io_uring: make sqpoll wakeup possible with getevents
[linux-block.git] / fs / io_uring.c
CommitLineData
2b188cc1
JA
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
5 *
6 * A note on the read/write ordering memory barriers that are matched between
1e84b97b
SB
7 * the application and kernel side.
8 *
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
16 * CQ entries.
17 *
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
23 * head will do).
24 *
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
28 * between.
2b188cc1
JA
29 *
30 * Also see the examples in the liburing library:
31 *
32 * git://git.kernel.dk/liburing
33 *
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
38 *
39 * Copyright (C) 2018-2019 Jens Axboe
c992fe29 40 * Copyright (c) 2018-2019 Christoph Hellwig
2b188cc1
JA
41 */
42#include <linux/kernel.h>
43#include <linux/init.h>
44#include <linux/errno.h>
45#include <linux/syscalls.h>
46#include <linux/compat.h>
47#include <linux/refcount.h>
48#include <linux/uio.h>
49
50#include <linux/sched/signal.h>
51#include <linux/fs.h>
52#include <linux/file.h>
53#include <linux/fdtable.h>
54#include <linux/mm.h>
55#include <linux/mman.h>
56#include <linux/mmu_context.h>
57#include <linux/percpu.h>
58#include <linux/slab.h>
59#include <linux/workqueue.h>
6c271ce2 60#include <linux/kthread.h>
2b188cc1 61#include <linux/blkdev.h>
edafccee 62#include <linux/bvec.h>
2b188cc1
JA
63#include <linux/net.h>
64#include <net/sock.h>
65#include <net/af_unix.h>
6b06314c 66#include <net/scm.h>
2b188cc1
JA
67#include <linux/anon_inodes.h>
68#include <linux/sched/mm.h>
69#include <linux/uaccess.h>
70#include <linux/nospec.h>
edafccee
JA
71#include <linux/sizes.h>
72#include <linux/hugetlb.h>
2b188cc1
JA
73
74#include <uapi/linux/io_uring.h>
75
76#include "internal.h"
77
78#define IORING_MAX_ENTRIES 4096
6b06314c 79#define IORING_MAX_FIXED_FILES 1024
2b188cc1
JA
80
81struct io_uring {
82 u32 head ____cacheline_aligned_in_smp;
83 u32 tail ____cacheline_aligned_in_smp;
84};
85
1e84b97b 86/*
75b28aff
HV
87 * This data is shared with the application through the mmap at offsets
88 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
1e84b97b
SB
89 *
90 * The offsets to the member fields are published through struct
91 * io_sqring_offsets when calling io_uring_setup.
92 */
75b28aff 93struct io_rings {
1e84b97b
SB
94 /*
95 * Head and tail offsets into the ring; the offsets need to be
96 * masked to get valid indices.
97 *
75b28aff
HV
98 * The kernel controls head of the sq ring and the tail of the cq ring,
99 * and the application controls tail of the sq ring and the head of the
100 * cq ring.
1e84b97b 101 */
75b28aff 102 struct io_uring sq, cq;
1e84b97b 103 /*
75b28aff 104 * Bitmasks to apply to head and tail offsets (constant, equals
1e84b97b
SB
105 * ring_entries - 1)
106 */
75b28aff
HV
107 u32 sq_ring_mask, cq_ring_mask;
108 /* Ring sizes (constant, power of 2) */
109 u32 sq_ring_entries, cq_ring_entries;
1e84b97b
SB
110 /*
111 * Number of invalid entries dropped by the kernel due to
112 * invalid index stored in array
113 *
114 * Written by the kernel, shouldn't be modified by the
115 * application (i.e. get number of "new events" by comparing to
116 * cached value).
117 *
118 * After a new SQ head value was read by the application this
119 * counter includes all submissions that were dropped reaching
120 * the new SQ head (and possibly more).
121 */
75b28aff 122 u32 sq_dropped;
1e84b97b
SB
123 /*
124 * Runtime flags
125 *
126 * Written by the kernel, shouldn't be modified by the
127 * application.
128 *
129 * The application needs a full memory barrier before checking
130 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
131 */
75b28aff 132 u32 sq_flags;
1e84b97b
SB
133 /*
134 * Number of completion events lost because the queue was full;
135 * this should be avoided by the application by making sure
136 * there are not more requests pending thatn there is space in
137 * the completion queue.
138 *
139 * Written by the kernel, shouldn't be modified by the
140 * application (i.e. get number of "new events" by comparing to
141 * cached value).
142 *
143 * As completion events come in out of order this counter is not
144 * ordered with any other data.
145 */
75b28aff 146 u32 cq_overflow;
1e84b97b
SB
147 /*
148 * Ring buffer of completion events.
149 *
150 * The kernel writes completion events fresh every time they are
151 * produced, so the application is allowed to modify pending
152 * entries.
153 */
75b28aff 154 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
2b188cc1
JA
155};
156
edafccee
JA
157struct io_mapped_ubuf {
158 u64 ubuf;
159 size_t len;
160 struct bio_vec *bvec;
161 unsigned int nr_bvecs;
162};
163
31b51510
JA
164struct async_list {
165 spinlock_t lock;
166 atomic_t cnt;
167 struct list_head list;
168
169 struct file *file;
6d5d5ac5 170 off_t io_start;
9310a7ba 171 size_t io_len;
31b51510
JA
172};
173
2b188cc1
JA
174struct io_ring_ctx {
175 struct {
176 struct percpu_ref refs;
177 } ____cacheline_aligned_in_smp;
178
179 struct {
180 unsigned int flags;
181 bool compat;
182 bool account_mem;
183
75b28aff
HV
184 /*
185 * Ring buffer of indices into array of io_uring_sqe, which is
186 * mmapped by the application using the IORING_OFF_SQES offset.
187 *
188 * This indirection could e.g. be used to assign fixed
189 * io_uring_sqe entries to operations and only submit them to
190 * the queue when needed.
191 *
192 * The kernel modifies neither the indices array nor the entries
193 * array.
194 */
195 u32 *sq_array;
2b188cc1
JA
196 unsigned cached_sq_head;
197 unsigned sq_entries;
198 unsigned sq_mask;
6c271ce2 199 unsigned sq_thread_idle;
2b188cc1 200 struct io_uring_sqe *sq_sqes;
de0617e4
JA
201
202 struct list_head defer_list;
2b188cc1
JA
203 } ____cacheline_aligned_in_smp;
204
205 /* IO offload */
54a91f3b 206 struct workqueue_struct *sqo_wq[2];
6c271ce2 207 struct task_struct *sqo_thread; /* if using sq thread polling */
2b188cc1 208 struct mm_struct *sqo_mm;
6c271ce2 209 wait_queue_head_t sqo_wait;
a4c0b3de 210 struct completion sqo_thread_started;
2b188cc1
JA
211
212 struct {
2b188cc1
JA
213 unsigned cached_cq_tail;
214 unsigned cq_entries;
215 unsigned cq_mask;
216 struct wait_queue_head cq_wait;
217 struct fasync_struct *cq_fasync;
9b402849 218 struct eventfd_ctx *cq_ev_fd;
2b188cc1
JA
219 } ____cacheline_aligned_in_smp;
220
75b28aff
HV
221 struct io_rings *rings;
222
6b06314c
JA
223 /*
224 * If used, fixed file set. Writers must ensure that ->refs is dead,
225 * readers must ensure that ->refs is alive as long as the file* is
226 * used. Only updated through io_uring_register(2).
227 */
228 struct file **user_files;
229 unsigned nr_user_files;
230
edafccee
JA
231 /* if used, fixed mapped user buffers */
232 unsigned nr_user_bufs;
233 struct io_mapped_ubuf *user_bufs;
234
2b188cc1
JA
235 struct user_struct *user;
236
237 struct completion ctx_done;
238
239 struct {
240 struct mutex uring_lock;
241 wait_queue_head_t wait;
242 } ____cacheline_aligned_in_smp;
243
244 struct {
245 spinlock_t completion_lock;
def596e9
JA
246 bool poll_multi_file;
247 /*
248 * ->poll_list is protected by the ctx->uring_lock for
249 * io_uring instances that don't use IORING_SETUP_SQPOLL.
250 * For SQPOLL, only the single threaded io_sq_thread() will
251 * manipulate the list, hence no extra locking is needed there.
252 */
253 struct list_head poll_list;
221c5eb2 254 struct list_head cancel_list;
2b188cc1
JA
255 } ____cacheline_aligned_in_smp;
256
31b51510
JA
257 struct async_list pending_async[2];
258
2b188cc1
JA
259#if defined(CONFIG_UNIX)
260 struct socket *ring_sock;
261#endif
262};
263
264struct sqe_submit {
265 const struct io_uring_sqe *sqe;
266 unsigned short index;
8776f3fa 267 u32 sequence;
2b188cc1 268 bool has_user;
def596e9 269 bool needs_lock;
6c271ce2 270 bool needs_fixed_file;
2b188cc1
JA
271};
272
09bb8394
JA
273/*
274 * First field must be the file pointer in all the
275 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
276 */
221c5eb2
JA
277struct io_poll_iocb {
278 struct file *file;
279 struct wait_queue_head *head;
280 __poll_t events;
8c838788 281 bool done;
221c5eb2
JA
282 bool canceled;
283 struct wait_queue_entry wait;
284};
285
09bb8394
JA
286/*
287 * NOTE! Each of the iocb union members has the file pointer
288 * as the first entry in their struct definition. So you can
289 * access the file pointer through any of the sub-structs,
290 * or directly as just 'ki_filp' in this struct.
291 */
2b188cc1 292struct io_kiocb {
221c5eb2 293 union {
09bb8394 294 struct file *file;
221c5eb2
JA
295 struct kiocb rw;
296 struct io_poll_iocb poll;
297 };
2b188cc1
JA
298
299 struct sqe_submit submit;
300
301 struct io_ring_ctx *ctx;
302 struct list_head list;
9e645e11 303 struct list_head link_list;
2b188cc1 304 unsigned int flags;
c16361c1 305 refcount_t refs;
8449eeda 306#define REQ_F_NOWAIT 1 /* must not punt to workers */
def596e9 307#define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */
6b06314c 308#define REQ_F_FIXED_FILE 4 /* ctx owns file */
31b51510 309#define REQ_F_SEQ_PREV 8 /* sequential with previous */
e2033e33
SB
310#define REQ_F_IO_DRAIN 16 /* drain existing IO first */
311#define REQ_F_IO_DRAINED 32 /* drain done */
9e645e11 312#define REQ_F_LINK 64 /* linked sqes */
f7b76ac9
ZL
313#define REQ_F_LINK_DONE 128 /* linked sqes done */
314#define REQ_F_FAIL_LINK 256 /* fail rest of links */
4fe2c963 315#define REQ_F_SHADOW_DRAIN 512 /* link-drain shadow req */
2b188cc1 316 u64 user_data;
9e645e11 317 u32 result;
de0617e4 318 u32 sequence;
2b188cc1
JA
319
320 struct work_struct work;
321};
322
323#define IO_PLUG_THRESHOLD 2
def596e9 324#define IO_IOPOLL_BATCH 8
2b188cc1 325
9a56a232
JA
326struct io_submit_state {
327 struct blk_plug plug;
328
2579f913
JA
329 /*
330 * io_kiocb alloc cache
331 */
332 void *reqs[IO_IOPOLL_BATCH];
333 unsigned int free_reqs;
334 unsigned int cur_req;
335
9a56a232
JA
336 /*
337 * File reference cache
338 */
339 struct file *file;
340 unsigned int fd;
341 unsigned int has_refs;
342 unsigned int used_refs;
343 unsigned int ios_left;
344};
345
de0617e4 346static void io_sq_wq_submit_work(struct work_struct *work);
4fe2c963 347static void __io_free_req(struct io_kiocb *req);
de0617e4 348
2b188cc1
JA
349static struct kmem_cache *req_cachep;
350
351static const struct file_operations io_uring_fops;
352
353struct sock *io_uring_get_socket(struct file *file)
354{
355#if defined(CONFIG_UNIX)
356 if (file->f_op == &io_uring_fops) {
357 struct io_ring_ctx *ctx = file->private_data;
358
359 return ctx->ring_sock->sk;
360 }
361#endif
362 return NULL;
363}
364EXPORT_SYMBOL(io_uring_get_socket);
365
366static void io_ring_ctx_ref_free(struct percpu_ref *ref)
367{
368 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
369
370 complete(&ctx->ctx_done);
371}
372
373static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
374{
375 struct io_ring_ctx *ctx;
31b51510 376 int i;
2b188cc1
JA
377
378 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
379 if (!ctx)
380 return NULL;
381
21482896
RG
382 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
383 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
2b188cc1
JA
384 kfree(ctx);
385 return NULL;
386 }
387
388 ctx->flags = p->flags;
389 init_waitqueue_head(&ctx->cq_wait);
390 init_completion(&ctx->ctx_done);
a4c0b3de 391 init_completion(&ctx->sqo_thread_started);
2b188cc1
JA
392 mutex_init(&ctx->uring_lock);
393 init_waitqueue_head(&ctx->wait);
31b51510
JA
394 for (i = 0; i < ARRAY_SIZE(ctx->pending_async); i++) {
395 spin_lock_init(&ctx->pending_async[i].lock);
396 INIT_LIST_HEAD(&ctx->pending_async[i].list);
397 atomic_set(&ctx->pending_async[i].cnt, 0);
398 }
2b188cc1 399 spin_lock_init(&ctx->completion_lock);
def596e9 400 INIT_LIST_HEAD(&ctx->poll_list);
221c5eb2 401 INIT_LIST_HEAD(&ctx->cancel_list);
de0617e4 402 INIT_LIST_HEAD(&ctx->defer_list);
2b188cc1
JA
403 return ctx;
404}
405
de0617e4
JA
406static inline bool io_sequence_defer(struct io_ring_ctx *ctx,
407 struct io_kiocb *req)
408{
409 if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) != REQ_F_IO_DRAIN)
410 return false;
411
75b28aff 412 return req->sequence != ctx->cached_cq_tail + ctx->rings->sq_dropped;
de0617e4
JA
413}
414
415static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
416{
417 struct io_kiocb *req;
418
419 if (list_empty(&ctx->defer_list))
420 return NULL;
421
422 req = list_first_entry(&ctx->defer_list, struct io_kiocb, list);
423 if (!io_sequence_defer(ctx, req)) {
424 list_del_init(&req->list);
425 return req;
426 }
427
428 return NULL;
429}
430
431static void __io_commit_cqring(struct io_ring_ctx *ctx)
2b188cc1 432{
75b28aff 433 struct io_rings *rings = ctx->rings;
2b188cc1 434
75b28aff 435 if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) {
2b188cc1 436 /* order cqe stores with ring update */
75b28aff 437 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
2b188cc1 438
2b188cc1
JA
439 if (wq_has_sleeper(&ctx->cq_wait)) {
440 wake_up_interruptible(&ctx->cq_wait);
441 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
442 }
443 }
444}
445
18d9be1a
JA
446static inline void io_queue_async_work(struct io_ring_ctx *ctx,
447 struct io_kiocb *req)
448{
54a91f3b
JA
449 int rw;
450
451 switch (req->submit.sqe->opcode) {
452 case IORING_OP_WRITEV:
453 case IORING_OP_WRITE_FIXED:
454 rw = !(req->rw.ki_flags & IOCB_DIRECT);
455 break;
456 default:
457 rw = 0;
458 break;
459 }
460
461 queue_work(ctx->sqo_wq[rw], &req->work);
18d9be1a
JA
462}
463
de0617e4
JA
464static void io_commit_cqring(struct io_ring_ctx *ctx)
465{
466 struct io_kiocb *req;
467
468 __io_commit_cqring(ctx);
469
470 while ((req = io_get_deferred_req(ctx)) != NULL) {
4fe2c963
JL
471 if (req->flags & REQ_F_SHADOW_DRAIN) {
472 /* Just for drain, free it. */
473 __io_free_req(req);
474 continue;
475 }
de0617e4 476 req->flags |= REQ_F_IO_DRAINED;
18d9be1a 477 io_queue_async_work(ctx, req);
de0617e4
JA
478 }
479}
480
2b188cc1
JA
481static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
482{
75b28aff 483 struct io_rings *rings = ctx->rings;
2b188cc1
JA
484 unsigned tail;
485
486 tail = ctx->cached_cq_tail;
115e12e5
SB
487 /*
488 * writes to the cq entry need to come after reading head; the
489 * control dependency is enough as we're using WRITE_ONCE to
490 * fill the cq entry
491 */
75b28aff 492 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
2b188cc1
JA
493 return NULL;
494
495 ctx->cached_cq_tail++;
75b28aff 496 return &rings->cqes[tail & ctx->cq_mask];
2b188cc1
JA
497}
498
499static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data,
c71ffb67 500 long res)
2b188cc1
JA
501{
502 struct io_uring_cqe *cqe;
503
504 /*
505 * If we can't get a cq entry, userspace overflowed the
506 * submission (by quite a lot). Increment the overflow count in
507 * the ring.
