]>
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 | |
7 | * the application and kernel side. When the application reads the CQ ring | |
8 | * tail, it must use an appropriate smp_rmb() to order with the smp_wmb() | |
9 | * the kernel uses after writing the tail. Failure to do so could cause a | |
10 | * delay in when the application notices that completion events available. | |
11 | * This isn't a fatal condition. Likewise, the application must use an | |
12 | * appropriate smp_wmb() both before writing the SQ tail, and after writing | |
13 | * the SQ tail. The first one orders the sqe writes with the tail write, and | |
14 | * the latter is paired with the smp_rmb() the kernel will issue before | |
15 | * reading the SQ tail on submission. | |
16 | * | |
17 | * Also see the examples in the liburing library: | |
18 | * | |
19 | * git://git.kernel.dk/liburing | |
20 | * | |
21 | * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens | |
22 | * from data shared between the kernel and application. This is done both | |
23 | * for ordering purposes, but also to ensure that once a value is loaded from | |
24 | * data that the application could potentially modify, it remains stable. | |
25 | * | |
26 | * Copyright (C) 2018-2019 Jens Axboe | |
c992fe29 | 27 | * Copyright (c) 2018-2019 Christoph Hellwig |
2b188cc1 JA |
28 | */ |
29 | #include <linux/kernel.h> | |
30 | #include <linux/init.h> | |
31 | #include <linux/errno.h> | |
32 | #include <linux/syscalls.h> | |
33 | #include <linux/compat.h> | |
34 | #include <linux/refcount.h> | |
35 | #include <linux/uio.h> | |
36 | ||
37 | #include <linux/sched/signal.h> | |
38 | #include <linux/fs.h> | |
39 | #include <linux/file.h> | |
40 | #include <linux/fdtable.h> | |
41 | #include <linux/mm.h> | |
42 | #include <linux/mman.h> | |
43 | #include <linux/mmu_context.h> | |
44 | #include <linux/percpu.h> | |
45 | #include <linux/slab.h> | |
46 | #include <linux/workqueue.h> | |
47 | #include <linux/blkdev.h> | |
edafccee | 48 | #include <linux/bvec.h> |
2b188cc1 JA |
49 | #include <linux/net.h> |
50 | #include <net/sock.h> | |
51 | #include <net/af_unix.h> | |
6b06314c | 52 | #include <net/scm.h> |
2b188cc1 JA |
53 | #include <linux/anon_inodes.h> |
54 | #include <linux/sched/mm.h> | |
55 | #include <linux/uaccess.h> | |
56 | #include <linux/nospec.h> | |
edafccee JA |
57 | #include <linux/sizes.h> |
58 | #include <linux/hugetlb.h> | |
2b188cc1 JA |
59 | |
60 | #include <uapi/linux/io_uring.h> | |
61 | ||
62 | #include "internal.h" | |
63 | ||
64 | #define IORING_MAX_ENTRIES 4096 | |
6b06314c | 65 | #define IORING_MAX_FIXED_FILES 1024 |
2b188cc1 JA |
66 | |
67 | struct io_uring { | |
68 | u32 head ____cacheline_aligned_in_smp; | |
69 | u32 tail ____cacheline_aligned_in_smp; | |
70 | }; | |
71 | ||
72 | struct io_sq_ring { | |
73 | struct io_uring r; | |
74 | u32 ring_mask; | |
75 | u32 ring_entries; | |
76 | u32 dropped; | |
77 | u32 flags; | |
78 | u32 array[]; | |
79 | }; | |
80 | ||
81 | struct io_cq_ring { | |
82 | struct io_uring r; | |
83 | u32 ring_mask; | |
84 | u32 ring_entries; | |
85 | u32 overflow; | |
86 | struct io_uring_cqe cqes[]; | |
87 | }; | |
88 | ||
edafccee JA |
89 | struct io_mapped_ubuf { |
90 | u64 ubuf; | |
91 | size_t len; | |
92 | struct bio_vec *bvec; | |
93 | unsigned int nr_bvecs; | |
94 | }; | |
95 | ||
2b188cc1 JA |
96 | struct io_ring_ctx { |
97 | struct { | |
98 | struct percpu_ref refs; | |
99 | } ____cacheline_aligned_in_smp; | |
100 | ||
101 | struct { | |
102 | unsigned int flags; | |
103 | bool compat; | |
104 | bool account_mem; | |
105 | ||
106 | /* SQ ring */ | |
107 | struct io_sq_ring *sq_ring; | |
108 | unsigned cached_sq_head; | |
109 | unsigned sq_entries; | |
110 | unsigned sq_mask; | |
111 | struct io_uring_sqe *sq_sqes; | |
112 | } ____cacheline_aligned_in_smp; | |
113 | ||
114 | /* IO offload */ | |
115 | struct workqueue_struct *sqo_wq; | |
116 | struct mm_struct *sqo_mm; | |
117 | ||
118 | struct { | |
119 | /* CQ ring */ | |
120 | struct io_cq_ring *cq_ring; | |
121 | unsigned cached_cq_tail; | |
122 | unsigned cq_entries; | |
123 | unsigned cq_mask; | |
124 | struct wait_queue_head cq_wait; | |
125 | struct fasync_struct *cq_fasync; | |
126 | } ____cacheline_aligned_in_smp; | |
127 | ||
6b06314c JA |
128 | /* |
129 | * If used, fixed file set. Writers must ensure that ->refs is dead, | |
130 | * readers must ensure that ->refs is alive as long as the file* is | |
131 | * used. Only updated through io_uring_register(2). | |
132 | */ | |
133 | struct file **user_files; | |
134 | unsigned nr_user_files; | |
135 | ||
edafccee JA |
136 | /* if used, fixed mapped user buffers */ |
137 | unsigned nr_user_bufs; | |
138 | struct io_mapped_ubuf *user_bufs; | |
139 | ||
2b188cc1 JA |
140 | struct user_struct *user; |
141 | ||
142 | struct completion ctx_done; | |
143 | ||
144 | struct { | |
145 | struct mutex uring_lock; | |
146 | wait_queue_head_t wait; | |
147 | } ____cacheline_aligned_in_smp; | |
148 | ||
149 | struct { | |
150 | spinlock_t completion_lock; | |
def596e9 JA |
151 | bool poll_multi_file; |
152 | /* | |
153 | * ->poll_list is protected by the ctx->uring_lock for | |
154 | * io_uring instances that don't use IORING_SETUP_SQPOLL. | |
155 | * For SQPOLL, only the single threaded io_sq_thread() will | |
156 | * manipulate the list, hence no extra locking is needed there. | |
157 | */ | |
158 | struct list_head poll_list; | |
2b188cc1 JA |
159 | } ____cacheline_aligned_in_smp; |
160 | ||
161 | #if defined(CONFIG_UNIX) | |
162 | struct socket *ring_sock; | |
163 | #endif | |
164 | }; | |
165 | ||
166 | struct sqe_submit { | |
167 | const struct io_uring_sqe *sqe; | |
168 | unsigned short index; | |
169 | bool has_user; | |
def596e9 | 170 | bool needs_lock; |
2b188cc1 JA |
171 | }; |
172 | ||
173 | struct io_kiocb { | |
174 | struct kiocb rw; | |
175 | ||
176 | struct sqe_submit submit; | |
177 | ||
178 | struct io_ring_ctx *ctx; | |
179 | struct list_head list; | |
180 | unsigned int flags; | |
181 | #define REQ_F_FORCE_NONBLOCK 1 /* inline submission attempt */ | |
def596e9 | 182 | #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */ |
6b06314c | 183 | #define REQ_F_FIXED_FILE 4 /* ctx owns file */ |
2b188cc1 | 184 | u64 user_data; |
def596e9 | 185 | u64 error; |
2b188cc1 JA |
186 | |
187 | struct work_struct work; | |
188 | }; | |
189 | ||
190 | #define IO_PLUG_THRESHOLD 2 | |
def596e9 | 191 | #define IO_IOPOLL_BATCH 8 |
2b188cc1 | 192 | |
9a56a232 JA |
193 | struct io_submit_state { |
194 | struct blk_plug plug; | |
195 | ||
2579f913 JA |
196 | /* |
197 | * io_kiocb alloc cache | |
198 | */ | |
199 | void *reqs[IO_IOPOLL_BATCH]; | |
200 | unsigned int free_reqs; | |
201 | unsigned int cur_req; | |
202 | ||
9a56a232 JA |
203 | /* |
204 | * File reference cache | |
205 | */ | |
206 | struct file *file; | |
207 | unsigned int fd; | |
208 | unsigned int has_refs; | |
209 | unsigned int used_refs; | |
210 | unsigned int ios_left; | |
211 | }; | |
212 | ||
2b188cc1 JA |
213 | static struct kmem_cache *req_cachep; |
214 | ||
215 | static const struct file_operations io_uring_fops; | |
216 | ||
217 | struct sock *io_uring_get_socket(struct file *file) | |
218 | { | |
219 | #if defined(CONFIG_UNIX) | |
220 | if (file->f_op == &io_uring_fops) { | |
221 | struct io_ring_ctx *ctx = file->private_data; | |
222 | ||
223 | return ctx->ring_sock->sk; | |
224 | } | |
225 | #endif | |
226 | return NULL; | |
227 | } | |
228 | EXPORT_SYMBOL(io_uring_get_socket); | |
229 | ||
230 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) | |
231 | { | |
232 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
233 | ||
234 | complete(&ctx->ctx_done); | |
235 | } | |
236 | ||
237 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) | |
238 | { | |
239 | struct io_ring_ctx *ctx; | |
240 | ||
241 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
242 | if (!ctx) | |
243 | return NULL; | |
244 | ||
245 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, 0, GFP_KERNEL)) { | |
246 | kfree(ctx); | |
247 | return NULL; | |
248 | } | |
249 | ||
250 | ctx->flags = p->flags; | |
251 | init_waitqueue_head(&ctx->cq_wait); | |
252 | init_completion(&ctx->ctx_done); | |
253 | mutex_init(&ctx->uring_lock); | |
254 | init_waitqueue_head(&ctx->wait); | |
255 | spin_lock_init(&ctx->completion_lock); | |
def596e9 | 256 | INIT_LIST_HEAD(&ctx->poll_list); |
2b188cc1 JA |
257 | return ctx; |
258 | } | |
259 | ||
260 | static void io_commit_cqring(struct io_ring_ctx *ctx) | |
261 | { | |
262 | struct io_cq_ring *ring = ctx->cq_ring; | |
263 | ||
264 | if (ctx->cached_cq_tail != READ_ONCE(ring->r.tail)) { | |
265 | /* order cqe stores with ring update */ | |
266 | smp_store_release(&ring->r.tail, ctx->cached_cq_tail); | |
267 | ||
268 | /* | |
269 | * Write sider barrier of tail update, app has read side. See | |
270 | * comment at the top of this file. | |
271 | */ | |
272 | smp_wmb(); | |
273 | ||
274 | if (wq_has_sleeper(&ctx->cq_wait)) { | |
275 | wake_up_interruptible(&ctx->cq_wait); | |
276 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
277 | } | |
278 | } | |
279 | } | |
280 | ||
281 | static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx) | |
282 | { | |
283 | struct io_cq_ring *ring = ctx->cq_ring; | |
284 | unsigned tail; | |
285 | ||
286 | tail = ctx->cached_cq_tail; | |
287 | /* See comment at the top of the file */ | |
