | |
2013 | if (s->needs_fixed_file) | |
2014 | return -EBADF; | |
2015 | req->file = io_file_get(state, fd); | |
2016 | if (unlikely(!req->file)) | |
2017 | return -EBADF; | |
2018 | } | |
2019 | ||
2020 | return 0; | |
2021 | } | |
2022 | ||
9e645e11 JA |
2023 | static int io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
2024 | struct sqe_submit *s) | |
2b188cc1 | 2025 | { |
e0c5c576 | 2026 | int ret; |
2b188cc1 | 2027 | |
8358e3a8 | 2028 | ret = __io_submit_sqe(ctx, req, s, true); |
8449eeda | 2029 | if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) { |
2b188cc1 JA |
2030 | struct io_uring_sqe *sqe_copy; |
2031 | ||
2032 | sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL); | |
2033 | if (sqe_copy) { | |
31b51510 JA |
2034 | struct async_list *list; |
2035 | ||
2b188cc1 JA |
2036 | memcpy(sqe_copy, s->sqe, sizeof(*sqe_copy)); |
2037 | s->sqe = sqe_copy; | |
2038 | ||
2039 | memcpy(&req->submit, s, sizeof(*s)); | |
31b51510 JA |
2040 | list = io_async_list_from_sqe(ctx, s->sqe); |
2041 | if (!io_add_to_prev_work(list, req)) { | |
2042 | if (list) | |
2043 | atomic_inc(&list->cnt); | |
2044 | INIT_WORK(&req->work, io_sq_wq_submit_work); | |
2045 | queue_work(ctx->sqo_wq, &req->work); | |
2046 | } | |
e65ef56d JA |
2047 | |
2048 | /* | |
2049 | * Queued up for async execution, worker will release | |
9e645e11 | 2050 | * submit reference when the iocb is actually submitted. |
e65ef56d JA |
2051 | */ |
2052 | return 0; | |
2b188cc1 JA |
2053 | } |
2054 | } | |
e65ef56d JA |
2055 | |
2056 | /* drop submission reference */ | |
2057 | io_put_req(req); | |
2058 | ||
2059 | /* and drop final reference, if we failed */ | |
9e645e11 JA |
2060 | if (ret) { |
2061 | io_cqring_add_event(ctx, req->user_data, ret); | |
2062 | if (req->flags & REQ_F_LINK) | |
2063 | req->flags |= REQ_F_FAIL_LINK; | |
e65ef56d | 2064 | io_put_req(req); |
9e645e11 | 2065 | } |
2b188cc1 JA |
2066 | |
2067 | return ret; | |
2068 | } | |
2069 | ||
9e645e11 JA |
2070 | #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK) |
2071 | ||
2072 | static void io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s, | |
2073 | struct io_submit_state *state, struct io_kiocb **link) | |
2074 | { | |
2075 | struct io_uring_sqe *sqe_copy; | |
2076 | struct io_kiocb *req; | |
2077 | int ret; | |
2078 | ||
2079 | /* enforce forwards compatibility on users */ | |
2080 | if (unlikely(s->sqe->flags & ~SQE_VALID_FLAGS)) { | |
2081 | ret = -EINVAL; | |
2082 | goto err; | |
2083 | } | |
2084 | ||
2085 | req = io_get_req(ctx, state); | |
2086 | if (unlikely(!req)) { | |
2087 | ret = -EAGAIN; | |
2088 | goto err; | |
2089 | } | |
2090 | ||
2091 | ret = io_req_set_file(ctx, s, state, req); | |
2092 | if (unlikely(ret)) { | |
2093 | err_req: | |
2094 | io_free_req(req); | |
2095 | err: | |
2096 | io_cqring_add_event(ctx, s->sqe->user_data, ret); | |
2097 | return; | |
2098 | } | |
2099 | ||
2100 | ret = io_req_defer(ctx, req, s->sqe); | |
2101 | if (ret) { | |
2102 | if (ret != -EIOCBQUEUED) | |
2103 | goto err_req; | |
2104 | return; | |
2105 | } | |
2106 | ||
2107 | /* | |
2108 | * If we already have a head request, queue this one for async | |
2109 | * submittal once the head completes. If we don't have a head but | |
2110 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
2111 | * submitted sync once the chain is complete. If none of those | |
2112 | * conditions are true (normal request), then just queue it. | |
2113 | */ | |
2114 | if (*link) { | |
2115 | struct io_kiocb *prev = *link; | |
2116 | ||
2117 | sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL); | |
2118 | if (!sqe_copy) { | |
2119 | ret = -EAGAIN; | |
2120 | goto err_req; | |
2121 | } | |
2122 | ||
2123 | s->sqe = sqe_copy; | |
2124 | memcpy(&req->submit, s, sizeof(*s)); | |
2125 | list_add_tail(&req->list, &prev->link_list); | |
2126 | } else if (s->sqe->flags & IOSQE_IO_LINK) { | |
2127 | req->flags |= REQ_F_LINK; | |
2128 | ||
2129 | memcpy(&req->submit, s, sizeof(*s)); | |
2130 | INIT_LIST_HEAD(&req->link_list); | |
2131 | *link = req; | |
2132 | } else { | |
2133 | io_queue_sqe(ctx, req, s); | |
2134 | } | |
2135 | } | |
2136 | ||
9a56a232 JA |
2137 | /* |
2138 | * Batched submission is done, ensure local IO is flushed out. | |
2139 | */ | |
2140 | static void io_submit_state_end(struct io_submit_state *state) | |
2141 | { | |
2142 | blk_finish_plug(&state->plug); | |
3d6770fb | 2143 | io_file_put(state); |
2579f913 JA |
2144 | if (state->free_reqs) |
2145 | kmem_cache_free_bulk(req_cachep, state->free_reqs, | |
2146 | &state->reqs[state->cur_req]); | |
9a56a232 JA |
2147 | } |
2148 | ||
2149 | /* | |
2150 | * Start submission side cache. | |
2151 | */ | |
2152 | static void io_submit_state_start(struct io_submit_state *state, | |
2153 | struct io_ring_ctx *ctx, unsigned max_ios) | |
2154 | { | |
2155 | blk_start_plug(&state->plug); | |
2579f913 | 2156 | state->free_reqs = 0; |
9a56a232 JA |
2157 | state->file = NULL; |
2158 | state->ios_left = max_ios; | |
2159 | } | |
2160 | ||
2b188cc1 JA |
2161 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
2162 | { | |
2163 | struct io_sq_ring *ring = ctx->sq_ring; | |
2164 | ||
2165 | if (ctx->cached_sq_head != READ_ONCE(ring->r.head)) { | |
2166 | /* | |
2167 | * Ensure any loads from the SQEs are done at this point, | |
2168 | * since once we write the new head, the application could | |
2169 | * write new data to them. | |
2170 | */ | |
2171 | smp_store_release(&ring->r.head, ctx->cached_sq_head); | |
2b188cc1 JA |
2172 | } |
2173 | } | |
2174 | ||
2b188cc1 JA |
2175 | /* |
2176 | * Fetch an sqe, if one is available. Note that s->sqe will point to memory | |
2177 | * that is mapped by userspace. This means that care needs to be taken to | |
2178 | * ensure that reads are stable, as we cannot rely on userspace always | |
2179 | * being a good citizen. If members of the sqe are validated and then later | |
2180 | * used, it's important that those reads are done through READ_ONCE() to | |
2181 | * prevent a re-load down the line. | |
2182 | */ | |
2183 | static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s) | |
2184 | { | |
2185 | struct io_sq_ring *ring = ctx->sq_ring; | |
2186 | unsigned head; | |
2187 | ||
2188 | /* | |
2189 | * The cached sq head (or cq tail) serves two purposes: | |