508 */
509 cqe = io_get_cqring(ctx);
510 if (cqe) {
511 WRITE_ONCE(cqe->user_data, ki_user_data);
512 WRITE_ONCE(cqe->res, res);
c71ffb67 513 WRITE_ONCE(cqe->flags, 0);
2b188cc1 514 } else {
75b28aff 515 unsigned overflow = READ_ONCE(ctx->rings->cq_overflow);
2b188cc1 516
75b28aff 517 WRITE_ONCE(ctx->rings->cq_overflow, overflow + 1);
2b188cc1
JA
518 }
519}
520
8c838788
JA
521static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
522{
523 if (waitqueue_active(&ctx->wait))
524 wake_up(&ctx->wait);
525 if (waitqueue_active(&ctx->sqo_wait))
526 wake_up(&ctx->sqo_wait);
9b402849
JA
527 if (ctx->cq_ev_fd)
528 eventfd_signal(ctx->cq_ev_fd, 1);
8c838788
JA
529}
530
531static void io_cqring_add_event(struct io_ring_ctx *ctx, u64 user_data,
c71ffb67 532 long res)
2b188cc1
JA
533{
534 unsigned long flags;
535
536 spin_lock_irqsave(&ctx->completion_lock, flags);
c71ffb67 537 io_cqring_fill_event(ctx, user_data, res);
2b188cc1
JA
538 io_commit_cqring(ctx);
539 spin_unlock_irqrestore(&ctx->completion_lock, flags);
540
8c838788 541 io_cqring_ev_posted(ctx);
2b188cc1
JA
542}
543
544static void io_ring_drop_ctx_refs(struct io_ring_ctx *ctx, unsigned refs)
545{
546 percpu_ref_put_many(&ctx->refs, refs);
547
548 if (waitqueue_active(&ctx->wait))
549 wake_up(&ctx->wait);
550}
551
2579f913
JA
552static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
553 struct io_submit_state *state)
2b188cc1 554{
fd6fab2c 555 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
2b188cc1
JA
556 struct io_kiocb *req;
557
558 if (!percpu_ref_tryget(&ctx->refs))
559 return NULL;
560
2579f913 561 if (!state) {
fd6fab2c 562 req = kmem_cache_alloc(req_cachep, gfp);
2579f913
JA
563 if (unlikely(!req))
564 goto out;
565 } else if (!state->free_reqs) {
566 size_t sz;
567 int ret;
568
569 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
fd6fab2c
JA
570 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
571
572 /*
573 * Bulk alloc is all-or-nothing. If we fail to get a batch,
574 * retry single alloc to be on the safe side.
575 */
576 if (unlikely(ret <= 0)) {
577 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
578 if (!state->reqs[0])
579 goto out;
580 ret = 1;
581 }
2579f913
JA
582 state->free_reqs = ret - 1;
583 state->cur_req = 1;
584 req = state->reqs[0];
585 } else {
586 req = state->reqs[state->cur_req];
587 state->free_reqs--;
588 state->cur_req++;
2b188cc1
JA
589 }
590
60c112b0 591 req->file = NULL;
2579f913
JA
592 req->ctx = ctx;
593 req->flags = 0;
e65ef56d
JA
594 /* one is dropped after submission, the other at completion */
595 refcount_set(&req->refs, 2);
9e645e11 596 req->result = 0;
2579f913
JA
597 return req;
598out:
2b188cc1
JA
599 io_ring_drop_ctx_refs(ctx, 1);
600 return NULL;
601}
602
def596e9
JA
603static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr)
604{
605 if (*nr) {
606 kmem_cache_free_bulk(req_cachep, *nr, reqs);
607 io_ring_drop_ctx_refs(ctx, *nr);
608 *nr = 0;
609 }
610}
611
9e645e11 612static void __io_free_req(struct io_kiocb *req)
2b188cc1 613{
09bb8394
JA
614 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
615 fput(req->file);
e65ef56d
JA
616 io_ring_drop_ctx_refs(req->ctx, 1);
617 kmem_cache_free(req_cachep, req);
618}
619
9e645e11
JA
620static void io_req_link_next(struct io_kiocb *req)
621{
622 struct io_kiocb *nxt;
623
624 /*
625 * The list should never be empty when we are called here. But could
626 * potentially happen if the chain is messed up, check to be on the
627 * safe side.
628 */
629 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list);
630 if (nxt) {
631 list_del(&nxt->list);
632 if (!list_empty(&req->link_list)) {
633 INIT_LIST_HEAD(&nxt->link_list);
634 list_splice(&req->link_list, &nxt->link_list);
635 nxt->flags |= REQ_F_LINK;
636 }
637
f7b76ac9 638 nxt->flags |= REQ_F_LINK_DONE;
9e645e11 639 INIT_WORK(&nxt->work, io_sq_wq_submit_work);
18d9be1a 640 io_queue_async_work(req->ctx, nxt);
9e645e11
JA
641 }
642}
643
644/*
645 * Called if REQ_F_LINK is set, and we fail the head request
646 */
647static void io_fail_links(struct io_kiocb *req)
648{
649 struct io_kiocb *link;
650
651 while (!list_empty(&req->link_list)) {
652 link = list_first_entry(&req->link_list, struct io_kiocb, list);
653 list_del(&link->list);
654
655 io_cqring_add_event(req->ctx, link->user_data, -ECANCELED);
656 __io_free_req(link);
657 }
658}
659
660static void io_free_req(struct io_kiocb *req)
661{
662 /*
663 * If LINK is set, we have dependent requests in this chain. If we
664 * didn't fail this request, queue the first one up, moving any other
665 * dependencies to the next request. In case of failure, fail the rest
666 * of the chain.
667 */
668 if (req->flags & REQ_F_LINK) {
669 if (req->flags & REQ_F_FAIL_LINK)
670 io_fail_links(req);
671 else
672 io_req_link_next(req);
673 }
674
675 __io_free_req(req);
676}
677
e65ef56d
JA
678static void io_put_req(struct io_kiocb *req)
679{
680 if (refcount_dec_and_test(&req->refs))
681 io_free_req(req);
2b188cc1
JA
682}
683
75b28aff 684static unsigned io_cqring_events(struct io_rings *rings)
a3a0e43f
JA
685{
686 /* See comment at the top of this file */
687 smp_rmb();
75b28aff 688 return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head);
a3a0e43f
JA
689}
690
def596e9
JA
691/*
692 * Find and free completed poll iocbs
693 */
694static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
695 struct list_head *done)
696{
697 void *reqs[IO_IOPOLL_BATCH];
698 struct io_kiocb *req;
09bb8394 699 int to_free;
def596e9 700
09bb8394 701 to_free = 0;
def596e9
JA
702 while (!list_empty(done)) {
703 req = list_first_entry(done, struct io_kiocb, list);
704 list_del(&req->list);
705
9e645e11 706 io_cqring_fill_event(ctx, req->user_data, req->result);
def596e9
JA
707 (*nr_events)++;
708
09bb8394
JA
709 if (refcount_dec_and_test(&req->refs)) {
710 /* If we're not using fixed files, we have to pair the
711 * completion part with the file put. Use regular
712 * completions for those, only batch free for fixed
9e645e11 713 * file and non-linked commands.
09bb8394 714 */
9e645e11
JA
715 if ((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) ==
716 REQ_F_FIXED_FILE) {
09bb8394
JA
717 reqs[to_free++] = req;
718 if (to_free == ARRAY_SIZE(reqs))
719 io_free_req_many(ctx, reqs, &to_free);
6b06314c 720 } else {
09bb8394 721 io_free_req(req);
6b06314c 722 }
9a56a232 723 }
def596e9 724 }
def596e9 725
09bb8394 726 io_commit_cqring(ctx);
def596e9
JA
727 io_free_req_many(ctx, reqs, &to_free);
728}
729
730static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
731 long min)
732{
733 struct io_kiocb *req, *tmp;
734 LIST_HEAD(done);
735 bool spin;
736 int ret;
737
738 /*
739 * Only spin for completions if we don't have multiple devices hanging
740 * off our complete list, and we're under the requested amount.
741 */
742 spin = !ctx->poll_multi_file && *nr_events < min;
743
744 ret = 0;
745 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
746 struct kiocb *kiocb = &req->rw;
747
748 /*
749 * Move completed entries to our local list. If we find a
750 * request that requires polling, break out and complete
751 * the done list first, if we have entries there.
752 */
753 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
754 list_move_tail(&req->list, &done);
755 continue;
756 }
757 if (!list_empty(&done))
758 break;
759
760 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
761 if (ret < 0)
762 break;
763
764 if (ret && spin)
765 spin = false;
766 ret = 0;
767 }
768
769 if (!list_empty(&done))
770 io_iopoll_complete(ctx, nr_events, &done);
771
772 return ret;
773}
774
775/*
776 * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a
777 * non-spinning poll check - we'll still enter the driver poll loop, but only
778 * as a non-spinning completion check.
779 */
780static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
781 long min)
782{
08f5439f 783 while (!list_empty(&ctx->poll_list) && !need_resched()) {
def596e9
JA
784 int ret;
785
786 ret = io_do_iopoll(ctx, nr_events, min);
787 if (ret < 0)
788 return ret;
789 if (!min || *nr_events >= min)
790 return 0;
791 }
792
793 return 1;
794}
795
796/*
797 * We can't just wait for polled events to come to us, we have to actively
798 * find and complete them.
799 */
800static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
801{
802 if (!(ctx->flags & IORING_SETUP_IOPOLL))
803 return;
804
805 mutex_lock(&ctx->uring_lock);
806 while (!list_empty(&ctx->poll_list)) {
807 unsigned int nr_events = 0;
808
809 io_iopoll_getevents(ctx, &nr_events, 1);
08f5439f
JA
810
811 /*
812 * Ensure we allow local-to-the-cpu processing to take place,
813 * in this case we need to ensure that we reap all events.
814 */
815 cond_resched();
def596e9
JA
816 }
817 mutex_unlock(&ctx->uring_lock);
818}
819
820static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
821 long min)
822{
500f9fba
JA
823 int iters, ret = 0;
824
825 /*
826 * We disallow the app entering submit/complete with polling, but we
827 * still need to lock the ring to prevent racing with polled issue
828 * that got punted to a workqueue.
829 */
830 mutex_lock(&ctx->uring_lock);
def596e9 831
500f9fba 832 iters = 0;
def596e9
JA
833 do {
834 int tmin = 0;
835
a3a0e43f
JA
836 /*
837 * Don't enter poll loop if we already have events pending.
838 * If we do, we can potentially be spinning for commands that
839 * already triggered a CQE (eg in error).
840 */
75b28aff 841 if (io_cqring_events(ctx->rings))
a3a0e43f
JA
842 break;
843
500f9fba
JA
844 /*
845 * If a submit got punted to a workqueue, we can have the
846 * application entering polling for a command before it gets
847 * issued. That app will hold the uring_lock for the duration
848 * of the poll right here, so we need to take a breather every
849 * now and then to ensure that the issue has a chance to add
850 * the poll to the issued list. Otherwise we can spin here
851 * forever, while the workqueue is stuck trying to acquire the
852 * very same mutex.
853 */
854 if (!(++iters & 7)) {
855 mutex_unlock(&ctx->uring_lock);
856 mutex_lock(&ctx->uring_lock);
857 }
858
def596e9
JA
859 if (*nr_events < min)
860 tmin = min - *nr_events;
861
862 ret = io_iopoll_getevents(ctx, nr_events, tmin);
863 if (ret <= 0)
864 break;
865 ret = 0;
866 } while (min && !*nr_events && !need_resched());
867
500f9fba 868 mutex_unlock(&ctx->uring_lock);
def596e9
JA
869 return ret;
870}
871
2b188cc1
JA
872static void kiocb_end_write(struct kiocb *kiocb)
873{
874 if (kiocb->ki_flags & IOCB_WRITE) {
875 struct inode *inode = file_inode(kiocb->ki_filp);
876
877 /*
878 * Tell lockdep we inherited freeze protection from submission
879 * thread.
880 */
881 if (S_ISREG(inode->i_mode))
882 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
883 file_end_write(kiocb->ki_filp);
884 }
885}
886
887static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
888{
889 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
890
891 kiocb_end_write(kiocb);
892
9e645e11
JA
893 if ((req->flags & REQ_F_LINK) && res != req->result)
894 req->flags |= REQ_F_FAIL_LINK;
c71ffb67 895 io_cqring_add_event(req->ctx, req->user_data, res);
e65ef56d 896 io_put_req(req);
2b188cc1
JA
897}
898
def596e9
JA
899static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
900{
901 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
902
903 kiocb_end_write(kiocb);
904
9e645e11
JA
905 if ((req->flags & REQ_F_LINK) && res != req->result)
906 req->flags |= REQ_F_FAIL_LINK;
907 req->result = res;
def596e9
JA
908 if (res != -EAGAIN)
909 req->flags |= REQ_F_IOPOLL_COMPLETED;
910}
911
912/*
913 * After the iocb has been issued, it's safe to be found on the poll list.
914 * Adding the kiocb to the list AFTER submission ensures that we don't
915 * find it from a io_iopoll_getevents() thread before the issuer is done
916 * accessing the kiocb cookie.
917 */
918static void io_iopoll_req_issued(struct io_kiocb *req)
919{
920 struct io_ring_ctx *ctx = req->ctx;
921
922 /*
923 * Track whether we have multiple files in our lists. This will impact
924 * how we do polling eventually, not spinning if we're on potentially
925 * different devices.
926 */
927 if (list_empty(&ctx->poll_list)) {
928 ctx->poll_multi_file = false;
929 } else if (!ctx->poll_multi_file) {
930 struct io_kiocb *list_req;
931
932 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
933 list);
934 if (list_req->rw.ki_filp != req->rw.ki_filp)
935 ctx->poll_multi_file = true;
936 }
937
938 /*
939 * For fast devices, IO may have already completed. If it has, add
940 * it to the front so we find it first.
941 */
942 if (req->flags & REQ_F_IOPOLL_COMPLETED)
943 list_add(&req->list, &ctx->poll_list);
944 else
945 list_add_tail(&req->list, &ctx->poll_list);
946}
947
3d6770fb 948static void io_file_put(struct io_submit_state *state)
9a56a232 949{
3d6770fb 950 if (state->file) {
9a56a232
JA
951 int diff = state->has_refs - state->used_refs;
952
953 if (diff)
954 fput_many(state->file, diff);
955 state->file = NULL;
956 }
957}
958
959/*
960 * Get as many references to a file as we have IOs left in this submission,
961 * assuming most submissions are for one file, or at least that each file
962 * has more than one submission.
963 */
964static struct file *io_file_get(struct io_submit_state *state, int fd)
965{
966 if (!state)
967 return fget(fd);
968
969 if (state->file) {
970 if (state->fd == fd) {
971 state->used_refs++;
972 state->ios_left--;
973 return state->file;
974 }
3d6770fb 975 io_file_put(state);
9a56a232
JA
976 }
977 state->file = fget_many(fd, state->ios_left);
978 if (!state->file)
979 return NULL;
980
981 state->fd = fd;
982 state->has_refs = state->ios_left;
983 state->used_refs = 1;
984 state->ios_left--;
985 return state->file;
986}
987
2b188cc1
JA
988/*
989 * If we tracked the file through the SCM inflight mechanism, we could support
990 * any file. For now, just ensure that anything potentially problematic is done
991 * inline.