288 | smp_rmb(); | |
289 | if (tail + 1 == READ_ONCE(ring->r.head)) | |
290 | return NULL; | |
291 | ||
292 | ctx->cached_cq_tail++; | |
293 | return &ring->cqes[tail & ctx->cq_mask]; | |
294 | } | |
295 | ||
296 | static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data, | |
297 | long res, unsigned ev_flags) | |
298 | { | |
299 | struct io_uring_cqe *cqe; | |
300 | ||
301 | /* | |
302 | * If we can't get a cq entry, userspace overflowed the | |
303 | * submission (by quite a lot). Increment the overflow count in | |
304 | * the ring. | |
305 | */ | |
306 | cqe = io_get_cqring(ctx); | |
307 | if (cqe) { | |
308 | WRITE_ONCE(cqe->user_data, ki_user_data); | |
309 | WRITE_ONCE(cqe->res, res); | |
310 | WRITE_ONCE(cqe->flags, ev_flags); | |
311 | } else { | |
312 | unsigned overflow = READ_ONCE(ctx->cq_ring->overflow); | |
313 | ||
314 | WRITE_ONCE(ctx->cq_ring->overflow, overflow + 1); | |
315 | } | |
316 | } | |
317 | ||
318 | static void io_cqring_add_event(struct io_ring_ctx *ctx, u64 ki_user_data, | |
319 | long res, unsigned ev_flags) | |
320 | { | |
321 | unsigned long flags; | |
322 | ||
323 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
324 | io_cqring_fill_event(ctx, ki_user_data, res, ev_flags); | |
325 | io_commit_cqring(ctx); | |
326 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
327 | ||
328 | if (waitqueue_active(&ctx->wait)) | |
329 | wake_up(&ctx->wait); | |
330 | } | |
331 | ||
332 | static void io_ring_drop_ctx_refs(struct io_ring_ctx *ctx, unsigned refs) | |
333 | { | |
334 | percpu_ref_put_many(&ctx->refs, refs); | |
335 | ||
336 | if (waitqueue_active(&ctx->wait)) | |
337 | wake_up(&ctx->wait); | |
338 | } | |
339 | ||
2579f913 JA |
340 | static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx, |
341 | struct io_submit_state *state) | |
2b188cc1 JA |
342 | { |
343 | struct io_kiocb *req; | |
344 | ||
345 | if (!percpu_ref_tryget(&ctx->refs)) | |
346 | return NULL; | |
347 | ||
2579f913 JA |
348 | if (!state) { |
349 | req = kmem_cache_alloc(req_cachep, __GFP_NOWARN); | |
350 | if (unlikely(!req)) | |
351 | goto out; | |
352 | } else if (!state->free_reqs) { | |
353 | size_t sz; | |
354 | int ret; | |
355 | ||
356 | sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs)); | |
357 | ret = kmem_cache_alloc_bulk(req_cachep, __GFP_NOWARN, sz, | |
358 | state->reqs); | |
359 | if (unlikely(ret <= 0)) | |
360 | goto out; | |
361 | state->free_reqs = ret - 1; | |
362 | state->cur_req = 1; | |
363 | req = state->reqs[0]; | |
364 | } else { | |
365 | req = state->reqs[state->cur_req]; | |
366 | state->free_reqs--; | |
367 | state->cur_req++; | |
2b188cc1 JA |
368 | } |
369 | ||
2579f913 JA |
370 | req->ctx = ctx; |
371 | req->flags = 0; | |
372 | return req; | |
373 | out: | |
2b188cc1 JA |
374 | io_ring_drop_ctx_refs(ctx, 1); |
375 | return NULL; | |
376 | } | |
377 | ||
def596e9 JA |
378 | static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr) |
379 | { | |
380 | if (*nr) { | |
381 | kmem_cache_free_bulk(req_cachep, *nr, reqs); | |
382 | io_ring_drop_ctx_refs(ctx, *nr); | |
383 | *nr = 0; | |
384 | } | |
385 | } | |
386 | ||
2b188cc1 JA |
387 | static void io_free_req(struct io_kiocb *req) |
388 | { | |
389 | io_ring_drop_ctx_refs(req->ctx, 1); | |
390 | kmem_cache_free(req_cachep, req); | |
391 | } | |
392 | ||
def596e9 JA |
393 | /* |
394 | * Find and free completed poll iocbs | |
395 | */ | |
396 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
397 | struct list_head *done) | |
398 | { | |
399 | void *reqs[IO_IOPOLL_BATCH]; | |
9a56a232 JA |
400 | int file_count, to_free; |
401 | struct file *file = NULL; | |
def596e9 | 402 | struct io_kiocb *req; |
def596e9 | 403 | |
9a56a232 | 404 | file_count = to_free = 0; |
def596e9 JA |
405 | while (!list_empty(done)) { |
406 | req = list_first_entry(done, struct io_kiocb, list); | |
407 | list_del(&req->list); | |
408 | ||
409 | io_cqring_fill_event(ctx, req->user_data, req->error, 0); | |
410 | ||
411 | reqs[to_free++] = req; | |
412 | (*nr_events)++; | |
413 | ||
9a56a232 JA |
414 | /* |
415 | * Batched puts of the same file, to avoid dirtying the | |
416 | * file usage count multiple times, if avoidable. | |
417 | */ | |
6b06314c JA |
418 | if (!(req->flags & REQ_F_FIXED_FILE)) { |
419 | if (!file) { | |
420 | file = req->rw.ki_filp; | |
421 | file_count = 1; | |
422 | } else if (file == req->rw.ki_filp) { | |
423 | file_count++; | |
424 | } else { | |
425 | fput_many(file, file_count); | |
426 | file = req->rw.ki_filp; | |
427 | file_count = 1; | |
428 | } | |
9a56a232 JA |
429 | } |
430 | ||
def596e9 JA |
431 | if (to_free == ARRAY_SIZE(reqs)) |
432 | io_free_req_many(ctx, reqs, &to_free); | |
433 | } | |
434 | io_commit_cqring(ctx); | |
435 | ||
9a56a232 JA |
436 | if (file) |
437 | fput_many(file, file_count); | |
def596e9 JA |
438 | io_free_req_many(ctx, reqs, &to_free); |
439 | } | |
440 | ||
441 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
442 | long min) | |
443 | { | |
444 | struct io_kiocb *req, *tmp; | |
445 | LIST_HEAD(done); | |
446 | bool spin; | |
447 | int ret; | |
448 | ||
449 | /* | |
450 | * Only spin for completions if we don't have multiple devices hanging | |
451 | * off our complete list, and we're under the requested amount. | |
452 | */ | |
453 | spin = !ctx->poll_multi_file && *nr_events < min; | |
454 | ||
455 | ret = 0; | |
456 | list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) { | |
457 | struct kiocb *kiocb = &req->rw; | |
458 | ||
459 | /* | |
460 | * Move completed entries to our local list. If we find a | |
461 | * request that requires polling, break out and complete | |
462 | * the done list first, if we have entries there. | |
463 | */ | |
464 | if (req->flags & REQ_F_IOPOLL_COMPLETED) { | |
465 | list_move_tail(&req->list, &done); | |
466 | continue; | |
467 | } | |
468 | if (!list_empty(&done)) | |
469 | break; | |
470 | ||
471 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
472 | if (ret < 0) | |
473 | break; | |
474 | ||
475 | if (ret && spin) | |
476 | spin = false; | |
477 | ret = 0; | |
478 | } | |
479 | ||
480 | if (!list_empty(&done)) | |
481 | io_iopoll_complete(ctx, nr_events, &done); | |
482 | ||
483 | return ret; | |
484 | } | |
485 | ||
486 | /* | |
487 | * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a | |
488 | * non-spinning poll check - we'll still enter the driver poll loop, but only | |
489 | * as a non-spinning completion check. | |
490 | */ | |
491 | static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
492 | long min) | |
493 | { | |
494 | while (!list_empty(&ctx->poll_list)) { | |
495 | int ret; | |
496 | ||
497 | ret = io_do_iopoll(ctx, nr_events, min); | |
498 | if (ret < 0) | |
499 | return ret; | |
500 | if (!min || *nr_events >= min) | |
501 | return 0; | |
502 | } | |
503 | ||
504 | return 1; | |
505 | } | |
506 | ||
507 | /* | |
508 | * We can't just wait for polled events to come to us, we have to actively | |
509 | * find and complete them. | |
510 | */ | |
511 | static void io_iopoll_reap_events(struct io_ring_ctx *ctx) | |
512 | { | |
513 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
514 | return; | |
515 | ||
516 | mutex_lock(&ctx->uring_lock); | |
517 | while (!list_empty(&ctx->poll_list)) { | |
518 | unsigned int nr_events = 0; | |
519 | ||
520 | io_iopoll_getevents(ctx, &nr_events, 1); | |
521 | } | |
522 | mutex_unlock(&ctx->uring_lock); | |
523 | } | |
524 | ||
525 | static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, | |
526 | long min) | |
527 | { | |
528 | int ret = 0; | |
529 | ||
530 | do { | |
531 | int tmin = 0; | |
532 | ||
533 | if (*nr_events < min) | |
534 | tmin = min - *nr_events; | |
535 | ||
536 | ret = io_iopoll_getevents(ctx, nr_events, tmin); | |
537 | if (ret <= 0) | |
538 | break; | |
539 | ret = 0; | |
540 | } while (min && !*nr_events && !need_resched()); | |
541 | ||
542 | return ret; | |
543 | } | |
544 | ||
2b188cc1 JA |
545 | static void kiocb_end_write(struct kiocb *kiocb) |
546 | { | |
547 | if (kiocb->ki_flags & IOCB_WRITE) { | |
548 | struct inode *inode = file_inode(kiocb->ki_filp); | |
549 | ||
550 | /* | |
551 | * Tell lockdep we inherited freeze protection from submission | |
552 | * thread. | |
553 | */ | |
554 | if (S_ISREG(inode->i_mode)) | |
555 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); | |
556 | file_end_write(kiocb->ki_filp); | |
557 | } | |
558 | } | |
559 | ||
6b06314c JA |
560 | static void io_fput(struct io_kiocb *req) |
561 | { | |
562 | if (!(req->flags & REQ_F_FIXED_FILE)) | |
563 | fput(req->rw.ki_filp); | |
564 | } | |
565 | ||
2b188cc1 JA |
566 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) |
567 | { | |
568 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
569 | ||
570 | kiocb_end_write(kiocb); | |
571 | ||
6b06314c | 572 | io_fput(req); |
2b188cc1 JA |
573 | io_cqring_add_event(req->ctx, req->user_data, res, 0); |
574 | io_free_req(req); | |
575 | } | |
576 | ||
def596e9 JA |
577 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
578 | { | |
579 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
580 | ||
581 | kiocb_end_write(kiocb); | |
582 | ||
583 | req->error = res; | |
584 | if (res != -EAGAIN) | |
585 | req->flags |= REQ_F_IOPOLL_COMPLETED; | |
586 | } | |
587 | ||
588 | /* | |
589 | * After the iocb has been issued, it's safe to be found on the poll list. | |