2190 | * | |
2191 | * 1) allows us to batch the cost of updating the user visible | |
2192 | * head updates. | |
2193 | * 2) allows the kernel side to track the head on its own, even | |
2194 | * though the application is the one updating it. | |
2195 | */ | |
2196 | head = ctx->cached_sq_head; | |
e523a29c SB |
2197 | /* make sure SQ entry isn't read before tail */ |
2198 | if (head == smp_load_acquire(&ring->r.tail)) | |
2b188cc1 JA |
2199 | return false; |
2200 | ||
2201 | head = READ_ONCE(ring->array[head & ctx->sq_mask]); | |
2202 | if (head < ctx->sq_entries) { | |
2203 | s->index = head; | |
2204 | s->sqe = &ctx->sq_sqes[head]; | |
2205 | ctx->cached_sq_head++; | |
2206 | return true; | |
2207 | } | |
2208 | ||
2209 | /* drop invalid entries */ | |
2210 | ctx->cached_sq_head++; | |
2211 | ring->dropped++; | |
2b188cc1 JA |
2212 | return false; |
2213 | } | |
2214 | ||
6c271ce2 JA |
2215 | static int io_submit_sqes(struct io_ring_ctx *ctx, struct sqe_submit *sqes, |
2216 | unsigned int nr, bool has_user, bool mm_fault) | |
2217 | { | |
2218 | struct io_submit_state state, *statep = NULL; | |
9e645e11 JA |
2219 | struct io_kiocb *link = NULL; |
2220 | bool prev_was_link = false; | |
2221 | int i, submitted = 0; | |
6c271ce2 JA |
2222 | |
2223 | if (nr > IO_PLUG_THRESHOLD) { | |
2224 | io_submit_state_start(&state, ctx, nr); | |
2225 | statep = &state; | |
2226 | } | |
2227 | ||
2228 | for (i = 0; i < nr; i++) { | |
9e645e11 JA |
2229 | /* |
2230 | * If previous wasn't linked and we have a linked command, | |
2231 | * that's the end of the chain. Submit the previous link. | |
2232 | */ | |
2233 | if (!prev_was_link && link) { | |
2234 | io_queue_sqe(ctx, link, &link->submit); | |
2235 | link = NULL; | |
2236 | } | |
2237 | prev_was_link = (sqes[i].sqe->flags & IOSQE_IO_LINK) != 0; | |
2238 | ||
6c271ce2 | 2239 | if (unlikely(mm_fault)) { |
9e645e11 JA |
2240 | io_cqring_add_event(ctx, sqes[i].sqe->user_data, |
2241 | -EFAULT); | |
6c271ce2 JA |
2242 | } else { |
2243 | sqes[i].has_user = has_user; | |
2244 | sqes[i].needs_lock = true; | |
2245 | sqes[i].needs_fixed_file = true; | |
9e645e11 | 2246 | io_submit_sqe(ctx, &sqes[i], statep, &link); |
6c271ce2 | 2247 | submitted++; |
6c271ce2 | 2248 | } |
6c271ce2 JA |
2249 | } |
2250 | ||
9e645e11 JA |
2251 | if (link) |
2252 | io_queue_sqe(ctx, link, &link->submit); | |
6c271ce2 JA |
2253 | if (statep) |
2254 | io_submit_state_end(&state); | |
2255 | ||
2256 | return submitted; | |
2257 | } | |
2258 | ||
2259 | static int io_sq_thread(void *data) | |
2260 | { | |
2261 | struct sqe_submit sqes[IO_IOPOLL_BATCH]; | |
2262 | struct io_ring_ctx *ctx = data; | |
2263 | struct mm_struct *cur_mm = NULL; | |
2264 | mm_segment_t old_fs; | |
2265 | DEFINE_WAIT(wait); | |
2266 | unsigned inflight; | |
2267 | unsigned long timeout; | |
2268 | ||
a4c0b3de JL |
2269 | complete(&ctx->sqo_thread_started); |
2270 | ||
6c271ce2 JA |
2271 | old_fs = get_fs(); |
2272 | set_fs(USER_DS); | |
2273 | ||
2274 | timeout = inflight = 0; | |
2bbcd6d3 | 2275 | while (!kthread_should_park()) { |
6c271ce2 JA |
2276 | bool all_fixed, mm_fault = false; |
2277 | int i; | |
2278 | ||
2279 | if (inflight) { | |
2280 | unsigned nr_events = 0; | |
2281 | ||
2282 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
2283 | /* | |
2284 | * We disallow the app entering submit/complete | |
2285 | * with polling, but we still need to lock the | |
2286 | * ring to prevent racing with polled issue | |
2287 | * that got punted to a workqueue. | |
2288 | */ | |
2289 | mutex_lock(&ctx->uring_lock); | |
2290 | io_iopoll_check(ctx, &nr_events, 0); | |
2291 | mutex_unlock(&ctx->uring_lock); | |
2292 | } else { | |
2293 | /* | |
2294 | * Normal IO, just pretend everything completed. | |
2295 | * We don't have to poll completions for that. | |
2296 | */ | |
2297 | nr_events = inflight; | |
2298 | } | |
2299 | ||
2300 | inflight -= nr_events; | |
2301 | if (!inflight) | |
2302 | timeout = jiffies + ctx->sq_thread_idle; | |
2303 | } | |
2304 | ||
2305 | if (!io_get_sqring(ctx, &sqes[0])) { | |
2306 | /* | |
2307 | * We're polling. If we're within the defined idle | |
2308 | * period, then let us spin without work before going | |
2309 | * to sleep. | |
2310 | */ | |
2311 | if (inflight || !time_after(jiffies, timeout)) { | |
2312 | cpu_relax(); | |
2313 | continue; | |
2314 | } | |
2315 | ||
2316 | /* | |
2317 | * Drop cur_mm before scheduling, we can't hold it for | |
2318 | * long periods (or over schedule()). Do this before | |
2319 | * adding ourselves to the waitqueue, as the unuse/drop | |
2320 | * may sleep. | |
2321 | */ | |
2322 | if (cur_mm) { | |
2323 | unuse_mm(cur_mm); | |
2324 | mmput(cur_mm); | |
2325 | cur_mm = NULL; | |
2326 | } | |
2327 | ||
2328 | prepare_to_wait(&ctx->sqo_wait, &wait, | |
2329 | TASK_INTERRUPTIBLE); | |
2330 | ||
2331 | /* Tell userspace we may need a wakeup call */ | |
2332 | ctx->sq_ring->flags |= IORING_SQ_NEED_WAKEUP; | |
0d7bae69 SB |
2333 | /* make sure to read SQ tail after writing flags */ |
2334 | smp_mb(); | |
6c271ce2 JA |
2335 | |
2336 | if (!io_get_sqring(ctx, &sqes[0])) { | |
2bbcd6d3 | 2337 | if (kthread_should_park()) { |
6c271ce2 JA |
2338 | finish_wait(&ctx->sqo_wait, &wait); |
2339 | break; | |
2340 | } | |
2341 | if (signal_pending(current)) | |
2342 | flush_signals(current); | |
2343 | schedule(); | |
2344 | finish_wait(&ctx->sqo_wait, &wait); | |
2345 | ||
2346 | ctx->sq_ring->flags &= ~IORING_SQ_NEED_WAKEUP; | |
6c271ce2 JA |
2347 | continue; |
2348 | } | |
2349 | finish_wait(&ctx->sqo_wait, &wait); | |
2350 | ||
2351 | ctx->sq_ring->flags &= ~IORING_SQ_NEED_WAKEUP; | |
6c271ce2 JA |
2352 | } |
2353 | ||
2354 | i = 0; | |
2355 | all_fixed = true; | |
2356 | do { | |
2357 | if (all_fixed && io_sqe_needs_user(sqes[i].sqe)) | |
2358 | all_fixed = false; | |
2359 | ||
2360 | i++; | |
2361 | if (i == ARRAY_SIZE(sqes)) | |
2362 | break; | |
2363 | } while (io_get_sqring(ctx, &sqes[i])); | |
2364 | ||
2365 | /* Unless all new commands are FIXED regions, grab mm */ | |
2366 | if (!all_fixed && !cur_mm) { | |
2367 | mm_fault = !mmget_not_zero(ctx->sqo_mm); | |
2368 | if (!mm_fault) { | |
2369 | use_mm(ctx->sqo_mm); | |
2370 | cur_mm = ctx->sqo_mm; | |
2371 | } | |
2372 | } | |