992 */
993static bool io_file_supports_async(struct file *file)
994{
995 umode_t mode = file_inode(file)->i_mode;
996
997 if (S_ISBLK(mode) || S_ISCHR(mode))
998 return true;
999 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
1000 return true;
1001
1002 return false;
1003}
1004
6c271ce2 1005static int io_prep_rw(struct io_kiocb *req, const struct sqe_submit *s,
8358e3a8 1006 bool force_nonblock)
2b188cc1 1007{
6c271ce2 1008 const struct io_uring_sqe *sqe = s->sqe;
def596e9 1009 struct io_ring_ctx *ctx = req->ctx;
2b188cc1 1010 struct kiocb *kiocb = &req->rw;
09bb8394
JA
1011 unsigned ioprio;
1012 int ret;
2b188cc1 1013
09bb8394
JA
1014 if (!req->file)
1015 return -EBADF;
2b188cc1 1016
09bb8394
JA
1017 if (force_nonblock && !io_file_supports_async(req->file))
1018 force_nonblock = false;
6b06314c 1019
2b188cc1
JA
1020 kiocb->ki_pos = READ_ONCE(sqe->off);
1021 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
1022 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
1023
1024 ioprio = READ_ONCE(sqe->ioprio);
1025 if (ioprio) {
1026 ret = ioprio_check_cap(ioprio);
1027 if (ret)
09bb8394 1028 return ret;
2b188cc1
JA
1029
1030 kiocb->ki_ioprio = ioprio;
1031 } else
1032 kiocb->ki_ioprio = get_current_ioprio();
1033
1034 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
1035 if (unlikely(ret))
09bb8394 1036 return ret;
8449eeda
SB
1037
1038 /* don't allow async punt if RWF_NOWAIT was requested */
1039 if (kiocb->ki_flags & IOCB_NOWAIT)
1040 req->flags |= REQ_F_NOWAIT;
1041
1042 if (force_nonblock)
2b188cc1 1043 kiocb->ki_flags |= IOCB_NOWAIT;
8449eeda 1044
def596e9 1045 if (ctx->flags & IORING_SETUP_IOPOLL) {
def596e9
JA
1046 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
1047 !kiocb->ki_filp->f_op->iopoll)
09bb8394 1048 return -EOPNOTSUPP;
2b188cc1 1049
def596e9
JA
1050 kiocb->ki_flags |= IOCB_HIPRI;
1051 kiocb->ki_complete = io_complete_rw_iopoll;
1052 } else {
09bb8394
JA
1053 if (kiocb->ki_flags & IOCB_HIPRI)
1054 return -EINVAL;
def596e9
JA
1055 kiocb->ki_complete = io_complete_rw;
1056 }
2b188cc1 1057 return 0;
2b188cc1
JA
1058}
1059
1060static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1061{
1062 switch (ret) {
1063 case -EIOCBQUEUED:
1064 break;
1065 case -ERESTARTSYS:
1066 case -ERESTARTNOINTR:
1067 case -ERESTARTNOHAND:
1068 case -ERESTART_RESTARTBLOCK:
1069 /*
1070 * We can't just restart the syscall, since previously
1071 * submitted sqes may already be in progress. Just fail this
1072 * IO with EINTR.
1073 */
1074 ret = -EINTR;
1075 /* fall through */
1076 default:
1077 kiocb->ki_complete(kiocb, ret, 0);
1078 }
1079}
1080
edafccee
JA
1081static int io_import_fixed(struct io_ring_ctx *ctx, int rw,
1082 const struct io_uring_sqe *sqe,
1083 struct iov_iter *iter)
1084{
1085 size_t len = READ_ONCE(sqe->len);
1086 struct io_mapped_ubuf *imu;
1087 unsigned index, buf_index;
1088 size_t offset;
1089 u64 buf_addr;
1090
1091 /* attempt to use fixed buffers without having provided iovecs */
1092 if (unlikely(!ctx->user_bufs))
1093 return -EFAULT;
1094
1095 buf_index = READ_ONCE(sqe->buf_index);
1096 if (unlikely(buf_index >= ctx->nr_user_bufs))
1097 return -EFAULT;
1098
1099 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
1100 imu = &ctx->user_bufs[index];
1101 buf_addr = READ_ONCE(sqe->addr);
1102
1103 /* overflow */
1104 if (buf_addr + len < buf_addr)
1105 return -EFAULT;
1106 /* not inside the mapped region */
1107 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
1108 return -EFAULT;
1109
1110 /*
1111 * May not be a start of buffer, set size appropriately
1112 * and advance us to the beginning.
1113 */
1114 offset = buf_addr - imu->ubuf;
1115 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
bd11b3a3
JA
1116
1117 if (offset) {
1118 /*
1119 * Don't use iov_iter_advance() here, as it's really slow for
1120 * using the latter parts of a big fixed buffer - it iterates
1121 * over each segment manually. We can cheat a bit here, because
1122 * we know that:
1123 *
1124 * 1) it's a BVEC iter, we set it up
1125 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1126 * first and last bvec
1127 *
1128 * So just find our index, and adjust the iterator afterwards.
1129 * If the offset is within the first bvec (or the whole first
1130 * bvec, just use iov_iter_advance(). This makes it easier
1131 * since we can just skip the first segment, which may not
1132 * be PAGE_SIZE aligned.
1133 */
1134 const struct bio_vec *bvec = imu->bvec;
1135
1136 if (offset <= bvec->bv_len) {
1137 iov_iter_advance(iter, offset);
1138 } else {
1139 unsigned long seg_skip;
1140
1141 /* skip first vec */
1142 offset -= bvec->bv_len;
1143 seg_skip = 1 + (offset >> PAGE_SHIFT);
1144
1145 iter->bvec = bvec + seg_skip;
1146 iter->nr_segs -= seg_skip;
99c79f66 1147 iter->count -= bvec->bv_len + offset;
bd11b3a3 1148 iter->iov_offset = offset & ~PAGE_MASK;
bd11b3a3
JA
1149 }
1150 }
1151
edafccee
JA
1152 return 0;
1153}
1154
87e5e6da
JA
1155static ssize_t io_import_iovec(struct io_ring_ctx *ctx, int rw,
1156 const struct sqe_submit *s, struct iovec **iovec,
1157 struct iov_iter *iter)
2b188cc1
JA
1158{
1159 const struct io_uring_sqe *sqe = s->sqe;
1160 void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1161 size_t sqe_len = READ_ONCE(sqe->len);
edafccee
JA
1162 u8 opcode;
1163
1164 /*
1165 * We're reading ->opcode for the second time, but the first read
1166 * doesn't care whether it's _FIXED or not, so it doesn't matter
1167 * whether ->opcode changes concurrently. The first read does care
1168 * about whether it is a READ or a WRITE, so we don't trust this read
1169 * for that purpose and instead let the caller pass in the read/write
1170 * flag.
1171 */
1172 opcode = READ_ONCE(sqe->opcode);
1173 if (opcode == IORING_OP_READ_FIXED ||
1174 opcode == IORING_OP_WRITE_FIXED) {
87e5e6da 1175 ssize_t ret = io_import_fixed(ctx, rw, sqe, iter);
edafccee
JA
1176 *iovec = NULL;
1177 return ret;
1178 }
2b188cc1
JA
1179
1180 if (!s->has_user)
1181 return -EFAULT;
1182
1183#ifdef CONFIG_COMPAT
1184 if (ctx->compat)
1185 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
1186 iovec, iter);
1187#endif
1188
1189 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
1190}
1191
6d5d5ac5
JA
1192static inline bool io_should_merge(struct async_list *al, struct kiocb *kiocb)
1193{
1194 if (al->file == kiocb->ki_filp) {
1195 off_t start, end;
1196
1197 /*
1198 * Allow merging if we're anywhere in the range of the same
1199 * page. Generally this happens for sub-page reads or writes,
1200 * and it's beneficial to allow the first worker to bring the
1201 * page in and the piggy backed work can then work on the
1202 * cached page.
1203 */
1204 start = al->io_start & PAGE_MASK;
1205 end = (al->io_start + al->io_len + PAGE_SIZE - 1) & PAGE_MASK;
1206 if (kiocb->ki_pos >= start && kiocb->ki_pos <= end)
1207 return true;
1208 }
1209
1210 al->file = NULL;
1211 return false;
1212}
1213
31b51510
JA
1214/*
1215 * Make a note of the last file/offset/direction we punted to async
1216 * context. We'll use this information to see if we can piggy back a
1217 * sequential request onto the previous one, if it's still hasn't been
1218 * completed by the async worker.
1219 */
1220static void io_async_list_note(int rw, struct io_kiocb *req, size_t len)
1221{
1222 struct async_list *async_list = &req->ctx->pending_async[rw];
1223 struct kiocb *kiocb = &req->rw;
1224 struct file *filp = kiocb->ki_filp;
31b51510 1225
6d5d5ac5 1226 if (io_should_merge(async_list, kiocb)) {
9310a7ba 1227 unsigned long max_bytes;
31b51510
JA
1228
1229 /* Use 8x RA size as a decent limiter for both reads/writes */
9310a7ba
ZL
1230 max_bytes = filp->f_ra.ra_pages << (PAGE_SHIFT + 3);
1231 if (!max_bytes)
1232 max_bytes = VM_READAHEAD_PAGES << (PAGE_SHIFT + 3);
1233
1234 /* If max len are exceeded, reset the state */
1235 if (async_list->io_len + len <= max_bytes) {
31b51510 1236 req->flags |= REQ_F_SEQ_PREV;
9310a7ba 1237 async_list->io_len += len;
31b51510 1238 } else {
6d5d5ac5 1239 async_list->file = NULL;
31b51510
JA
1240 }
1241 }
1242
1243 /* New file? Reset state. */
1244 if (async_list->file != filp) {
6d5d5ac5
JA
1245 async_list->io_start = kiocb->ki_pos;
1246 async_list->io_len = len;
31b51510
JA
1247 async_list->file = filp;
1248 }
31b51510
JA
1249}
1250
e0c5c576 1251static int io_read(struct io_kiocb *req, const struct sqe_submit *s,
8358e3a8 1252 bool force_nonblock)
2b188cc1
JA
1253{
1254 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1255 struct kiocb *kiocb = &req->rw;
1256 struct iov_iter iter;
1257 struct file *file;
31b51510 1258 size_t iov_count;
9d93a3f5 1259 ssize_t read_size, ret;
2b188cc1 1260
8358e3a8 1261 ret = io_prep_rw(req, s, force_nonblock);
2b188cc1
JA
1262 if (ret)
1263 return ret;
1264 file = kiocb->ki_filp;
1265
2b188cc1 1266 if (unlikely(!(file->f_mode & FMODE_READ)))
09bb8394 1267 return -EBADF;
2b188cc1 1268 if (unlikely(!file->f_op->read_iter))
09bb8394 1269 return -EINVAL;
2b188cc1
JA
1270
1271 ret = io_import_iovec(req->ctx, READ, s, &iovec, &iter);
87e5e6da 1272 if (ret < 0)
09bb8394 1273 return ret;
2b188cc1 1274
9d93a3f5 1275 read_size = ret;
9e645e11
JA
1276 if (req->flags & REQ_F_LINK)
1277 req->result = read_size;
1278
31b51510
JA
1279 iov_count = iov_iter_count(&iter);
1280 ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count);
2b188cc1
JA
1281 if (!ret) {
1282 ssize_t ret2;
1283
2b188cc1 1284 ret2 = call_read_iter(file, kiocb, &iter);
9d93a3f5
JA
1285 /*
1286 * In case of a short read, punt to async. This can happen
1287 * if we have data partially cached. Alternatively we can
1288 * return the short read, in which case the application will
1289 * need to issue another SQE and wait for it. That SQE will
1290 * need async punt anyway, so it's more efficient to do it
1291 * here.
1292 */
1293 if (force_nonblock && ret2 > 0 && ret2 < read_size)
1294 ret2 = -EAGAIN;
1295 /* Catch -EAGAIN return for forced non-blocking submission */
31b51510 1296 if (!force_nonblock || ret2 != -EAGAIN) {
2b188cc1 1297 io_rw_done(kiocb, ret2);
31b51510
JA
1298 } else {
1299 /*
1300 * If ->needs_lock is true, we're already in async
1301 * context.
1302 */
1303 if (!s->needs_lock)
1304 io_async_list_note(READ, req, iov_count);
2b188cc1 1305 ret = -EAGAIN;
31b51510 1306 }
2b188cc1
JA
1307 }
1308 kfree(iovec);
2b188cc1
JA
1309 return ret;
1310}
1311
e0c5c576 1312static int io_write(struct io_kiocb *req, const struct sqe_submit *s,
8358e3a8 1313 bool force_nonblock)
2b188cc1
JA
1314{
1315 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1316 struct kiocb *kiocb = &req->rw;
1317 struct iov_iter iter;
1318 struct file *file;
31b51510 1319 size_t iov_count;
87e5e6da 1320 ssize_t ret;
2b188cc1 1321
8358e3a8 1322 ret = io_prep_rw(req, s, force_nonblock);
2b188cc1
JA
1323 if (ret)
1324 return ret;
2b188cc1 1325
2b188cc1
JA
1326 file = kiocb->ki_filp;
1327 if (unlikely(!(file->f_mode & FMODE_WRITE)))
09bb8394 1328 return -EBADF;
2b188cc1 1329 if (unlikely(!file->f_op->write_iter))
09bb8394 1330 return -EINVAL;
2b188cc1
JA
1331
1332 ret = io_import_iovec(req->ctx, WRITE, s, &iovec, &iter);
87e5e6da 1333 if (ret < 0)
09bb8394 1334 return ret;
2b188cc1 1335
9e645e11
JA
1336 if (req->flags & REQ_F_LINK)
1337 req->result = ret;
1338
31b51510
JA
1339 iov_count = iov_iter_count(&iter);
1340
1341 ret = -EAGAIN;
1342 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT)) {
1343 /* If ->needs_lock is true, we're already in async context. */
1344 if (!s->needs_lock)
1345 io_async_list_note(WRITE, req, iov_count);
1346 goto out_free;
1347 }
1348
1349 ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count);
2b188cc1 1350 if (!ret) {
9bf7933f
RP
1351 ssize_t ret2;
1352
2b188cc1
JA
1353 /*
1354 * Open-code file_start_write here to grab freeze protection,
1355 * which will be released by another thread in
1356 * io_complete_rw(). Fool lockdep by telling it the lock got
1357 * released so that it doesn't complain about the held lock when
1358 * we return to userspace.
1359 */
1360 if (S_ISREG(file_inode(file)->i_mode)) {
1361 __sb_start_write(file_inode(file)->i_sb,
1362 SB_FREEZE_WRITE, true);
1363 __sb_writers_release(file_inode(file)->i_sb,
1364 SB_FREEZE_WRITE);
1365 }
1366 kiocb->ki_flags |= IOCB_WRITE;
9bf7933f
RP
1367
1368 ret2 = call_write_iter(file, kiocb, &iter);
1369 if (!force_nonblock || ret2 != -EAGAIN) {
1370 io_rw_done(kiocb, ret2);
1371 } else {
1372 /*
1373 * If ->needs_lock is true, we're already in async
1374 * context.
1375 */
1376 if (!s->needs_lock)
1377 io_async_list_note(WRITE, req, iov_count);
1378 ret = -EAGAIN;
1379 }
2b188cc1 1380 }
31b51510 1381out_free:
2b188cc1 1382 kfree(iovec);
2b188cc1
JA
1383 return ret;
1384}
1385
1386/*
1387 * IORING_OP_NOP just posts a completion event, nothing else.