590 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
591 | * find it from a io_iopoll_getevents() thread before the issuer is done | |
592 | * accessing the kiocb cookie. | |
593 | */ | |
594 | static void io_iopoll_req_issued(struct io_kiocb *req) | |
595 | { | |
596 | struct io_ring_ctx *ctx = req->ctx; | |
597 | ||
598 | /* | |
599 | * Track whether we have multiple files in our lists. This will impact | |
600 | * how we do polling eventually, not spinning if we're on potentially | |
601 | * different devices. | |
602 | */ | |
603 | if (list_empty(&ctx->poll_list)) { | |
604 | ctx->poll_multi_file = false; | |
605 | } else if (!ctx->poll_multi_file) { | |
606 | struct io_kiocb *list_req; | |
607 | ||
608 | list_req = list_first_entry(&ctx->poll_list, struct io_kiocb, | |
609 | list); | |
610 | if (list_req->rw.ki_filp != req->rw.ki_filp) | |
611 | ctx->poll_multi_file = true; | |
612 | } | |
613 | ||
614 | /* | |
615 | * For fast devices, IO may have already completed. If it has, add | |
616 | * it to the front so we find it first. | |
617 | */ | |
618 | if (req->flags & REQ_F_IOPOLL_COMPLETED) | |
619 | list_add(&req->list, &ctx->poll_list); | |
620 | else | |
621 | list_add_tail(&req->list, &ctx->poll_list); | |
622 | } | |
623 | ||
9a56a232 JA |
624 | static void io_file_put(struct io_submit_state *state, struct file *file) |
625 | { | |
626 | if (!state) { | |
627 | fput(file); | |
628 | } else if (state->file) { | |
629 | int diff = state->has_refs - state->used_refs; | |
630 | ||
631 | if (diff) | |
632 | fput_many(state->file, diff); | |
633 | state->file = NULL; | |
634 | } | |
635 | } | |
636 | ||
637 | /* | |
638 | * Get as many references to a file as we have IOs left in this submission, | |
639 | * assuming most submissions are for one file, or at least that each file | |
640 | * has more than one submission. | |
641 | */ | |
642 | static struct file *io_file_get(struct io_submit_state *state, int fd) | |
643 | { | |
644 | if (!state) | |
645 | return fget(fd); | |
646 | ||
647 | if (state->file) { | |
648 | if (state->fd == fd) { | |
649 | state->used_refs++; | |
650 | state->ios_left--; | |
651 | return state->file; | |
652 | } | |
653 | io_file_put(state, NULL); | |
654 | } | |
655 | state->file = fget_many(fd, state->ios_left); | |
656 | if (!state->file) | |
657 | return NULL; | |
658 | ||
659 | state->fd = fd; | |
660 | state->has_refs = state->ios_left; | |
661 | state->used_refs = 1; | |
662 | state->ios_left--; | |
663 | return state->file; | |
664 | } | |
665 | ||
2b188cc1 JA |
666 | /* |
667 | * If we tracked the file through the SCM inflight mechanism, we could support | |
668 | * any file. For now, just ensure that anything potentially problematic is done | |
669 | * inline. | |
670 | */ | |
671 | static bool io_file_supports_async(struct file *file) | |
672 | { | |
673 | umode_t mode = file_inode(file)->i_mode; | |
674 | ||
675 | if (S_ISBLK(mode) || S_ISCHR(mode)) | |
676 | return true; | |
677 | if (S_ISREG(mode) && file->f_op != &io_uring_fops) | |
678 | return true; | |
679 | ||
680 | return false; | |
681 | } | |
682 | ||
683 | static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
9a56a232 | 684 | bool force_nonblock, struct io_submit_state *state) |
2b188cc1 | 685 | { |
def596e9 | 686 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 | 687 | struct kiocb *kiocb = &req->rw; |
6b06314c | 688 | unsigned ioprio, flags; |
2b188cc1 JA |
689 | int fd, ret; |
690 | ||
691 | /* For -EAGAIN retry, everything is already prepped */ | |
692 | if (kiocb->ki_filp) | |
693 | return 0; | |
694 | ||
6b06314c | 695 | flags = READ_ONCE(sqe->flags); |
2b188cc1 | 696 | fd = READ_ONCE(sqe->fd); |
6b06314c JA |
697 | |
698 | if (flags & IOSQE_FIXED_FILE) { | |
699 | if (unlikely(!ctx->user_files || | |
700 | (unsigned) fd >= ctx->nr_user_files)) | |
701 | return -EBADF; | |
702 | kiocb->ki_filp = ctx->user_files[fd]; | |
703 | req->flags |= REQ_F_FIXED_FILE; | |
704 | } else { | |
705 | kiocb->ki_filp = io_file_get(state, fd); | |
706 | if (unlikely(!kiocb->ki_filp)) | |
707 | return -EBADF; | |
708 | if (force_nonblock && !io_file_supports_async(kiocb->ki_filp)) | |
709 | force_nonblock = false; | |
710 | } | |
2b188cc1 JA |
711 | kiocb->ki_pos = READ_ONCE(sqe->off); |
712 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); | |
713 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); | |
714 | ||
715 | ioprio = READ_ONCE(sqe->ioprio); | |
716 | if (ioprio) { | |
717 | ret = ioprio_check_cap(ioprio); | |
718 | if (ret) | |
719 | goto out_fput; | |
720 | ||
721 | kiocb->ki_ioprio = ioprio; | |
722 | } else | |
723 | kiocb->ki_ioprio = get_current_ioprio(); | |
724 | ||
725 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
726 | if (unlikely(ret)) | |
727 | goto out_fput; | |
728 | if (force_nonblock) { | |
729 | kiocb->ki_flags |= IOCB_NOWAIT; | |
730 | req->flags |= REQ_F_FORCE_NONBLOCK; | |
731 | } | |
def596e9 JA |
732 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
733 | ret = -EOPNOTSUPP; | |
734 | if (!(kiocb->ki_flags & IOCB_DIRECT) || | |
735 | !kiocb->ki_filp->f_op->iopoll) | |
736 | goto out_fput; | |
2b188cc1 | 737 | |
def596e9 JA |
738 | req->error = 0; |
739 | kiocb->ki_flags |= IOCB_HIPRI; | |
740 | kiocb->ki_complete = io_complete_rw_iopoll; | |
741 | } else { | |
742 | if (kiocb->ki_flags & IOCB_HIPRI) { | |
743 | ret = -EINVAL; | |
744 | goto out_fput; | |
745 | } | |
746 | kiocb->ki_complete = io_complete_rw; | |
747 | } | |
2b188cc1 JA |
748 | return 0; |
749 | out_fput: | |
6b06314c JA |
750 | if (!(flags & IOSQE_FIXED_FILE)) { |
751 | /* | |
752 | * in case of error, we didn't use this file reference. drop it. | |
753 | */ | |
754 | if (state) | |
755 | state->used_refs--; | |
756 | io_file_put(state, kiocb->ki_filp); | |
757 | } | |
2b188cc1 JA |
758 | return ret; |
759 | } | |
760 | ||
761 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
762 | { | |
763 | switch (ret) { | |
764 | case -EIOCBQUEUED: | |
765 | break; | |
766 | case -ERESTARTSYS: | |
767 | case -ERESTARTNOINTR: | |
768 | case -ERESTARTNOHAND: | |
769 | case -ERESTART_RESTARTBLOCK: | |
770 | /* | |
771 | * We can't just restart the syscall, since previously | |
772 | * submitted sqes may already be in progress. Just fail this | |
773 | * IO with EINTR. | |
774 | */ | |
775 | ret = -EINTR; | |
776 | /* fall through */ | |
777 | default: | |
778 | kiocb->ki_complete(kiocb, ret, 0); | |
779 | } | |
780 | } | |
781 | ||
edafccee JA |
782 | static int io_import_fixed(struct io_ring_ctx *ctx, int rw, |
783 | const struct io_uring_sqe *sqe, | |
784 | struct iov_iter *iter) | |
785 | { | |
786 | size_t len = READ_ONCE(sqe->len); | |
787 | struct io_mapped_ubuf *imu; | |
788 | unsigned index, buf_index; | |
789 | size_t offset; | |
790 | u64 buf_addr; | |
791 | ||
792 | /* attempt to use fixed buffers without having provided iovecs */ | |
793 | if (unlikely(!ctx->user_bufs)) | |
794 | return -EFAULT; | |
795 | ||
796 | buf_index = READ_ONCE(sqe->buf_index); | |
797 | if (unlikely(buf_index >= ctx->nr_user_bufs)) | |
798 | return -EFAULT; | |
799 | ||
800 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); | |
801 | imu = &ctx->user_bufs[index]; | |
802 | buf_addr = READ_ONCE(sqe->addr); | |
803 | ||
804 | /* overflow */ | |
805 | if (buf_addr + len < buf_addr) | |
806 | return -EFAULT; | |
807 | /* not inside the mapped region */ | |
808 | if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len) | |
809 | return -EFAULT; | |
810 | ||
811 | /* | |
812 | * May not be a start of buffer, set size appropriately | |
813 | * and advance us to the beginning. | |
814 | */ | |
815 | offset = buf_addr - imu->ubuf; | |
816 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
817 | if (offset) | |
818 | iov_iter_advance(iter, offset); | |
819 | return 0; | |
820 | } | |
821 | ||
2b188cc1 JA |
822 | static int io_import_iovec(struct io_ring_ctx *ctx, int rw, |
823 | const struct sqe_submit *s, struct iovec **iovec, | |
824 | struct iov_iter *iter) | |
825 | { | |
826 | const struct io_uring_sqe *sqe = s->sqe; | |
827 | void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
828 | size_t sqe_len = READ_ONCE(sqe->len); | |
edafccee JA |
829 | u8 opcode; |
830 | ||
831 | /* | |
832 | * We're reading ->opcode for the second time, but the first read | |
833 | * doesn't care whether it's _FIXED or not, so it doesn't matter | |
834 | * whether ->opcode changes concurrently. The first read does care | |
835 | * about whether it is a READ or a WRITE, so we don't trust this read | |
836 | * for that purpose and instead let the caller pass in the read/write | |
837 | * flag. | |
838 | */ | |
839 | opcode = READ_ONCE(sqe->opcode); | |
840 | if (opcode == IORING_OP_READ_FIXED || | |
841 | opcode == IORING_OP_WRITE_FIXED) { | |
842 | ssize_t ret = io_import_fixed(ctx, rw, sqe, iter); | |
843 | *iovec = NULL; | |
844 | return ret; | |
845 | } | |
2b188cc1 JA |
846 | |
847 | if (!s->has_user) | |
848 | return -EFAULT; | |
849 | ||
850 | #ifdef CONFIG_COMPAT | |
851 | if (ctx->compat) | |
852 | return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV, | |
853 | iovec, iter); | |
854 | #endif | |
855 | ||
856 | return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter); | |
857 | } | |
858 | ||
859 | static ssize_t io_read(struct io_kiocb *req, const struct sqe_submit *s, | |
