]> git.ipfire.org Git - thirdparty/kernel/stable.git/blob - fs/io_uring.c
projects / thirdparty / kernel / stable.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
io_uring: add io_kiocb ref count
[thirdparty/kernel/stable.git] / fs / io_uring.c
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
27 * Copyright (c) 2018-2019 Christoph Hellwig
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/kthread.h>
48 #include <linux/blkdev.h>
49 #include <linux/bvec.h>
50 #include <linux/net.h>
51 #include <net/sock.h>
52 #include <net/af_unix.h>
53 #include <net/scm.h>
54 #include <linux/anon_inodes.h>
55 #include <linux/sched/mm.h>
56 #include <linux/uaccess.h>
57 #include <linux/nospec.h>
58 #include <linux/sizes.h>
59 #include <linux/hugetlb.h>
60
61 #include <uapi/linux/io_uring.h>
62
63 #include "internal.h"
64
65 #define IORING_MAX_ENTRIES 4096
66 #define IORING_MAX_FIXED_FILES 1024
67
68 struct io_uring {
69 u32 head ____cacheline_aligned_in_smp;
70 u32 tail ____cacheline_aligned_in_smp;
71 };
72
73 struct io_sq_ring {
74 struct io_uring r;
75 u32 ring_mask;
76 u32 ring_entries;
77 u32 dropped;
78 u32 flags;
79 u32 array[];
80 };
81
82 struct io_cq_ring {
83 struct io_uring r;
84 u32 ring_mask;
85 u32 ring_entries;
86 u32 overflow;
87 struct io_uring_cqe cqes[];
88 };
89
90 struct io_mapped_ubuf {
91 u64 ubuf;
92 size_t len;
93 struct bio_vec *bvec;
94 unsigned int nr_bvecs;
95 };
96
97 struct io_ring_ctx {
98 struct {
99 struct percpu_ref refs;
100 } ____cacheline_aligned_in_smp;
101
102 struct {
103 unsigned int flags;
104 bool compat;
105 bool account_mem;
106
107 /* SQ ring */
108 struct io_sq_ring *sq_ring;
109 unsigned cached_sq_head;
110 unsigned sq_entries;
111 unsigned sq_mask;
112 unsigned sq_thread_idle;
113 struct io_uring_sqe *sq_sqes;
114 } ____cacheline_aligned_in_smp;
115
116 /* IO offload */
117 struct workqueue_struct *sqo_wq;
118 struct task_struct *sqo_thread; /* if using sq thread polling */
119 struct mm_struct *sqo_mm;
120 wait_queue_head_t sqo_wait;
121 unsigned sqo_stop;
122
123 struct {
124 /* CQ ring */
125 struct io_cq_ring *cq_ring;
126 unsigned cached_cq_tail;
127 unsigned cq_entries;
128 unsigned cq_mask;
129 struct wait_queue_head cq_wait;
130 struct fasync_struct *cq_fasync;
131 } ____cacheline_aligned_in_smp;
132
133 /*
134 * If used, fixed file set. Writers must ensure that ->refs is dead,
135 * readers must ensure that ->refs is alive as long as the file* is
136 * used. Only updated through io_uring_register(2).
137 */
138 struct file **user_files;
139 unsigned nr_user_files;
140
141 /* if used, fixed mapped user buffers */
142 unsigned nr_user_bufs;
143 struct io_mapped_ubuf *user_bufs;
144
145 struct user_struct *user;
146
147 struct completion ctx_done;
148
149 struct {
150 struct mutex uring_lock;
151 wait_queue_head_t wait;
152 } ____cacheline_aligned_in_smp;
153
154 struct {
155 spinlock_t completion_lock;
156 bool poll_multi_file;
157 /*
158 * ->poll_list is protected by the ctx->uring_lock for
159 * io_uring instances that don't use IORING_SETUP_SQPOLL.
160 * For SQPOLL, only the single threaded io_sq_thread() will
161 * manipulate the list, hence no extra locking is needed there.
162 */
163 struct list_head poll_list;
164 } ____cacheline_aligned_in_smp;
165
166 #if defined(CONFIG_UNIX)
167 struct socket *ring_sock;
168 #endif
169 };
170
171 struct sqe_submit {
172 const struct io_uring_sqe *sqe;
173 unsigned short index;
174 bool has_user;
175 bool needs_lock;
176 bool needs_fixed_file;
177 };
178
179 struct io_kiocb {
180 struct kiocb rw;
181
182 struct sqe_submit submit;
183
184 struct io_ring_ctx *ctx;
185 struct list_head list;
186 unsigned int flags;
187 refcount_t refs;
188 #define REQ_F_FORCE_NONBLOCK 1 /* inline submission attempt */
189 #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */
190 #define REQ_F_FIXED_FILE 4 /* ctx owns file */
191 u64 user_data;
192 u64 error;
193
194 struct work_struct work;
195 };
196
197 #define IO_PLUG_THRESHOLD 2
198 #define IO_IOPOLL_BATCH 8
199
200 struct io_submit_state {
201 struct blk_plug plug;
202
203 /*
204 * io_kiocb alloc cache
205 */
206 void *reqs[IO_IOPOLL_BATCH];
207 unsigned int free_reqs;
208 unsigned int cur_req;
209
210 /*
211 * File reference cache
212 */
213 struct file *file;
214 unsigned int fd;
215 unsigned int has_refs;
216 unsigned int used_refs;
217 unsigned int ios_left;
218 };
219
220 static struct kmem_cache *req_cachep;
221
222 static const struct file_operations io_uring_fops;
223
224 struct sock *io_uring_get_socket(struct file *file)
225 {
226 #if defined(CONFIG_UNIX)
227 if (file->f_op == &io_uring_fops) {
228 struct io_ring_ctx *ctx = file->private_data;
229
230 return ctx->ring_sock->sk;
231 }
232 #endif
233 return NULL;
234 }
235 EXPORT_SYMBOL(io_uring_get_socket);
236
237 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
238 {
239 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
240
241 complete(&ctx->ctx_done);
242 }
243
244 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
245 {
246 struct io_ring_ctx *ctx;
247
248 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
249 if (!ctx)
250 return NULL;
251
252 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, 0, GFP_KERNEL)) {
253 kfree(ctx);
254 return NULL;
255 }
256
257 ctx->flags = p->flags;
258 init_waitqueue_head(&ctx->cq_wait);
259 init_completion(&ctx->ctx_done);
260 mutex_init(&ctx->uring_lock);
261 init_waitqueue_head(&ctx->wait);
262 spin_lock_init(&ctx->completion_lock);
263 INIT_LIST_HEAD(&ctx->poll_list);
264 return ctx;
265 }
266
267 static void io_commit_cqring(struct io_ring_ctx *ctx)
268 {
269 struct io_cq_ring *ring = ctx->cq_ring;
270
271 if (ctx->cached_cq_tail != READ_ONCE(ring->r.tail)) {
272 /* order cqe stores with ring update */
273 smp_store_release(&ring->r.tail, ctx->cached_cq_tail);
274
275 /*
276 * Write sider barrier of tail update, app has read side. See
277 * comment at the top of this file.
278 */
279 smp_wmb();
280
281 if (wq_has_sleeper(&ctx->cq_wait)) {
282 wake_up_interruptible(&ctx->cq_wait);
283 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
284 }
285 }
286 }
287
288 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
289 {
290 struct io_cq_ring *ring = ctx->cq_ring;
291 unsigned tail;
292
293 tail = ctx->cached_cq_tail;
294 /* See comment at the top of the file */
295 smp_rmb();
296 if (tail + 1 == READ_ONCE(ring->r.head))
297 return NULL;
298
299 ctx->cached_cq_tail++;
300 return &ring->cqes[tail & ctx->cq_mask];
301 }
302
303 static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data,
304 long res, unsigned ev_flags)
305 {
306 struct io_uring_cqe *cqe;
307
308 /*
309 * If we can't get a cq entry, userspace overflowed the
310 * submission (by quite a lot). Increment the overflow count in
311 * the ring.