1388 */
1389static int io_nop(struct io_kiocb *req, u64 user_data)
1390{
1391 struct io_ring_ctx *ctx = req->ctx;
1392 long err = 0;
1393
def596e9
JA
1394 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1395 return -EINVAL;
1396
c71ffb67 1397 io_cqring_add_event(ctx, user_data, err);
e65ef56d 1398 io_put_req(req);
2b188cc1
JA
1399 return 0;
1400}
1401
c992fe29
CH
1402static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1403{
6b06314c 1404 struct io_ring_ctx *ctx = req->ctx;
c992fe29 1405
09bb8394
JA
1406 if (!req->file)
1407 return -EBADF;
c992fe29 1408
6b06314c 1409 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
def596e9 1410 return -EINVAL;
edafccee 1411 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
c992fe29
CH
1412 return -EINVAL;
1413
c992fe29
CH
1414 return 0;
1415}
1416
1417static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1418 bool force_nonblock)
1419{
1420 loff_t sqe_off = READ_ONCE(sqe->off);
1421 loff_t sqe_len = READ_ONCE(sqe->len);
1422 loff_t end = sqe_off + sqe_len;
1423 unsigned fsync_flags;
1424 int ret;
1425
1426 fsync_flags = READ_ONCE(sqe->fsync_flags);
1427 if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC))
1428 return -EINVAL;
1429
1430 ret = io_prep_fsync(req, sqe);
1431 if (ret)
1432 return ret;
1433
1434 /* fsync always requires a blocking context */
1435 if (force_nonblock)
1436 return -EAGAIN;
1437
1438 ret = vfs_fsync_range(req->rw.ki_filp, sqe_off,
1439 end > 0 ? end : LLONG_MAX,
1440 fsync_flags & IORING_FSYNC_DATASYNC);
1441
9e645e11
JA
1442 if (ret < 0 && (req->flags & REQ_F_LINK))
1443 req->flags |= REQ_F_FAIL_LINK;
c71ffb67 1444 io_cqring_add_event(req->ctx, sqe->user_data, ret);
e65ef56d 1445 io_put_req(req);
c992fe29
CH
1446 return 0;
1447}
1448
5d17b4a4
JA
1449static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1450{
1451 struct io_ring_ctx *ctx = req->ctx;
1452 int ret = 0;
1453
1454 if (!req->file)
1455 return -EBADF;
5d17b4a4
JA
1456
1457 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1458 return -EINVAL;
1459 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1460 return -EINVAL;
1461
5d17b4a4
JA
1462 return ret;
1463}
1464
1465static int io_sync_file_range(struct io_kiocb *req,
1466 const struct io_uring_sqe *sqe,
1467 bool force_nonblock)
1468{
1469 loff_t sqe_off;
1470 loff_t sqe_len;
1471 unsigned flags;
1472 int ret;
1473
1474 ret = io_prep_sfr(req, sqe);
1475 if (ret)
1476 return ret;
1477
1478 /* sync_file_range always requires a blocking context */
1479 if (force_nonblock)
1480 return -EAGAIN;
1481
1482 sqe_off = READ_ONCE(sqe->off);
1483 sqe_len = READ_ONCE(sqe->len);
1484 flags = READ_ONCE(sqe->sync_range_flags);
1485
1486 ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags);
1487
9e645e11
JA
1488 if (ret < 0 && (req->flags & REQ_F_LINK))
1489 req->flags |= REQ_F_FAIL_LINK;
c71ffb67 1490 io_cqring_add_event(req->ctx, sqe->user_data, ret);
5d17b4a4
JA
1491 io_put_req(req);
1492 return 0;
1493}
1494
0fa03c62 1495#if defined(CONFIG_NET)
aa1fa28f
JA
1496static int io_send_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1497 bool force_nonblock,
1498 long (*fn)(struct socket *, struct user_msghdr __user *,
1499 unsigned int))
1500{
0fa03c62
JA
1501 struct socket *sock;
1502 int ret;
1503
1504 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1505 return -EINVAL;
1506
1507 sock = sock_from_file(req->file, &ret);
1508 if (sock) {
1509 struct user_msghdr __user *msg;
1510 unsigned flags;
1511
1512 flags = READ_ONCE(sqe->msg_flags);
1513 if (flags & MSG_DONTWAIT)
1514 req->flags |= REQ_F_NOWAIT;
1515 else if (force_nonblock)
1516 flags |= MSG_DONTWAIT;
1517
1518 msg = (struct user_msghdr __user *) (unsigned long)
1519 READ_ONCE(sqe->addr);
1520
aa1fa28f 1521 ret = fn(sock, msg, flags);
0fa03c62
JA
1522 if (force_nonblock && ret == -EAGAIN)
1523 return ret;
1524 }
1525
c71ffb67 1526 io_cqring_add_event(req->ctx, sqe->user_data, ret);
5d17b4a4
JA
1527 io_put_req(req);
1528 return 0;
1529}
aa1fa28f
JA
1530#endif
1531
1532static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1533 bool force_nonblock)
1534{
1535#if defined(CONFIG_NET)
1536 return io_send_recvmsg(req, sqe, force_nonblock, __sys_sendmsg_sock);
1537#else
1538 return -EOPNOTSUPP;
1539#endif
1540}
1541
1542static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1543 bool force_nonblock)
1544{
1545#if defined(CONFIG_NET)
1546 return io_send_recvmsg(req, sqe, force_nonblock, __sys_recvmsg_sock);
0fa03c62
JA
1547#else
1548 return -EOPNOTSUPP;
1549#endif
1550}
5d17b4a4 1551
221c5eb2
JA
1552static void io_poll_remove_one(struct io_kiocb *req)
1553{
1554 struct io_poll_iocb *poll = &req->poll;
1555
1556 spin_lock(&poll->head->lock);
1557 WRITE_ONCE(poll->canceled, true);
1558 if (!list_empty(&poll->wait.entry)) {
1559 list_del_init(&poll->wait.entry);
18d9be1a 1560 io_queue_async_work(req->ctx, req);
221c5eb2
JA
1561 }
1562 spin_unlock(&poll->head->lock);
1563
1564 list_del_init(&req->list);
1565}
1566
1567static void io_poll_remove_all(struct io_ring_ctx *ctx)
1568{
1569 struct io_kiocb *req;
1570
1571 spin_lock_irq(&ctx->completion_lock);
1572 while (!list_empty(&ctx->cancel_list)) {
1573 req = list_first_entry(&ctx->cancel_list, struct io_kiocb,list);
1574 io_poll_remove_one(req);
1575 }
1576 spin_unlock_irq(&ctx->completion_lock);
1577}
1578
1579/*
1580 * Find a running poll command that matches one specified in sqe->addr,
1581 * and remove it if found.
1582 */
1583static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1584{
1585 struct io_ring_ctx *ctx = req->ctx;
1586 struct io_kiocb *poll_req, *next;
1587 int ret = -ENOENT;
1588
1589 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1590 return -EINVAL;
1591 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
1592 sqe->poll_events)
1593 return -EINVAL;
1594
1595 spin_lock_irq(&ctx->completion_lock);
1596 list_for_each_entry_safe(poll_req, next, &ctx->cancel_list, list) {
1597 if (READ_ONCE(sqe->addr) == poll_req->user_data) {
1598 io_poll_remove_one(poll_req);
1599 ret = 0;
1600 break;
1601 }
1602 }
1603 spin_unlock_irq(&ctx->completion_lock);
1604
c71ffb67 1605 io_cqring_add_event(req->ctx, sqe->user_data, ret);
e65ef56d 1606 io_put_req(req);
221c5eb2
JA
1607 return 0;
1608}
1609
8c838788
JA
1610static void io_poll_complete(struct io_ring_ctx *ctx, struct io_kiocb *req,
1611 __poll_t mask)
221c5eb2 1612{
8c838788 1613 req->poll.done = true;
c71ffb67 1614 io_cqring_fill_event(ctx, req->user_data, mangle_poll(mask));
8c838788 1615 io_commit_cqring(ctx);
221c5eb2
JA
1616}
1617
1618static void io_poll_complete_work(struct work_struct *work)
1619{
1620 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1621 struct io_poll_iocb *poll = &req->poll;
1622 struct poll_table_struct pt = { ._key = poll->events };
1623 struct io_ring_ctx *ctx = req->ctx;
1624 __poll_t mask = 0;
1625
1626 if (!READ_ONCE(poll->canceled))
1627 mask = vfs_poll(poll->file, &pt) & poll->events;
1628
1629 /*
1630 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1631 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1632 * synchronize with them. In the cancellation case the list_del_init
1633 * itself is not actually needed, but harmless so we keep it in to
1634 * avoid further branches in the fast path.
1635 */
1636 spin_lock_irq(&ctx->completion_lock);
1637 if (!mask && !READ_ONCE(poll->canceled)) {
1638 add_wait_queue(poll->head, &poll->wait);
1639 spin_unlock_irq(&ctx->completion_lock);
1640 return;
1641 }
1642 list_del_init(&req->list);
8c838788 1643 io_poll_complete(ctx, req, mask);
221c5eb2
JA
1644 spin_unlock_irq(&ctx->completion_lock);
1645
8c838788
JA
1646 io_cqring_ev_posted(ctx);
1647 io_put_req(req);
221c5eb2
JA
1648}
1649
1650static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1651 void *key)
1652{
1653 struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb,
1654 wait);
1655 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
1656 struct io_ring_ctx *ctx = req->ctx;
1657 __poll_t mask = key_to_poll(key);
8c838788 1658 unsigned long flags;
221c5eb2
JA
1659
1660 /* for instances that support it check for an event match first: */
8c838788
JA
1661 if (mask && !(mask & poll->events))
1662 return 0;
221c5eb2 1663
8c838788 1664 list_del_init(&poll->wait.entry);
221c5eb2 1665
8c838788
JA
1666 if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) {
1667 list_del(&req->list);
1668 io_poll_complete(ctx, req, mask);
1669 spin_unlock_irqrestore(&ctx->completion_lock, flags);
221c5eb2 1670
8c838788
JA
1671 io_cqring_ev_posted(ctx);
1672 io_put_req(req);
1673 } else {
18d9be1a 1674 io_queue_async_work(ctx, req);
221c5eb2
JA
1675 }
1676
221c5eb2
JA
1677 return 1;
1678}
1679
1680struct io_poll_table {
1681 struct poll_table_struct pt;
1682 struct io_kiocb *req;
1683 int error;
1684};
1685
1686static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1687 struct poll_table_struct *p)
1688{
1689 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
1690
1691 if (unlikely(pt->req->poll.head)) {
1692 pt->error = -EINVAL;
1693 return;
1694 }
1695
1696 pt->error = 0;
1697 pt->req->poll.head = head;
1698 add_wait_queue(head, &pt->req->poll.wait);
1699}
1700
1701static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1702{
1703 struct io_poll_iocb *poll = &req->poll;
1704 struct io_ring_ctx *ctx = req->ctx;
1705 struct io_poll_table ipt;
8c838788 1706 bool cancel = false;
221c5eb2
JA
1707 __poll_t mask;
1708 u16 events;
221c5eb2
JA
1709
1710 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1711 return -EINVAL;
1712 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
1713 return -EINVAL;
09bb8394
JA
1714 if (!poll->file)
1715 return -EBADF;
221c5eb2
JA
1716
1717 INIT_WORK(&req->work, io_poll_complete_work);
1718 events = READ_ONCE(sqe->poll_events);
1719 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
1720
221c5eb2 1721 poll->head = NULL;
8c838788 1722 poll->done = false;
221c5eb2
JA
1723 poll->canceled = false;
1724
1725 ipt.pt._qproc = io_poll_queue_proc;
1726 ipt.pt._key = poll->events;
1727 ipt.req = req;
1728 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1729
1730 /* initialized the list so that we can do list_empty checks */
1731 INIT_LIST_HEAD(&poll->wait.entry);
1732 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
1733
36703247
JA
1734 INIT_LIST_HEAD(&req->list);
1735
221c5eb2 1736 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
221c5eb2
JA
1737
1738 spin_lock_irq(&ctx->completion_lock);
8c838788
JA
1739 if (likely(poll->head)) {
1740 spin_lock(&poll->head->lock);
1741 if (unlikely(list_empty(&poll->wait.entry))) {
1742 if (ipt.error)
1743 cancel = true;
1744 ipt.error = 0;
1745 mask = 0;
1746 }
1747 if (mask || ipt.error)
1748 list_del_init(&poll->wait.entry);
1749 else if (cancel)
1750 WRITE_ONCE(poll->canceled, true);
1751 else if (!poll->done) /* actually waiting for an event */
1752 list_add_tail(&req->list, &ctx->cancel_list);
1753 spin_unlock(&poll->head->lock);
1754 }
1755 if (mask) { /* no async, we'd stolen it */
221c5eb2 1756 ipt.error = 0;
8c838788 1757 io_poll_complete(ctx, req, mask);
221c5eb2 1758 }
221c5eb2
JA
1759 spin_unlock_irq(&ctx->completion_lock);
1760
8c838788
JA
1761 if (mask) {
1762 io_cqring_ev_posted(ctx);
e65ef56d 1763 io_put_req(req);
221c5eb2 1764 }
8c838788 1765 return ipt.error;
221c5eb2
JA
1766}
1767
de0617e4
JA
1768static int io_req_defer(struct io_ring_ctx *ctx, struct io_kiocb *req,
1769 const struct io_uring_sqe *sqe)
1770{
1771 struct io_uring_sqe *sqe_copy;
1772
1773 if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list))
1774 return 0;
1775
1776 sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL);
1777 if (!sqe_copy)
1778 return -EAGAIN;
1779
1780 spin_lock_irq(&ctx->completion_lock);
1781 if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) {
1782 spin_unlock_irq(&ctx->completion_lock);
1783 kfree(sqe_copy);
1784 return 0;
1785 }
1786
1787 memcpy(sqe_copy, sqe, sizeof(*sqe_copy));
1788 req->submit.sqe = sqe_copy;
1789
1790 INIT_WORK(&req->work, io_sq_wq_submit_work);
1791 list_add_tail(&req->list, &ctx->defer_list);
1792 spin_unlock_irq(&ctx->completion_lock);
1793 return -EIOCBQUEUED;
1794}
1795
2b188cc1 1796static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
8358e3a8 1797 const struct sqe_submit *s, bool force_nonblock)
2b188cc1 1798{
e0c5c576 1799 int ret, opcode;
2b188cc1 1800
9e645e11
JA
1801 req->user_data = READ_ONCE(s->sqe->user_data);
1802
2b188cc1
JA
1803 if (unlikely(s->index >= ctx->sq_entries))
1804 return -EINVAL;
2b188cc1
JA
1805
1806 opcode = READ_ONCE(s->sqe->opcode);
1807 switch (opcode) {
1808 case IORING_OP_NOP:
1809 ret = io_nop(req, req->user_data);
1810 break;
1811 case IORING_OP_READV:
edafccee
JA
1812 if (unlikely(s->sqe->buf_index))
1813 return -EINVAL;
8358e3a8 1814 ret = io_read(req, s, force_nonblock);
2b188cc1
JA
1815 break;
1816 case IORING_OP_WRITEV:
edafccee
JA
1817 if (unlikely(s->sqe->buf_index))
1818 return -EINVAL;
8358e3a8 1819 ret = io_write(req, s, force_nonblock);
edafccee
JA
1820 break;
1821 case IORING_OP_READ_FIXED:
8358e3a8 1822 ret = io_read(req, s, force_nonblock);
edafccee
JA
1823 break;
1824 case IORING_OP_WRITE_FIXED:
8358e3a8 1825 ret = io_write(req, s, force_nonblock);
2b188cc1 1826 break;
c992fe29
CH
1827 case IORING_OP_FSYNC:
1828 ret = io_fsync(req, s->sqe, force_nonblock);
1829 break;
221c5eb2
JA
1830 case IORING_OP_POLL_ADD:
1831 ret = io_poll_add(req, s->sqe);
1832 break;
1833 case IORING_OP_POLL_REMOVE:
1834 ret = io_poll_remove(req, s->sqe);
1835 break;
5d17b4a4
JA
1836 case IORING_OP_SYNC_FILE_RANGE:
1837 ret = io_sync_file_range(req, s->sqe, force_nonblock);
1838 break;
0fa03c62
JA
1839 case IORING_OP_SENDMSG:
1840 ret = io_sendmsg(req, s->sqe, force_nonblock);
1841 break;
aa1fa28f
JA
1842 case IORING_OP_RECVMSG:
1843 ret = io_recvmsg(req, s->sqe, force_nonblock);
1844 break;
2b188cc1
JA
1845 default:
1846 ret = -EINVAL;
1847 break;
1848 }
1849
def596e9
JA
1850 if (ret)
1851 return ret;
1852
1853 if (ctx->flags & IORING_SETUP_IOPOLL) {
9e645e11 1854 if (req->result == -EAGAIN)
def596e9
JA
1855 return -EAGAIN;
1856
1857 /* workqueue context doesn't hold uring_lock, grab it now */
1858 if (s->needs_lock)
1859 mutex_lock(&ctx->uring_lock);
1860 io_iopoll_req_issued(req);
1861 if (s->needs_lock)
1862 mutex_unlock(&ctx->uring_lock);
1863 }
1864
1865 return 0;
2b188cc1
JA
1866}
1867
31b51510
JA
1868static struct async_list *io_async_list_from_sqe(struct io_ring_ctx *ctx,
1869 const struct io_uring_sqe *sqe)
1870{
1871 switch (sqe->opcode) {
1872 case IORING_OP_READV:
1873 case IORING_OP_READ_FIXED:
1874 return &ctx->pending_async[READ];
1875 case IORING_OP_WRITEV:
1876 case IORING_OP_WRITE_FIXED:
1877 return &ctx->pending_async[WRITE];
1878 default:
1879 return NULL;
1880 }
1881}
1882
edafccee
JA
1883static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe)
1884{
1885 u8 opcode = READ_ONCE(sqe->opcode);
1886
1887 return !(opcode == IORING_OP_READ_FIXED ||
1888 opcode == IORING_OP_WRITE_FIXED);
1889}
1890
2b188cc1
JA
1891static void io_sq_wq_submit_work(struct work_struct *work)
1892{
1893 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2b188cc1 1894 struct io_ring_ctx *ctx = req->ctx;
31b51510
JA
1895 struct mm_struct *cur_mm = NULL;
1896 struct async_list *async_list;
1897 LIST_HEAD(req_list);
edafccee 1898 mm_segment_t old_fs;
2b188cc1
JA
1899 int ret;
1900
31b51510
JA
1901 async_list = io_async_list_from_sqe(ctx, req->submit.sqe);
1902restart:
1903 do {
1904 struct sqe_submit *s = &req->submit;
1905 const struct io_uring_sqe *sqe = s->sqe;
d0ee8791 1906 unsigned int flags = req->flags;
2b188cc1 1907
8449eeda 1908 /* Ensure we clear previously set non-block flag */
31b51510
JA
1909 req->rw.ki_flags &= ~IOCB_NOWAIT;
1910
1911 ret = 0;
1912 if (io_sqe_needs_user(sqe) && !cur_mm) {
1913 if (!mmget_not_zero(ctx->sqo_mm)) {
1914 ret = -EFAULT;
1915 } else {
1916 cur_mm = ctx->sqo_mm;
1917 use_mm(cur_mm);
1918 old_fs = get_fs();
1919 set_fs(USER_DS);
1920 }
1921 }
1922
1923 if (!ret) {
1924 s->has_user = cur_mm != NULL;
1925 s->needs_lock = true;
1926 do {
8358e3a8 1927 ret = __io_submit_sqe(ctx, req, s, false);
31b51510
JA
1928 /*
1929 * We can get EAGAIN for polled IO even though
1930 * we're forcing a sync submission from here,
1931 * since we can't wait for request slots on the
1932 * block side.