2373 | ||
2374 | inflight += io_submit_sqes(ctx, sqes, i, cur_mm != NULL, | |
2375 | mm_fault); | |
2376 | ||
2377 | /* Commit SQ ring head once we've consumed all SQEs */ | |
2378 | io_commit_sqring(ctx); | |
2379 | } | |
2380 | ||
2381 | set_fs(old_fs); | |
2382 | if (cur_mm) { | |
2383 | unuse_mm(cur_mm); | |
2384 | mmput(cur_mm); | |
2385 | } | |
06058632 | 2386 | |
2bbcd6d3 | 2387 | kthread_parkme(); |
06058632 | 2388 | |
6c271ce2 JA |
2389 | return 0; |
2390 | } | |
2391 | ||
2b188cc1 JA |
2392 | static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit) |
2393 | { | |
9a56a232 | 2394 | struct io_submit_state state, *statep = NULL; |
9e645e11 JA |
2395 | struct io_kiocb *link = NULL; |
2396 | bool prev_was_link = false; | |
5c8b0b54 | 2397 | int i, submit = 0; |
2b188cc1 | 2398 | |
9a56a232 JA |
2399 | if (to_submit > IO_PLUG_THRESHOLD) { |
2400 | io_submit_state_start(&state, ctx, to_submit); | |
2401 | statep = &state; | |
2402 | } | |
2b188cc1 JA |
2403 | |
2404 | for (i = 0; i < to_submit; i++) { | |
2405 | struct sqe_submit s; | |
2406 | ||
2407 | if (!io_get_sqring(ctx, &s)) | |
2408 | break; | |
2409 | ||
9e645e11 JA |
2410 | /* |
2411 | * If previous wasn't linked and we have a linked command, | |
2412 | * that's the end of the chain. Submit the previous link. | |
2413 | */ | |
2414 | if (!prev_was_link && link) { | |
2415 | io_queue_sqe(ctx, link, &link->submit); | |
2416 | link = NULL; | |
2417 | } | |
2418 | prev_was_link = (s.sqe->flags & IOSQE_IO_LINK) != 0; | |
2419 | ||
2b188cc1 | 2420 | s.has_user = true; |
def596e9 | 2421 | s.needs_lock = false; |
6c271ce2 | 2422 | s.needs_fixed_file = false; |
5c8b0b54 | 2423 | submit++; |
9e645e11 | 2424 | io_submit_sqe(ctx, &s, statep, &link); |
2b188cc1 JA |
2425 | } |
2426 | io_commit_sqring(ctx); | |
2427 | ||
9e645e11 JA |
2428 | if (link) |
2429 | io_queue_sqe(ctx, link, &link->submit); | |
9a56a232 JA |
2430 | if (statep) |
2431 | io_submit_state_end(statep); | |
2b188cc1 | 2432 | |
5c8b0b54 | 2433 | return submit; |
2b188cc1 JA |
2434 | } |
2435 | ||
2436 | static unsigned io_cqring_events(struct io_cq_ring *ring) | |
2437 | { | |
dc6ce4bc JL |
2438 | /* See comment at the top of this file */ |
2439 | smp_rmb(); | |
2b188cc1 JA |
2440 | return READ_ONCE(ring->r.tail) - READ_ONCE(ring->r.head); |
2441 | } | |
2442 | ||
2443 | /* | |
2444 | * Wait until events become available, if we don't already have some. The | |
2445 | * application must reap them itself, as they reside on the shared cq ring. | |
2446 | */ | |
2447 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
2448 | const sigset_t __user *sig, size_t sigsz) | |
2449 | { | |
2450 | struct io_cq_ring *ring = ctx->cq_ring; | |
2b188cc1 JA |
2451 | int ret; |
2452 | ||
2b188cc1 JA |
2453 | if (io_cqring_events(ring) >= min_events) |
2454 | return 0; | |
2455 | ||
2456 | if (sig) { | |
9e75ad5d AB |
2457 | #ifdef CONFIG_COMPAT |
2458 | if (in_compat_syscall()) | |
2459 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 2460 | sigsz); |
9e75ad5d AB |
2461 | else |
2462 | #endif | |
b772434b | 2463 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 2464 | |
2b188cc1 JA |
2465 | if (ret) |
2466 | return ret; | |
2467 | } | |
2468 | ||
fdb288a6 | 2469 | ret = wait_event_interruptible(ctx->wait, io_cqring_events(ring) >= min_events); |
b772434b | 2470 | restore_saved_sigmask_unless(ret == -ERESTARTSYS); |
97abc889 ON |
2471 | if (ret == -ERESTARTSYS) |
2472 | ret = -EINTR; | |
2b188cc1 JA |
2473 | |
2474 | return READ_ONCE(ring->r.head) == READ_ONCE(ring->r.tail) ? ret : 0; | |
2475 | } | |
2476 | ||
6b06314c JA |
2477 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
2478 | { | |
2479 | #if defined(CONFIG_UNIX) | |
2480 | if (ctx->ring_sock) { | |
2481 | struct sock *sock = ctx->ring_sock->sk; | |
2482 | struct sk_buff *skb; | |
2483 | ||
2484 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
2485 | kfree_skb(skb); | |
2486 | } | |
2487 | #else | |
2488 | int i; | |
2489 | ||
2490 | for (i = 0; i < ctx->nr_user_files; i++) | |
2491 | fput(ctx->user_files[i]); | |
2492 | #endif | |
2493 | } | |
2494 | ||
2495 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) | |
2496 | { | |
2497 | if (!ctx->user_files) | |
2498 | return -ENXIO; | |
2499 | ||
2500 | __io_sqe_files_unregister(ctx); | |
2501 | kfree(ctx->user_files); | |
2502 | ctx->user_files = NULL; | |
2503 | ctx->nr_user_files = 0; | |
2504 | return 0; | |
2505 | } | |
2506 | ||
6c271ce2 JA |
2507 | static void io_sq_thread_stop(struct io_ring_ctx *ctx) |
2508 | { | |
2509 | if (ctx->sqo_thread) { | |
a4c0b3de | 2510 | wait_for_completion(&ctx->sqo_thread_started); |
2bbcd6d3 RP |
2511 | /* |
2512 | * The park is a bit of a work-around, without it we get | |
2513 | * warning spews on shutdown with SQPOLL set and affinity | |
2514 | * set to a single CPU. | |
2515 | */ | |
06058632 | 2516 | kthread_park(ctx->sqo_thread); |
6c271ce2 JA |
2517 | kthread_stop(ctx->sqo_thread); |
2518 | ctx->sqo_thread = NULL; | |
2519 | } | |
2520 | } | |
2521 | ||
6b06314c JA |
2522 | static void io_finish_async(struct io_ring_ctx *ctx) |
2523 | { | |
6c271ce2 JA |
2524 | io_sq_thread_stop(ctx); |
2525 | ||
6b06314c JA |
2526 | if (ctx->sqo_wq) { |
2527 | destroy_workqueue(ctx->sqo_wq); | |
2528 | ctx->sqo_wq = NULL; | |
2529 | } | |
2530 | } | |
2531 | ||
2532 | #if defined(CONFIG_UNIX) | |
2533 | static void io_destruct_skb(struct sk_buff *skb) | |
2534 | { | |
2535 | struct io_ring_ctx *ctx = skb->sk->sk_user_data; | |
2536 | ||
2537 | io_finish_async(ctx); | |
2538 | unix_destruct_scm(skb); | |
2539 | } | |
2540 | ||
2541 | /* | |
2542 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
2543 | * the io_uring can be safely unregistered on process exit, even if we have | |
2544 | * loops in the file referencing. | |
2545 | */ | |
2546 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
2547 | { | |
2548 | struct sock *sk = ctx->ring_sock->sk; | |
2549 | struct scm_fp_list *fpl; | |
2550 | struct sk_buff *skb; | |
2551 | int i; | |
2552 | ||
2553 | if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) { | |
2554 | unsigned long inflight = ctx->user->unix_inflight + nr; | |
2555 | ||
2556 | if (inflight > task_rlimit(current, RLIMIT_NOFILE)) | |
2557 | return -EMFILE; | |