312 */
313 cqe = io_get_cqring(ctx);
314 if (cqe) {
315 WRITE_ONCE(cqe->user_data, ki_user_data);
316 WRITE_ONCE(cqe->res, res);
317 WRITE_ONCE(cqe->flags, ev_flags);
318 } else {
319 unsigned overflow = READ_ONCE(ctx->cq_ring->overflow);
320
321 WRITE_ONCE(ctx->cq_ring->overflow, overflow + 1);
322 }
323 }
324
325 static void io_cqring_add_event(struct io_ring_ctx *ctx, u64 ki_user_data,
326 long res, unsigned ev_flags)
327 {
328 unsigned long flags;
329
330 spin_lock_irqsave(&ctx->completion_lock, flags);
331 io_cqring_fill_event(ctx, ki_user_data, res, ev_flags);
332 io_commit_cqring(ctx);
333 spin_unlock_irqrestore(&ctx->completion_lock, flags);
334
335 if (waitqueue_active(&ctx->wait))
336 wake_up(&ctx->wait);
337 if (waitqueue_active(&ctx->sqo_wait))
338 wake_up(&ctx->sqo_wait);
339 }
340
341 static void io_ring_drop_ctx_refs(struct io_ring_ctx *ctx, unsigned refs)
342 {
343 percpu_ref_put_many(&ctx->refs, refs);
344
345 if (waitqueue_active(&ctx->wait))
346 wake_up(&ctx->wait);
347 }
348
349 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
350 struct io_submit_state *state)
351 {
352 struct io_kiocb *req;
353
354 if (!percpu_ref_tryget(&ctx->refs))
355 return NULL;
356
357 if (!state) {
358 req = kmem_cache_alloc(req_cachep, __GFP_NOWARN);
359 if (unlikely(!req))
360 goto out;
361 } else if (!state->free_reqs) {
362 size_t sz;
363 int ret;
364
365 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
366 ret = kmem_cache_alloc_bulk(req_cachep, __GFP_NOWARN, sz,
367 state->reqs);
368 if (unlikely(ret <= 0))
369 goto out;
370 state->free_reqs = ret - 1;
371 state->cur_req = 1;
372 req = state->reqs[0];
373 } else {
374 req = state->reqs[state->cur_req];
375 state->free_reqs--;
376 state->cur_req++;
377 }
378
379 req->ctx = ctx;
380 req->flags = 0;
381 refcount_set(&req->refs, 0);
382 return req;
383 out:
384 io_ring_drop_ctx_refs(ctx, 1);
385 return NULL;
386 }
387
388 static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr)
389 {
390 if (*nr) {
391 kmem_cache_free_bulk(req_cachep, *nr, reqs);
392 io_ring_drop_ctx_refs(ctx, *nr);
393 *nr = 0;
394 }
395 }
396
397 static void io_free_req(struct io_kiocb *req)
398 {
399 if (!refcount_read(&req->refs) || refcount_dec_and_test(&req->refs)) {
400 io_ring_drop_ctx_refs(req->ctx, 1);
401 kmem_cache_free(req_cachep, req);
402 }
403 }
404
405 /*
406 * Find and free completed poll iocbs
407 */
408 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
409 struct list_head *done)
410 {
411 void *reqs[IO_IOPOLL_BATCH];
412 int file_count, to_free;
413 struct file *file = NULL;
414 struct io_kiocb *req;
415
416 file_count = to_free = 0;
417 while (!list_empty(done)) {
418 req = list_first_entry(done, struct io_kiocb, list);
419 list_del(&req->list);
420
421 io_cqring_fill_event(ctx, req->user_data, req->error, 0);
422
423 reqs[to_free++] = req;
424 (*nr_events)++;
425
426 /*
427 * Batched puts of the same file, to avoid dirtying the
428 * file usage count multiple times, if avoidable.
429 */
430 if (!(req->flags & REQ_F_FIXED_FILE)) {
431 if (!file) {
432 file = req->rw.ki_filp;
433 file_count = 1;
434 } else if (file == req->rw.ki_filp) {
435 file_count++;
436 } else {
437 fput_many(file, file_count);
438 file = req->rw.ki_filp;
439 file_count = 1;
440 }
441 }
442
443 if (to_free == ARRAY_SIZE(reqs))
444 io_free_req_many(ctx, reqs, &to_free);
445 }
446 io_commit_cqring(ctx);
447
448 if (file)
449 fput_many(file, file_count);
450 io_free_req_many(ctx, reqs, &to_free);
451 }
452
453 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
454 long min)
455 {
456 struct io_kiocb *req, *tmp;
457 LIST_HEAD(done);
458 bool spin;
459 int ret;
460
461 /*
462 * Only spin for completions if we don't have multiple devices hanging
463 * off our complete list, and we're under the requested amount.
464 */
465 spin = !ctx->poll_multi_file && *nr_events < min;
466
467 ret = 0;
468 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
469 struct kiocb *kiocb = &req->rw;
470
471 /*
472 * Move completed entries to our local list. If we find a
473 * request that requires polling, break out and complete
474 * the done list first, if we have entries there.
475 */
476 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
477 list_move_tail(&req->list, &done);
478 continue;
479 }
480 if (!list_empty(&done))
481 break;
482
483 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
484 if (ret < 0)
485 break;
486
487 if (ret && spin)
488 spin = false;
489 ret = 0;
490 }
491
492 if (!list_empty(&done))
493 io_iopoll_complete(ctx, nr_events, &done);
494
495 return ret;
496 }
497
498 /*
499 * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a
500 * non-spinning poll check - we'll still enter the driver poll loop, but only
501 * as a non-spinning completion check.
502 */
503 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
504 long min)
505 {
506 while (!list_empty(&ctx->poll_list)) {
507 int ret;
508
509 ret = io_do_iopoll(ctx, nr_events, min);
510 if (ret < 0)
511 return ret;
512 if (!min || *nr_events >= min)
513 return 0;
514 }
515
516 return 1;
517 }
518
519 /*
520 * We can't just wait for polled events to come to us, we have to actively
521 * find and complete them.
522 */
523 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
524 {
525 if (!(ctx->flags & IORING_SETUP_IOPOLL))
526 return;
527
528 mutex_lock(&ctx->uring_lock);
529 while (!list_empty(&ctx->poll_list)) {
530 unsigned int nr_events = 0;
531
532 io_iopoll_getevents(ctx, &nr_events, 1);
533 }
534 mutex_unlock(&ctx->uring_lock);
535 }
536
537 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
538 long min)
539 {
540 int ret = 0;
541
542 do {
543 int tmin = 0;
544
545 if (*nr_events < min)
546 tmin = min - *nr_events;
547
548 ret = io_iopoll_getevents(ctx, nr_events, tmin);
549 if (ret <= 0)
550 break;
551 ret = 0;
552 } while (min && !*nr_events && !need_resched());
553
554 return ret;
555 }
556
557 static void kiocb_end_write(struct kiocb *kiocb)
558 {
559 if (kiocb->ki_flags & IOCB_WRITE) {
560 struct inode *inode = file_inode(kiocb->ki_filp);
561
562 /*
563 * Tell lockdep we inherited freeze protection from submission
564 * thread.
565 */
566 if (S_ISREG(inode->i_mode))
567 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
568 file_end_write(kiocb->ki_filp);
569 }
570 }
571
572 static void io_fput(struct io_kiocb *req)
573 {
574 if (!(req->flags & REQ_F_FIXED_FILE))
575 fput(req->rw.ki_filp);
576 }
577
578 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
579 {
580 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
581
582 kiocb_end_write(kiocb);
583
584 io_fput(req);
585 io_cqring_add_event(req->ctx, req->user_data, res, 0);
586 io_free_req(req);
587 }
588
589 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
590 {
591 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
592
593 kiocb_end_write(kiocb);
594
595 req->error = res;
596 if (res != -EAGAIN)
597 req->flags |= REQ_F_IOPOLL_COMPLETED;
598 }
599
600 /*
601 * After the iocb has been issued, it's safe to be found on the poll list.
602 * Adding the kiocb to the list AFTER submission ensures that we don't
603 * find it from a io_iopoll_getevents() thread before the issuer is done
604 * accessing the kiocb cookie.
605 */
606 static void io_iopoll_req_issued(struct io_kiocb *req)
607 {
608 struct io_ring_ctx *ctx = req->ctx;
609
610 /*
611 * Track whether we have multiple files in our lists. This will impact
612 * how we do polling eventually, not spinning if we're on potentially
613 * different devices.
614 */
615 if (list_empty(&ctx->poll_list)) {
616 ctx->poll_multi_file = false;
617 } else if (!ctx->poll_multi_file) {
618 struct io_kiocb *list_req;
619
620 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
621 list);
622 if (list_req->rw.ki_filp != req->rw.ki_filp)
623 ctx->poll_multi_file = true;
624 }
625
626 /*
627 * For fast devices, IO may have already completed. If it has, add
628 * it to the front so we find it first.
629 */
630 if (req->flags & REQ_F_IOPOLL_COMPLETED)
631 list_add(&req->list, &ctx->poll_list);
632 else
633 list_add_tail(&req->list, &ctx->poll_list);
634 }
635
636 static void io_file_put(struct io_submit_state *state, struct file *file)
637 {
638 if (!state) {
639 fput(file);
640 } else if (state->file) {
641 int diff = state->has_refs - state->used_refs;
642
643 if (diff)
644 fput_many(state->file, diff);
645 state->file = NULL;
646 }
647 }
648
649 /*
650 * Get as many references to a file as we have IOs left in this submission,
651 * assuming most submissions are for one file, or at least that each file
652 * has more than one submission.
653 */
654 static struct file *io_file_get(struct io_submit_state *state, int fd)
655 {
656 if (!state)
657 return fget(fd);
658
659 if (state->file) {
660 if (state->fd == fd) {
661 state->used_refs++;
662 state->ios_left--;
663 return state->file;
664 }
665 io_file_put(state, NULL);
666 }
667 state->file = fget_many(fd, state->ios_left);
668 if (!state->file)
669 return NULL;
670
671 state->fd = fd;
672 state->has_refs = state->ios_left;
673 state->used_refs = 1;
674 state->ios_left--;
675 return state->file;
676 }
677
678 /*
679 * If we tracked the file through the SCM inflight mechanism, we could support
680 * any file. For now, just ensure that anything potentially problematic is done
681 * inline.