1933 */
1934 if (ret != -EAGAIN)
1935 break;
1936 cond_resched();
1937 } while (1);
1938 }
817869d2
JA
1939
1940 /* drop submission reference */
1941 io_put_req(req);
1942
31b51510 1943 if (ret) {
c71ffb67 1944 io_cqring_add_event(ctx, sqe->user_data, ret);
e65ef56d 1945 io_put_req(req);
31b51510
JA
1946 }
1947
1948 /* async context always use a copy of the sqe */
1949 kfree(sqe);
1950
f7b76ac9 1951 /* req from defer and link list needn't decrease async cnt */
d0ee8791 1952 if (flags & (REQ_F_IO_DRAINED | REQ_F_LINK_DONE))
f7b76ac9
ZL
1953 goto out;
1954
31b51510
JA
1955 if (!async_list)
1956 break;
1957 if (!list_empty(&req_list)) {
1958 req = list_first_entry(&req_list, struct io_kiocb,
1959 list);
1960 list_del(&req->list);
1961 continue;
1962 }
1963 if (list_empty(&async_list->list))
1964 break;
1965
1966 req = NULL;
1967 spin_lock(&async_list->lock);
1968 if (list_empty(&async_list->list)) {
1969 spin_unlock(&async_list->lock);
1970 break;
1971 }
1972 list_splice_init(&async_list->list, &req_list);
1973 spin_unlock(&async_list->lock);
1974
1975 req = list_first_entry(&req_list, struct io_kiocb, list);
1976 list_del(&req->list);
1977 } while (req);
edafccee
JA
1978
1979 /*
31b51510
JA
1980 * Rare case of racing with a submitter. If we find the count has
1981 * dropped to zero AND we have pending work items, then restart
1982 * the processing. This is a tiny race window.
edafccee 1983 */
31b51510
JA
1984 if (async_list) {
1985 ret = atomic_dec_return(&async_list->cnt);
1986 while (!ret && !list_empty(&async_list->list)) {
1987 spin_lock(&async_list->lock);
1988 atomic_inc(&async_list->cnt);
1989 list_splice_init(&async_list->list, &req_list);
1990 spin_unlock(&async_list->lock);
1991
1992 if (!list_empty(&req_list)) {
1993 req = list_first_entry(&req_list,
1994 struct io_kiocb, list);
1995 list_del(&req->list);
1996 goto restart;
1997 }
1998 ret = atomic_dec_return(&async_list->cnt);
edafccee 1999 }
edafccee 2000 }
2b188cc1 2001
f7b76ac9 2002out:
31b51510 2003 if (cur_mm) {
edafccee 2004 set_fs(old_fs);
31b51510
JA
2005 unuse_mm(cur_mm);
2006 mmput(cur_mm);
2b188cc1 2007 }
31b51510 2008}
2b188cc1 2009
31b51510
JA
2010/*
2011 * See if we can piggy back onto previously submitted work, that is still
2012 * running. We currently only allow this if the new request is sequential
2013 * to the previous one we punted.
2014 */
2015static bool io_add_to_prev_work(struct async_list *list, struct io_kiocb *req)
2016{
6d5d5ac5 2017 bool ret;
31b51510
JA
2018
2019 if (!list)
2020 return false;
2021 if (!(req->flags & REQ_F_SEQ_PREV))
2022 return false;
2023 if (!atomic_read(&list->cnt))
2024 return false;
2025
2026 ret = true;
2027 spin_lock(&list->lock);
2028 list_add_tail(&req->list, &list->list);
c0e48f9d
ZL
2029 /*
2030 * Ensure we see a simultaneous modification from io_sq_wq_submit_work()
2031 */
2032 smp_mb();
31b51510
JA
2033 if (!atomic_read(&list->cnt)) {
2034 list_del_init(&req->list);
2035 ret = false;
2036 }
2037 spin_unlock(&list->lock);
2038 return ret;
2b188cc1
JA
2039}
2040
09bb8394
JA
2041static bool io_op_needs_file(const struct io_uring_sqe *sqe)
2042{
2043 int op = READ_ONCE(sqe->opcode);
2044
2045 switch (op) {
2046 case IORING_OP_NOP:
2047 case IORING_OP_POLL_REMOVE:
2048 return false;
2049 default:
2050 return true;
2051 }
2052}
2053
2054static int io_req_set_file(struct io_ring_ctx *ctx, const struct sqe_submit *s,
2055 struct io_submit_state *state, struct io_kiocb *req)
2056{
2057 unsigned flags;
2058 int fd;
2059
2060 flags = READ_ONCE(s->sqe->flags);
2061 fd = READ_ONCE(s->sqe->fd);
2062
4fe2c963 2063 if (flags & IOSQE_IO_DRAIN)
de0617e4 2064 req->flags |= REQ_F_IO_DRAIN;
4fe2c963
JL
2065 /*
2066 * All io need record the previous position, if LINK vs DARIN,
2067 * it can be used to mark the position of the first IO in the
2068 * link list.
2069 */
2070 req->sequence = s->sequence;
de0617e4 2071
60c112b0 2072 if (!io_op_needs_file(s->sqe))
09bb8394 2073 return 0;
09bb8394
JA
2074
2075 if (flags & IOSQE_FIXED_FILE) {
2076 if (unlikely(!ctx->user_files ||
2077 (unsigned) fd >= ctx->nr_user_files))
2078 return -EBADF;
2079 req->file = ctx->user_files[fd];
2080 req->flags |= REQ_F_FIXED_FILE;
2081 } else {
2082 if (s->needs_fixed_file)
2083 return -EBADF;
2084 req->file = io_file_get(state, fd);
2085 if (unlikely(!req->file))
2086 return -EBADF;
2087 }
2088
2089 return 0;
2090}
2091
4fe2c963 2092static int __io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
c5766668 2093 struct sqe_submit *s, bool force_nonblock)
2b188cc1 2094{
e0c5c576 2095 int ret;
2b188cc1 2096
c5766668 2097 ret = __io_submit_sqe(ctx, req, s, force_nonblock);
8449eeda 2098 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
2b188cc1
JA
2099 struct io_uring_sqe *sqe_copy;
2100
2101 sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL);
2102 if (sqe_copy) {
31b51510
JA
2103 struct async_list *list;
2104
2b188cc1
JA
2105 memcpy(sqe_copy, s->sqe, sizeof(*sqe_copy));
2106 s->sqe = sqe_copy;
2107
2108 memcpy(&req->submit, s, sizeof(*s));
31b51510
JA
2109 list = io_async_list_from_sqe(ctx, s->sqe);
2110 if (!io_add_to_prev_work(list, req)) {
2111 if (list)
2112 atomic_inc(&list->cnt);
2113 INIT_WORK(&req->work, io_sq_wq_submit_work);
18d9be1a 2114 io_queue_async_work(ctx, req);
31b51510 2115 }
e65ef56d
JA
2116
2117 /*
2118 * Queued up for async execution, worker will release
9e645e11 2119 * submit reference when the iocb is actually submitted.
e65ef56d
JA
2120 */
2121 return 0;
2b188cc1
JA
2122 }
2123 }
e65ef56d
JA
2124
2125 /* drop submission reference */
2126 io_put_req(req);
2127
2128 /* and drop final reference, if we failed */
9e645e11
JA
2129 if (ret) {
2130 io_cqring_add_event(ctx, req->user_data, ret);
2131 if (req->flags & REQ_F_LINK)
2132 req->flags |= REQ_F_FAIL_LINK;
e65ef56d 2133 io_put_req(req);
9e645e11 2134 }
2b188cc1
JA
2135
2136 return ret;
2137}
2138
4fe2c963 2139static int io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
c5766668 2140 struct sqe_submit *s, bool force_nonblock)
4fe2c963
JL
2141{
2142 int ret;
2143
2144 ret = io_req_defer(ctx, req, s->sqe);
2145 if (ret) {
2146 if (ret != -EIOCBQUEUED) {
2147 io_free_req(req);
2148 io_cqring_add_event(ctx, s->sqe->user_data, ret);
2149 }
2150 return 0;
2151 }
2152
c5766668 2153 return __io_queue_sqe(ctx, req, s, force_nonblock);
4fe2c963
JL
2154}
2155
2156static int io_queue_link_head(struct io_ring_ctx *ctx, struct io_kiocb *req,
c5766668
JA
2157 struct sqe_submit *s, struct io_kiocb *shadow,
2158 bool force_nonblock)
4fe2c963
JL
2159{
2160 int ret;
2161 int need_submit = false;
2162
2163 if (!shadow)
c5766668 2164 return io_queue_sqe(ctx, req, s, force_nonblock);
4fe2c963
JL
2165
2166 /*
2167 * Mark the first IO in link list as DRAIN, let all the following
2168 * IOs enter the defer list. all IO needs to be completed before link
2169 * list.
2170 */
2171 req->flags |= REQ_F_IO_DRAIN;
2172 ret = io_req_defer(ctx, req, s->sqe);
2173 if (ret) {
2174 if (ret != -EIOCBQUEUED) {
2175 io_free_req(req);
2176 io_cqring_add_event(ctx, s->sqe->user_data, ret);
2177 return 0;
2178 }
2179 } else {
2180 /*
2181 * If ret == 0 means that all IOs in front of link io are
2182 * running done. let's queue link head.
2183 */
2184 need_submit = true;
2185 }
2186
2187 /* Insert shadow req to defer_list, blocking next IOs */
2188 spin_lock_irq(&ctx->completion_lock);
2189 list_add_tail(&shadow->list, &ctx->defer_list);
2190 spin_unlock_irq(&ctx->completion_lock);
2191
2192 if (need_submit)
c5766668 2193 return __io_queue_sqe(ctx, req, s, force_nonblock);
4fe2c963
JL
2194
2195 return 0;
2196}
2197
9e645e11
JA
2198#define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK)
2199
2200static void io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s,
c5766668
JA
2201 struct io_submit_state *state, struct io_kiocb **link,
2202 bool force_nonblock)
9e645e11
JA
2203{
2204 struct io_uring_sqe *sqe_copy;
2205 struct io_kiocb *req;
2206 int ret;
2207
2208 /* enforce forwards compatibility on users */
2209 if (unlikely(s->sqe->flags & ~SQE_VALID_FLAGS)) {
2210 ret = -EINVAL;
2211 goto err;
2212 }
2213
2214 req = io_get_req(ctx, state);
2215 if (unlikely(!req)) {
2216 ret = -EAGAIN;
2217 goto err;
2218 }
2219
2220 ret = io_req_set_file(ctx, s, state, req);
2221 if (unlikely(ret)) {
2222err_req:
2223 io_free_req(req);
2224err:
2225 io_cqring_add_event(ctx, s->sqe->user_data, ret);
2226 return;
2227 }
2228
9e645e11
JA
2229 /*
2230 * If we already have a head request, queue this one for async
2231 * submittal once the head completes. If we don't have a head but
2232 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2233 * submitted sync once the chain is complete. If none of those
2234 * conditions are true (normal request), then just queue it.
2235 */
2236 if (*link) {
2237 struct io_kiocb *prev = *link;
2238
2239 sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL);
2240 if (!sqe_copy) {
2241 ret = -EAGAIN;
2242 goto err_req;
2243 }
2244
2245 s->sqe = sqe_copy;
2246 memcpy(&req->submit, s, sizeof(*s));
2247 list_add_tail(&req->list, &prev->link_list);
2248 } else if (s->sqe->flags & IOSQE_IO_LINK) {
2249 req->flags |= REQ_F_LINK;
2250
2251 memcpy(&req->submit, s, sizeof(*s));
2252 INIT_LIST_HEAD(&req->link_list);
2253 *link = req;
2254 } else {
c5766668 2255 io_queue_sqe(ctx, req, s, force_nonblock);
9e645e11
JA
2256 }
2257}
2258
9a56a232
JA
2259/*
2260 * Batched submission is done, ensure local IO is flushed out.
2261 */
2262static void io_submit_state_end(struct io_submit_state *state)
2263{
2264 blk_finish_plug(&state->plug);
3d6770fb 2265 io_file_put(state);
2579f913
JA
2266 if (state->free_reqs)
2267 kmem_cache_free_bulk(req_cachep, state->free_reqs,
2268 &state->reqs[state->cur_req]);
9a56a232
JA
2269}
2270
2271/*
2272 * Start submission side cache.
2273 */
2274static void io_submit_state_start(struct io_submit_state *state,
2275 struct io_ring_ctx *ctx, unsigned max_ios)
2276{
2277 blk_start_plug(&state->plug);
2579f913 2278 state->free_reqs = 0;
9a56a232
JA
2279 state->file = NULL;
2280 state->ios_left = max_ios;
2281}
2282
2b188cc1
JA
2283static void io_commit_sqring(struct io_ring_ctx *ctx)
2284{
75b28aff 2285 struct io_rings *rings = ctx->rings;
2b188cc1 2286
75b28aff 2287 if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) {
2b188cc1
JA
2288 /*
2289 * Ensure any loads from the SQEs are done at this point,
2290 * since once we write the new head, the application could
2291 * write new data to them.