9a56a232 | 860 | bool force_nonblock, struct io_submit_state *state) |
2b188cc1 JA |
861 | { |
862 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
863 | struct kiocb *kiocb = &req->rw; | |
864 | struct iov_iter iter; | |
865 | struct file *file; | |
866 | ssize_t ret; | |
867 | ||
9a56a232 | 868 | ret = io_prep_rw(req, s->sqe, force_nonblock, state); |
2b188cc1 JA |
869 | if (ret) |
870 | return ret; | |
871 | file = kiocb->ki_filp; | |
872 | ||
873 | ret = -EBADF; | |
874 | if (unlikely(!(file->f_mode & FMODE_READ))) | |
875 | goto out_fput; | |
876 | ret = -EINVAL; | |
877 | if (unlikely(!file->f_op->read_iter)) | |
878 | goto out_fput; | |
879 | ||
880 | ret = io_import_iovec(req->ctx, READ, s, &iovec, &iter); | |
881 | if (ret) | |
882 | goto out_fput; | |
883 | ||
884 | ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_iter_count(&iter)); | |
885 | if (!ret) { | |
886 | ssize_t ret2; | |
887 | ||
888 | /* Catch -EAGAIN return for forced non-blocking submission */ | |
889 | ret2 = call_read_iter(file, kiocb, &iter); | |
890 | if (!force_nonblock || ret2 != -EAGAIN) | |
891 | io_rw_done(kiocb, ret2); | |
892 | else | |
893 | ret = -EAGAIN; | |
894 | } | |
895 | kfree(iovec); | |
896 | out_fput: | |
897 | /* Hold on to the file for -EAGAIN */ | |
898 | if (unlikely(ret && ret != -EAGAIN)) | |
6b06314c | 899 | io_fput(req); |
2b188cc1 JA |
900 | return ret; |
901 | } | |
902 | ||
903 | static ssize_t io_write(struct io_kiocb *req, const struct sqe_submit *s, | |
9a56a232 | 904 | bool force_nonblock, struct io_submit_state *state) |
2b188cc1 JA |
905 | { |
906 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
907 | struct kiocb *kiocb = &req->rw; | |
908 | struct iov_iter iter; | |
909 | struct file *file; | |
910 | ssize_t ret; | |
911 | ||
9a56a232 | 912 | ret = io_prep_rw(req, s->sqe, force_nonblock, state); |
2b188cc1 JA |
913 | if (ret) |
914 | return ret; | |
915 | /* Hold on to the file for -EAGAIN */ | |
916 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT)) | |
917 | return -EAGAIN; | |
918 | ||
919 | ret = -EBADF; | |
920 | file = kiocb->ki_filp; | |
921 | if (unlikely(!(file->f_mode & FMODE_WRITE))) | |
922 | goto out_fput; | |
923 | ret = -EINVAL; | |
924 | if (unlikely(!file->f_op->write_iter)) | |
925 | goto out_fput; | |
926 | ||
927 | ret = io_import_iovec(req->ctx, WRITE, s, &iovec, &iter); | |
928 | if (ret) | |
929 | goto out_fput; | |
930 | ||
931 | ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, | |
932 | iov_iter_count(&iter)); | |
933 | if (!ret) { | |
934 | /* | |
935 | * Open-code file_start_write here to grab freeze protection, | |
936 | * which will be released by another thread in | |
937 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
938 | * released so that it doesn't complain about the held lock when | |
939 | * we return to userspace. | |
940 | */ | |
941 | if (S_ISREG(file_inode(file)->i_mode)) { | |
942 | __sb_start_write(file_inode(file)->i_sb, | |
943 | SB_FREEZE_WRITE, true); | |
944 | __sb_writers_release(file_inode(file)->i_sb, | |
945 | SB_FREEZE_WRITE); | |
946 | } | |
947 | kiocb->ki_flags |= IOCB_WRITE; | |
948 | io_rw_done(kiocb, call_write_iter(file, kiocb, &iter)); | |
949 | } | |
950 | kfree(iovec); | |
951 | out_fput: | |
952 | if (unlikely(ret)) | |
6b06314c | 953 | io_fput(req); |
2b188cc1 JA |
954 | return ret; |
955 | } | |
956 | ||
957 | /* | |
958 | * IORING_OP_NOP just posts a completion event, nothing else. | |
959 | */ | |
960 | static int io_nop(struct io_kiocb *req, u64 user_data) | |
961 | { | |
962 | struct io_ring_ctx *ctx = req->ctx; | |
963 | long err = 0; | |
964 | ||
def596e9 JA |
965 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
966 | return -EINVAL; | |
967 | ||
2b188cc1 JA |
968 | /* |
969 | * Twilight zone - it's possible that someone issued an opcode that | |
970 | * has a file attached, then got -EAGAIN on submission, and changed | |
971 | * the sqe before we retried it from async context. Avoid dropping | |
972 | * a file reference for this malicious case, and flag the error. | |
973 | */ | |
974 | if (req->rw.ki_filp) { | |
975 | err = -EBADF; | |
6b06314c | 976 | io_fput(req); |
2b188cc1 JA |
977 | } |
978 | io_cqring_add_event(ctx, user_data, err, 0); | |
979 | io_free_req(req); | |
980 | return 0; | |
981 | } | |
982 | ||
c992fe29 CH |
983 | static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
984 | { | |
6b06314c JA |
985 | struct io_ring_ctx *ctx = req->ctx; |
986 | unsigned flags; | |
c992fe29 CH |
987 | int fd; |
988 | ||
989 | /* Prep already done */ | |
990 | if (req->rw.ki_filp) | |
991 | return 0; | |
992 | ||
6b06314c | 993 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 994 | return -EINVAL; |
edafccee | 995 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
996 | return -EINVAL; |
997 | ||
998 | fd = READ_ONCE(sqe->fd); | |
6b06314c JA |
999 | flags = READ_ONCE(sqe->flags); |
1000 | ||
1001 | if (flags & IOSQE_FIXED_FILE) { | |
1002 | if (unlikely(!ctx->user_files || fd >= ctx->nr_user_files)) | |
1003 | return -EBADF; | |
1004 | req->rw.ki_filp = ctx->user_files[fd]; | |
1005 | req->flags |= REQ_F_FIXED_FILE; | |
1006 | } else { | |
1007 | req->rw.ki_filp = fget(fd); | |
1008 | if (unlikely(!req->rw.ki_filp)) | |
1009 | return -EBADF; | |
1010 | } | |
c992fe29 CH |
1011 | |
1012 | return 0; | |
1013 | } | |
1014 | ||
1015 | static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
1016 | bool force_nonblock) | |
1017 | { | |
1018 | loff_t sqe_off = READ_ONCE(sqe->off); | |
1019 | loff_t sqe_len = READ_ONCE(sqe->len); | |
1020 | loff_t end = sqe_off + sqe_len; | |
1021 | unsigned fsync_flags; | |
1022 | int ret; | |
1023 | ||
1024 | fsync_flags = READ_ONCE(sqe->fsync_flags); | |
1025 | if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC)) | |
1026 | return -EINVAL; | |
1027 | ||
1028 | ret = io_prep_fsync(req, sqe); | |
1029 | if (ret) | |
1030 | return ret; | |
1031 | ||
1032 | /* fsync always requires a blocking context */ | |
1033 | if (force_nonblock) | |
1034 | return -EAGAIN; | |
1035 | ||
1036 | ret = vfs_fsync_range(req->rw.ki_filp, sqe_off, | |
1037 | end > 0 ? end : LLONG_MAX, | |
1038 | fsync_flags & IORING_FSYNC_DATASYNC); | |
1039 | ||
6b06314c | 1040 | io_fput(req); |
c992fe29 CH |
1041 | io_cqring_add_event(req->ctx, sqe->user_data, ret, 0); |
1042 | io_free_req(req); | |
1043 | return 0; | |
1044 | } | |
1045 | ||
2b188cc1 | 1046 | static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
9a56a232 JA |
1047 | const struct sqe_submit *s, bool force_nonblock, |
1048 | struct io_submit_state *state) | |
2b188cc1 JA |
1049 | { |
1050 | ssize_t ret; | |
1051 | int opcode; | |
1052 | ||
1053 | if (unlikely(s->index >= ctx->sq_entries)) | |
1054 | return -EINVAL; | |
1055 | req->user_data = READ_ONCE(s->sqe->user_data); | |
1056 | ||
1057 | opcode = READ_ONCE(s->sqe->opcode); | |
1058 | switch (opcode) { | |
1059 | case IORING_OP_NOP: | |
1060 | ret = io_nop(req, req->user_data); | |
1061 | break; | |
1062 | case IORING_OP_READV: | |
edafccee JA |
1063 | if (unlikely(s->sqe->buf_index)) |
1064 | return -EINVAL; | |
9a56a232 | 1065 | ret = io_read(req, s, force_nonblock, state); |
2b188cc1 JA |
1066 | break; |
1067 | case IORING_OP_WRITEV: | |
edafccee JA |
1068 | if (unlikely(s->sqe->buf_index)) |
1069 | return -EINVAL; | |
1070 | ret = io_write(req, s, force_nonblock, state); | |
1071 | break; | |
1072 | case IORING_OP_READ_FIXED: | |
1073 | ret = io_read(req, s, force_nonblock, state); | |
1074 | break; | |
1075 | case IORING_OP_WRITE_FIXED: | |
9a56a232 | 1076 | ret = io_write(req, s, force_nonblock, state); |
2b188cc1 | 1077 | break; |
c992fe29 CH |
1078 | case IORING_OP_FSYNC: |
1079 | ret = io_fsync(req, s->sqe, force_nonblock); | |
1080 | break; | |
2b188cc1 JA |
1081 | default: |
1082 | ret = -EINVAL; | |
1083 | break; | |
1084 | } | |
1085 | ||
def596e9 JA |
1086 | if (ret) |
1087 | return ret; | |
1088 | ||
1089 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
1090 | if (req->error == -EAGAIN) | |
1091 | return -EAGAIN; | |
1092 | ||
1093 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
1094 | if (s->needs_lock) | |
1095 | mutex_lock(&ctx->uring_lock); | |
1096 | io_iopoll_req_issued(req); | |
1097 | if (s->needs_lock) | |
1098 | mutex_unlock(&ctx->uring_lock); | |
1099 | } | |
1100 | ||
1101 | return 0; | |
2b188cc1 JA |
1102 | } |
1103 | ||
edafccee JA |
1104 | static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe) |
1105 | { | |
1106 | u8 opcode = READ_ONCE(sqe->opcode); | |
1107 | ||
1108 | return !(opcode == IORING_OP_READ_FIXED || | |
1109 | opcode == IORING_OP_WRITE_FIXED); | |
1110 | } | |
1111 | ||
2b188cc1 JA |
1112 | static void io_sq_wq_submit_work(struct work_struct *work) |
1113 | { | |
1114 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
1115 | struct sqe_submit *s = &req->submit; | |
1116 | const struct io_uring_sqe *sqe = s->sqe; | |
1117 | struct io_ring_ctx *ctx = req->ctx; | |
edafccee JA |
1118 | mm_segment_t old_fs; |
1119 | bool needs_user; | |
2b188cc1 JA |
1120 | int ret; |
1121 | ||
1122 | /* Ensure we clear previously set forced non-block flag */ | |
1123 | req->flags &= ~REQ_F_FORCE_NONBLOCK; | |
1124 | req->rw.ki_flags &= ~IOCB_NOWAIT; | |
1125 | ||
def596e9 | 1126 | s->needs_lock = true; |
edafccee JA |
1127 | s->has_user = false; |
1128 | ||
1129 | /* | |
1130 | * If we're doing IO to fixed buffers, we don't need to get/set | |