682 */
683 static bool io_file_supports_async(struct file *file)
684 {
685 umode_t mode = file_inode(file)->i_mode;
686
687 if (S_ISBLK(mode) || S_ISCHR(mode))
688 return true;
689 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
690 return true;
691
692 return false;
693 }
694
695 static int io_prep_rw(struct io_kiocb *req, const struct sqe_submit *s,
696 bool force_nonblock, struct io_submit_state *state)
697 {
698 const struct io_uring_sqe *sqe = s->sqe;
699 struct io_ring_ctx *ctx = req->ctx;
700 struct kiocb *kiocb = &req->rw;
701 unsigned ioprio, flags;
702 int fd, ret;
703
704 /* For -EAGAIN retry, everything is already prepped */
705 if (kiocb->ki_filp)
706 return 0;
707
708 flags = READ_ONCE(sqe->flags);
709 fd = READ_ONCE(sqe->fd);
710
711 if (flags & IOSQE_FIXED_FILE) {
712 if (unlikely(!ctx->user_files ||
713 (unsigned) fd >= ctx->nr_user_files))
714 return -EBADF;
715 kiocb->ki_filp = ctx->user_files[fd];
716 req->flags |= REQ_F_FIXED_FILE;
717 } else {
718 if (s->needs_fixed_file)
719 return -EBADF;
720 kiocb->ki_filp = io_file_get(state, fd);
721 if (unlikely(!kiocb->ki_filp))
722 return -EBADF;
723 if (force_nonblock && !io_file_supports_async(kiocb->ki_filp))
724 force_nonblock = false;
725 }
726 kiocb->ki_pos = READ_ONCE(sqe->off);
727 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
728 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
729
730 ioprio = READ_ONCE(sqe->ioprio);
731 if (ioprio) {
732 ret = ioprio_check_cap(ioprio);
733 if (ret)
734 goto out_fput;
735
736 kiocb->ki_ioprio = ioprio;
737 } else
738 kiocb->ki_ioprio = get_current_ioprio();
739
740 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
741 if (unlikely(ret))
742 goto out_fput;
743 if (force_nonblock) {
744 kiocb->ki_flags |= IOCB_NOWAIT;
745 req->flags |= REQ_F_FORCE_NONBLOCK;
746 }
747 if (ctx->flags & IORING_SETUP_IOPOLL) {
748 ret = -EOPNOTSUPP;
749 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
750 !kiocb->ki_filp->f_op->iopoll)
751 goto out_fput;
752
753 req->error = 0;
754 kiocb->ki_flags |= IOCB_HIPRI;
755 kiocb->ki_complete = io_complete_rw_iopoll;
756 } else {
757 if (kiocb->ki_flags & IOCB_HIPRI) {
758 ret = -EINVAL;
759 goto out_fput;
760 }
761 kiocb->ki_complete = io_complete_rw;
762 }
763 return 0;
764 out_fput:
765 if (!(flags & IOSQE_FIXED_FILE)) {
766 /*
767 * in case of error, we didn't use this file reference. drop it.
768 */
769 if (state)
770 state->used_refs--;
771 io_file_put(state, kiocb->ki_filp);
772 }
773 return ret;
774 }
775
776 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
777 {
778 switch (ret) {
779 case -EIOCBQUEUED:
780 break;
781 case -ERESTARTSYS:
782 case -ERESTARTNOINTR:
783 case -ERESTARTNOHAND:
784 case -ERESTART_RESTARTBLOCK:
785 /*
786 * We can't just restart the syscall, since previously
787 * submitted sqes may already be in progress. Just fail this
788 * IO with EINTR.
789 */
790 ret = -EINTR;
791 /* fall through */
792 default:
793 kiocb->ki_complete(kiocb, ret, 0);
794 }
795 }
796
797 static int io_import_fixed(struct io_ring_ctx *ctx, int rw,
798 const struct io_uring_sqe *sqe,
799 struct iov_iter *iter)
800 {
801 size_t len = READ_ONCE(sqe->len);
802 struct io_mapped_ubuf *imu;
803 unsigned index, buf_index;
804 size_t offset;
805 u64 buf_addr;
806
807 /* attempt to use fixed buffers without having provided iovecs */
808 if (unlikely(!ctx->user_bufs))
809 return -EFAULT;
810
811 buf_index = READ_ONCE(sqe->buf_index);
812 if (unlikely(buf_index >= ctx->nr_user_bufs))
813 return -EFAULT;
814
815 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
816 imu = &ctx->user_bufs[index];
817 buf_addr = READ_ONCE(sqe->addr);
818
819 /* overflow */
820 if (buf_addr + len < buf_addr)
821 return -EFAULT;
822 /* not inside the mapped region */
823 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
824 return -EFAULT;
825
826 /*
827 * May not be a start of buffer, set size appropriately
828 * and advance us to the beginning.
829 */
830 offset = buf_addr - imu->ubuf;
831 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
832 if (offset)
833 iov_iter_advance(iter, offset);
834 return 0;
835 }
836
837 static int io_import_iovec(struct io_ring_ctx *ctx, int rw,
838 const struct sqe_submit *s, struct iovec **iovec,
839 struct iov_iter *iter)
840 {
841 const struct io_uring_sqe *sqe = s->sqe;
842 void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
843 size_t sqe_len = READ_ONCE(sqe->len);
844 u8 opcode;
845
846 /*
847 * We're reading ->opcode for the second time, but the first read
848 * doesn't care whether it's _FIXED or not, so it doesn't matter
849 * whether ->opcode changes concurrently. The first read does care
850 * about whether it is a READ or a WRITE, so we don't trust this read
851 * for that purpose and instead let the caller pass in the read/write
852 * flag.
853 */
854 opcode = READ_ONCE(sqe->opcode);
855 if (opcode == IORING_OP_READ_FIXED ||
856 opcode == IORING_OP_WRITE_FIXED) {
857 ssize_t ret = io_import_fixed(ctx, rw, sqe, iter);
858 *iovec = NULL;
859 return ret;
860 }
861
862 if (!s->has_user)
863 return -EFAULT;
864
865 #ifdef CONFIG_COMPAT
866 if (ctx->compat)
867 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
868 iovec, iter);
869 #endif
870
871 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
872 }
873
874 static ssize_t io_read(struct io_kiocb *req, const struct sqe_submit *s,
875 bool force_nonblock, struct io_submit_state *state)
876 {
877 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
878 struct kiocb *kiocb = &req->rw;
879 struct iov_iter iter;
880 struct file *file;
881 ssize_t ret;
882
883 ret = io_prep_rw(req, s, force_nonblock, state);
884 if (ret)
885 return ret;
886 file = kiocb->ki_filp;
887
888 ret = -EBADF;
889 if (unlikely(!(file->f_mode & FMODE_READ)))
890 goto out_fput;
891 ret = -EINVAL;
892 if (unlikely(!file->f_op->read_iter))
893 goto out_fput;
894
895 ret = io_import_iovec(req->ctx, READ, s, &iovec, &iter);
896 if (ret)
897 goto out_fput;
898
899 ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_iter_count(&iter));
900 if (!ret) {
901 ssize_t ret2;
902
903 /* Catch -EAGAIN return for forced non-blocking submission */
904 ret2 = call_read_iter(file, kiocb, &iter);
905 if (!force_nonblock || ret2 != -EAGAIN)
906 io_rw_done(kiocb, ret2);
907 else
908 ret = -EAGAIN;
909 }
910 kfree(iovec);
911 out_fput:
912 /* Hold on to the file for -EAGAIN */
913 if (unlikely(ret && ret != -EAGAIN))
914 io_fput(req);
915 return ret;
916 }
917
918 static ssize_t io_write(struct io_kiocb *req, const struct sqe_submit *s,
919 bool force_nonblock, struct io_submit_state *state)
920 {
921 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
922 struct kiocb *kiocb = &req->rw;
923 struct iov_iter iter;
924 struct file *file;
925 ssize_t ret;
926
927 ret = io_prep_rw(req, s, force_nonblock, state);
928 if (ret)
929 return ret;
930 /* Hold on to the file for -EAGAIN */
931 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT))
932 return -EAGAIN;
933
934 ret = -EBADF;
935 file = kiocb->ki_filp;
936 if (unlikely(!(file->f_mode & FMODE_WRITE)))
937 goto out_fput;
938 ret = -EINVAL;
939 if (unlikely(!file->f_op->write_iter))
940 goto out_fput;
941
942 ret = io_import_iovec(req->ctx, WRITE, s, &iovec, &iter);
943 if (ret)
944 goto out_fput;
945
946 ret = rw_verify_area(WRITE, file, &kiocb->ki_pos,
947 iov_iter_count(&iter));
948 if (!ret) {
949 /*
950 * Open-code file_start_write here to grab freeze protection,
951 * which will be released by another thread in
952 * io_complete_rw(). Fool lockdep by telling it the lock got
953 * released so that it doesn't complain about the held lock when
954 * we return to userspace.
955 */
956 if (S_ISREG(file_inode(file)->i_mode)) {
957 __sb_start_write(file_inode(file)->i_sb,
958 SB_FREEZE_WRITE, true);
959 __sb_writers_release(file_inode(file)->i_sb,
960 SB_FREEZE_WRITE);
961 }
962 kiocb->ki_flags |= IOCB_WRITE;
963 io_rw_done(kiocb, call_write_iter(file, kiocb, &iter));
964 }
965 kfree(iovec);
966 out_fput:
967 if (unlikely(ret))
968 io_fput(req);
969 return ret;
970 }
971
972 /*
973 * IORING_OP_NOP just posts a completion event, nothing else.
974 */
975 static int io_nop(struct io_kiocb *req, u64 user_data)
976 {
977 struct io_ring_ctx *ctx = req->ctx;
978 long err = 0;
979
980 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
981 return -EINVAL;
982
983 /*
984 * Twilight zone - it's possible that someone issued an opcode that
985 * has a file attached, then got -EAGAIN on submission, and changed
986 * the sqe before we retried it from async context. Avoid dropping
987 * a file reference for this malicious case, and flag the error.