2292 */
75b28aff 2293 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2b188cc1
JA
2294 }
2295}
2296
2b188cc1
JA
2297/*
2298 * Fetch an sqe, if one is available. Note that s->sqe will point to memory
2299 * that is mapped by userspace. This means that care needs to be taken to
2300 * ensure that reads are stable, as we cannot rely on userspace always
2301 * being a good citizen. If members of the sqe are validated and then later
2302 * used, it's important that those reads are done through READ_ONCE() to
2303 * prevent a re-load down the line.
2304 */
2305static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s)
2306{
75b28aff
HV
2307 struct io_rings *rings = ctx->rings;
2308 u32 *sq_array = ctx->sq_array;
2b188cc1
JA
2309 unsigned head;
2310
2311 /*
2312 * The cached sq head (or cq tail) serves two purposes:
2313 *
2314 * 1) allows us to batch the cost of updating the user visible
2315 * head updates.
2316 * 2) allows the kernel side to track the head on its own, even
2317 * though the application is the one updating it.
2318 */
2319 head = ctx->cached_sq_head;
e523a29c 2320 /* make sure SQ entry isn't read before tail */
75b28aff 2321 if (head == smp_load_acquire(&rings->sq.tail))
2b188cc1
JA
2322 return false;
2323
75b28aff 2324 head = READ_ONCE(sq_array[head & ctx->sq_mask]);
2b188cc1
JA
2325 if (head < ctx->sq_entries) {
2326 s->index = head;
2327 s->sqe = &ctx->sq_sqes[head];
8776f3fa 2328 s->sequence = ctx->cached_sq_head;
2b188cc1
JA
2329 ctx->cached_sq_head++;
2330 return true;
2331 }
2332
2333 /* drop invalid entries */
2334 ctx->cached_sq_head++;
75b28aff 2335 rings->sq_dropped++;
2b188cc1
JA
2336 return false;
2337}
2338
6c271ce2
JA
2339static int io_submit_sqes(struct io_ring_ctx *ctx, struct sqe_submit *sqes,
2340 unsigned int nr, bool has_user, bool mm_fault)
2341{
2342 struct io_submit_state state, *statep = NULL;
9e645e11 2343 struct io_kiocb *link = NULL;
4fe2c963 2344 struct io_kiocb *shadow_req = NULL;
9e645e11
JA
2345 bool prev_was_link = false;
2346 int i, submitted = 0;
6c271ce2
JA
2347
2348 if (nr > IO_PLUG_THRESHOLD) {
2349 io_submit_state_start(&state, ctx, nr);
2350 statep = &state;
2351 }
2352
2353 for (i = 0; i < nr; i++) {
9e645e11
JA
2354 /*
2355 * If previous wasn't linked and we have a linked command,
2356 * that's the end of the chain. Submit the previous link.
2357 */
2358 if (!prev_was_link && link) {
c5766668
JA
2359 io_queue_link_head(ctx, link, &link->submit, shadow_req,
2360 true);
9e645e11
JA
2361 link = NULL;
2362 }
2363 prev_was_link = (sqes[i].sqe->flags & IOSQE_IO_LINK) != 0;
2364
4fe2c963
JL
2365 if (link && (sqes[i].sqe->flags & IOSQE_IO_DRAIN)) {
2366 if (!shadow_req) {
2367 shadow_req = io_get_req(ctx, NULL);
2368 shadow_req->flags |= (REQ_F_IO_DRAIN | REQ_F_SHADOW_DRAIN);
2369 refcount_dec(&shadow_req->refs);
2370 }
2371 shadow_req->sequence = sqes[i].sequence;
2372 }
2373
6c271ce2 2374 if (unlikely(mm_fault)) {
9e645e11
JA
2375 io_cqring_add_event(ctx, sqes[i].sqe->user_data,
2376 -EFAULT);
6c271ce2
JA
2377 } else {
2378 sqes[i].has_user = has_user;
2379 sqes[i].needs_lock = true;
2380 sqes[i].needs_fixed_file = true;
c5766668 2381 io_submit_sqe(ctx, &sqes[i], statep, &link, true);
6c271ce2 2382 submitted++;
6c271ce2 2383 }
6c271ce2
JA
2384 }
2385
9e645e11 2386 if (link)
c5766668 2387 io_queue_link_head(ctx, link, &link->submit, shadow_req, true);
6c271ce2
JA
2388 if (statep)
2389 io_submit_state_end(&state);
2390
2391 return submitted;
2392}
2393
2394static int io_sq_thread(void *data)
2395{
2396 struct sqe_submit sqes[IO_IOPOLL_BATCH];
2397 struct io_ring_ctx *ctx = data;
2398 struct mm_struct *cur_mm = NULL;
2399 mm_segment_t old_fs;
2400 DEFINE_WAIT(wait);
2401 unsigned inflight;
2402 unsigned long timeout;
2403
a4c0b3de
JL
2404 complete(&ctx->sqo_thread_started);
2405
6c271ce2
JA
2406 old_fs = get_fs();
2407 set_fs(USER_DS);
2408
2409 timeout = inflight = 0;
2bbcd6d3 2410 while (!kthread_should_park()) {
6c271ce2
JA
2411 bool all_fixed, mm_fault = false;
2412 int i;
2413
2414 if (inflight) {
2415 unsigned nr_events = 0;
2416
2417 if (ctx->flags & IORING_SETUP_IOPOLL) {
6c271ce2 2418 io_iopoll_check(ctx, &nr_events, 0);
6c271ce2
JA
2419 } else {
2420 /*
2421 * Normal IO, just pretend everything completed.
2422 * We don't have to poll completions for that.
2423 */
2424 nr_events = inflight;
2425 }
2426
2427 inflight -= nr_events;
2428 if (!inflight)
2429 timeout = jiffies + ctx->sq_thread_idle;
2430 }
2431
2432 if (!io_get_sqring(ctx, &sqes[0])) {
2433 /*
2434 * We're polling. If we're within the defined idle
2435 * period, then let us spin without work before going
2436 * to sleep.
2437 */
2438 if (inflight || !time_after(jiffies, timeout)) {
2439 cpu_relax();
2440 continue;
2441 }
2442
2443 /*
2444 * Drop cur_mm before scheduling, we can't hold it for
2445 * long periods (or over schedule()). Do this before
2446 * adding ourselves to the waitqueue, as the unuse/drop
2447 * may sleep.
2448 */
2449 if (cur_mm) {
2450 unuse_mm(cur_mm);
2451 mmput(cur_mm);
2452 cur_mm = NULL;
2453 }
2454
2455 prepare_to_wait(&ctx->sqo_wait, &wait,
2456 TASK_INTERRUPTIBLE);
2457
2458 /* Tell userspace we may need a wakeup call */
75b28aff 2459 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
0d7bae69
SB
2460 /* make sure to read SQ tail after writing flags */
2461 smp_mb();
6c271ce2
JA
2462
2463 if (!io_get_sqring(ctx, &sqes[0])) {
2bbcd6d3 2464 if (kthread_should_park()) {
6c271ce2
JA
2465 finish_wait(&ctx->sqo_wait, &wait);
2466 break;
2467 }
2468 if (signal_pending(current))
2469 flush_signals(current);
2470 schedule();
2471 finish_wait(&ctx->sqo_wait, &wait);
2472
75b28aff 2473 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6c271ce2
JA
2474 continue;
2475 }
2476 finish_wait(&ctx->sqo_wait, &wait);
2477
75b28aff 2478 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6c271ce2
JA
2479 }
2480
2481 i = 0;
2482 all_fixed = true;
2483 do {
2484 if (all_fixed && io_sqe_needs_user(sqes[i].sqe))
2485 all_fixed = false;
2486
2487 i++;
2488 if (i == ARRAY_SIZE(sqes))
2489 break;
2490 } while (io_get_sqring(ctx, &sqes[i]));
2491
2492 /* Unless all new commands are FIXED regions, grab mm */
2493 if (!all_fixed && !cur_mm) {
2494 mm_fault = !mmget_not_zero(ctx->sqo_mm);
2495 if (!mm_fault) {
2496 use_mm(ctx->sqo_mm);
2497 cur_mm = ctx->sqo_mm;
2498 }
2499 }
2500
2501 inflight += io_submit_sqes(ctx, sqes, i, cur_mm != NULL,
2502 mm_fault);
2503
2504 /* Commit SQ ring head once we've consumed all SQEs */
2505 io_commit_sqring(ctx);
2506 }
2507
2508 set_fs(old_fs);
2509 if (cur_mm) {
2510 unuse_mm(cur_mm);
2511 mmput(cur_mm);
2512 }
06058632 2513
2bbcd6d3 2514 kthread_parkme();
06058632 2515
6c271ce2
JA
2516 return 0;
2517}
2518
c5766668
JA
2519static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit,
2520 bool block_for_last)
2b188cc1 2521{
9a56a232 2522 struct io_submit_state state, *statep = NULL;
9e645e11 2523 struct io_kiocb *link = NULL;
4fe2c963 2524 struct io_kiocb *shadow_req = NULL;
9e645e11 2525 bool prev_was_link = false;
5c8b0b54 2526 int i, submit = 0;
2b188cc1 2527
9a56a232
JA
2528 if (to_submit > IO_PLUG_THRESHOLD) {
2529 io_submit_state_start(&state, ctx, to_submit);
2530 statep = &state;
2531 }
2b188cc1
JA
2532
2533 for (i = 0; i < to_submit; i++) {
c5766668 2534 bool force_nonblock = true;
2b188cc1
JA
2535 struct sqe_submit s;
2536
2537 if (!io_get_sqring(ctx, &s))
2538 break;
2539
9e645e11
JA
2540 /*
2541 * If previous wasn't linked and we have a linked command,
2542 * that's the end of the chain. Submit the previous link.
2543 */
2544 if (!prev_was_link && link) {
c5766668
JA
2545 io_queue_link_head(ctx, link, &link->submit, shadow_req,
2546 force_nonblock);
9e645e11
JA
2547 link = NULL;
2548 }
2549 prev_was_link = (s.sqe->flags & IOSQE_IO_LINK) != 0;
2550
4fe2c963
JL
2551 if (link && (s.sqe->flags & IOSQE_IO_DRAIN)) {
2552 if (!shadow_req) {
2553 shadow_req = io_get_req(ctx, NULL);
2554 shadow_req->flags |= (REQ_F_IO_DRAIN | REQ_F_SHADOW_DRAIN);
2555 refcount_dec(&shadow_req->refs);
2556 }
2557 shadow_req->sequence = s.sequence;
2558 }
2559
2b188cc1 2560 s.has_user = true;
def596e9 2561 s.needs_lock = false;
6c271ce2 2562 s.needs_fixed_file = false;
5c8b0b54 2563 submit++;
c5766668
JA
2564
2565 /*
2566 * The caller will block for events after submit, submit the
2567 * last IO non-blocking. This is either the only IO it's
2568 * submitting, or it already submitted the previous ones. This
2569 * improves performance by avoiding an async punt that we don't
2570 * need to do.
2571 */
2572 if (block_for_last && submit == to_submit)
2573 force_nonblock = false;
2574
2575 io_submit_sqe(ctx, &s, statep, &link, force_nonblock);
2b188cc1
JA
2576 }
2577 io_commit_sqring(ctx);
2578
9e645e11 2579 if (link)
c5766668
JA
2580 io_queue_link_head(ctx, link, &link->submit, shadow_req,
2581 block_for_last);
9a56a232
JA
2582 if (statep)
2583 io_submit_state_end(statep);
2b188cc1 2584
5c8b0b54 2585 return submit;
2b188cc1
JA
2586}
2587
2b188cc1
JA
2588/*
2589 * Wait until events become available, if we don't already have some. The
2590 * application must reap them itself, as they reside on the shared cq ring.
2591 */
2592static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2593 const sigset_t __user *sig, size_t sigsz)
2594{
75b28aff 2595 struct io_rings *rings = ctx->rings;
2b188cc1
JA
2596 int ret;
2597
75b28aff 2598 if (io_cqring_events(rings) >= min_events)
2b188cc1
JA
2599 return 0;
2600
2601 if (sig) {
9e75ad5d
AB
2602#ifdef CONFIG_COMPAT
2603 if (in_compat_syscall())
2604 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
b772434b 2605 sigsz);
9e75ad5d
AB
2606 else
2607#endif
b772434b 2608 ret = set_user_sigmask(sig, sigsz);
9e75ad5d 2609
2b188cc1
JA
2610 if (ret)
2611 return ret;
2612 }
2613
75b28aff 2614 ret = wait_event_interruptible(ctx->wait, io_cqring_events(rings) >= min_events);
b772434b 2615 restore_saved_sigmask_unless(ret == -ERESTARTSYS);
97abc889
ON
2616 if (ret == -ERESTARTSYS)
2617 ret = -EINTR;
2b188cc1 2618
75b28aff 2619 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2b188cc1
JA
2620}
2621
6b06314c
JA
2622static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
2623{
2624#if defined(CONFIG_UNIX)
2625 if (ctx->ring_sock) {
2626 struct sock *sock = ctx->ring_sock->sk;
2627 struct sk_buff *skb;
2628
2629 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
2630 kfree_skb(skb);
2631 }
2632#else
2633 int i;
2634
2635 for (i = 0; i < ctx->nr_user_files; i++)
2636 fput(ctx->user_files[i]);
2637#endif
2638}
2639
2640static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
2641{
2642 if (!ctx->user_files)
2643 return -ENXIO;
2644
2645 __io_sqe_files_unregister(ctx);
2646 kfree(ctx->user_files);
2647 ctx->user_files = NULL;
2648 ctx->nr_user_files = 0;
2649 return 0;
2650}
2651
6c271ce2
JA
2652static void io_sq_thread_stop(struct io_ring_ctx *ctx)
2653{
2654 if (ctx->sqo_thread) {
a4c0b3de 2655 wait_for_completion(&ctx->sqo_thread_started);
2bbcd6d3
RP
2656 /*
2657 * The park is a bit of a work-around, without it we get
2658 * warning spews on shutdown with SQPOLL set and affinity
2659 * set to a single CPU.
2660 */
06058632 2661 kthread_park(ctx->sqo_thread);
6c271ce2
JA
2662 kthread_stop(ctx->sqo_thread);
2663 ctx->sqo_thread = NULL;
2664 }
2665}
2666
6b06314c
JA
2667static void io_finish_async(struct io_ring_ctx *ctx)
2668{
54a91f3b
JA
2669 int i;
2670
6c271ce2
JA
2671 io_sq_thread_stop(ctx);
2672
54a91f3b
JA
2673 for (i = 0; i < ARRAY_SIZE(ctx->sqo_wq); i++) {
2674 if (ctx->sqo_wq[i]) {
2675 destroy_workqueue(ctx->sqo_wq[i]);
2676 ctx->sqo_wq[i] = NULL;
2677 }
6b06314c
JA
2678 }
2679}
2680
2681#if defined(CONFIG_UNIX)
2682static void io_destruct_skb(struct sk_buff *skb)
2683{
2684 struct io_ring_ctx *ctx = skb->sk->sk_user_data;
2685
2686 io_finish_async(ctx);
2687 unix_destruct_scm(skb);
2688}
2689
2690/*
2691 * Ensure the UNIX gc is aware of our file set, so we are certain that
2692 * the io_uring can be safely unregistered on process exit, even if we have
2693 * loops in the file referencing.