1131 | * user context | |
1132 | */ | |
1133 | needs_user = io_sqe_needs_user(s->sqe); | |
1134 | if (needs_user) { | |
1135 | if (!mmget_not_zero(ctx->sqo_mm)) { | |
1136 | ret = -EFAULT; | |
1137 | goto err; | |
1138 | } | |
1139 | use_mm(ctx->sqo_mm); | |
1140 | old_fs = get_fs(); | |
1141 | set_fs(USER_DS); | |
1142 | s->has_user = true; | |
1143 | } | |
2b188cc1 | 1144 | |
def596e9 | 1145 | do { |
9a56a232 | 1146 | ret = __io_submit_sqe(ctx, req, s, false, NULL); |
def596e9 JA |
1147 | /* |
1148 | * We can get EAGAIN for polled IO even though we're forcing | |
1149 | * a sync submission from here, since we can't wait for | |
1150 | * request slots on the block side. | |
1151 | */ | |
1152 | if (ret != -EAGAIN) | |
1153 | break; | |
1154 | cond_resched(); | |
1155 | } while (1); | |
2b188cc1 | 1156 | |
edafccee JA |
1157 | if (needs_user) { |
1158 | set_fs(old_fs); | |
1159 | unuse_mm(ctx->sqo_mm); | |
1160 | mmput(ctx->sqo_mm); | |
1161 | } | |
2b188cc1 JA |
1162 | err: |
1163 | if (ret) { | |
1164 | io_cqring_add_event(ctx, sqe->user_data, ret, 0); | |
1165 | io_free_req(req); | |
1166 | } | |
1167 | ||
1168 | /* async context always use a copy of the sqe */ | |
1169 | kfree(sqe); | |
1170 | } | |
1171 | ||
9a56a232 JA |
1172 | static int io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s, |
1173 | struct io_submit_state *state) | |
2b188cc1 JA |
1174 | { |
1175 | struct io_kiocb *req; | |
1176 | ssize_t ret; | |
1177 | ||
1178 | /* enforce forwards compatibility on users */ | |
6b06314c | 1179 | if (unlikely(s->sqe->flags & ~IOSQE_FIXED_FILE)) |
2b188cc1 JA |
1180 | return -EINVAL; |
1181 | ||
2579f913 | 1182 | req = io_get_req(ctx, state); |
2b188cc1 JA |
1183 | if (unlikely(!req)) |
1184 | return -EAGAIN; | |
1185 | ||
1186 | req->rw.ki_filp = NULL; | |
1187 | ||
9a56a232 | 1188 | ret = __io_submit_sqe(ctx, req, s, true, state); |
2b188cc1 JA |
1189 | if (ret == -EAGAIN) { |
1190 | struct io_uring_sqe *sqe_copy; | |
1191 | ||
1192 | sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL); | |
1193 | if (sqe_copy) { | |
1194 | memcpy(sqe_copy, s->sqe, sizeof(*sqe_copy)); | |
1195 | s->sqe = sqe_copy; | |
1196 | ||
1197 | memcpy(&req->submit, s, sizeof(*s)); | |
1198 | INIT_WORK(&req->work, io_sq_wq_submit_work); | |
1199 | queue_work(ctx->sqo_wq, &req->work); | |
1200 | ret = 0; | |
1201 | } | |
1202 | } | |
1203 | if (ret) | |
1204 | io_free_req(req); | |
1205 | ||
1206 | return ret; | |
1207 | } | |
1208 | ||
9a56a232 JA |
1209 | /* |
1210 | * Batched submission is done, ensure local IO is flushed out. | |
1211 | */ | |
1212 | static void io_submit_state_end(struct io_submit_state *state) | |
1213 | { | |
1214 | blk_finish_plug(&state->plug); | |
1215 | io_file_put(state, NULL); | |
2579f913 JA |
1216 | if (state->free_reqs) |
1217 | kmem_cache_free_bulk(req_cachep, state->free_reqs, | |
1218 | &state->reqs[state->cur_req]); | |
9a56a232 JA |
1219 | } |
1220 | ||
1221 | /* | |
1222 | * Start submission side cache. | |
1223 | */ | |
1224 | static void io_submit_state_start(struct io_submit_state *state, | |
1225 | struct io_ring_ctx *ctx, unsigned max_ios) | |
1226 | { | |
1227 | blk_start_plug(&state->plug); | |
2579f913 | 1228 | state->free_reqs = 0; |
9a56a232 JA |
1229 | state->file = NULL; |
1230 | state->ios_left = max_ios; | |
1231 | } | |
1232 | ||
2b188cc1 JA |
1233 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
1234 | { | |
1235 | struct io_sq_ring *ring = ctx->sq_ring; | |
1236 | ||
1237 | if (ctx->cached_sq_head != READ_ONCE(ring->r.head)) { | |
1238 | /* | |
1239 | * Ensure any loads from the SQEs are done at this point, | |
1240 | * since once we write the new head, the application could | |
1241 | * write new data to them. | |
1242 | */ | |
1243 | smp_store_release(&ring->r.head, ctx->cached_sq_head); | |
1244 | ||
1245 | /* | |
1246 | * write side barrier of head update, app has read side. See | |
1247 | * comment at the top of this file | |
1248 | */ | |
1249 | smp_wmb(); | |
1250 | } | |
1251 | } | |
1252 | ||
1253 | /* | |
1254 | * Undo last io_get_sqring() | |
1255 | */ | |
1256 | static void io_drop_sqring(struct io_ring_ctx *ctx) | |
1257 | { | |
1258 | ctx->cached_sq_head--; | |
1259 | } | |
1260 | ||
1261 | /* | |
1262 | * Fetch an sqe, if one is available. Note that s->sqe will point to memory | |
1263 | * that is mapped by userspace. This means that care needs to be taken to | |
1264 | * ensure that reads are stable, as we cannot rely on userspace always | |
1265 | * being a good citizen. If members of the sqe are validated and then later | |
1266 | * used, it's important that those reads are done through READ_ONCE() to | |
1267 | * prevent a re-load down the line. | |
1268 | */ | |
1269 | static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s) | |
1270 | { | |
1271 | struct io_sq_ring *ring = ctx->sq_ring; | |
1272 | unsigned head; | |
1273 | ||
1274 | /* | |
1275 | * The cached sq head (or cq tail) serves two purposes: | |
1276 | * | |
1277 | * 1) allows us to batch the cost of updating the user visible | |
1278 | * head updates. | |
1279 | * 2) allows the kernel side to track the head on its own, even | |
1280 | * though the application is the one updating it. | |
1281 | */ | |
1282 | head = ctx->cached_sq_head; | |
1283 | /* See comment at the top of this file */ | |
1284 | smp_rmb(); | |
1285 | if (head == READ_ONCE(ring->r.tail)) | |
1286 | return false; | |
1287 | ||
1288 | head = READ_ONCE(ring->array[head & ctx->sq_mask]); | |
1289 | if (head < ctx->sq_entries) { | |
1290 | s->index = head; | |
1291 | s->sqe = &ctx->sq_sqes[head]; | |
1292 | ctx->cached_sq_head++; | |
1293 | return true; | |
1294 | } | |
1295 | ||
1296 | /* drop invalid entries */ | |
1297 | ctx->cached_sq_head++; | |
1298 | ring->dropped++; | |
1299 | /* See comment at the top of this file */ | |
1300 | smp_wmb(); | |
1301 | return false; | |
1302 | } | |
1303 | ||
1304 | static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit) | |
1305 | { | |
9a56a232 | 1306 | struct io_submit_state state, *statep = NULL; |
2b188cc1 | 1307 | int i, ret = 0, submit = 0; |
2b188cc1 | 1308 | |
9a56a232 JA |
1309 | if (to_submit > IO_PLUG_THRESHOLD) { |
1310 | io_submit_state_start(&state, ctx, to_submit); | |
1311 | statep = &state; | |
1312 | } | |
2b188cc1 JA |
1313 | |
1314 | for (i = 0; i < to_submit; i++) { | |
1315 | struct sqe_submit s; | |
1316 | ||
1317 | if (!io_get_sqring(ctx, &s)) | |
1318 | break; | |
1319 | ||
1320 | s.has_user = true; | |
def596e9 JA |
1321 | s.needs_lock = false; |
1322 | ||
9a56a232 | 1323 | ret = io_submit_sqe(ctx, &s, statep); |
2b188cc1 JA |
1324 | if (ret) { |
1325 | io_drop_sqring(ctx); | |
1326 | break; | |
1327 | } | |
1328 | ||
1329 | submit++; | |
1330 | } | |
1331 | io_commit_sqring(ctx); | |
1332 | ||
9a56a232 JA |
1333 | if (statep) |
1334 | io_submit_state_end(statep); | |
2b188cc1 JA |
1335 | |
1336 | return submit ? submit : ret; | |
1337 | } | |
1338 | ||
1339 | static unsigned io_cqring_events(struct io_cq_ring *ring) | |
1340 | { | |
1341 | return READ_ONCE(ring->r.tail) - READ_ONCE(ring->r.head); | |
1342 | } | |
1343 | ||
1344 | /* | |
1345 | * Wait until events become available, if we don't already have some. The | |
1346 | * application must reap them itself, as they reside on the shared cq ring. | |
1347 | */ | |
1348 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
1349 | const sigset_t __user *sig, size_t sigsz) | |
1350 | { | |
1351 | struct io_cq_ring *ring = ctx->cq_ring; | |
1352 | sigset_t ksigmask, sigsaved; | |
1353 | DEFINE_WAIT(wait); | |
1354 | int ret; | |
1355 | ||
1356 | /* See comment at the top of this file */ | |
1357 | smp_rmb(); | |
1358 | if (io_cqring_events(ring) >= min_events) | |
1359 | return 0; | |
1360 | ||
1361 | if (sig) { | |
1362 | ret = set_user_sigmask(sig, &ksigmask, &sigsaved, sigsz); | |
1363 | if (ret) | |
1364 | return ret; | |
1365 | } | |
1366 | ||
1367 | do { | |
1368 | prepare_to_wait(&ctx->wait, &wait, TASK_INTERRUPTIBLE); | |
1369 | ||
1370 | ret = 0; | |
1371 | /* See comment at the top of this file */ | |
1372 | smp_rmb(); | |
1373 | if (io_cqring_events(ring) >= min_events) | |
1374 | break; | |
1375 | ||
1376 | schedule(); | |
1377 | ||
1378 | ret = -EINTR; | |
1379 | if (signal_pending(current)) | |
1380 | break; | |
1381 | } while (1); | |
1382 | ||
1383 | finish_wait(&ctx->wait, &wait); | |
1384 | ||
1385 | if (sig) | |
1386 | restore_user_sigmask(sig, &sigsaved); | |
1387 | ||
1388 | return READ_ONCE(ring->r.head) == READ_ONCE(ring->r.tail) ? ret : 0; | |
1389 | } | |
1390 | ||
6b06314c JA |
1391 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
1392 | { | |
1393 | #if defined(CONFIG_UNIX) | |
1394 | if (ctx->ring_sock) { | |
1395 | struct sock *sock = ctx->ring_sock->sk; | |
1396 | struct sk_buff *skb; | |
1397 | ||
1398 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
1399 | kfree_skb(skb); | |
1400 | } | |
1401 | #else | |
1402 | int i; | |
1403 | ||
1404 | for (i = 0; i < ctx->nr_user_files; i++) | |
1405 | fput(ctx->user_files[i]); | |
1406 | #endif | |
1407 | } | |
1408 | ||
1409 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) | |
1410 | { | |
1411 | if (!ctx->user_files) | |
1412 | return -ENXIO; | |
1413 | ||
1414 | __io_sqe_files_unregister(ctx); | |
1415 | kfree(ctx->user_files); | |
1416 | ctx->user_files = NULL; | |