988 */
989 if (req->rw.ki_filp) {
990 err = -EBADF;
991 io_fput(req);
992 }
993 io_cqring_add_event(ctx, user_data, err, 0);
994 io_free_req(req);
995 return 0;
996 }
997
998 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
999 {
1000 struct io_ring_ctx *ctx = req->ctx;
1001 unsigned flags;
1002 int fd;
1003
1004 /* Prep already done */
1005 if (req->rw.ki_filp)
1006 return 0;
1007
1008 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1009 return -EINVAL;
1010 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1011 return -EINVAL;
1012
1013 fd = READ_ONCE(sqe->fd);
1014 flags = READ_ONCE(sqe->flags);
1015
1016 if (flags & IOSQE_FIXED_FILE) {
1017 if (unlikely(!ctx->user_files || fd >= ctx->nr_user_files))
1018 return -EBADF;
1019 req->rw.ki_filp = ctx->user_files[fd];
1020 req->flags |= REQ_F_FIXED_FILE;
1021 } else {
1022 req->rw.ki_filp = fget(fd);
1023 if (unlikely(!req->rw.ki_filp))
1024 return -EBADF;
1025 }
1026
1027 return 0;
1028 }
1029
1030 static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1031 bool force_nonblock)
1032 {
1033 loff_t sqe_off = READ_ONCE(sqe->off);
1034 loff_t sqe_len = READ_ONCE(sqe->len);
1035 loff_t end = sqe_off + sqe_len;
1036 unsigned fsync_flags;
1037 int ret;
1038
1039 fsync_flags = READ_ONCE(sqe->fsync_flags);
1040 if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC))
1041 return -EINVAL;
1042
1043 ret = io_prep_fsync(req, sqe);
1044 if (ret)
1045 return ret;
1046
1047 /* fsync always requires a blocking context */
1048 if (force_nonblock)
1049 return -EAGAIN;
1050
1051 ret = vfs_fsync_range(req->rw.ki_filp, sqe_off,
1052 end > 0 ? end : LLONG_MAX,
1053 fsync_flags & IORING_FSYNC_DATASYNC);
1054
1055 io_fput(req);
1056 io_cqring_add_event(req->ctx, sqe->user_data, ret, 0);
1057 io_free_req(req);
1058 return 0;
1059 }
1060
1061 static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
1062 const struct sqe_submit *s, bool force_nonblock,
1063 struct io_submit_state *state)
1064 {
1065 ssize_t ret;
1066 int opcode;
1067
1068 if (unlikely(s->index >= ctx->sq_entries))
1069 return -EINVAL;
1070 req->user_data = READ_ONCE(s->sqe->user_data);
1071
1072 opcode = READ_ONCE(s->sqe->opcode);
1073 switch (opcode) {
1074 case IORING_OP_NOP:
1075 ret = io_nop(req, req->user_data);
1076 break;
1077 case IORING_OP_READV:
1078 if (unlikely(s->sqe->buf_index))
1079 return -EINVAL;
1080 ret = io_read(req, s, force_nonblock, state);
1081 break;
1082 case IORING_OP_WRITEV:
1083 if (unlikely(s->sqe->buf_index))
1084 return -EINVAL;
1085 ret = io_write(req, s, force_nonblock, state);
1086 break;
1087 case IORING_OP_READ_FIXED:
1088 ret = io_read(req, s, force_nonblock, state);
1089 break;
1090 case IORING_OP_WRITE_FIXED:
1091 ret = io_write(req, s, force_nonblock, state);
1092 break;
1093 case IORING_OP_FSYNC:
1094 ret = io_fsync(req, s->sqe, force_nonblock);
1095 break;
1096 default:
1097 ret = -EINVAL;
1098 break;
1099 }
1100
1101 if (ret)
1102 return ret;
1103
1104 if (ctx->flags & IORING_SETUP_IOPOLL) {
1105 if (req->error == -EAGAIN)
1106 return -EAGAIN;
1107
1108 /* workqueue context doesn't hold uring_lock, grab it now */
1109 if (s->needs_lock)
1110 mutex_lock(&ctx->uring_lock);
1111 io_iopoll_req_issued(req);
1112 if (s->needs_lock)
1113 mutex_unlock(&ctx->uring_lock);
1114 }
1115
1116 return 0;
1117 }
1118
1119 static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe)
1120 {
1121 u8 opcode = READ_ONCE(sqe->opcode);
1122
1123 return !(opcode == IORING_OP_READ_FIXED ||
1124 opcode == IORING_OP_WRITE_FIXED);
1125 }
1126
1127 static void io_sq_wq_submit_work(struct work_struct *work)
1128 {
1129 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1130 struct sqe_submit *s = &req->submit;
1131 const struct io_uring_sqe *sqe = s->sqe;
1132 struct io_ring_ctx *ctx = req->ctx;
1133 mm_segment_t old_fs;
1134 bool needs_user;
1135 int ret;
1136
1137 /* Ensure we clear previously set forced non-block flag */
1138 req->flags &= ~REQ_F_FORCE_NONBLOCK;
1139 req->rw.ki_flags &= ~IOCB_NOWAIT;
1140
1141 s->needs_lock = true;
1142 s->has_user = false;
1143
1144 /*
1145 * If we're doing IO to fixed buffers, we don't need to get/set
1146 * user context
1147 */
1148 needs_user = io_sqe_needs_user(s->sqe);
1149 if (needs_user) {
1150 if (!mmget_not_zero(ctx->sqo_mm)) {
1151 ret = -EFAULT;
1152 goto err;
1153 }
1154 use_mm(ctx->sqo_mm);
1155 old_fs = get_fs();
1156 set_fs(USER_DS);
1157 s->has_user = true;
1158 }
1159
1160 do {
1161 ret = __io_submit_sqe(ctx, req, s, false, NULL);
1162 /*
1163 * We can get EAGAIN for polled IO even though we're forcing
1164 * a sync submission from here, since we can't wait for
1165 * request slots on the block side.
1166 */
1167 if (ret != -EAGAIN)
1168 break;
1169 cond_resched();
1170 } while (1);
1171
1172 if (needs_user) {
1173 set_fs(old_fs);
1174 unuse_mm(ctx->sqo_mm);
1175 mmput(ctx->sqo_mm);
1176 }
1177 err:
1178 if (ret) {
1179 io_cqring_add_event(ctx, sqe->user_data, ret, 0);
1180 io_free_req(req);
1181 }
1182
1183 /* async context always use a copy of the sqe */
1184 kfree(sqe);
1185 }
1186
1187 static int io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s,
1188 struct io_submit_state *state)
1189 {
1190 struct io_kiocb *req;
1191 ssize_t ret;
1192
1193 /* enforce forwards compatibility on users */
1194 if (unlikely(s->sqe->flags & ~IOSQE_FIXED_FILE))
1195 return -EINVAL;
1196
1197 req = io_get_req(ctx, state);
1198 if (unlikely(!req))
1199 return -EAGAIN;
1200
1201 req->rw.ki_filp = NULL;
1202
1203 ret = __io_submit_sqe(ctx, req, s, true, state);
1204 if (ret == -EAGAIN) {
1205 struct io_uring_sqe *sqe_copy;
1206
1207 sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL);
1208 if (sqe_copy) {
1209 memcpy(sqe_copy, s->sqe, sizeof(*sqe_copy));
1210 s->sqe = sqe_copy;
1211
1212 memcpy(&req->submit, s, sizeof(*s));
1213 INIT_WORK(&req->work, io_sq_wq_submit_work);
1214 queue_work(ctx->sqo_wq, &req->work);
1215 ret = 0;
1216 }
1217 }
1218 if (ret)
1219 io_free_req(req);
1220
1221 return ret;
1222 }
1223
1224 /*
1225 * Batched submission is done, ensure local IO is flushed out.
1226 */
1227 static void io_submit_state_end(struct io_submit_state *state)
1228 {
1229 blk_finish_plug(&state->plug);
1230 io_file_put(state, NULL);
1231 if (state->free_reqs)
1232 kmem_cache_free_bulk(req_cachep, state->free_reqs,
1233 &state->reqs[state->cur_req]);
1234 }
1235
1236 /*
1237 * Start submission side cache.
1238 */
1239 static void io_submit_state_start(struct io_submit_state *state,
1240 struct io_ring_ctx *ctx, unsigned max_ios)
1241 {
1242 blk_start_plug(&state->plug);
1243 state->free_reqs = 0;
1244 state->file = NULL;
1245 state->ios_left = max_ios;
1246 }
1247
1248 static void io_commit_sqring(struct io_ring_ctx *ctx)
1249 {
1250 struct io_sq_ring *ring = ctx->sq_ring;
1251
1252 if (ctx->cached_sq_head != READ_ONCE(ring->r.head)) {
1253 /*
1254 * Ensure any loads from the SQEs are done at this point,
1255 * since once we write the new head, the application could
1256 * write new data to them.
1257 */
1258 smp_store_release(&ring->r.head, ctx->cached_sq_head);
1259
1260 /*
1261 * write side barrier of head update, app has read side. See
1262 * comment at the top of this file
1263 */
1264 smp_wmb();
1265 }
1266 }
1267
1268 /*
1269 * Undo last io_get_sqring()
1270 */
1271 static void io_drop_sqring(struct io_ring_ctx *ctx)
1272 {
1273 ctx->cached_sq_head--;
1274 }
1275
1276 /*
1277 * Fetch an sqe, if one is available. Note that s->sqe will point to memory
1278 * that is mapped by userspace. This means that care needs to be taken to
1279 * ensure that reads are stable, as we cannot rely on userspace always
1280 * being a good citizen. If members of the sqe are validated and then later
1281 * used, it's important that those reads are done through READ_ONCE() to
1282 * prevent a re-load down the line.
1283 */
1284 static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s)
1285 {
1286 struct io_sq_ring *ring = ctx->sq_ring;
1287 unsigned head;
1288
1289 /*
1290 * The cached sq head (or cq tail) serves two purposes:
1291 *
1292 * 1) allows us to batch the cost of updating the user visible
1293 * head updates.
1294 * 2) allows the kernel side to track the head on its own, even
1295 * though the application is the one updating it.