2558 | } | |
2559 | ||
2560 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); | |
2561 | if (!fpl) | |
2562 | return -ENOMEM; | |
2563 | ||
2564 | skb = alloc_skb(0, GFP_KERNEL); | |
2565 | if (!skb) { | |
2566 | kfree(fpl); | |
2567 | return -ENOMEM; | |
2568 | } | |
2569 | ||
2570 | skb->sk = sk; | |
2571 | skb->destructor = io_destruct_skb; | |
2572 | ||
2573 | fpl->user = get_uid(ctx->user); | |
2574 | for (i = 0; i < nr; i++) { | |
2575 | fpl->fp[i] = get_file(ctx->user_files[i + offset]); | |
2576 | unix_inflight(fpl->user, fpl->fp[i]); | |
2577 | } | |
2578 | ||
2579 | fpl->max = fpl->count = nr; | |
2580 | UNIXCB(skb).fp = fpl; | |
2581 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); | |
2582 | skb_queue_head(&sk->sk_receive_queue, skb); | |
2583 | ||
2584 | for (i = 0; i < nr; i++) | |
2585 | fput(fpl->fp[i]); | |
2586 | ||
2587 | return 0; | |
2588 | } | |
2589 | ||
2590 | /* | |
2591 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
2592 | * causes regular reference counting to break down. We rely on the UNIX | |
2593 | * garbage collection to take care of this problem for us. | |
2594 | */ | |
2595 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
2596 | { | |
2597 | unsigned left, total; | |
2598 | int ret = 0; | |
2599 | ||
2600 | total = 0; | |
2601 | left = ctx->nr_user_files; | |
2602 | while (left) { | |
2603 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
2604 | |
2605 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
2606 | if (ret) | |
2607 | break; | |
2608 | left -= this_files; | |
2609 | total += this_files; | |
2610 | } | |
2611 | ||
2612 | if (!ret) | |
2613 | return 0; | |
2614 | ||
2615 | while (total < ctx->nr_user_files) { | |
2616 | fput(ctx->user_files[total]); | |
2617 | total++; | |
2618 | } | |
2619 | ||
2620 | return ret; | |
2621 | } | |
2622 | #else | |
2623 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
2624 | { | |
2625 | return 0; | |
2626 | } | |
2627 | #endif | |
2628 | ||
2629 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, | |
2630 | unsigned nr_args) | |
2631 | { | |
2632 | __s32 __user *fds = (__s32 __user *) arg; | |
2633 | int fd, ret = 0; | |
2634 | unsigned i; | |
2635 | ||
2636 | if (ctx->user_files) | |
2637 | return -EBUSY; | |
2638 | if (!nr_args) | |
2639 | return -EINVAL; | |
2640 | if (nr_args > IORING_MAX_FIXED_FILES) | |
2641 | return -EMFILE; | |
2642 | ||
2643 | ctx->user_files = kcalloc(nr_args, sizeof(struct file *), GFP_KERNEL); | |
2644 | if (!ctx->user_files) | |
2645 | return -ENOMEM; | |
2646 | ||
2647 | for (i = 0; i < nr_args; i++) { | |
2648 | ret = -EFAULT; | |
2649 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) | |
2650 | break; | |
2651 | ||
2652 | ctx->user_files[i] = fget(fd); | |
2653 | ||
2654 | ret = -EBADF; | |
2655 | if (!ctx->user_files[i]) | |
2656 | break; | |
2657 | /* | |
2658 | * Don't allow io_uring instances to be registered. If UNIX | |
2659 | * isn't enabled, then this causes a reference cycle and this | |
2660 | * instance can never get freed. If UNIX is enabled we'll | |
2661 | * handle it just fine, but there's still no point in allowing | |
2662 | * a ring fd as it doesn't support regular read/write anyway. | |
2663 | */ | |
2664 | if (ctx->user_files[i]->f_op == &io_uring_fops) { | |
2665 | fput(ctx->user_files[i]); | |
2666 | break; | |
2667 | } | |
2668 | ctx->nr_user_files++; | |
2669 | ret = 0; | |
2670 | } | |
2671 | ||
2672 | if (ret) { | |
2673 | for (i = 0; i < ctx->nr_user_files; i++) | |
2674 | fput(ctx->user_files[i]); | |
2675 | ||
2676 | kfree(ctx->user_files); | |
25adf50f | 2677 | ctx->user_files = NULL; |
6b06314c JA |
2678 | ctx->nr_user_files = 0; |
2679 | return ret; | |
2680 | } | |
2681 | ||
2682 | ret = io_sqe_files_scm(ctx); | |
2683 | if (ret) | |
2684 | io_sqe_files_unregister(ctx); | |
2685 | ||
2686 | return ret; | |
2687 | } | |
2688 | ||
6c271ce2 JA |
2689 | static int io_sq_offload_start(struct io_ring_ctx *ctx, |
2690 | struct io_uring_params *p) | |
2b188cc1 JA |
2691 | { |
2692 | int ret; | |
2693 | ||
6c271ce2 | 2694 | init_waitqueue_head(&ctx->sqo_wait); |
2b188cc1 JA |
2695 | mmgrab(current->mm); |
2696 | ctx->sqo_mm = current->mm; | |
2697 | ||
6c271ce2 | 2698 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3ec482d1 JA |
2699 | ret = -EPERM; |
2700 | if (!capable(CAP_SYS_ADMIN)) | |
2701 | goto err; | |
2702 | ||
917257da JA |
2703 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
2704 | if (!ctx->sq_thread_idle) | |
2705 | ctx->sq_thread_idle = HZ; | |
2706 | ||
6c271ce2 | 2707 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 2708 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 2709 | |
917257da | 2710 | ret = -EINVAL; |
44a9bd18 JA |
2711 | if (cpu >= nr_cpu_ids) |
2712 | goto err; | |
7889f44d | 2713 | if (!cpu_online(cpu)) |
917257da JA |
2714 | goto err; |
2715 | ||
6c271ce2 JA |
2716 | ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread, |
2717 | ctx, cpu, | |
2718 | "io_uring-sq"); | |
2719 | } else { | |
2720 | ctx->sqo_thread = kthread_create(io_sq_thread, ctx, | |
2721 | "io_uring-sq"); | |
2722 | } | |
2723 | if (IS_ERR(ctx->sqo_thread)) { | |
2724 | ret = PTR_ERR(ctx->sqo_thread); | |
2725 | ctx->sqo_thread = NULL; | |
2726 | goto err; | |
2727 | } | |
2728 | wake_up_process(ctx->sqo_thread); | |
2729 | } else if (p->flags & IORING_SETUP_SQ_AFF) { | |
2730 | /* Can't have SQ_AFF without SQPOLL */ | |
2731 | ret = -EINVAL; | |
2732 | goto err; | |
2733 | } | |
2734 | ||
2b188cc1 JA |
2735 | /* Do QD, or 2 * CPUS, whatever is smallest */ |
2736 | ctx->sqo_wq = alloc_workqueue("io_ring-wq", WQ_UNBOUND | WQ_FREEZABLE, | |
2737 | min(ctx->sq_entries - 1, 2 * num_online_cpus())); | |
2738 | if (!ctx->sqo_wq) { | |
2739 | ret = -ENOMEM; | |
2740 | goto err; | |
2741 | } | |
2742 | ||
2743 | return 0; | |
2744 | err: | |
6c271ce2 | 2745 | io_sq_thread_stop(ctx); |
2b188cc1 JA |
2746 | mmdrop(ctx->sqo_mm); |
2747 | ctx->sqo_mm = NULL; | |
2748 | return ret; | |
2749 | } | |
2750 | ||
2751 | static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages) | |
2752 | { | |
2753 | atomic_long_sub(nr_pages, &user->locked_vm); | |
2754 | } | |
2755 | ||
2756 | static int io_account_mem(struct user_struct *user, unsigned long nr_pages) | |
2757 | { | |