1417 | ctx->nr_user_files = 0; | |
1418 | return 0; | |
1419 | } | |
1420 | ||
1421 | static void io_finish_async(struct io_ring_ctx *ctx) | |
1422 | { | |
1423 | if (ctx->sqo_wq) { | |
1424 | destroy_workqueue(ctx->sqo_wq); | |
1425 | ctx->sqo_wq = NULL; | |
1426 | } | |
1427 | } | |
1428 | ||
1429 | #if defined(CONFIG_UNIX) | |
1430 | static void io_destruct_skb(struct sk_buff *skb) | |
1431 | { | |
1432 | struct io_ring_ctx *ctx = skb->sk->sk_user_data; | |
1433 | ||
1434 | io_finish_async(ctx); | |
1435 | unix_destruct_scm(skb); | |
1436 | } | |
1437 | ||
1438 | /* | |
1439 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
1440 | * the io_uring can be safely unregistered on process exit, even if we have | |
1441 | * loops in the file referencing. | |
1442 | */ | |
1443 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
1444 | { | |
1445 | struct sock *sk = ctx->ring_sock->sk; | |
1446 | struct scm_fp_list *fpl; | |
1447 | struct sk_buff *skb; | |
1448 | int i; | |
1449 | ||
1450 | if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) { | |
1451 | unsigned long inflight = ctx->user->unix_inflight + nr; | |
1452 | ||
1453 | if (inflight > task_rlimit(current, RLIMIT_NOFILE)) | |
1454 | return -EMFILE; | |
1455 | } | |
1456 | ||
1457 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); | |
1458 | if (!fpl) | |
1459 | return -ENOMEM; | |
1460 | ||
1461 | skb = alloc_skb(0, GFP_KERNEL); | |
1462 | if (!skb) { | |
1463 | kfree(fpl); | |
1464 | return -ENOMEM; | |
1465 | } | |
1466 | ||
1467 | skb->sk = sk; | |
1468 | skb->destructor = io_destruct_skb; | |
1469 | ||
1470 | fpl->user = get_uid(ctx->user); | |
1471 | for (i = 0; i < nr; i++) { | |
1472 | fpl->fp[i] = get_file(ctx->user_files[i + offset]); | |
1473 | unix_inflight(fpl->user, fpl->fp[i]); | |
1474 | } | |
1475 | ||
1476 | fpl->max = fpl->count = nr; | |
1477 | UNIXCB(skb).fp = fpl; | |
1478 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); | |
1479 | skb_queue_head(&sk->sk_receive_queue, skb); | |
1480 | ||
1481 | for (i = 0; i < nr; i++) | |
1482 | fput(fpl->fp[i]); | |
1483 | ||
1484 | return 0; | |
1485 | } | |
1486 | ||
1487 | /* | |
1488 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
1489 | * causes regular reference counting to break down. We rely on the UNIX | |
1490 | * garbage collection to take care of this problem for us. | |
1491 | */ | |
1492 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
1493 | { | |
1494 | unsigned left, total; | |
1495 | int ret = 0; | |
1496 | ||
1497 | total = 0; | |
1498 | left = ctx->nr_user_files; | |
1499 | while (left) { | |
1500 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
1501 | int ret; | |
1502 | ||
1503 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
1504 | if (ret) | |
1505 | break; | |
1506 | left -= this_files; | |
1507 | total += this_files; | |
1508 | } | |
1509 | ||
1510 | if (!ret) | |
1511 | return 0; | |
1512 | ||
1513 | while (total < ctx->nr_user_files) { | |
1514 | fput(ctx->user_files[total]); | |
1515 | total++; | |
1516 | } | |
1517 | ||
1518 | return ret; | |
1519 | } | |
1520 | #else | |
1521 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
1522 | { | |
1523 | return 0; | |
1524 | } | |
1525 | #endif | |
1526 | ||
1527 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, | |
1528 | unsigned nr_args) | |
1529 | { | |
1530 | __s32 __user *fds = (__s32 __user *) arg; | |
1531 | int fd, ret = 0; | |
1532 | unsigned i; | |
1533 | ||
1534 | if (ctx->user_files) | |
1535 | return -EBUSY; | |
1536 | if (!nr_args) | |
1537 | return -EINVAL; | |
1538 | if (nr_args > IORING_MAX_FIXED_FILES) | |
1539 | return -EMFILE; | |
1540 | ||
1541 | ctx->user_files = kcalloc(nr_args, sizeof(struct file *), GFP_KERNEL); | |
1542 | if (!ctx->user_files) | |
1543 | return -ENOMEM; | |
1544 | ||
1545 | for (i = 0; i < nr_args; i++) { | |
1546 | ret = -EFAULT; | |
1547 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) | |
1548 | break; | |
1549 | ||
1550 | ctx->user_files[i] = fget(fd); | |
1551 | ||
1552 | ret = -EBADF; | |
1553 | if (!ctx->user_files[i]) | |
1554 | break; | |
1555 | /* | |
1556 | * Don't allow io_uring instances to be registered. If UNIX | |
1557 | * isn't enabled, then this causes a reference cycle and this | |
1558 | * instance can never get freed. If UNIX is enabled we'll | |
1559 | * handle it just fine, but there's still no point in allowing | |
1560 | * a ring fd as it doesn't support regular read/write anyway. | |
1561 | */ | |
1562 | if (ctx->user_files[i]->f_op == &io_uring_fops) { | |
1563 | fput(ctx->user_files[i]); | |
1564 | break; | |
1565 | } | |
1566 | ctx->nr_user_files++; | |
1567 | ret = 0; | |
1568 | } | |
1569 | ||
1570 | if (ret) { | |
1571 | for (i = 0; i < ctx->nr_user_files; i++) | |
1572 | fput(ctx->user_files[i]); | |
1573 | ||
1574 | kfree(ctx->user_files); | |
1575 | ctx->nr_user_files = 0; | |
1576 | return ret; | |
1577 | } | |
1578 | ||
1579 | ret = io_sqe_files_scm(ctx); | |
1580 | if (ret) | |
1581 | io_sqe_files_unregister(ctx); | |
1582 | ||
1583 | return ret; | |
1584 | } | |
1585 | ||
2b188cc1 JA |
1586 | static int io_sq_offload_start(struct io_ring_ctx *ctx) |
1587 | { | |
1588 | int ret; | |
1589 | ||
1590 | mmgrab(current->mm); | |
1591 | ctx->sqo_mm = current->mm; | |
1592 | ||
1593 | /* Do QD, or 2 * CPUS, whatever is smallest */ | |
1594 | ctx->sqo_wq = alloc_workqueue("io_ring-wq", WQ_UNBOUND | WQ_FREEZABLE, | |
1595 | min(ctx->sq_entries - 1, 2 * num_online_cpus())); | |
1596 | if (!ctx->sqo_wq) { | |
1597 | ret = -ENOMEM; | |
1598 | goto err; | |
1599 | } | |
1600 | ||
1601 | return 0; | |
1602 | err: | |
1603 | mmdrop(ctx->sqo_mm); | |
1604 | ctx->sqo_mm = NULL; | |
1605 | return ret; | |
1606 | } | |
1607 | ||
1608 | static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages) | |
1609 | { | |
1610 | atomic_long_sub(nr_pages, &user->locked_vm); | |
1611 | } | |
1612 | ||
1613 | static int io_account_mem(struct user_struct *user, unsigned long nr_pages) | |
1614 | { | |
1615 | unsigned long page_limit, cur_pages, new_pages; | |
1616 | ||
1617 | /* Don't allow more pages than we can safely lock */ | |
1618 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
1619 | ||
1620 | do { | |
1621 | cur_pages = atomic_long_read(&user->locked_vm); | |
1622 | new_pages = cur_pages + nr_pages; | |
1623 | if (new_pages > page_limit) | |
1624 | return -ENOMEM; | |
1625 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
1626 | new_pages) != cur_pages); | |
1627 | ||
1628 | return 0; | |
1629 | } | |
1630 | ||
1631 | static void io_mem_free(void *ptr) | |
1632 | { | |
1633 | struct page *page = virt_to_head_page(ptr); | |
1634 | ||
1635 | if (put_page_testzero(page)) | |
1636 | free_compound_page(page); | |
1637 | } | |
1638 | ||
1639 | static void *io_mem_alloc(size_t size) | |
1640 | { | |
1641 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
1642 | __GFP_NORETRY; | |
1643 | ||
1644 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
1645 | } | |
1646 | ||
1647 | static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries) | |
1648 | { | |
1649 | struct io_sq_ring *sq_ring; | |
1650 | struct io_cq_ring *cq_ring; | |
1651 | size_t bytes; | |
1652 | ||
1653 | bytes = struct_size(sq_ring, array, sq_entries); | |
1654 | bytes += array_size(sizeof(struct io_uring_sqe), sq_entries); | |
1655 | bytes += struct_size(cq_ring, cqes, cq_entries); | |
1656 | ||
1657 | return (bytes + PAGE_SIZE - 1) / PAGE_SIZE; | |
1658 | } | |
1659 | ||
edafccee JA |
1660 | static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx) |
1661 | { | |
1662 | int i, j; | |
1663 | ||
1664 | if (!ctx->user_bufs) | |
1665 | return -ENXIO; | |
1666 | ||
1667 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
1668 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
1669 | ||
1670 | for (j = 0; j < imu->nr_bvecs; j++) | |
1671 | put_page(imu->bvec[j].bv_page); | |
1672 | ||
1673 | if (ctx->account_mem) | |
1674 | io_unaccount_mem(ctx->user, imu->nr_bvecs); | |
1675 | kfree(imu->bvec); | |
1676 | imu->nr_bvecs = 0; | |
1677 | } | |
1678 | ||
1679 | kfree(ctx->user_bufs); | |
1680 | ctx->user_bufs = NULL; | |
1681 | ctx->nr_user_bufs = 0; | |
1682 | return 0; | |
1683 | } | |
1684 | ||
1685 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
1686 | void __user *arg, unsigned index) | |
1687 | { | |
1688 | struct iovec __user *src; | |
1689 | ||
1690 | #ifdef CONFIG_COMPAT | |
1691 | if (ctx->compat) { | |
1692 | struct compat_iovec __user *ciovs; | |
1693 | struct compat_iovec ciov; | |
1694 | ||
1695 | ciovs = (struct compat_iovec __user *) arg; | |
1696 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
1697 | return -EFAULT; | |
1698 | ||
1699 | dst->iov_base = (void __user *) (unsigned long) ciov.iov_base; | |
1700 | dst->iov_len = ciov.iov_len; | |
1701 | return 0; | |
1702 | } | |
1703 | #endif | |
1704 | src = (struct iovec __user *) arg; | |
1705 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
1706 | return -EFAULT; | |
1707 | return 0; | |
1708 | } | |
1709 | ||
1710 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, | |
1711 | unsigned nr_args) | |
1712 | { | |
1713 | struct vm_area_struct **vmas = NULL; | |
1714 | struct page **pages = NULL; | |
1715 | int i, j, got_pages = 0; | |
1716 | int ret = -EINVAL; | |
1717 | ||
1718 | if (ctx->user_bufs) | |
1719 | return -EBUSY; | |