1296 */
1297 head = ctx->cached_sq_head;
1298 /* See comment at the top of this file */
1299 smp_rmb();
1300 if (head == READ_ONCE(ring->r.tail))
1301 return false;
1302
1303 head = READ_ONCE(ring->array[head & ctx->sq_mask]);
1304 if (head < ctx->sq_entries) {
1305 s->index = head;
1306 s->sqe = &ctx->sq_sqes[head];
1307 ctx->cached_sq_head++;
1308 return true;
1309 }
1310
1311 /* drop invalid entries */
1312 ctx->cached_sq_head++;
1313 ring->dropped++;
1314 /* See comment at the top of this file */
1315 smp_wmb();
1316 return false;
1317 }
1318
1319 static int io_submit_sqes(struct io_ring_ctx *ctx, struct sqe_submit *sqes,
1320 unsigned int nr, bool has_user, bool mm_fault)
1321 {
1322 struct io_submit_state state, *statep = NULL;
1323 int ret, i, submitted = 0;
1324
1325 if (nr > IO_PLUG_THRESHOLD) {
1326 io_submit_state_start(&state, ctx, nr);
1327 statep = &state;
1328 }
1329
1330 for (i = 0; i < nr; i++) {
1331 if (unlikely(mm_fault)) {
1332 ret = -EFAULT;
1333 } else {
1334 sqes[i].has_user = has_user;
1335 sqes[i].needs_lock = true;
1336 sqes[i].needs_fixed_file = true;
1337 ret = io_submit_sqe(ctx, &sqes[i], statep);
1338 }
1339 if (!ret) {
1340 submitted++;
1341 continue;
1342 }
1343
1344 io_cqring_add_event(ctx, sqes[i].sqe->user_data, ret, 0);
1345 }
1346
1347 if (statep)
1348 io_submit_state_end(&state);
1349
1350 return submitted;
1351 }
1352
1353 static int io_sq_thread(void *data)
1354 {
1355 struct sqe_submit sqes[IO_IOPOLL_BATCH];
1356 struct io_ring_ctx *ctx = data;
1357 struct mm_struct *cur_mm = NULL;
1358 mm_segment_t old_fs;
1359 DEFINE_WAIT(wait);
1360 unsigned inflight;
1361 unsigned long timeout;
1362
1363 old_fs = get_fs();
1364 set_fs(USER_DS);
1365
1366 timeout = inflight = 0;
1367 while (!kthread_should_stop() && !ctx->sqo_stop) {
1368 bool all_fixed, mm_fault = false;
1369 int i;
1370
1371 if (inflight) {
1372 unsigned nr_events = 0;
1373
1374 if (ctx->flags & IORING_SETUP_IOPOLL) {
1375 /*
1376 * We disallow the app entering submit/complete
1377 * with polling, but we still need to lock the
1378 * ring to prevent racing with polled issue
1379 * that got punted to a workqueue.
1380 */
1381 mutex_lock(&ctx->uring_lock);
1382 io_iopoll_check(ctx, &nr_events, 0);
1383 mutex_unlock(&ctx->uring_lock);
1384 } else {
1385 /*
1386 * Normal IO, just pretend everything completed.
1387 * We don't have to poll completions for that.
1388 */
1389 nr_events = inflight;
1390 }
1391
1392 inflight -= nr_events;
1393 if (!inflight)
1394 timeout = jiffies + ctx->sq_thread_idle;
1395 }
1396
1397 if (!io_get_sqring(ctx, &sqes[0])) {
1398 /*
1399 * We're polling. If we're within the defined idle
1400 * period, then let us spin without work before going
1401 * to sleep.
1402 */
1403 if (inflight || !time_after(jiffies, timeout)) {
1404 cpu_relax();
1405 continue;
1406 }
1407
1408 /*
1409 * Drop cur_mm before scheduling, we can't hold it for
1410 * long periods (or over schedule()). Do this before
1411 * adding ourselves to the waitqueue, as the unuse/drop
1412 * may sleep.
1413 */
1414 if (cur_mm) {
1415 unuse_mm(cur_mm);
1416 mmput(cur_mm);
1417 cur_mm = NULL;
1418 }
1419
1420 prepare_to_wait(&ctx->sqo_wait, &wait,
1421 TASK_INTERRUPTIBLE);
1422
1423 /* Tell userspace we may need a wakeup call */
1424 ctx->sq_ring->flags |= IORING_SQ_NEED_WAKEUP;
1425 smp_wmb();
1426
1427 if (!io_get_sqring(ctx, &sqes[0])) {
1428 if (kthread_should_stop()) {
1429 finish_wait(&ctx->sqo_wait, &wait);
1430 break;
1431 }
1432 if (signal_pending(current))
1433 flush_signals(current);
1434 schedule();
1435 finish_wait(&ctx->sqo_wait, &wait);
1436
1437 ctx->sq_ring->flags &= ~IORING_SQ_NEED_WAKEUP;
1438 smp_wmb();
1439 continue;
1440 }
1441 finish_wait(&ctx->sqo_wait, &wait);
1442
1443 ctx->sq_ring->flags &= ~IORING_SQ_NEED_WAKEUP;
1444 smp_wmb();
1445 }
1446
1447 i = 0;
1448 all_fixed = true;
1449 do {
1450 if (all_fixed && io_sqe_needs_user(sqes[i].sqe))
1451 all_fixed = false;
1452
1453 i++;
1454 if (i == ARRAY_SIZE(sqes))
1455 break;
1456 } while (io_get_sqring(ctx, &sqes[i]));
1457
1458 /* Unless all new commands are FIXED regions, grab mm */
1459 if (!all_fixed && !cur_mm) {
1460 mm_fault = !mmget_not_zero(ctx->sqo_mm);
1461 if (!mm_fault) {
1462 use_mm(ctx->sqo_mm);
1463 cur_mm = ctx->sqo_mm;
1464 }
1465 }
1466
1467 inflight += io_submit_sqes(ctx, sqes, i, cur_mm != NULL,
1468 mm_fault);
1469
1470 /* Commit SQ ring head once we've consumed all SQEs */
1471 io_commit_sqring(ctx);
1472 }
1473
1474 set_fs(old_fs);
1475 if (cur_mm) {
1476 unuse_mm(cur_mm);
1477 mmput(cur_mm);
1478 }
1479 return 0;
1480 }
1481
1482 static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit)
1483 {
1484 struct io_submit_state state, *statep = NULL;
1485 int i, ret = 0, submit = 0;
1486
1487 if (to_submit > IO_PLUG_THRESHOLD) {
1488 io_submit_state_start(&state, ctx, to_submit);
1489 statep = &state;
1490 }
1491
1492 for (i = 0; i < to_submit; i++) {
1493 struct sqe_submit s;
1494
1495 if (!io_get_sqring(ctx, &s))
1496 break;
1497
1498 s.has_user = true;
1499 s.needs_lock = false;
1500 s.needs_fixed_file = false;
1501
1502 ret = io_submit_sqe(ctx, &s, statep);
1503 if (ret) {
1504 io_drop_sqring(ctx);
1505 break;
1506 }
1507
1508 submit++;
1509 }
1510 io_commit_sqring(ctx);
1511
1512 if (statep)
1513 io_submit_state_end(statep);
1514
1515 return submit ? submit : ret;
1516 }
1517
1518 static unsigned io_cqring_events(struct io_cq_ring *ring)
1519 {
1520 return READ_ONCE(ring->r.tail) - READ_ONCE(ring->r.head);
1521 }
1522
1523 /*
1524 * Wait until events become available, if we don't already have some. The
1525 * application must reap them itself, as they reside on the shared cq ring.
1526 */
1527 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
1528 const sigset_t __user *sig, size_t sigsz)
1529 {
1530 struct io_cq_ring *ring = ctx->cq_ring;
1531 sigset_t ksigmask, sigsaved;
1532 DEFINE_WAIT(wait);
1533 int ret;
1534
1535 /* See comment at the top of this file */
1536 smp_rmb();
1537 if (io_cqring_events(ring) >= min_events)
1538 return 0;
1539
1540 if (sig) {
1541 ret = set_user_sigmask(sig, &ksigmask, &sigsaved, sigsz);
1542 if (ret)
1543 return ret;
1544 }
1545
1546 do {
1547 prepare_to_wait(&ctx->wait, &wait, TASK_INTERRUPTIBLE);
1548
1549 ret = 0;
1550 /* See comment at the top of this file */
1551 smp_rmb();
1552 if (io_cqring_events(ring) >= min_events)
1553 break;
1554
1555 schedule();
1556
1557 ret = -EINTR;
1558 if (signal_pending(current))
1559 break;
1560 } while (1);
1561
1562 finish_wait(&ctx->wait, &wait);
1563
1564 if (sig)
1565 restore_user_sigmask(sig, &sigsaved);
1566
1567 return READ_ONCE(ring->r.head) == READ_ONCE(ring->r.tail) ? ret : 0;
1568 }
1569
1570 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
1571 {
1572 #if defined(CONFIG_UNIX)
1573 if (ctx->ring_sock) {
1574 struct sock *sock = ctx->ring_sock->sk;
1575 struct sk_buff *skb;
1576
1577 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
1578 kfree_skb(skb);
1579 }
1580 #else
1581 int i;
1582
1583 for (i = 0; i < ctx->nr_user_files; i++)
1584 fput(ctx->user_files[i]);
1585 #endif
1586 }
1587
1588 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
1589 {
1590 if (!ctx->user_files)
1591 return -ENXIO;
1592
1593 __io_sqe_files_unregister(ctx);
1594 kfree(ctx->user_files);
1595 ctx->user_files = NULL;
1596 ctx->nr_user_files = 0;
1597 return 0;
1598 }
1599
1600 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
1601 {
1602 if (ctx->sqo_thread) {
1603 ctx->sqo_stop = 1;
1604 mb();
1605 kthread_stop(ctx->sqo_thread);
1606 ctx->sqo_thread = NULL;
1607 }
1608 }
1609
1610 static void io_finish_async(struct io_ring_ctx *ctx)
1611 {
1612 io_sq_thread_stop(ctx);
1613
1614 if (ctx->sqo_wq) {
1615 destroy_workqueue(ctx->sqo_wq);
1616 ctx->sqo_wq = NULL;
1617 }
1618 }
1619
1620 #if defined(CONFIG_UNIX)
1621 static void io_destruct_skb(struct sk_buff *skb)
1622 {
1623 struct io_ring_ctx *ctx = skb->sk->sk_user_data;
1624
1625 io_finish_async(ctx);
1626 unix_destruct_scm(skb);
1627 }
1628
1629 /*
1630 * Ensure the UNIX gc is aware of our file set, so we are certain that
1631 * the io_uring can be safely unregistered on process exit, even if we have
1632 * loops in the file referencing.