2758 | unsigned long page_limit, cur_pages, new_pages; | |
2759 | ||
2760 | /* Don't allow more pages than we can safely lock */ | |
2761 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
2762 | ||
2763 | do { | |
2764 | cur_pages = atomic_long_read(&user->locked_vm); | |
2765 | new_pages = cur_pages + nr_pages; | |
2766 | if (new_pages > page_limit) | |
2767 | return -ENOMEM; | |
2768 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
2769 | new_pages) != cur_pages); | |
2770 | ||
2771 | return 0; | |
2772 | } | |
2773 | ||
2774 | static void io_mem_free(void *ptr) | |
2775 | { | |
52e04ef4 MR |
2776 | struct page *page; |
2777 | ||
2778 | if (!ptr) | |
2779 | return; | |
2b188cc1 | 2780 | |
52e04ef4 | 2781 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
2782 | if (put_page_testzero(page)) |
2783 | free_compound_page(page); | |
2784 | } | |
2785 | ||
2786 | static void *io_mem_alloc(size_t size) | |
2787 | { | |
2788 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
2789 | __GFP_NORETRY; | |
2790 | ||
2791 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
2792 | } | |
2793 | ||
2794 | static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries) | |
2795 | { | |
2796 | struct io_sq_ring *sq_ring; | |
2797 | struct io_cq_ring *cq_ring; | |
2798 | size_t bytes; | |
2799 | ||
2800 | bytes = struct_size(sq_ring, array, sq_entries); | |
2801 | bytes += array_size(sizeof(struct io_uring_sqe), sq_entries); | |
2802 | bytes += struct_size(cq_ring, cqes, cq_entries); | |
2803 | ||
2804 | return (bytes + PAGE_SIZE - 1) / PAGE_SIZE; | |
2805 | } | |
2806 | ||
edafccee JA |
2807 | static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx) |
2808 | { | |
2809 | int i, j; | |
2810 | ||
2811 | if (!ctx->user_bufs) | |
2812 | return -ENXIO; | |
2813 | ||
2814 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
2815 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
2816 | ||
2817 | for (j = 0; j < imu->nr_bvecs; j++) | |
2818 | put_page(imu->bvec[j].bv_page); | |
2819 | ||
2820 | if (ctx->account_mem) | |
2821 | io_unaccount_mem(ctx->user, imu->nr_bvecs); | |
d4ef6475 | 2822 | kvfree(imu->bvec); |
edafccee JA |
2823 | imu->nr_bvecs = 0; |
2824 | } | |
2825 | ||
2826 | kfree(ctx->user_bufs); | |
2827 | ctx->user_bufs = NULL; | |
2828 | ctx->nr_user_bufs = 0; | |
2829 | return 0; | |
2830 | } | |
2831 | ||
2832 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
2833 | void __user *arg, unsigned index) | |
2834 | { | |
2835 | struct iovec __user *src; | |
2836 | ||
2837 | #ifdef CONFIG_COMPAT | |
2838 | if (ctx->compat) { | |
2839 | struct compat_iovec __user *ciovs; | |
2840 | struct compat_iovec ciov; | |
2841 | ||
2842 | ciovs = (struct compat_iovec __user *) arg; | |
2843 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
2844 | return -EFAULT; | |
2845 | ||
2846 | dst->iov_base = (void __user *) (unsigned long) ciov.iov_base; | |
2847 | dst->iov_len = ciov.iov_len; | |
2848 | return 0; | |
2849 | } | |
2850 | #endif | |
2851 | src = (struct iovec __user *) arg; | |
2852 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
2853 | return -EFAULT; | |
2854 | return 0; | |
2855 | } | |
2856 | ||
2857 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, | |
2858 | unsigned nr_args) | |
2859 | { | |
2860 | struct vm_area_struct **vmas = NULL; | |
2861 | struct page **pages = NULL; | |
2862 | int i, j, got_pages = 0; | |
2863 | int ret = -EINVAL; | |
2864 | ||
2865 | if (ctx->user_bufs) | |
2866 | return -EBUSY; | |
2867 | if (!nr_args || nr_args > UIO_MAXIOV) | |
2868 | return -EINVAL; | |
2869 | ||
2870 | ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf), | |
2871 | GFP_KERNEL); | |
2872 | if (!ctx->user_bufs) | |
2873 | return -ENOMEM; | |
2874 | ||
2875 | for (i = 0; i < nr_args; i++) { | |
2876 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
2877 | unsigned long off, start, end, ubuf; | |
2878 | int pret, nr_pages; | |
2879 | struct iovec iov; | |
2880 | size_t size; | |
2881 | ||
2882 | ret = io_copy_iov(ctx, &iov, arg, i); | |
2883 | if (ret) | |
a278682d | 2884 | goto err; |
edafccee JA |
2885 | |
2886 | /* | |
2887 | * Don't impose further limits on the size and buffer | |
2888 | * constraints here, we'll -EINVAL later when IO is | |
2889 | * submitted if they are wrong. | |
2890 | */ | |
2891 | ret = -EFAULT; | |
2892 | if (!iov.iov_base || !iov.iov_len) | |
2893 | goto err; | |
2894 | ||
2895 | /* arbitrary limit, but we need something */ | |
2896 | if (iov.iov_len > SZ_1G) | |
2897 | goto err; | |
2898 | ||
2899 | ubuf = (unsigned long) iov.iov_base; | |
2900 | end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
2901 | start = ubuf >> PAGE_SHIFT; | |
2902 | nr_pages = end - start; | |
2903 | ||
2904 | if (ctx->account_mem) { | |
2905 | ret = io_account_mem(ctx->user, nr_pages); | |
2906 | if (ret) | |
2907 | goto err; | |
2908 | } | |
2909 | ||
2910 | ret = 0; | |
2911 | if (!pages || nr_pages > got_pages) { | |
2912 | kfree(vmas); | |
2913 | kfree(pages); | |
d4ef6475 | 2914 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), |
edafccee | 2915 | GFP_KERNEL); |
d4ef6475 | 2916 | vmas = kvmalloc_array(nr_pages, |
edafccee JA |
2917 | sizeof(struct vm_area_struct *), |
2918 | GFP_KERNEL); | |
2919 | if (!pages || !vmas) { | |
2920 | ret = -ENOMEM; | |
2921 | if (ctx->account_mem) | |
2922 | io_unaccount_mem(ctx->user, nr_pages); | |
2923 | goto err; | |
2924 | } | |
2925 | got_pages = nr_pages; | |
2926 | } | |
2927 | ||
d4ef6475 | 2928 | imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec), |
edafccee JA |
2929 | GFP_KERNEL); |
2930 | ret = -ENOMEM; | |
2931 | if (!imu->bvec) { | |
2932 | if (ctx->account_mem) | |
2933 | io_unaccount_mem(ctx->user, nr_pages); | |
2934 | goto err; | |
2935 | } | |
2936 | ||
2937 | ret = 0; | |
2938 | down_read(¤t->mm->mmap_sem); | |
932f4a63 IW |
2939 | pret = get_user_pages(ubuf, nr_pages, |
2940 | FOLL_WRITE | FOLL_LONGTERM, | |
2941 | pages, vmas); | |
edafccee JA |
2942 | if (pret == nr_pages) { |
2943 | /* don't support file backed memory */ | |
2944 | for (j = 0; j < nr_pages; j++) { | |
2945 | struct vm_area_struct *vma = vmas[j]; | |
2946 | ||
2947 | if (vma->vm_file && | |
2948 | !is_file_hugepages(vma->vm_file)) { | |
2949 | ret = -EOPNOTSUPP; | |