1720 | if (!nr_args || nr_args > UIO_MAXIOV) | |
1721 | return -EINVAL; | |
1722 | ||
1723 | ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf), | |
1724 | GFP_KERNEL); | |
1725 | if (!ctx->user_bufs) | |
1726 | return -ENOMEM; | |
1727 | ||
1728 | for (i = 0; i < nr_args; i++) { | |
1729 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
1730 | unsigned long off, start, end, ubuf; | |
1731 | int pret, nr_pages; | |
1732 | struct iovec iov; | |
1733 | size_t size; | |
1734 | ||
1735 | ret = io_copy_iov(ctx, &iov, arg, i); | |
1736 | if (ret) | |
1737 | break; | |
1738 | ||
1739 | /* | |
1740 | * Don't impose further limits on the size and buffer | |
1741 | * constraints here, we'll -EINVAL later when IO is | |
1742 | * submitted if they are wrong. | |
1743 | */ | |
1744 | ret = -EFAULT; | |
1745 | if (!iov.iov_base || !iov.iov_len) | |
1746 | goto err; | |
1747 | ||
1748 | /* arbitrary limit, but we need something */ | |
1749 | if (iov.iov_len > SZ_1G) | |
1750 | goto err; | |
1751 | ||
1752 | ubuf = (unsigned long) iov.iov_base; | |
1753 | end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
1754 | start = ubuf >> PAGE_SHIFT; | |
1755 | nr_pages = end - start; | |
1756 | ||
1757 | if (ctx->account_mem) { | |
1758 | ret = io_account_mem(ctx->user, nr_pages); | |
1759 | if (ret) | |
1760 | goto err; | |
1761 | } | |
1762 | ||
1763 | ret = 0; | |
1764 | if (!pages || nr_pages > got_pages) { | |
1765 | kfree(vmas); | |
1766 | kfree(pages); | |
1767 | pages = kmalloc_array(nr_pages, sizeof(struct page *), | |
1768 | GFP_KERNEL); | |
1769 | vmas = kmalloc_array(nr_pages, | |
1770 | sizeof(struct vm_area_struct *), | |
1771 | GFP_KERNEL); | |
1772 | if (!pages || !vmas) { | |
1773 | ret = -ENOMEM; | |
1774 | if (ctx->account_mem) | |
1775 | io_unaccount_mem(ctx->user, nr_pages); | |
1776 | goto err; | |
1777 | } | |
1778 | got_pages = nr_pages; | |
1779 | } | |
1780 | ||
1781 | imu->bvec = kmalloc_array(nr_pages, sizeof(struct bio_vec), | |
1782 | GFP_KERNEL); | |
1783 | ret = -ENOMEM; | |
1784 | if (!imu->bvec) { | |
1785 | if (ctx->account_mem) | |
1786 | io_unaccount_mem(ctx->user, nr_pages); | |
1787 | goto err; | |
1788 | } | |
1789 | ||
1790 | ret = 0; | |
1791 | down_read(¤t->mm->mmap_sem); | |
1792 | pret = get_user_pages_longterm(ubuf, nr_pages, FOLL_WRITE, | |
1793 | pages, vmas); | |
1794 | if (pret == nr_pages) { | |
1795 | /* don't support file backed memory */ | |
1796 | for (j = 0; j < nr_pages; j++) { | |
1797 | struct vm_area_struct *vma = vmas[j]; | |
1798 | ||
1799 | if (vma->vm_file && | |
1800 | !is_file_hugepages(vma->vm_file)) { | |
1801 | ret = -EOPNOTSUPP; | |
1802 | break; | |
1803 | } | |
1804 | } | |
1805 | } else { | |
1806 | ret = pret < 0 ? pret : -EFAULT; | |
1807 | } | |
1808 | up_read(¤t->mm->mmap_sem); | |
1809 | if (ret) { | |
1810 | /* | |
1811 | * if we did partial map, or found file backed vmas, | |
1812 | * release any pages we did get | |
1813 | */ | |
1814 | if (pret > 0) { | |
1815 | for (j = 0; j < pret; j++) | |
1816 | put_page(pages[j]); | |
1817 | } | |
1818 | if (ctx->account_mem) | |
1819 | io_unaccount_mem(ctx->user, nr_pages); | |
1820 | goto err; | |
1821 | } | |
1822 | ||
1823 | off = ubuf & ~PAGE_MASK; | |
1824 | size = iov.iov_len; | |
1825 | for (j = 0; j < nr_pages; j++) { | |
1826 | size_t vec_len; | |
1827 | ||
1828 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
1829 | imu->bvec[j].bv_page = pages[j]; | |
1830 | imu->bvec[j].bv_len = vec_len; | |
1831 | imu->bvec[j].bv_offset = off; | |
1832 | off = 0; | |
1833 | size -= vec_len; | |
1834 | } | |
1835 | /* store original address for later verification */ | |
1836 | imu->ubuf = ubuf; | |
1837 | imu->len = iov.iov_len; | |
1838 | imu->nr_bvecs = nr_pages; | |
1839 | ||
1840 | ctx->nr_user_bufs++; | |
1841 | } | |
1842 | kfree(pages); | |
1843 | kfree(vmas); | |
1844 | return 0; | |
1845 | err: | |
1846 | kfree(pages); | |
1847 | kfree(vmas); | |
1848 | io_sqe_buffer_unregister(ctx); | |
1849 | return ret; | |
1850 | } | |
1851 | ||
2b188cc1 JA |
1852 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
1853 | { | |
6b06314c | 1854 | io_finish_async(ctx); |
2b188cc1 JA |
1855 | if (ctx->sqo_mm) |
1856 | mmdrop(ctx->sqo_mm); | |
def596e9 JA |
1857 | |
1858 | io_iopoll_reap_events(ctx); | |
edafccee | 1859 | io_sqe_buffer_unregister(ctx); |
6b06314c | 1860 | io_sqe_files_unregister(ctx); |
def596e9 | 1861 | |
2b188cc1 JA |
1862 | #if defined(CONFIG_UNIX) |
1863 | if (ctx->ring_sock) | |
1864 | sock_release(ctx->ring_sock); | |
1865 | #endif | |
1866 | ||
1867 | io_mem_free(ctx->sq_ring); | |
1868 | io_mem_free(ctx->sq_sqes); | |
1869 | io_mem_free(ctx->cq_ring); | |
1870 | ||
1871 | percpu_ref_exit(&ctx->refs); | |
1872 | if (ctx->account_mem) | |
1873 | io_unaccount_mem(ctx->user, | |
1874 | ring_pages(ctx->sq_entries, ctx->cq_entries)); | |
1875 | free_uid(ctx->user); | |
1876 | kfree(ctx); | |
1877 | } | |
1878 | ||
1879 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
1880 | { | |
1881 | struct io_ring_ctx *ctx = file->private_data; | |
1882 | __poll_t mask = 0; | |
1883 | ||
1884 | poll_wait(file, &ctx->cq_wait, wait); | |
1885 | /* See comment at the top of this file */ | |
1886 | smp_rmb(); | |
1887 | if (READ_ONCE(ctx->sq_ring->r.tail) + 1 != ctx->cached_sq_head) | |
1888 | mask |= EPOLLOUT | EPOLLWRNORM; | |
1889 | if (READ_ONCE(ctx->cq_ring->r.head) != ctx->cached_cq_tail) | |
1890 | mask |= EPOLLIN | EPOLLRDNORM; | |
1891 | ||
1892 | return mask; | |
1893 | } | |
1894 | ||
1895 | static int io_uring_fasync(int fd, struct file *file, int on) | |
1896 | { | |
1897 | struct io_ring_ctx *ctx = file->private_data; | |
1898 | ||
1899 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
1900 | } | |
1901 | ||
1902 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) | |
1903 | { | |
1904 | mutex_lock(&ctx->uring_lock); | |
1905 | percpu_ref_kill(&ctx->refs); | |
1906 | mutex_unlock(&ctx->uring_lock); | |
1907 | ||
def596e9 | 1908 | io_iopoll_reap_events(ctx); |
2b188cc1 JA |
1909 | wait_for_completion(&ctx->ctx_done); |
1910 | io_ring_ctx_free(ctx); | |
1911 | } | |
1912 | ||
1913 | static int io_uring_release(struct inode *inode, struct file *file) | |
1914 | { | |
1915 | struct io_ring_ctx *ctx = file->private_data; | |
1916 | ||
1917 | file->private_data = NULL; | |
1918 | io_ring_ctx_wait_and_kill(ctx); | |
1919 | return 0; | |
1920 | } | |
1921 | ||
1922 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
1923 | { | |
1924 | loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT; | |
1925 | unsigned long sz = vma->vm_end - vma->vm_start; | |
1926 | struct io_ring_ctx *ctx = file->private_data; | |
1927 | unsigned long pfn; | |
1928 | struct page *page; | |
1929 | void *ptr; | |
1930 | ||
1931 | switch (offset) { | |
1932 | case IORING_OFF_SQ_RING: | |
1933 | ptr = ctx->sq_ring; | |
1934 | break; | |
1935 | case IORING_OFF_SQES: | |
1936 | ptr = ctx->sq_sqes; | |
1937 | break; | |
1938 | case IORING_OFF_CQ_RING: | |
1939 | ptr = ctx->cq_ring; | |
1940 | break; | |
1941 | default: | |
1942 | return -EINVAL; | |
1943 | } | |
1944 | ||
1945 | page = virt_to_head_page(ptr); | |
1946 | if (sz > (PAGE_SIZE << compound_order(page))) | |
1947 | return -EINVAL; | |
1948 | ||
1949 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
1950 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
1951 | } | |
1952 | ||
1953 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, | |
1954 | u32, min_complete, u32, flags, const sigset_t __user *, sig, | |
1955 | size_t, sigsz) | |
1956 | { | |
1957 | struct io_ring_ctx *ctx; | |
1958 | long ret = -EBADF; | |
1959 | int submitted = 0; | |
1960 | struct fd f; | |
1961 | ||
1962 | if (flags & ~IORING_ENTER_GETEVENTS) | |
1963 | return -EINVAL; | |
1964 | ||
1965 | f = fdget(fd); | |
1966 | if (!f.file) | |
1967 | return -EBADF; | |
1968 | ||
1969 | ret = -EOPNOTSUPP; | |
1970 | if (f.file->f_op != &io_uring_fops) | |
1971 | goto out_fput; | |
1972 | ||
1973 | ret = -ENXIO; | |
1974 | ctx = f.file->private_data; | |
1975 | if (!percpu_ref_tryget(&ctx->refs)) | |
1976 | goto out_fput; | |
1977 | ||
1978 | ret = 0; | |
1979 | if (to_submit) { | |
1980 | to_submit = min(to_submit, ctx->sq_entries); | |
1981 | ||
1982 | mutex_lock(&ctx->uring_lock); | |
1983 | submitted = io_ring_submit(ctx, to_submit); | |
1984 | mutex_unlock(&ctx->uring_lock); | |
1985 | ||
1986 | if (submitted < 0) | |
1987 | goto out_ctx; | |
1988 | } | |
1989 | if (flags & IORING_ENTER_GETEVENTS) { | |
def596e9 JA |
1990 | unsigned nr_events = 0; |
1991 | ||
2b188cc1 JA |
1992 | min_complete = min(min_complete, ctx->cq_entries); |
1993 | ||
1994 | /* | |
1995 | * The application could have included the 'to_submit' count | |
1996 | * in how many events it wanted to wait for. If we failed to | |
1997 | * submit the desired count, we may need to adjust the number | |
1998 | * of events to poll/wait for. | |
1999 | */ | |
2000 | if (submitted < to_submit) | |
2001 | min_complete = min_t(unsigned, submitted, min_complete); | |
2002 | ||
def596e9 JA |
2003 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
2004 | mutex_lock(&ctx->uring_lock); | |
2005 | ret = io_iopoll_check(ctx, &nr_events, min_complete); | |
2006 | mutex_unlock(&ctx->uring_lock); | |
2007 | } else { | |
2008 | ret = io_cqring_wait(ctx, min_complete, sig, sigsz); | |
2009 | } | |
2b188cc1 JA |