1633 */
1634 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
1635 {
1636 struct sock *sk = ctx->ring_sock->sk;
1637 struct scm_fp_list *fpl;
1638 struct sk_buff *skb;
1639 int i;
1640
1641 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
1642 unsigned long inflight = ctx->user->unix_inflight + nr;
1643
1644 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
1645 return -EMFILE;
1646 }
1647
1648 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
1649 if (!fpl)
1650 return -ENOMEM;
1651
1652 skb = alloc_skb(0, GFP_KERNEL);
1653 if (!skb) {
1654 kfree(fpl);
1655 return -ENOMEM;
1656 }
1657
1658 skb->sk = sk;
1659 skb->destructor = io_destruct_skb;
1660
1661 fpl->user = get_uid(ctx->user);
1662 for (i = 0; i < nr; i++) {
1663 fpl->fp[i] = get_file(ctx->user_files[i + offset]);
1664 unix_inflight(fpl->user, fpl->fp[i]);
1665 }
1666
1667 fpl->max = fpl->count = nr;
1668 UNIXCB(skb).fp = fpl;
1669 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1670 skb_queue_head(&sk->sk_receive_queue, skb);
1671
1672 for (i = 0; i < nr; i++)
1673 fput(fpl->fp[i]);
1674
1675 return 0;
1676 }
1677
1678 /*
1679 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
1680 * causes regular reference counting to break down. We rely on the UNIX
1681 * garbage collection to take care of this problem for us.
1682 */
1683 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
1684 {
1685 unsigned left, total;
1686 int ret = 0;
1687
1688 total = 0;
1689 left = ctx->nr_user_files;
1690 while (left) {
1691 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
1692 int ret;
1693
1694 ret = __io_sqe_files_scm(ctx, this_files, total);
1695 if (ret)
1696 break;
1697 left -= this_files;
1698 total += this_files;
1699 }
1700
1701 if (!ret)
1702 return 0;
1703
1704 while (total < ctx->nr_user_files) {
1705 fput(ctx->user_files[total]);
1706 total++;
1707 }
1708
1709 return ret;
1710 }
1711 #else
1712 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
1713 {
1714 return 0;
1715 }
1716 #endif
1717
1718 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
1719 unsigned nr_args)
1720 {
1721 __s32 __user *fds = (__s32 __user *) arg;
1722 int fd, ret = 0;
1723 unsigned i;
1724
1725 if (ctx->user_files)
1726 return -EBUSY;
1727 if (!nr_args)
1728 return -EINVAL;
1729 if (nr_args > IORING_MAX_FIXED_FILES)
1730 return -EMFILE;
1731
1732 ctx->user_files = kcalloc(nr_args, sizeof(struct file *), GFP_KERNEL);
1733 if (!ctx->user_files)
1734 return -ENOMEM;
1735
1736 for (i = 0; i < nr_args; i++) {
1737 ret = -EFAULT;
1738 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
1739 break;
1740
1741 ctx->user_files[i] = fget(fd);
1742
1743 ret = -EBADF;
1744 if (!ctx->user_files[i])
1745 break;
1746 /*
1747 * Don't allow io_uring instances to be registered. If UNIX
1748 * isn't enabled, then this causes a reference cycle and this
1749 * instance can never get freed. If UNIX is enabled we'll
1750 * handle it just fine, but there's still no point in allowing
1751 * a ring fd as it doesn't support regular read/write anyway.
1752 */
1753 if (ctx->user_files[i]->f_op == &io_uring_fops) {
1754 fput(ctx->user_files[i]);
1755 break;
1756 }
1757 ctx->nr_user_files++;
1758 ret = 0;
1759 }
1760
1761 if (ret) {
1762 for (i = 0; i < ctx->nr_user_files; i++)
1763 fput(ctx->user_files[i]);
1764
1765 kfree(ctx->user_files);
1766 ctx->nr_user_files = 0;
1767 return ret;
1768 }
1769
1770 ret = io_sqe_files_scm(ctx);
1771 if (ret)
1772 io_sqe_files_unregister(ctx);
1773
1774 return ret;
1775 }
1776
1777 static int io_sq_offload_start(struct io_ring_ctx *ctx,
1778 struct io_uring_params *p)
1779 {
1780 int ret;
1781
1782 init_waitqueue_head(&ctx->sqo_wait);
1783 mmgrab(current->mm);
1784 ctx->sqo_mm = current->mm;
1785
1786 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
1787 if (!ctx->sq_thread_idle)
1788 ctx->sq_thread_idle = HZ;
1789
1790 ret = -EINVAL;
1791 if (!cpu_possible(p->sq_thread_cpu))
1792 goto err;
1793
1794 if (ctx->flags & IORING_SETUP_SQPOLL) {
1795 if (p->flags & IORING_SETUP_SQ_AFF) {
1796 int cpu;
1797
1798 cpu = array_index_nospec(p->sq_thread_cpu, NR_CPUS);
1799 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
1800 ctx, cpu,
1801 "io_uring-sq");
1802 } else {
1803 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
1804 "io_uring-sq");
1805 }
1806 if (IS_ERR(ctx->sqo_thread)) {
1807 ret = PTR_ERR(ctx->sqo_thread);
1808 ctx->sqo_thread = NULL;
1809 goto err;
1810 }
1811 wake_up_process(ctx->sqo_thread);
1812 } else if (p->flags & IORING_SETUP_SQ_AFF) {
1813 /* Can't have SQ_AFF without SQPOLL */
1814 ret = -EINVAL;
1815 goto err;
1816 }
1817
1818 /* Do QD, or 2 * CPUS, whatever is smallest */
1819 ctx->sqo_wq = alloc_workqueue("io_ring-wq", WQ_UNBOUND | WQ_FREEZABLE,
1820 min(ctx->sq_entries - 1, 2 * num_online_cpus()));
1821 if (!ctx->sqo_wq) {
1822 ret = -ENOMEM;
1823 goto err;
1824 }
1825
1826 return 0;
1827 err:
1828 io_sq_thread_stop(ctx);
1829 mmdrop(ctx->sqo_mm);
1830 ctx->sqo_mm = NULL;
1831 return ret;
1832 }
1833
1834 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
1835 {
1836 atomic_long_sub(nr_pages, &user->locked_vm);
1837 }
1838
1839 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
1840 {
1841 unsigned long page_limit, cur_pages, new_pages;
1842
1843 /* Don't allow more pages than we can safely lock */
1844 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1845
1846 do {
1847 cur_pages = atomic_long_read(&user->locked_vm);
1848 new_pages = cur_pages + nr_pages;
1849 if (new_pages > page_limit)
1850 return -ENOMEM;
1851 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
1852 new_pages) != cur_pages);
1853
1854 return 0;
1855 }
1856
1857 static void io_mem_free(void *ptr)
1858 {
1859 struct page *page = virt_to_head_page(ptr);
1860
1861 if (put_page_testzero(page))
1862 free_compound_page(page);
1863 }
1864
1865 static void *io_mem_alloc(size_t size)
1866 {
1867 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
1868 __GFP_NORETRY;
1869
1870 return (void *) __get_free_pages(gfp_flags, get_order(size));
1871 }
1872
1873 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
1874 {
1875 struct io_sq_ring *sq_ring;
1876 struct io_cq_ring *cq_ring;
1877 size_t bytes;
1878
1879 bytes = struct_size(sq_ring, array, sq_entries);
1880 bytes += array_size(sizeof(struct io_uring_sqe), sq_entries);
1881 bytes += struct_size(cq_ring, cqes, cq_entries);
1882
1883 return (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1884 }
1885
1886 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
1887 {
1888 int i, j;
1889
1890 if (!ctx->user_bufs)
1891 return -ENXIO;
1892
1893 for (i = 0; i < ctx->nr_user_bufs; i++) {
1894 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
1895
1896 for (j = 0; j < imu->nr_bvecs; j++)
1897 put_page(imu->bvec[j].bv_page);
1898
1899 if (ctx->account_mem)
1900 io_unaccount_mem(ctx->user, imu->nr_bvecs);
1901 kfree(imu->bvec);
1902 imu->nr_bvecs = 0;
1903 }
1904
1905 kfree(ctx->user_bufs);
1906 ctx->user_bufs = NULL;
1907 ctx->nr_user_bufs = 0;
1908 return 0;
1909 }
1910
1911 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
1912 void __user *arg, unsigned index)
1913 {
1914 struct iovec __user *src;
1915
1916 #ifdef CONFIG_COMPAT
1917 if (ctx->compat) {
1918 struct compat_iovec __user *ciovs;
1919 struct compat_iovec ciov;
1920
1921 ciovs = (struct compat_iovec __user *) arg;
1922 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
1923 return -EFAULT;
1924
1925 dst->iov_base = (void __user *) (unsigned long) ciov.iov_base;
1926 dst->iov_len = ciov.iov_len;
1927 return 0;
1928 }
1929 #endif
1930 src = (struct iovec __user *) arg;
1931 if (copy_from_user(dst, &src[index], sizeof(*dst)))
1932 return -EFAULT;
1933 return 0;
1934 }
1935
1936 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
1937 unsigned nr_args)
1938 {
1939 struct vm_area_struct **vmas = NULL;
1940 struct page **pages = NULL;
1941 int i, j, got_pages = 0;
1942 int ret = -EINVAL;
1943
1944 if (ctx->user_bufs)
1945 return -EBUSY;
1946 if (!nr_args || nr_args > UIO_MAXIOV)
1947 return -EINVAL;
1948
1949 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
1950 GFP_KERNEL);
1951 if (!ctx->user_bufs)
1952 return -ENOMEM;
1953
1954 for (i = 0; i < nr_args; i++) {
1955 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
1956 unsigned long off, start, end, ubuf;
1957 int pret, nr_pages;
1958 struct iovec iov;
1959 size_t size;
1960
1961 ret = io_copy_iov(ctx, &iov, arg, i);
1962 if (ret)
1963 break;
1964
1965 /*
1966 * Don't impose further limits on the size and buffer
1967 * constraints here, we'll -EINVAL later when IO is
1968 * submitted if they are wrong.