2950 | break; | |
2951 | } | |
2952 | } | |
2953 | } else { | |
2954 | ret = pret < 0 ? pret : -EFAULT; | |
2955 | } | |
2956 | up_read(¤t->mm->mmap_sem); | |
2957 | if (ret) { | |
2958 | /* | |
2959 | * if we did partial map, or found file backed vmas, | |
2960 | * release any pages we did get | |
2961 | */ | |
2962 | if (pret > 0) { | |
2963 | for (j = 0; j < pret; j++) | |
2964 | put_page(pages[j]); | |
2965 | } | |
2966 | if (ctx->account_mem) | |
2967 | io_unaccount_mem(ctx->user, nr_pages); | |
d4ef6475 | 2968 | kvfree(imu->bvec); |
edafccee JA |
2969 | goto err; |
2970 | } | |
2971 | ||
2972 | off = ubuf & ~PAGE_MASK; | |
2973 | size = iov.iov_len; | |
2974 | for (j = 0; j < nr_pages; j++) { | |
2975 | size_t vec_len; | |
2976 | ||
2977 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
2978 | imu->bvec[j].bv_page = pages[j]; | |
2979 | imu->bvec[j].bv_len = vec_len; | |
2980 | imu->bvec[j].bv_offset = off; | |
2981 | off = 0; | |
2982 | size -= vec_len; | |
2983 | } | |
2984 | /* store original address for later verification */ | |
2985 | imu->ubuf = ubuf; | |
2986 | imu->len = iov.iov_len; | |
2987 | imu->nr_bvecs = nr_pages; | |
2988 | ||
2989 | ctx->nr_user_bufs++; | |
2990 | } | |
d4ef6475 MR |
2991 | kvfree(pages); |
2992 | kvfree(vmas); | |
edafccee JA |
2993 | return 0; |
2994 | err: | |
d4ef6475 MR |
2995 | kvfree(pages); |
2996 | kvfree(vmas); | |
edafccee JA |
2997 | io_sqe_buffer_unregister(ctx); |
2998 | return ret; | |
2999 | } | |
3000 | ||
9b402849 JA |
3001 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
3002 | { | |
3003 | __s32 __user *fds = arg; | |
3004 | int fd; | |
3005 | ||
3006 | if (ctx->cq_ev_fd) | |
3007 | return -EBUSY; | |
3008 | ||
3009 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
3010 | return -EFAULT; | |
3011 | ||
3012 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
3013 | if (IS_ERR(ctx->cq_ev_fd)) { | |
3014 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
3015 | ctx->cq_ev_fd = NULL; | |
3016 | return ret; | |
3017 | } | |
3018 | ||
3019 | return 0; | |
3020 | } | |
3021 | ||
3022 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
3023 | { | |
3024 | if (ctx->cq_ev_fd) { | |
3025 | eventfd_ctx_put(ctx->cq_ev_fd); | |
3026 | ctx->cq_ev_fd = NULL; | |
3027 | return 0; | |
3028 | } | |
3029 | ||
3030 | return -ENXIO; | |
3031 | } | |
3032 | ||
2b188cc1 JA |
3033 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
3034 | { | |
6b06314c | 3035 | io_finish_async(ctx); |
2b188cc1 JA |
3036 | if (ctx->sqo_mm) |
3037 | mmdrop(ctx->sqo_mm); | |
def596e9 JA |
3038 | |
3039 | io_iopoll_reap_events(ctx); | |
edafccee | 3040 | io_sqe_buffer_unregister(ctx); |
6b06314c | 3041 | io_sqe_files_unregister(ctx); |
9b402849 | 3042 | io_eventfd_unregister(ctx); |
def596e9 | 3043 | |
2b188cc1 | 3044 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
3045 | if (ctx->ring_sock) { |
3046 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 3047 | sock_release(ctx->ring_sock); |
355e8d26 | 3048 | } |
2b188cc1 JA |
3049 | #endif |
3050 | ||
3051 | io_mem_free(ctx->sq_ring); | |
3052 | io_mem_free(ctx->sq_sqes); | |
3053 | io_mem_free(ctx->cq_ring); | |
3054 | ||
3055 | percpu_ref_exit(&ctx->refs); | |
3056 | if (ctx->account_mem) | |
3057 | io_unaccount_mem(ctx->user, | |
3058 | ring_pages(ctx->sq_entries, ctx->cq_entries)); | |
3059 | free_uid(ctx->user); | |
3060 | kfree(ctx); | |
3061 | } | |
3062 | ||
3063 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
3064 | { | |
3065 | struct io_ring_ctx *ctx = file->private_data; | |
3066 | __poll_t mask = 0; | |
3067 | ||
3068 | poll_wait(file, &ctx->cq_wait, wait); | |
4f7067c3 SB |
3069 | /* |
3070 | * synchronizes with barrier from wq_has_sleeper call in | |
3071 | * io_commit_cqring | |
3072 | */ | |
2b188cc1 | 3073 | smp_rmb(); |
fb775faa SB |
3074 | if (READ_ONCE(ctx->sq_ring->r.tail) - ctx->cached_sq_head != |
3075 | ctx->sq_ring->ring_entries) | |
2b188cc1 JA |
3076 | mask |= EPOLLOUT | EPOLLWRNORM; |
3077 | if (READ_ONCE(ctx->cq_ring->r.head) != ctx->cached_cq_tail) | |
3078 | mask |= EPOLLIN | EPOLLRDNORM; | |
3079 | ||
3080 | return mask; | |
3081 | } | |
3082 | ||
3083 | static int io_uring_fasync(int fd, struct file *file, int on) | |
3084 | { | |
3085 | struct io_ring_ctx *ctx = file->private_data; | |
3086 | ||
3087 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
3088 | } | |
3089 | ||
3090 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) | |
3091 | { | |
3092 | mutex_lock(&ctx->uring_lock); | |
3093 | percpu_ref_kill(&ctx->refs); | |
3094 | mutex_unlock(&ctx->uring_lock); | |
3095 | ||
221c5eb2 | 3096 | io_poll_remove_all(ctx); |
def596e9 | 3097 | io_iopoll_reap_events(ctx); |
2b188cc1 JA |
3098 | wait_for_completion(&ctx->ctx_done); |
3099 | io_ring_ctx_free(ctx); | |
3100 | } | |
3101 | ||
3102 | static int io_uring_release(struct inode *inode, struct file *file) | |
3103 | { | |
3104 | struct io_ring_ctx *ctx = file->private_data; | |
3105 | ||
3106 | file->private_data = NULL; | |
3107 | io_ring_ctx_wait_and_kill(ctx); | |
3108 | return 0; | |
3109 | } | |
3110 | ||
3111 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
3112 | { | |
3113 | loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT; | |
3114 | unsigned long sz = vma->vm_end - vma->vm_start; | |
3115 | struct io_ring_ctx *ctx = file->private_data; | |
3116 | unsigned long pfn; | |
3117 | struct page *page; | |
3118 | void *ptr; | |
3119 | ||
3120 | switch (offset) { | |
3121 | case IORING_OFF_SQ_RING: | |
3122 | ptr = ctx->sq_ring; | |
3123 | break; | |
3124 | case IORING_OFF_SQES: | |
3125 | ptr = ctx->sq_sqes; | |
3126 | break; | |
3127 | case IORING_OFF_CQ_RING: | |
3128 | ptr = ctx->cq_ring; | |
3129 | break; | |
3130 | default: | |
3131 | return -EINVAL; | |
3132 | } | |
3133 | ||
3134 | page = virt_to_head_page(ptr); | |
3135 | if (sz > (PAGE_SIZE << compound_order(page))) | |
3136 | return -EINVAL; | |
3137 | ||
3138 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
3139 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
3140 | } | |
3141 | ||
3142 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, | |
3143 | u32, min_complete, u32, flags, const sigset_t __user *, sig, | |
3144 | size_t, sigsz) | |
3145 | { | |
3146 | struct io_ring_ctx *ctx; | |