2010 | } |
2011 | ||
2012 | out_ctx: | |
2013 | io_ring_drop_ctx_refs(ctx, 1); | |
2014 | out_fput: | |
2015 | fdput(f); | |
2016 | return submitted ? submitted : ret; | |
2017 | } | |
2018 | ||
2019 | static const struct file_operations io_uring_fops = { | |
2020 | .release = io_uring_release, | |
2021 | .mmap = io_uring_mmap, | |
2022 | .poll = io_uring_poll, | |
2023 | .fasync = io_uring_fasync, | |
2024 | }; | |
2025 | ||
2026 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
2027 | struct io_uring_params *p) | |
2028 | { | |
2029 | struct io_sq_ring *sq_ring; | |
2030 | struct io_cq_ring *cq_ring; | |
2031 | size_t size; | |
2032 | ||
2033 | sq_ring = io_mem_alloc(struct_size(sq_ring, array, p->sq_entries)); | |
2034 | if (!sq_ring) | |
2035 | return -ENOMEM; | |
2036 | ||
2037 | ctx->sq_ring = sq_ring; | |
2038 | sq_ring->ring_mask = p->sq_entries - 1; | |
2039 | sq_ring->ring_entries = p->sq_entries; | |
2040 | ctx->sq_mask = sq_ring->ring_mask; | |
2041 | ctx->sq_entries = sq_ring->ring_entries; | |
2042 | ||
2043 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
2044 | if (size == SIZE_MAX) | |
2045 | return -EOVERFLOW; | |
2046 | ||
2047 | ctx->sq_sqes = io_mem_alloc(size); | |
2048 | if (!ctx->sq_sqes) { | |
2049 | io_mem_free(ctx->sq_ring); | |
2050 | return -ENOMEM; | |
2051 | } | |
2052 | ||
2053 | cq_ring = io_mem_alloc(struct_size(cq_ring, cqes, p->cq_entries)); | |
2054 | if (!cq_ring) { | |
2055 | io_mem_free(ctx->sq_ring); | |
2056 | io_mem_free(ctx->sq_sqes); | |
2057 | return -ENOMEM; | |
2058 | } | |
2059 | ||
2060 | ctx->cq_ring = cq_ring; | |
2061 | cq_ring->ring_mask = p->cq_entries - 1; | |
2062 | cq_ring->ring_entries = p->cq_entries; | |
2063 | ctx->cq_mask = cq_ring->ring_mask; | |
2064 | ctx->cq_entries = cq_ring->ring_entries; | |
2065 | return 0; | |
2066 | } | |
2067 | ||
2068 | /* | |
2069 | * Allocate an anonymous fd, this is what constitutes the application | |
2070 | * visible backing of an io_uring instance. The application mmaps this | |
2071 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
2072 | * we have to tie this fd to a socket for file garbage collection purposes. | |
2073 | */ | |
2074 | static int io_uring_get_fd(struct io_ring_ctx *ctx) | |
2075 | { | |
2076 | struct file *file; | |
2077 | int ret; | |
2078 | ||
2079 | #if defined(CONFIG_UNIX) | |
2080 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, | |
2081 | &ctx->ring_sock); | |
2082 | if (ret) | |
2083 | return ret; | |
2084 | #endif | |
2085 | ||
2086 | ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
2087 | if (ret < 0) | |
2088 | goto err; | |
2089 | ||
2090 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, | |
2091 | O_RDWR | O_CLOEXEC); | |
2092 | if (IS_ERR(file)) { | |
2093 | put_unused_fd(ret); | |
2094 | ret = PTR_ERR(file); | |
2095 | goto err; | |
2096 | } | |
2097 | ||
2098 | #if defined(CONFIG_UNIX) | |
2099 | ctx->ring_sock->file = file; | |
6b06314c | 2100 | ctx->ring_sock->sk->sk_user_data = ctx; |
2b188cc1 JA |
2101 | #endif |
2102 | fd_install(ret, file); | |
2103 | return ret; | |
2104 | err: | |
2105 | #if defined(CONFIG_UNIX) | |
2106 | sock_release(ctx->ring_sock); | |
2107 | ctx->ring_sock = NULL; | |
2108 | #endif | |
2109 | return ret; | |
2110 | } | |
2111 | ||
2112 | static int io_uring_create(unsigned entries, struct io_uring_params *p) | |
2113 | { | |
2114 | struct user_struct *user = NULL; | |
2115 | struct io_ring_ctx *ctx; | |
2116 | bool account_mem; | |
2117 | int ret; | |
2118 | ||
2119 | if (!entries || entries > IORING_MAX_ENTRIES) | |
2120 | return -EINVAL; | |
2121 | ||
2122 | /* | |
2123 | * Use twice as many entries for the CQ ring. It's possible for the | |
2124 | * application to drive a higher depth than the size of the SQ ring, | |
2125 | * since the sqes are only used at submission time. This allows for | |
2126 | * some flexibility in overcommitting a bit. | |
2127 | */ | |
2128 | p->sq_entries = roundup_pow_of_two(entries); | |
2129 | p->cq_entries = 2 * p->sq_entries; | |
2130 | ||
2131 | user = get_uid(current_user()); | |
2132 | account_mem = !capable(CAP_IPC_LOCK); | |
2133 | ||
2134 | if (account_mem) { | |
2135 | ret = io_account_mem(user, | |
2136 | ring_pages(p->sq_entries, p->cq_entries)); | |
2137 | if (ret) { | |
2138 | free_uid(user); | |
2139 | return ret; | |
2140 | } | |
2141 | } | |
2142 | ||
2143 | ctx = io_ring_ctx_alloc(p); | |
2144 | if (!ctx) { | |
2145 | if (account_mem) | |
2146 | io_unaccount_mem(user, ring_pages(p->sq_entries, | |
2147 | p->cq_entries)); | |
2148 | free_uid(user); | |
2149 | return -ENOMEM; | |
2150 | } | |
2151 | ctx->compat = in_compat_syscall(); | |
2152 | ctx->account_mem = account_mem; | |
2153 | ctx->user = user; | |
2154 | ||
2155 | ret = io_allocate_scq_urings(ctx, p); | |
2156 | if (ret) | |
2157 | goto err; | |
2158 | ||
2159 | ret = io_sq_offload_start(ctx); | |
2160 | if (ret) | |
2161 | goto err; | |
2162 | ||
2163 | ret = io_uring_get_fd(ctx); | |
2164 | if (ret < 0) | |
2165 | goto err; | |
2166 | ||
2167 | memset(&p->sq_off, 0, sizeof(p->sq_off)); | |
2168 | p->sq_off.head = offsetof(struct io_sq_ring, r.head); | |
2169 | p->sq_off.tail = offsetof(struct io_sq_ring, r.tail); | |
2170 | p->sq_off.ring_mask = offsetof(struct io_sq_ring, ring_mask); | |
2171 | p->sq_off.ring_entries = offsetof(struct io_sq_ring, ring_entries); | |
2172 | p->sq_off.flags = offsetof(struct io_sq_ring, flags); | |
2173 | p->sq_off.dropped = offsetof(struct io_sq_ring, dropped); | |
2174 | p->sq_off.array = offsetof(struct io_sq_ring, array); | |
2175 | ||
2176 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
2177 | p->cq_off.head = offsetof(struct io_cq_ring, r.head); | |
2178 | p->cq_off.tail = offsetof(struct io_cq_ring, r.tail); | |
2179 | p->cq_off.ring_mask = offsetof(struct io_cq_ring, ring_mask); | |
2180 | p->cq_off.ring_entries = offsetof(struct io_cq_ring, ring_entries); | |
2181 | p->cq_off.overflow = offsetof(struct io_cq_ring, overflow); | |
2182 | p->cq_off.cqes = offsetof(struct io_cq_ring, cqes); | |
2183 | return ret; | |
2184 | err: | |
2185 | io_ring_ctx_wait_and_kill(ctx); | |
2186 | return ret; | |
2187 | } | |
2188 | ||
2189 | /* | |
2190 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
2191 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
2192 | * params structure passed in. | |
2193 | */ | |
2194 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
2195 | { | |
2196 | struct io_uring_params p; | |
2197 | long ret; | |
2198 | int i; | |
2199 | ||
2200 | if (copy_from_user(&p, params, sizeof(p))) | |
2201 | return -EFAULT; | |
2202 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
2203 | if (p.resv[i]) | |
2204 | return -EINVAL; | |
2205 | } | |
2206 | ||
def596e9 | 2207 | if (p.flags & ~IORING_SETUP_IOPOLL) |
2b188cc1 JA |
2208 | return -EINVAL; |
2209 | ||
2210 | ret = io_uring_create(entries, &p); | |
2211 | if (ret < 0) | |
2212 | return ret; | |
2213 | ||
2214 | if (copy_to_user(params, &p, sizeof(p))) | |
2215 | return -EFAULT; | |
2216 | ||
2217 | return ret; | |
2218 | } | |
2219 | ||
2220 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
2221 | struct io_uring_params __user *, params) | |
2222 | { | |
2223 | return io_uring_setup(entries, params); | |
2224 | } | |
2225 | ||
edafccee JA |
2226 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
2227 | void __user *arg, unsigned nr_args) | |
2228 | { | |
2229 | int ret; | |
2230 | ||
2231 | percpu_ref_kill(&ctx->refs); | |
2232 | wait_for_completion(&ctx->ctx_done); | |
2233 | ||
2234 | switch (opcode) { | |
2235 | case IORING_REGISTER_BUFFERS: | |
2236 | ret = io_sqe_buffer_register(ctx, arg, nr_args); | |
2237 | break; | |
2238 | case IORING_UNREGISTER_BUFFERS: | |
2239 | ret = -EINVAL; | |
2240 | if (arg || nr_args) | |
2241 | break; | |
2242 | ret = io_sqe_buffer_unregister(ctx); | |
2243 | break; | |
6b06314c JA |
2244 | case IORING_REGISTER_FILES: |
2245 | ret = io_sqe_files_register(ctx, arg, nr_args); | |
2246 | break; | |
2247 | case IORING_UNREGISTER_FILES: | |
2248 | ret = -EINVAL; | |
2249 | if (arg || nr_args) | |
2250 | break; | |
2251 | ret = io_sqe_files_unregister(ctx); | |
2252 | break; | |
edafccee JA |
2253 | default: |
2254 | ret = -EINVAL; | |
2255 | break; | |
2256 | } | |
2257 | ||
2258 | /* bring the ctx back to life */ | |
2259 | reinit_completion(&ctx->ctx_done); | |
2260 | percpu_ref_reinit(&ctx->refs); | |
2261 | return ret; | |
2262 | } | |
2263 | ||
2264 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
2265 | void __user *, arg, unsigned int, nr_args) | |
2266 | { | |
2267 | struct io_ring_ctx *ctx; | |
2268 | long ret = -EBADF; | |
2269 | struct fd f; | |
2270 | ||
2271 | f = fdget(fd); | |
2272 | if (!f.file) | |
2273 | return -EBADF; | |
2274 | ||
2275 | ret = -EOPNOTSUPP; | |
2276 | if (f.file->f_op != &io_uring_fops) | |
2277 | goto out_fput; | |
2278 | ||
2279 | ctx = f.file->private_data; | |
2280 | ||
2281 | mutex_lock(&ctx->uring_lock); | |
2282 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
2283 | mutex_unlock(&ctx->uring_lock); | |
2284 | out_fput: | |
2285 | fdput(f); | |
2286 | return ret; | |
2287 | } | |
2288 | ||
2b188cc1 JA |
2289 | static int __init io_uring_init(void) |
2290 | { | |
2291 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC); | |
2292 | return 0; | |
2293 | }; | |
2294 | __initcall(io_uring_init); |