1969 */
1970 ret = -EFAULT;
1971 if (!iov.iov_base || !iov.iov_len)
1972 goto err;
1973
1974 /* arbitrary limit, but we need something */
1975 if (iov.iov_len > SZ_1G)
1976 goto err;
1977
1978 ubuf = (unsigned long) iov.iov_base;
1979 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1980 start = ubuf >> PAGE_SHIFT;
1981 nr_pages = end - start;
1982
1983 if (ctx->account_mem) {
1984 ret = io_account_mem(ctx->user, nr_pages);
1985 if (ret)
1986 goto err;
1987 }
1988
1989 ret = 0;
1990 if (!pages || nr_pages > got_pages) {
1991 kfree(vmas);
1992 kfree(pages);
1993 pages = kmalloc_array(nr_pages, sizeof(struct page *),
1994 GFP_KERNEL);
1995 vmas = kmalloc_array(nr_pages,
1996 sizeof(struct vm_area_struct *),
1997 GFP_KERNEL);
1998 if (!pages || !vmas) {
1999 ret = -ENOMEM;
2000 if (ctx->account_mem)
2001 io_unaccount_mem(ctx->user, nr_pages);
2002 goto err;
2003 }
2004 got_pages = nr_pages;
2005 }
2006
2007 imu->bvec = kmalloc_array(nr_pages, sizeof(struct bio_vec),
2008 GFP_KERNEL);
2009 ret = -ENOMEM;
2010 if (!imu->bvec) {
2011 if (ctx->account_mem)
2012 io_unaccount_mem(ctx->user, nr_pages);
2013 goto err;
2014 }
2015
2016 ret = 0;
2017 down_read(&current->mm->mmap_sem);
2018 pret = get_user_pages_longterm(ubuf, nr_pages, FOLL_WRITE,
2019 pages, vmas);
2020 if (pret == nr_pages) {
2021 /* don't support file backed memory */
2022 for (j = 0; j < nr_pages; j++) {
2023 struct vm_area_struct *vma = vmas[j];
2024
2025 if (vma->vm_file &&
2026 !is_file_hugepages(vma->vm_file)) {
2027 ret = -EOPNOTSUPP;
2028 break;
2029 }
2030 }
2031 } else {
2032 ret = pret < 0 ? pret : -EFAULT;
2033 }
2034 up_read(&current->mm->mmap_sem);
2035 if (ret) {
2036 /*
2037 * if we did partial map, or found file backed vmas,
2038 * release any pages we did get
2039 */
2040 if (pret > 0) {
2041 for (j = 0; j < pret; j++)
2042 put_page(pages[j]);
2043 }
2044 if (ctx->account_mem)
2045 io_unaccount_mem(ctx->user, nr_pages);
2046 goto err;
2047 }
2048
2049 off = ubuf & ~PAGE_MASK;
2050 size = iov.iov_len;
2051 for (j = 0; j < nr_pages; j++) {
2052 size_t vec_len;
2053
2054 vec_len = min_t(size_t, size, PAGE_SIZE - off);
2055 imu->bvec[j].bv_page = pages[j];
2056 imu->bvec[j].bv_len = vec_len;
2057 imu->bvec[j].bv_offset = off;
2058 off = 0;
2059 size -= vec_len;
2060 }
2061 /* store original address for later verification */
2062 imu->ubuf = ubuf;
2063 imu->len = iov.iov_len;
2064 imu->nr_bvecs = nr_pages;
2065
2066 ctx->nr_user_bufs++;
2067 }
2068 kfree(pages);
2069 kfree(vmas);
2070 return 0;
2071 err:
2072 kfree(pages);
2073 kfree(vmas);
2074 io_sqe_buffer_unregister(ctx);
2075 return ret;
2076 }
2077
2078 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
2079 {
2080 io_finish_async(ctx);
2081 if (ctx->sqo_mm)
2082 mmdrop(ctx->sqo_mm);
2083
2084 io_iopoll_reap_events(ctx);
2085 io_sqe_buffer_unregister(ctx);
2086 io_sqe_files_unregister(ctx);
2087
2088 #if defined(CONFIG_UNIX)
2089 if (ctx->ring_sock)
2090 sock_release(ctx->ring_sock);
2091 #endif
2092
2093 io_mem_free(ctx->sq_ring);
2094 io_mem_free(ctx->sq_sqes);
2095 io_mem_free(ctx->cq_ring);
2096
2097 percpu_ref_exit(&ctx->refs);
2098 if (ctx->account_mem)
2099 io_unaccount_mem(ctx->user,
2100 ring_pages(ctx->sq_entries, ctx->cq_entries));
2101 free_uid(ctx->user);
2102 kfree(ctx);
2103 }
2104
2105 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2106 {
2107 struct io_ring_ctx *ctx = file->private_data;
2108 __poll_t mask = 0;
2109
2110 poll_wait(file, &ctx->cq_wait, wait);
2111 /* See comment at the top of this file */
2112 smp_rmb();
2113 if (READ_ONCE(ctx->sq_ring->r.tail) + 1 != ctx->cached_sq_head)
2114 mask |= EPOLLOUT | EPOLLWRNORM;
2115 if (READ_ONCE(ctx->cq_ring->r.head) != ctx->cached_cq_tail)
2116 mask |= EPOLLIN | EPOLLRDNORM;
2117
2118 return mask;
2119 }
2120
2121 static int io_uring_fasync(int fd, struct file *file, int on)
2122 {
2123 struct io_ring_ctx *ctx = file->private_data;
2124
2125 return fasync_helper(fd, file, on, &ctx->cq_fasync);
2126 }
2127
2128 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2129 {
2130 mutex_lock(&ctx->uring_lock);
2131 percpu_ref_kill(&ctx->refs);
2132 mutex_unlock(&ctx->uring_lock);
2133
2134 io_iopoll_reap_events(ctx);
2135 wait_for_completion(&ctx->ctx_done);
2136 io_ring_ctx_free(ctx);
2137 }
2138
2139 static int io_uring_release(struct inode *inode, struct file *file)
2140 {
2141 struct io_ring_ctx *ctx = file->private_data;
2142
2143 file->private_data = NULL;
2144 io_ring_ctx_wait_and_kill(ctx);
2145 return 0;
2146 }
2147
2148 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
2149 {
2150 loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT;
2151 unsigned long sz = vma->vm_end - vma->vm_start;
2152 struct io_ring_ctx *ctx = file->private_data;
2153 unsigned long pfn;
2154 struct page *page;
2155 void *ptr;
2156
2157 switch (offset) {
2158 case IORING_OFF_SQ_RING:
2159 ptr = ctx->sq_ring;
2160 break;
2161 case IORING_OFF_SQES:
2162 ptr = ctx->sq_sqes;
2163 break;
2164 case IORING_OFF_CQ_RING:
2165 ptr = ctx->cq_ring;
2166 break;
2167 default:
2168 return -EINVAL;
2169 }
2170
2171 page = virt_to_head_page(ptr);
2172 if (sz > (PAGE_SIZE << compound_order(page)))
2173 return -EINVAL;
2174
2175 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
2176 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
2177 }
2178
2179 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
2180 u32, min_complete, u32, flags, const sigset_t __user *, sig,
2181 size_t, sigsz)
2182 {
2183 struct io_ring_ctx *ctx;
2184 long ret = -EBADF;
2185 int submitted = 0;
2186 struct fd f;
2187
2188 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
2189 return -EINVAL;
2190
2191 f = fdget(fd);
2192 if (!f.file)
2193 return -EBADF;
2194
2195 ret = -EOPNOTSUPP;
2196 if (f.file->f_op != &io_uring_fops)
2197 goto out_fput;
2198
2199 ret = -ENXIO;
2200 ctx = f.file->private_data;
2201 if (!percpu_ref_tryget(&ctx->refs))
2202 goto out_fput;
2203
2204 /*
2205 * For SQ polling, the thread will do all submissions and completions.
2206 * Just return the requested submit count, and wake the thread if
2207 * we were asked to.
2208 */
2209 if (ctx->flags & IORING_SETUP_SQPOLL) {
2210 if (flags & IORING_ENTER_SQ_WAKEUP)
2211 wake_up(&ctx->sqo_wait);
2212 submitted = to_submit;
2213 goto out_ctx;
2214 }
2215
2216 ret = 0;
2217 if (to_submit) {
2218 to_submit = min(to_submit, ctx->sq_entries);
2219
2220 mutex_lock(&ctx->uring_lock);
2221 submitted = io_ring_submit(ctx, to_submit);
2222 mutex_unlock(&ctx->uring_lock);
2223
2224 if (submitted < 0)
2225 goto out_ctx;
2226 }
2227 if (flags & IORING_ENTER_GETEVENTS) {
2228 unsigned nr_events = 0;
2229
2230 min_complete = min(min_complete, ctx->cq_entries);
2231
2232 /*
2233 * The application could have included the 'to_submit' count
2234 * in how many events it wanted to wait for. If we failed to
2235 * submit the desired count, we may need to adjust the number
2236 * of events to poll/wait for.