2694 */
2695static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
2696{
2697 struct sock *sk = ctx->ring_sock->sk;
2698 struct scm_fp_list *fpl;
2699 struct sk_buff *skb;
2700 int i;
2701
2702 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
2703 unsigned long inflight = ctx->user->unix_inflight + nr;
2704
2705 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
2706 return -EMFILE;
2707 }
2708
2709 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
2710 if (!fpl)
2711 return -ENOMEM;
2712
2713 skb = alloc_skb(0, GFP_KERNEL);
2714 if (!skb) {
2715 kfree(fpl);
2716 return -ENOMEM;
2717 }
2718
2719 skb->sk = sk;
2720 skb->destructor = io_destruct_skb;
2721
2722 fpl->user = get_uid(ctx->user);
2723 for (i = 0; i < nr; i++) {
2724 fpl->fp[i] = get_file(ctx->user_files[i + offset]);
2725 unix_inflight(fpl->user, fpl->fp[i]);
2726 }
2727
2728 fpl->max = fpl->count = nr;
2729 UNIXCB(skb).fp = fpl;
2730 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
2731 skb_queue_head(&sk->sk_receive_queue, skb);
2732
2733 for (i = 0; i < nr; i++)
2734 fput(fpl->fp[i]);
2735
2736 return 0;
2737}
2738
2739/*
2740 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
2741 * causes regular reference counting to break down. We rely on the UNIX
2742 * garbage collection to take care of this problem for us.
2743 */
2744static int io_sqe_files_scm(struct io_ring_ctx *ctx)
2745{
2746 unsigned left, total;
2747 int ret = 0;
2748
2749 total = 0;
2750 left = ctx->nr_user_files;
2751 while (left) {
2752 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6b06314c
JA
2753
2754 ret = __io_sqe_files_scm(ctx, this_files, total);
2755 if (ret)
2756 break;
2757 left -= this_files;
2758 total += this_files;
2759 }
2760
2761 if (!ret)
2762 return 0;
2763
2764 while (total < ctx->nr_user_files) {
2765 fput(ctx->user_files[total]);
2766 total++;
2767 }
2768
2769 return ret;
2770}
2771#else
2772static int io_sqe_files_scm(struct io_ring_ctx *ctx)
2773{
2774 return 0;
2775}
2776#endif
2777
2778static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
2779 unsigned nr_args)
2780{
2781 __s32 __user *fds = (__s32 __user *) arg;
2782 int fd, ret = 0;
2783 unsigned i;
2784
2785 if (ctx->user_files)
2786 return -EBUSY;
2787 if (!nr_args)
2788 return -EINVAL;
2789 if (nr_args > IORING_MAX_FIXED_FILES)
2790 return -EMFILE;
2791
2792 ctx->user_files = kcalloc(nr_args, sizeof(struct file *), GFP_KERNEL);
2793 if (!ctx->user_files)
2794 return -ENOMEM;
2795
2796 for (i = 0; i < nr_args; i++) {
2797 ret = -EFAULT;
2798 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
2799 break;
2800
2801 ctx->user_files[i] = fget(fd);
2802
2803 ret = -EBADF;
2804 if (!ctx->user_files[i])
2805 break;
2806 /*
2807 * Don't allow io_uring instances to be registered. If UNIX
2808 * isn't enabled, then this causes a reference cycle and this
2809 * instance can never get freed. If UNIX is enabled we'll
2810 * handle it just fine, but there's still no point in allowing
2811 * a ring fd as it doesn't support regular read/write anyway.
2812 */
2813 if (ctx->user_files[i]->f_op == &io_uring_fops) {
2814 fput(ctx->user_files[i]);
2815 break;
2816 }
2817 ctx->nr_user_files++;
2818 ret = 0;
2819 }
2820
2821 if (ret) {
2822 for (i = 0; i < ctx->nr_user_files; i++)
2823 fput(ctx->user_files[i]);
2824
2825 kfree(ctx->user_files);
25adf50f 2826 ctx->user_files = NULL;
6b06314c
JA
2827 ctx->nr_user_files = 0;
2828 return ret;
2829 }
2830
2831 ret = io_sqe_files_scm(ctx);
2832 if (ret)
2833 io_sqe_files_unregister(ctx);
2834
2835 return ret;
2836}
2837
6c271ce2
JA
2838static int io_sq_offload_start(struct io_ring_ctx *ctx,
2839 struct io_uring_params *p)
2b188cc1
JA
2840{
2841 int ret;
2842
6c271ce2 2843 init_waitqueue_head(&ctx->sqo_wait);
2b188cc1
JA
2844 mmgrab(current->mm);
2845 ctx->sqo_mm = current->mm;
2846
6c271ce2 2847 if (ctx->flags & IORING_SETUP_SQPOLL) {
3ec482d1
JA
2848 ret = -EPERM;
2849 if (!capable(CAP_SYS_ADMIN))
2850 goto err;
2851
917257da
JA
2852 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
2853 if (!ctx->sq_thread_idle)
2854 ctx->sq_thread_idle = HZ;
2855
6c271ce2 2856 if (p->flags & IORING_SETUP_SQ_AFF) {
44a9bd18 2857 int cpu = p->sq_thread_cpu;
6c271ce2 2858
917257da 2859 ret = -EINVAL;
44a9bd18
JA
2860 if (cpu >= nr_cpu_ids)
2861 goto err;
7889f44d 2862 if (!cpu_online(cpu))
917257da
JA
2863 goto err;
2864
6c271ce2
JA
2865 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
2866 ctx, cpu,
2867 "io_uring-sq");
2868 } else {
2869 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
2870 "io_uring-sq");
2871 }
2872 if (IS_ERR(ctx->sqo_thread)) {
2873 ret = PTR_ERR(ctx->sqo_thread);
2874 ctx->sqo_thread = NULL;
2875 goto err;
2876 }
2877 wake_up_process(ctx->sqo_thread);
2878 } else if (p->flags & IORING_SETUP_SQ_AFF) {
2879 /* Can't have SQ_AFF without SQPOLL */
2880 ret = -EINVAL;
2881 goto err;
2882 }
2883
2b188cc1 2884 /* Do QD, or 2 * CPUS, whatever is smallest */
54a91f3b
JA
2885 ctx->sqo_wq[0] = alloc_workqueue("io_ring-wq",
2886 WQ_UNBOUND | WQ_FREEZABLE,
2b188cc1 2887 min(ctx->sq_entries - 1, 2 * num_online_cpus()));
54a91f3b
JA
2888 if (!ctx->sqo_wq[0]) {
2889 ret = -ENOMEM;
2890 goto err;
2891 }
2892
2893 /*
2894 * This is for buffered writes, where we want to limit the parallelism
2895 * due to file locking in file systems. As "normal" buffered writes
2896 * should parellelize on writeout quite nicely, limit us to having 2
2897 * pending. This avoids massive contention on the inode when doing
2898 * buffered async writes.
2899 */
2900 ctx->sqo_wq[1] = alloc_workqueue("io_ring-write-wq",
2901 WQ_UNBOUND | WQ_FREEZABLE, 2);
2902 if (!ctx->sqo_wq[1]) {
2b188cc1
JA
2903 ret = -ENOMEM;
2904 goto err;
2905 }
2906
2907 return 0;
2908err:
54a91f3b 2909 io_finish_async(ctx);
2b188cc1
JA
2910 mmdrop(ctx->sqo_mm);
2911 ctx->sqo_mm = NULL;
2912 return ret;
2913}
2914
2915static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
2916{
2917 atomic_long_sub(nr_pages, &user->locked_vm);
2918}
2919
2920static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
2921{
2922 unsigned long page_limit, cur_pages, new_pages;
2923
2924 /* Don't allow more pages than we can safely lock */
2925 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
2926
2927 do {
2928 cur_pages = atomic_long_read(&user->locked_vm);
2929 new_pages = cur_pages + nr_pages;
2930 if (new_pages > page_limit)
2931 return -ENOMEM;
2932 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
2933 new_pages) != cur_pages);
2934
2935 return 0;
2936}
2937
2938static void io_mem_free(void *ptr)
2939{
52e04ef4
MR
2940 struct page *page;
2941
2942 if (!ptr)
2943 return;
2b188cc1 2944
52e04ef4 2945 page = virt_to_head_page(ptr);
2b188cc1
JA
2946 if (put_page_testzero(page))
2947 free_compound_page(page);
2948}
2949
2950static void *io_mem_alloc(size_t size)
2951{
2952 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
2953 __GFP_NORETRY;
2954
2955 return (void *) __get_free_pages(gfp_flags, get_order(size));
2956}
2957
75b28aff
HV
2958static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
2959 size_t *sq_offset)
2960{
2961 struct io_rings *rings;
2962 size_t off, sq_array_size;
2963
2964 off = struct_size(rings, cqes, cq_entries);
2965 if (off == SIZE_MAX)
2966 return SIZE_MAX;
2967
2968#ifdef CONFIG_SMP
2969 off = ALIGN(off, SMP_CACHE_BYTES);
2970 if (off == 0)
2971 return SIZE_MAX;
2972#endif
2973
2974 sq_array_size = array_size(sizeof(u32), sq_entries);
2975 if (sq_array_size == SIZE_MAX)
2976 return SIZE_MAX;
2977
2978 if (check_add_overflow(off, sq_array_size, &off))
2979 return SIZE_MAX;
2980
2981 if (sq_offset)
2982 *sq_offset = off;
2983
2984 return off;
2985}
2986
2b188cc1
JA
2987static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
2988{
75b28aff 2989 size_t pages;
2b188cc1 2990
75b28aff
HV
2991 pages = (size_t)1 << get_order(
2992 rings_size(sq_entries, cq_entries, NULL));
2993 pages += (size_t)1 << get_order(
2994 array_size(sizeof(struct io_uring_sqe), sq_entries));
2b188cc1 2995
75b28aff 2996 return pages;
2b188cc1
JA
2997}
2998
edafccee
JA
2999static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
3000{
3001 int i, j;
3002
3003 if (!ctx->user_bufs)
3004 return -ENXIO;
3005
3006 for (i = 0; i < ctx->nr_user_bufs; i++) {
3007 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
3008
3009 for (j = 0; j < imu->nr_bvecs; j++)
27c4d3a3 3010 put_user_page(imu->bvec[j].bv_page);
edafccee
JA
3011
3012 if (ctx->account_mem)
3013 io_unaccount_mem(ctx->user, imu->nr_bvecs);
d4ef6475 3014 kvfree(imu->bvec);
edafccee
JA
3015 imu->nr_bvecs = 0;
3016 }
3017
3018 kfree(ctx->user_bufs);
3019 ctx->user_bufs = NULL;
3020 ctx->nr_user_bufs = 0;
3021 return 0;
3022}
3023
3024static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
3025 void __user *arg, unsigned index)
3026{
3027 struct iovec __user *src;
3028
3029#ifdef CONFIG_COMPAT
3030 if (ctx->compat) {
3031 struct compat_iovec __user *ciovs;
3032 struct compat_iovec ciov;
3033
3034 ciovs = (struct compat_iovec __user *) arg;
3035 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
3036 return -EFAULT;
3037
3038 dst->iov_base = (void __user *) (unsigned long) ciov.iov_base;
3039 dst->iov_len = ciov.iov_len;
3040 return 0;
3041 }
3042#endif
3043 src = (struct iovec __user *) arg;
3044 if (copy_from_user(dst, &src[index], sizeof(*dst)))
3045 return -EFAULT;
3046 return 0;
3047}
3048
3049static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
3050 unsigned nr_args)
3051{
3052 struct vm_area_struct **vmas = NULL;
3053 struct page **pages = NULL;
3054 int i, j, got_pages = 0;
3055 int ret = -EINVAL;
3056
3057 if (ctx->user_bufs)
3058 return -EBUSY;
3059 if (!nr_args || nr_args > UIO_MAXIOV)
3060 return -EINVAL;
3061
3062 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
3063 GFP_KERNEL);
3064 if (!ctx->user_bufs)
3065 return -ENOMEM;
3066
3067 for (i = 0; i < nr_args; i++) {
3068 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
3069 unsigned long off, start, end, ubuf;
3070 int pret, nr_pages;
3071 struct iovec iov;
3072 size_t size;
3073
3074 ret = io_copy_iov(ctx, &iov, arg, i);
3075 if (ret)
a278682d 3076 goto err;
edafccee
JA
3077
3078 /*
3079 * Don't impose further limits on the size and buffer
3080 * constraints here, we'll -EINVAL later when IO is
3081 * submitted if they are wrong.