3147 | long ret = -EBADF; | |
3148 | int submitted = 0; | |
3149 | struct fd f; | |
3150 | ||
6c271ce2 | 3151 | if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP)) |
2b188cc1 JA |
3152 | return -EINVAL; |
3153 | ||
3154 | f = fdget(fd); | |
3155 | if (!f.file) | |
3156 | return -EBADF; | |
3157 | ||
3158 | ret = -EOPNOTSUPP; | |
3159 | if (f.file->f_op != &io_uring_fops) | |
3160 | goto out_fput; | |
3161 | ||
3162 | ret = -ENXIO; | |
3163 | ctx = f.file->private_data; | |
3164 | if (!percpu_ref_tryget(&ctx->refs)) | |
3165 | goto out_fput; | |
3166 | ||
6c271ce2 JA |
3167 | /* |
3168 | * For SQ polling, the thread will do all submissions and completions. | |
3169 | * Just return the requested submit count, and wake the thread if | |
3170 | * we were asked to. | |
3171 | */ | |
3172 | if (ctx->flags & IORING_SETUP_SQPOLL) { | |
3173 | if (flags & IORING_ENTER_SQ_WAKEUP) | |
3174 | wake_up(&ctx->sqo_wait); | |
3175 | submitted = to_submit; | |
3176 | goto out_ctx; | |
3177 | } | |
3178 | ||
2b188cc1 JA |
3179 | ret = 0; |
3180 | if (to_submit) { | |
3181 | to_submit = min(to_submit, ctx->sq_entries); | |
3182 | ||
3183 | mutex_lock(&ctx->uring_lock); | |
3184 | submitted = io_ring_submit(ctx, to_submit); | |
3185 | mutex_unlock(&ctx->uring_lock); | |
2b188cc1 JA |
3186 | } |
3187 | if (flags & IORING_ENTER_GETEVENTS) { | |
def596e9 JA |
3188 | unsigned nr_events = 0; |
3189 | ||
2b188cc1 JA |
3190 | min_complete = min(min_complete, ctx->cq_entries); |
3191 | ||
def596e9 JA |
3192 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
3193 | mutex_lock(&ctx->uring_lock); | |
3194 | ret = io_iopoll_check(ctx, &nr_events, min_complete); | |
3195 | mutex_unlock(&ctx->uring_lock); | |
3196 | } else { | |
3197 | ret = io_cqring_wait(ctx, min_complete, sig, sigsz); | |
3198 | } | |
2b188cc1 JA |
3199 | } |
3200 | ||
3201 | out_ctx: | |
3202 | io_ring_drop_ctx_refs(ctx, 1); | |
3203 | out_fput: | |
3204 | fdput(f); | |
3205 | return submitted ? submitted : ret; | |
3206 | } | |
3207 | ||
3208 | static const struct file_operations io_uring_fops = { | |
3209 | .release = io_uring_release, | |
3210 | .mmap = io_uring_mmap, | |
3211 | .poll = io_uring_poll, | |
3212 | .fasync = io_uring_fasync, | |
3213 | }; | |
3214 | ||
3215 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
3216 | struct io_uring_params *p) | |
3217 | { | |
3218 | struct io_sq_ring *sq_ring; | |
3219 | struct io_cq_ring *cq_ring; | |
3220 | size_t size; | |
3221 | ||
3222 | sq_ring = io_mem_alloc(struct_size(sq_ring, array, p->sq_entries)); | |
3223 | if (!sq_ring) | |
3224 | return -ENOMEM; | |
3225 | ||
3226 | ctx->sq_ring = sq_ring; | |
3227 | sq_ring->ring_mask = p->sq_entries - 1; | |
3228 | sq_ring->ring_entries = p->sq_entries; | |
3229 | ctx->sq_mask = sq_ring->ring_mask; | |
3230 | ctx->sq_entries = sq_ring->ring_entries; | |
3231 | ||
3232 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
3233 | if (size == SIZE_MAX) | |
3234 | return -EOVERFLOW; | |
3235 | ||
3236 | ctx->sq_sqes = io_mem_alloc(size); | |
52e04ef4 | 3237 | if (!ctx->sq_sqes) |
2b188cc1 | 3238 | return -ENOMEM; |
2b188cc1 JA |
3239 | |
3240 | cq_ring = io_mem_alloc(struct_size(cq_ring, cqes, p->cq_entries)); | |
52e04ef4 | 3241 | if (!cq_ring) |
2b188cc1 | 3242 | return -ENOMEM; |
2b188cc1 JA |
3243 | |
3244 | ctx->cq_ring = cq_ring; | |
3245 | cq_ring->ring_mask = p->cq_entries - 1; | |
3246 | cq_ring->ring_entries = p->cq_entries; | |
3247 | ctx->cq_mask = cq_ring->ring_mask; | |
3248 | ctx->cq_entries = cq_ring->ring_entries; | |
3249 | return 0; | |
3250 | } | |
3251 | ||
3252 | /* | |
3253 | * Allocate an anonymous fd, this is what constitutes the application | |
3254 | * visible backing of an io_uring instance. The application mmaps this | |
3255 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
3256 | * we have to tie this fd to a socket for file garbage collection purposes. | |
3257 | */ | |
3258 | static int io_uring_get_fd(struct io_ring_ctx *ctx) | |
3259 | { | |
3260 | struct file *file; | |
3261 | int ret; | |
3262 | ||
3263 | #if defined(CONFIG_UNIX) | |
3264 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, | |
3265 | &ctx->ring_sock); | |
3266 | if (ret) | |
3267 | return ret; | |
3268 | #endif | |
3269 | ||
3270 | ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
3271 | if (ret < 0) | |
3272 | goto err; | |
3273 | ||
3274 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, | |
3275 | O_RDWR | O_CLOEXEC); | |
3276 | if (IS_ERR(file)) { | |
3277 | put_unused_fd(ret); | |
3278 | ret = PTR_ERR(file); | |
3279 | goto err; | |
3280 | } | |
3281 | ||
3282 | #if defined(CONFIG_UNIX) | |
3283 | ctx->ring_sock->file = file; | |
6b06314c | 3284 | ctx->ring_sock->sk->sk_user_data = ctx; |
2b188cc1 JA |
3285 | #endif |
3286 | fd_install(ret, file); | |
3287 | return ret; | |
3288 | err: | |
3289 | #if defined(CONFIG_UNIX) | |
3290 | sock_release(ctx->ring_sock); | |
3291 | ctx->ring_sock = NULL; | |
3292 | #endif | |
3293 | return ret; | |
3294 | } | |
3295 | ||
3296 | static int io_uring_create(unsigned entries, struct io_uring_params *p) | |
3297 | { | |
3298 | struct user_struct *user = NULL; | |
3299 | struct io_ring_ctx *ctx; | |
3300 | bool account_mem; | |
3301 | int ret; | |
3302 | ||
3303 | if (!entries || entries > IORING_MAX_ENTRIES) | |
3304 | return -EINVAL; | |
3305 | ||
3306 | /* | |
3307 | * Use twice as many entries for the CQ ring. It's possible for the | |
3308 | * application to drive a higher depth than the size of the SQ ring, | |
3309 | * since the sqes are only used at submission time. This allows for | |
3310 | * some flexibility in overcommitting a bit. | |
3311 | */ | |
3312 | p->sq_entries = roundup_pow_of_two(entries); | |
3313 | p->cq_entries = 2 * p->sq_entries; | |
3314 | ||
3315 | user = get_uid(current_user()); | |
3316 | account_mem = !capable(CAP_IPC_LOCK); | |
3317 | ||
3318 | if (account_mem) { | |
3319 | ret = io_account_mem(user, | |
3320 | ring_pages(p->sq_entries, p->cq_entries)); | |
3321 | if (ret) { | |
3322 | free_uid(user); | |
3323 | return ret; | |
3324 | } | |
3325 | } | |
3326 | ||
3327 | ctx = io_ring_ctx_alloc(p); | |
3328 | if (!ctx) { | |
3329 | if (account_mem) | |
3330 | io_unaccount_mem(user, ring_pages(p->sq_entries, | |