2237 */
2238 if (submitted < to_submit)
2239 min_complete = min_t(unsigned, submitted, min_complete);
2240
2241 if (ctx->flags & IORING_SETUP_IOPOLL) {
2242 mutex_lock(&ctx->uring_lock);
2243 ret = io_iopoll_check(ctx, &nr_events, min_complete);
2244 mutex_unlock(&ctx->uring_lock);
2245 } else {
2246 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
2247 }
2248 }
2249
2250 out_ctx:
2251 io_ring_drop_ctx_refs(ctx, 1);
2252 out_fput:
2253 fdput(f);
2254 return submitted ? submitted : ret;
2255 }
2256
2257 static const struct file_operations io_uring_fops = {
2258 .release = io_uring_release,
2259 .mmap = io_uring_mmap,
2260 .poll = io_uring_poll,
2261 .fasync = io_uring_fasync,
2262 };
2263
2264 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
2265 struct io_uring_params *p)
2266 {
2267 struct io_sq_ring *sq_ring;
2268 struct io_cq_ring *cq_ring;
2269 size_t size;
2270
2271 sq_ring = io_mem_alloc(struct_size(sq_ring, array, p->sq_entries));
2272 if (!sq_ring)
2273 return -ENOMEM;
2274
2275 ctx->sq_ring = sq_ring;
2276 sq_ring->ring_mask = p->sq_entries - 1;
2277 sq_ring->ring_entries = p->sq_entries;
2278 ctx->sq_mask = sq_ring->ring_mask;
2279 ctx->sq_entries = sq_ring->ring_entries;
2280
2281 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
2282 if (size == SIZE_MAX)
2283 return -EOVERFLOW;
2284
2285 ctx->sq_sqes = io_mem_alloc(size);
2286 if (!ctx->sq_sqes) {
2287 io_mem_free(ctx->sq_ring);
2288 return -ENOMEM;
2289 }
2290
2291 cq_ring = io_mem_alloc(struct_size(cq_ring, cqes, p->cq_entries));
2292 if (!cq_ring) {
2293 io_mem_free(ctx->sq_ring);
2294 io_mem_free(ctx->sq_sqes);
2295 return -ENOMEM;
2296 }
2297
2298 ctx->cq_ring = cq_ring;
2299 cq_ring->ring_mask = p->cq_entries - 1;
2300 cq_ring->ring_entries = p->cq_entries;
2301 ctx->cq_mask = cq_ring->ring_mask;
2302 ctx->cq_entries = cq_ring->ring_entries;
2303 return 0;
2304 }
2305
2306 /*
2307 * Allocate an anonymous fd, this is what constitutes the application
2308 * visible backing of an io_uring instance. The application mmaps this
2309 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
2310 * we have to tie this fd to a socket for file garbage collection purposes.
2311 */
2312 static int io_uring_get_fd(struct io_ring_ctx *ctx)
2313 {
2314 struct file *file;
2315 int ret;
2316
2317 #if defined(CONFIG_UNIX)
2318 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
2319 &ctx->ring_sock);
2320 if (ret)
2321 return ret;
2322 #endif
2323
2324 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
2325 if (ret < 0)
2326 goto err;
2327
2328 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
2329 O_RDWR | O_CLOEXEC);
2330 if (IS_ERR(file)) {
2331 put_unused_fd(ret);
2332 ret = PTR_ERR(file);
2333 goto err;
2334 }
2335
2336 #if defined(CONFIG_UNIX)
2337 ctx->ring_sock->file = file;
2338 ctx->ring_sock->sk->sk_user_data = ctx;
2339 #endif
2340 fd_install(ret, file);
2341 return ret;
2342 err:
2343 #if defined(CONFIG_UNIX)
2344 sock_release(ctx->ring_sock);
2345 ctx->ring_sock = NULL;
2346 #endif
2347 return ret;
2348 }
2349
2350 static int io_uring_create(unsigned entries, struct io_uring_params *p)
2351 {
2352 struct user_struct *user = NULL;
2353 struct io_ring_ctx *ctx;
2354 bool account_mem;
2355 int ret;
2356
2357 if (!entries || entries > IORING_MAX_ENTRIES)
2358 return -EINVAL;
2359
2360 /*
2361 * Use twice as many entries for the CQ ring. It's possible for the
2362 * application to drive a higher depth than the size of the SQ ring,
2363 * since the sqes are only used at submission time. This allows for
2364 * some flexibility in overcommitting a bit.
2365 */
2366 p->sq_entries = roundup_pow_of_two(entries);
2367 p->cq_entries = 2 * p->sq_entries;
2368
2369 user = get_uid(current_user());
2370 account_mem = !capable(CAP_IPC_LOCK);
2371
2372 if (account_mem) {
2373 ret = io_account_mem(user,
2374 ring_pages(p->sq_entries, p->cq_entries));
2375 if (ret) {
2376 free_uid(user);
2377 return ret;
2378 }
2379 }
2380
2381 ctx = io_ring_ctx_alloc(p);
2382 if (!ctx) {
2383 if (account_mem)
2384 io_unaccount_mem(user, ring_pages(p->sq_entries,
2385 p->cq_entries));
2386 free_uid(user);
2387 return -ENOMEM;
2388 }
2389 ctx->compat = in_compat_syscall();
2390 ctx->account_mem = account_mem;
2391 ctx->user = user;
2392
2393 ret = io_allocate_scq_urings(ctx, p);
2394 if (ret)
2395 goto err;
2396
2397 ret = io_sq_offload_start(ctx, p);
2398 if (ret)
2399 goto err;
2400
2401 ret = io_uring_get_fd(ctx);
2402 if (ret < 0)
2403 goto err;
2404
2405 memset(&p->sq_off, 0, sizeof(p->sq_off));
2406 p->sq_off.head = offsetof(struct io_sq_ring, r.head);
2407 p->sq_off.tail = offsetof(struct io_sq_ring, r.tail);
2408 p->sq_off.ring_mask = offsetof(struct io_sq_ring, ring_mask);
2409 p->sq_off.ring_entries = offsetof(struct io_sq_ring, ring_entries);
2410 p->sq_off.flags = offsetof(struct io_sq_ring, flags);
2411 p->sq_off.dropped = offsetof(struct io_sq_ring, dropped);
2412 p->sq_off.array = offsetof(struct io_sq_ring, array);
2413
2414 memset(&p->cq_off, 0, sizeof(p->cq_off));
2415 p->cq_off.head = offsetof(struct io_cq_ring, r.head);
2416 p->cq_off.tail = offsetof(struct io_cq_ring, r.tail);
2417 p->cq_off.ring_mask = offsetof(struct io_cq_ring, ring_mask);
2418 p->cq_off.ring_entries = offsetof(struct io_cq_ring, ring_entries);
2419 p->cq_off.overflow = offsetof(struct io_cq_ring, overflow);
2420 p->cq_off.cqes = offsetof(struct io_cq_ring, cqes);
2421 return ret;
2422 err:
2423 io_ring_ctx_wait_and_kill(ctx);
2424 return ret;
2425 }
2426
2427 /*
2428 * Sets up an aio uring context, and returns the fd. Applications asks for a
2429 * ring size, we return the actual sq/cq ring sizes (among other things) in the
2430 * params structure passed in.
2431 */
2432 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
2433 {
2434 struct io_uring_params p;
2435 long ret;
2436 int i;
2437
2438 if (copy_from_user(&p, params, sizeof(p)))
2439 return -EFAULT;
2440 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
2441 if (p.resv[i])
2442 return -EINVAL;
2443 }
2444
2445 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
2446 IORING_SETUP_SQ_AFF))
2447 return -EINVAL;
2448
2449 ret = io_uring_create(entries, &p);
2450 if (ret < 0)
2451 return ret;
2452
2453 if (copy_to_user(params, &p, sizeof(p)))
2454 return -EFAULT;
2455
2456 return ret;
2457 }
2458
2459 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
2460 struct io_uring_params __user *, params)
2461 {
2462 return io_uring_setup(entries, params);
2463 }
2464
2465 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
2466 void __user *arg, unsigned nr_args)
2467 {
2468 int ret;
2469
2470 percpu_ref_kill(&ctx->refs);
2471 wait_for_completion(&ctx->ctx_done);
2472
2473 switch (opcode) {
2474 case IORING_REGISTER_BUFFERS:
2475 ret = io_sqe_buffer_register(ctx, arg, nr_args);
2476 break;
2477 case IORING_UNREGISTER_BUFFERS:
2478 ret = -EINVAL;
2479 if (arg || nr_args)
2480 break;
2481 ret = io_sqe_buffer_unregister(ctx);
2482 break;
2483 case IORING_REGISTER_FILES:
2484 ret = io_sqe_files_register(ctx, arg, nr_args);
2485 break;
2486 case IORING_UNREGISTER_FILES:
2487 ret = -EINVAL;
2488 if (arg || nr_args)
2489 break;
2490 ret = io_sqe_files_unregister(ctx);
2491 break;
2492 default:
2493 ret = -EINVAL;
2494 break;
2495 }
2496
2497 /* bring the ctx back to life */
2498 reinit_completion(&ctx->ctx_done);
2499 percpu_ref_reinit(&ctx->refs);
2500 return ret;
2501 }
2502
2503 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
2504 void __user *, arg, unsigned int, nr_args)
2505 {
2506 struct io_ring_ctx *ctx;
2507 long ret = -EBADF;
2508 struct fd f;
2509
2510 f = fdget(fd);
2511 if (!f.file)
2512 return -EBADF;
2513
2514 ret = -EOPNOTSUPP;
2515 if (f.file->f_op != &io_uring_fops)
2516 goto out_fput;
2517
2518 ctx = f.file->private_data;
2519
2520 mutex_lock(&ctx->uring_lock);
2521 ret = __io_uring_register(ctx, opcode, arg, nr_args);
2522 mutex_unlock(&ctx->uring_lock);
2523 out_fput:
2524 fdput(f);
2525 return ret;
2526 }
2527
2528 static int __init io_uring_init(void)
2529 {
2530 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
2531 return 0;
2532 };
2533 __initcall(io_uring_init);
Mirror https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git