3331 | p->cq_entries)); | |
3332 | free_uid(user); | |
3333 | return -ENOMEM; | |
3334 | } | |
3335 | ctx->compat = in_compat_syscall(); | |
3336 | ctx->account_mem = account_mem; | |
3337 | ctx->user = user; | |
3338 | ||
3339 | ret = io_allocate_scq_urings(ctx, p); | |
3340 | if (ret) | |
3341 | goto err; | |
3342 | ||
6c271ce2 | 3343 | ret = io_sq_offload_start(ctx, p); |
2b188cc1 JA |
3344 | if (ret) |
3345 | goto err; | |
3346 | ||
3347 | ret = io_uring_get_fd(ctx); | |
3348 | if (ret < 0) | |
3349 | goto err; | |
3350 | ||
3351 | memset(&p->sq_off, 0, sizeof(p->sq_off)); | |
3352 | p->sq_off.head = offsetof(struct io_sq_ring, r.head); | |
3353 | p->sq_off.tail = offsetof(struct io_sq_ring, r.tail); | |
3354 | p->sq_off.ring_mask = offsetof(struct io_sq_ring, ring_mask); | |
3355 | p->sq_off.ring_entries = offsetof(struct io_sq_ring, ring_entries); | |
3356 | p->sq_off.flags = offsetof(struct io_sq_ring, flags); | |
3357 | p->sq_off.dropped = offsetof(struct io_sq_ring, dropped); | |
3358 | p->sq_off.array = offsetof(struct io_sq_ring, array); | |
3359 | ||
3360 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
3361 | p->cq_off.head = offsetof(struct io_cq_ring, r.head); | |
3362 | p->cq_off.tail = offsetof(struct io_cq_ring, r.tail); | |
3363 | p->cq_off.ring_mask = offsetof(struct io_cq_ring, ring_mask); | |
3364 | p->cq_off.ring_entries = offsetof(struct io_cq_ring, ring_entries); | |
3365 | p->cq_off.overflow = offsetof(struct io_cq_ring, overflow); | |
3366 | p->cq_off.cqes = offsetof(struct io_cq_ring, cqes); | |
3367 | return ret; | |
3368 | err: | |
3369 | io_ring_ctx_wait_and_kill(ctx); | |
3370 | return ret; | |
3371 | } | |
3372 | ||
3373 | /* | |
3374 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
3375 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
3376 | * params structure passed in. | |
3377 | */ | |
3378 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
3379 | { | |
3380 | struct io_uring_params p; | |
3381 | long ret; | |
3382 | int i; | |
3383 | ||
3384 | if (copy_from_user(&p, params, sizeof(p))) | |
3385 | return -EFAULT; | |
3386 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
3387 | if (p.resv[i]) | |
3388 | return -EINVAL; | |
3389 | } | |
3390 | ||
6c271ce2 JA |
3391 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
3392 | IORING_SETUP_SQ_AFF)) | |
2b188cc1 JA |
3393 | return -EINVAL; |
3394 | ||
3395 | ret = io_uring_create(entries, &p); | |
3396 | if (ret < 0) | |
3397 | return ret; | |
3398 | ||
3399 | if (copy_to_user(params, &p, sizeof(p))) | |
3400 | return -EFAULT; | |
3401 | ||
3402 | return ret; | |
3403 | } | |
3404 | ||
3405 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
3406 | struct io_uring_params __user *, params) | |
3407 | { | |
3408 | return io_uring_setup(entries, params); | |
3409 | } | |
3410 | ||
edafccee JA |
3411 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
3412 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
3413 | __releases(ctx->uring_lock) |
3414 | __acquires(ctx->uring_lock) | |
edafccee JA |
3415 | { |
3416 | int ret; | |
3417 | ||
35fa71a0 JA |
3418 | /* |
3419 | * We're inside the ring mutex, if the ref is already dying, then | |
3420 | * someone else killed the ctx or is already going through | |
3421 | * io_uring_register(). | |
3422 | */ | |
3423 | if (percpu_ref_is_dying(&ctx->refs)) | |
3424 | return -ENXIO; | |
3425 | ||
edafccee | 3426 | percpu_ref_kill(&ctx->refs); |
b19062a5 JA |
3427 | |
3428 | /* | |
3429 | * Drop uring mutex before waiting for references to exit. If another | |
3430 | * thread is currently inside io_uring_enter() it might need to grab | |
3431 | * the uring_lock to make progress. If we hold it here across the drain | |
3432 | * wait, then we can deadlock. It's safe to drop the mutex here, since | |
3433 | * no new references will come in after we've killed the percpu ref. | |
3434 | */ | |
3435 | mutex_unlock(&ctx->uring_lock); | |
edafccee | 3436 | wait_for_completion(&ctx->ctx_done); |
b19062a5 | 3437 | mutex_lock(&ctx->uring_lock); |
edafccee JA |
3438 | |
3439 | switch (opcode) { | |
3440 | case IORING_REGISTER_BUFFERS: | |
3441 | ret = io_sqe_buffer_register(ctx, arg, nr_args); | |
3442 | break; | |
3443 | case IORING_UNREGISTER_BUFFERS: | |
3444 | ret = -EINVAL; | |
3445 | if (arg || nr_args) | |
3446 | break; | |
3447 | ret = io_sqe_buffer_unregister(ctx); | |
3448 | break; | |
6b06314c JA |
3449 | case IORING_REGISTER_FILES: |
3450 | ret = io_sqe_files_register(ctx, arg, nr_args); | |
3451 | break; | |
3452 | case IORING_UNREGISTER_FILES: | |
3453 | ret = -EINVAL; | |
3454 | if (arg || nr_args) | |
3455 | break; | |
3456 | ret = io_sqe_files_unregister(ctx); | |
3457 | break; | |
9b402849 JA |
3458 | case IORING_REGISTER_EVENTFD: |
3459 | ret = -EINVAL; | |
3460 | if (nr_args != 1) | |
3461 | break; | |
3462 | ret = io_eventfd_register(ctx, arg); | |
3463 | break; | |
3464 | case IORING_UNREGISTER_EVENTFD: | |
3465 | ret = -EINVAL; | |
3466 | if (arg || nr_args) | |
3467 | break; | |
3468 | ret = io_eventfd_unregister(ctx); | |
3469 | break; | |
edafccee JA |
3470 | default: |
3471 | ret = -EINVAL; | |
3472 | break; | |
3473 | } | |
3474 | ||
3475 | /* bring the ctx back to life */ | |
3476 | reinit_completion(&ctx->ctx_done); | |
3477 | percpu_ref_reinit(&ctx->refs); | |
3478 | return ret; | |
3479 | } | |
3480 | ||
3481 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
3482 | void __user *, arg, unsigned int, nr_args) | |
3483 | { | |
3484 | struct io_ring_ctx *ctx; | |
3485 | long ret = -EBADF; | |
3486 | struct fd f; | |
3487 | ||
3488 | f = fdget(fd); | |
3489 | if (!f.file) | |
3490 | return -EBADF; | |
3491 | ||
3492 | ret = -EOPNOTSUPP; | |
3493 | if (f.file->f_op != &io_uring_fops) | |
3494 | goto out_fput; | |
3495 | ||
3496 | ctx = f.file->private_data; | |
3497 | ||
3498 | mutex_lock(&ctx->uring_lock); | |
3499 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
3500 | mutex_unlock(&ctx->uring_lock); | |
3501 | out_fput: | |
3502 | fdput(f); | |
3503 | return ret; | |
3504 | } | |
3505 | ||
2b188cc1 JA |
3506 | static int __init io_uring_init(void) |
3507 | { | |
3508 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC); | |
3509 | return 0; | |
3510 | }; | |
3511 | __initcall(io_uring_init); |