3082 */
3083 ret = -EFAULT;
3084 if (!iov.iov_base || !iov.iov_len)
3085 goto err;
3086
3087 /* arbitrary limit, but we need something */
3088 if (iov.iov_len > SZ_1G)
3089 goto err;
3090
3091 ubuf = (unsigned long) iov.iov_base;
3092 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
3093 start = ubuf >> PAGE_SHIFT;
3094 nr_pages = end - start;
3095
3096 if (ctx->account_mem) {
3097 ret = io_account_mem(ctx->user, nr_pages);
3098 if (ret)
3099 goto err;
3100 }
3101
3102 ret = 0;
3103 if (!pages || nr_pages > got_pages) {
3104 kfree(vmas);
3105 kfree(pages);
d4ef6475 3106 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
edafccee 3107 GFP_KERNEL);
d4ef6475 3108 vmas = kvmalloc_array(nr_pages,
edafccee
JA
3109 sizeof(struct vm_area_struct *),
3110 GFP_KERNEL);
3111 if (!pages || !vmas) {
3112 ret = -ENOMEM;
3113 if (ctx->account_mem)
3114 io_unaccount_mem(ctx->user, nr_pages);
3115 goto err;
3116 }
3117 got_pages = nr_pages;
3118 }
3119
d4ef6475 3120 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
edafccee
JA
3121 GFP_KERNEL);
3122 ret = -ENOMEM;
3123 if (!imu->bvec) {
3124 if (ctx->account_mem)
3125 io_unaccount_mem(ctx->user, nr_pages);
3126 goto err;
3127 }
3128
3129 ret = 0;
3130 down_read(&current->mm->mmap_sem);
932f4a63
IW
3131 pret = get_user_pages(ubuf, nr_pages,
3132 FOLL_WRITE | FOLL_LONGTERM,
3133 pages, vmas);
edafccee
JA
3134 if (pret == nr_pages) {
3135 /* don't support file backed memory */
3136 for (j = 0; j < nr_pages; j++) {
3137 struct vm_area_struct *vma = vmas[j];
3138
3139 if (vma->vm_file &&
3140 !is_file_hugepages(vma->vm_file)) {
3141 ret = -EOPNOTSUPP;
3142 break;
3143 }
3144 }
3145 } else {
3146 ret = pret < 0 ? pret : -EFAULT;
3147 }
3148 up_read(&current->mm->mmap_sem);
3149 if (ret) {
3150 /*
3151 * if we did partial map, or found file backed vmas,
3152 * release any pages we did get
3153 */
27c4d3a3
JH
3154 if (pret > 0)
3155 put_user_pages(pages, pret);
edafccee
JA
3156 if (ctx->account_mem)
3157 io_unaccount_mem(ctx->user, nr_pages);
d4ef6475 3158 kvfree(imu->bvec);
edafccee
JA
3159 goto err;
3160 }
3161
3162 off = ubuf & ~PAGE_MASK;
3163 size = iov.iov_len;
3164 for (j = 0; j < nr_pages; j++) {
3165 size_t vec_len;
3166
3167 vec_len = min_t(size_t, size, PAGE_SIZE - off);
3168 imu->bvec[j].bv_page = pages[j];
3169 imu->bvec[j].bv_len = vec_len;
3170 imu->bvec[j].bv_offset = off;
3171 off = 0;
3172 size -= vec_len;
3173 }
3174 /* store original address for later verification */
3175 imu->ubuf = ubuf;
3176 imu->len = iov.iov_len;
3177 imu->nr_bvecs = nr_pages;
3178
3179 ctx->nr_user_bufs++;
3180 }
d4ef6475
MR
3181 kvfree(pages);
3182 kvfree(vmas);
edafccee
JA
3183 return 0;
3184err:
d4ef6475
MR
3185 kvfree(pages);
3186 kvfree(vmas);
edafccee
JA
3187 io_sqe_buffer_unregister(ctx);
3188 return ret;
3189}
3190
9b402849
JA
3191static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
3192{
3193 __s32 __user *fds = arg;
3194 int fd;
3195
3196 if (ctx->cq_ev_fd)
3197 return -EBUSY;
3198
3199 if (copy_from_user(&fd, fds, sizeof(*fds)))
3200 return -EFAULT;
3201
3202 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
3203 if (IS_ERR(ctx->cq_ev_fd)) {
3204 int ret = PTR_ERR(ctx->cq_ev_fd);
3205 ctx->cq_ev_fd = NULL;
3206 return ret;
3207 }
3208
3209 return 0;
3210}
3211
3212static int io_eventfd_unregister(struct io_ring_ctx *ctx)
3213{
3214 if (ctx->cq_ev_fd) {
3215 eventfd_ctx_put(ctx->cq_ev_fd);
3216 ctx->cq_ev_fd = NULL;
3217 return 0;
3218 }
3219
3220 return -ENXIO;
3221}
3222
2b188cc1
JA
3223static void io_ring_ctx_free(struct io_ring_ctx *ctx)
3224{
6b06314c 3225 io_finish_async(ctx);
2b188cc1
JA
3226 if (ctx->sqo_mm)
3227 mmdrop(ctx->sqo_mm);
def596e9
JA
3228
3229 io_iopoll_reap_events(ctx);
edafccee 3230 io_sqe_buffer_unregister(ctx);
6b06314c 3231 io_sqe_files_unregister(ctx);
9b402849 3232 io_eventfd_unregister(ctx);
def596e9 3233
2b188cc1 3234#if defined(CONFIG_UNIX)
355e8d26
EB
3235 if (ctx->ring_sock) {
3236 ctx->ring_sock->file = NULL; /* so that iput() is called */
2b188cc1 3237 sock_release(ctx->ring_sock);
355e8d26 3238 }
2b188cc1
JA
3239#endif
3240
75b28aff 3241 io_mem_free(ctx->rings);
2b188cc1 3242 io_mem_free(ctx->sq_sqes);
2b188cc1
JA
3243
3244 percpu_ref_exit(&ctx->refs);
3245 if (ctx->account_mem)
3246 io_unaccount_mem(ctx->user,
3247 ring_pages(ctx->sq_entries, ctx->cq_entries));
3248 free_uid(ctx->user);
3249 kfree(ctx);
3250}
3251
3252static __poll_t io_uring_poll(struct file *file, poll_table *wait)
3253{
3254 struct io_ring_ctx *ctx = file->private_data;
3255 __poll_t mask = 0;
3256
3257 poll_wait(file, &ctx->cq_wait, wait);
4f7067c3
SB
3258 /*
3259 * synchronizes with barrier from wq_has_sleeper call in
3260 * io_commit_cqring
3261 */
2b188cc1 3262 smp_rmb();
75b28aff
HV
3263 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
3264 ctx->rings->sq_ring_entries)
2b188cc1 3265 mask |= EPOLLOUT | EPOLLWRNORM;
75b28aff 3266 if (READ_ONCE(ctx->rings->sq.head) != ctx->cached_cq_tail)
2b188cc1
JA
3267 mask |= EPOLLIN | EPOLLRDNORM;
3268
3269 return mask;
3270}
3271
3272static int io_uring_fasync(int fd, struct file *file, int on)
3273{
3274 struct io_ring_ctx *ctx = file->private_data;
3275
3276 return fasync_helper(fd, file, on, &ctx->cq_fasync);
3277}
3278
3279static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3280{
3281 mutex_lock(&ctx->uring_lock);
3282 percpu_ref_kill(&ctx->refs);
3283 mutex_unlock(&ctx->uring_lock);
3284
221c5eb2 3285 io_poll_remove_all(ctx);
def596e9 3286 io_iopoll_reap_events(ctx);
2b188cc1
JA
3287 wait_for_completion(&ctx->ctx_done);
3288 io_ring_ctx_free(ctx);
3289}
3290
3291static int io_uring_release(struct inode *inode, struct file *file)
3292{
3293 struct io_ring_ctx *ctx = file->private_data;
3294
3295 file->private_data = NULL;
3296 io_ring_ctx_wait_and_kill(ctx);
3297 return 0;
3298}
3299
3300static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3301{
3302 loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT;
3303 unsigned long sz = vma->vm_end - vma->vm_start;
3304 struct io_ring_ctx *ctx = file->private_data;
3305 unsigned long pfn;
3306 struct page *page;
3307 void *ptr;
3308
3309 switch (offset) {
3310 case IORING_OFF_SQ_RING:
75b28aff
HV
3311 case IORING_OFF_CQ_RING:
3312 ptr = ctx->rings;
2b188cc1
JA
3313 break;
3314 case IORING_OFF_SQES:
3315 ptr = ctx->sq_sqes;
3316 break;
2b188cc1
JA
3317 default:
3318 return -EINVAL;
3319 }
3320
3321 page = virt_to_head_page(ptr);
3322 if (sz > (PAGE_SIZE << compound_order(page)))
3323 return -EINVAL;
3324
3325 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3326 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3327}
3328
3329SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3330 u32, min_complete, u32, flags, const sigset_t __user *, sig,
3331 size_t, sigsz)
3332{
3333 struct io_ring_ctx *ctx;
3334 long ret = -EBADF;
3335 int submitted = 0;
3336 struct fd f;
3337
6c271ce2 3338 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
2b188cc1
JA
3339 return -EINVAL;
3340
3341 f = fdget(fd);
3342 if (!f.file)
3343 return -EBADF;
3344
3345 ret = -EOPNOTSUPP;
3346 if (f.file->f_op != &io_uring_fops)
3347 goto out_fput;
3348
3349 ret = -ENXIO;
3350 ctx = f.file->private_data;
3351 if (!percpu_ref_tryget(&ctx->refs))
3352 goto out_fput;
3353
6c271ce2
JA
3354 /*
3355 * For SQ polling, the thread will do all submissions and completions.
3356 * Just return the requested submit count, and wake the thread if
3357 * we were asked to.
3358 */
b2a9eada 3359 ret = 0;
6c271ce2
JA
3360 if (ctx->flags & IORING_SETUP_SQPOLL) {
3361 if (flags & IORING_ENTER_SQ_WAKEUP)
3362 wake_up(&ctx->sqo_wait);
3363 submitted = to_submit;
b2a9eada 3364 } else if (to_submit) {
c5766668
JA
3365 bool block_for_last = false;
3366
2b188cc1
JA
3367 to_submit = min(to_submit, ctx->sq_entries);
3368
c5766668
JA
3369 /*
3370 * Allow last submission to block in a series, IFF the caller
3371 * asked to wait for events and we don't currently have
3372 * enough. This potentially avoids an async punt.
3373 */
3374 if (to_submit == min_complete &&
3375 io_cqring_events(ctx->rings) < min_complete)
3376 block_for_last = true;
3377
2b188cc1 3378 mutex_lock(&ctx->uring_lock);
c5766668 3379 submitted = io_ring_submit(ctx, to_submit, block_for_last);
2b188cc1 3380 mutex_unlock(&ctx->uring_lock);
2b188cc1
JA
3381 }
3382 if (flags & IORING_ENTER_GETEVENTS) {
def596e9
JA
3383 unsigned nr_events = 0;
3384
2b188cc1
JA
3385 min_complete = min(min_complete, ctx->cq_entries);
3386
def596e9 3387 if (ctx->flags & IORING_SETUP_IOPOLL) {
def596e9 3388 ret = io_iopoll_check(ctx, &nr_events, min_complete);
def596e9
JA
3389 } else {
3390 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
3391 }
2b188cc1
JA
3392 }
3393
2b188cc1
JA
3394 io_ring_drop_ctx_refs(ctx, 1);
3395out_fput:
3396 fdput(f);
3397 return submitted ? submitted : ret;
3398}
3399
3400static const struct file_operations io_uring_fops = {
3401 .release = io_uring_release,
3402 .mmap = io_uring_mmap,
3403 .poll = io_uring_poll,
3404 .fasync = io_uring_fasync,
3405};
3406
3407static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3408 struct io_uring_params *p)
3409{
75b28aff
HV
3410 struct io_rings *rings;
3411 size_t size, sq_array_offset;
2b188cc1 3412
75b28aff
HV
3413 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
3414 if (size == SIZE_MAX)
3415 return -EOVERFLOW;
3416
3417 rings = io_mem_alloc(size);
3418 if (!rings)
2b188cc1
JA
3419 return -ENOMEM;
3420
75b28aff
HV
3421 ctx->rings = rings;
3422 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3423 rings->sq_ring_mask = p->sq_entries - 1;
3424 rings->cq_ring_mask = p->cq_entries - 1;
3425 rings->sq_ring_entries = p->sq_entries;
3426 rings->cq_ring_entries = p->cq_entries;
3427 ctx->sq_mask = rings->sq_ring_mask;
3428 ctx->cq_mask = rings->cq_ring_mask;
3429 ctx->sq_entries = rings->sq_ring_entries;
3430 ctx->cq_entries = rings->cq_ring_entries;
2b188cc1
JA
3431
3432 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3433 if (size == SIZE_MAX)
3434 return -EOVERFLOW;
3435
3436 ctx->sq_sqes = io_mem_alloc(size);
52e04ef4 3437 if (!ctx->sq_sqes)
2b188cc1 3438 return -ENOMEM;
2b188cc1 3439
2b188cc1
JA
3440 return 0;
3441}
3442
3443/*
3444 * Allocate an anonymous fd, this is what constitutes the application
3445 * visible backing of an io_uring instance. The application mmaps this
3446 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3447 * we have to tie this fd to a socket for file garbage collection purposes.
3448 */
3449static int io_uring_get_fd(struct io_ring_ctx *ctx)
3450{
3451 struct file *file;
3452 int ret;
3453
3454#if defined(CONFIG_UNIX)
3455 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3456 &ctx->ring_sock);
3457 if (ret)
3458 return ret;
3459#endif
3460
3461 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3462 if (ret < 0)
3463 goto err;
3464
3465 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
3466 O_RDWR | O_CLOEXEC);
3467 if (IS_ERR(file)) {
3468 put_unused_fd(ret);
3469 ret = PTR_ERR(file);
3470 goto err;
3471 }
3472
3473#if defined(CONFIG_UNIX)
3474 ctx->ring_sock->file = file;
6b06314c 3475 ctx->ring_sock->sk->sk_user_data = ctx;
2b188cc1
JA
3476#endif
3477 fd_install(ret, file);
3478 return ret;
3479err:
3480#if defined(CONFIG_UNIX)
3481 sock_release(ctx->ring_sock);
3482 ctx->ring_sock = NULL;
3483#endif
3484 return ret;
3485}
3486
3487static int io_uring_create(unsigned entries, struct io_uring_params *p)
3488{
3489 struct user_struct *user = NULL;
3490 struct io_ring_ctx *ctx;
3491 bool account_mem;
3492 int ret;
3493
3494 if (!entries || entries > IORING_MAX_ENTRIES)
3495 return -EINVAL;
3496
3497 /*
3498 * Use twice as many entries for the CQ ring. It's possible for the
3499 * application to drive a higher depth than the size of the SQ ring,
3500 * since the sqes are only used at submission time. This allows for
3501 * some flexibility in overcommitting a bit.
3502 */
3503 p->sq_entries = roundup_pow_of_two(entries);
3504 p->cq_entries = 2 * p->sq_entries;
3505
3506 user = get_uid(current_user());
3507 account_mem = !capable(CAP_IPC_LOCK);
3508
3509 if (account_mem) {
3510 ret = io_account_mem(user,
3511 ring_pages(p->sq_entries, p->cq_entries));
3512 if (ret) {
3513 free_uid(user);
3514 return ret;
3515 }
3516 }
3517
3518 ctx = io_ring_ctx_alloc(p);
3519 if (!ctx) {
3520 if (account_mem)
3521 io_unaccount_mem(user, ring_pages(p->sq_entries,
3522 p->cq_entries));
3523 free_uid(user);
3524 return -ENOMEM;
3525 }
3526 ctx->compat = in_compat_syscall();
3527 ctx->account_mem = account_mem;
3528 ctx->user = user;
3529
3530 ret = io_allocate_scq_urings(ctx, p);
3531 if (ret)
3532 goto err;
3533
6c271ce2 3534 ret = io_sq_offload_start(ctx, p);
2b188cc1
JA
3535 if (ret)
3536 goto err;
3537
3538 ret = io_uring_get_fd(ctx);
3539 if (ret < 0)
3540 goto err;
3541
3542 memset(&p->sq_off, 0, sizeof(p->sq_off));
75b28aff
HV
3543 p->sq_off.head = offsetof(struct io_rings, sq.head);
3544 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3545 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3546 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3547 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3548 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3549 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
2b188cc1
JA
3550
3551 memset(&p->cq_off, 0, sizeof(p->cq_off));
75b28aff
HV
3552 p->cq_off.head = offsetof(struct io_rings, cq.head);
3553 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3554 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3555 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3556 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3557 p->cq_off.cqes = offsetof(struct io_rings, cqes);
ac90f249
JA
3558
3559 p->features = IORING_FEAT_SINGLE_MMAP;
2b188cc1
JA
3560 return ret;
3561err:
3562 io_ring_ctx_wait_and_kill(ctx);
3563 return ret;
3564}
3565
3566/*
3567 * Sets up an aio uring context, and returns the fd. Applications asks for a
3568 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3569 * params structure passed in.
3570 */
3571static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3572{
3573 struct io_uring_params p;
3574 long ret;
3575 int i;
3576
3577 if (copy_from_user(&p, params, sizeof(p)))
3578 return -EFAULT;
3579 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3580 if (p.resv[i])
3581 return -EINVAL;
3582 }
3583
6c271ce2
JA
3584 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3585 IORING_SETUP_SQ_AFF))
2b188cc1
JA
3586 return -EINVAL;
3587
3588 ret = io_uring_create(entries, &p);
3589 if (ret < 0)
3590 return ret;
3591
3592 if (copy_to_user(params, &p, sizeof(p)))
3593 return -EFAULT;
3594
3595 return ret;
3596}
3597
3598SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3599 struct io_uring_params __user *, params)
3600{
3601 return io_uring_setup(entries, params);
3602}
3603
edafccee
JA
3604static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3605 void __user *arg, unsigned nr_args)
b19062a5
JA
3606 __releases(ctx->uring_lock)
3607 __acquires(ctx->uring_lock)
edafccee
JA
3608{
3609 int ret;
3610
35fa71a0
JA
3611 /*
3612 * We're inside the ring mutex, if the ref is already dying, then
3613 * someone else killed the ctx or is already going through
3614 * io_uring_register().
3615 */
3616 if (percpu_ref_is_dying(&ctx->refs))
3617 return -ENXIO;
3618
edafccee 3619 percpu_ref_kill(&ctx->refs);
b19062a5
JA
3620
3621 /*
3622 * Drop uring mutex before waiting for references to exit. If another
3623 * thread is currently inside io_uring_enter() it might need to grab
3624 * the uring_lock to make progress. If we hold it here across the drain
3625 * wait, then we can deadlock. It's safe to drop the mutex here, since
3626 * no new references will come in after we've killed the percpu ref.
3627 */
3628 mutex_unlock(&ctx->uring_lock);
edafccee 3629 wait_for_completion(&ctx->ctx_done);
b19062a5 3630 mutex_lock(&ctx->uring_lock);
edafccee
JA
3631
3632 switch (opcode) {
3633 case IORING_REGISTER_BUFFERS:
3634 ret = io_sqe_buffer_register(ctx, arg, nr_args);
3635 break;
3636 case IORING_UNREGISTER_BUFFERS:
3637 ret = -EINVAL;
3638 if (arg || nr_args)
3639 break;
3640 ret = io_sqe_buffer_unregister(ctx);
3641 break;
6b06314c
JA
3642 case IORING_REGISTER_FILES:
3643 ret = io_sqe_files_register(ctx, arg, nr_args);
3644 break;
3645 case IORING_UNREGISTER_FILES:
3646 ret = -EINVAL;
3647 if (arg || nr_args)
3648 break;
3649 ret = io_sqe_files_unregister(ctx);
3650 break;
9b402849
JA
3651 case IORING_REGISTER_EVENTFD:
3652 ret = -EINVAL;
3653 if (nr_args != 1)
3654 break;
3655 ret = io_eventfd_register(ctx, arg);
3656 break;
3657 case IORING_UNREGISTER_EVENTFD:
3658 ret = -EINVAL;
3659 if (arg || nr_args)
3660 break;
3661 ret = io_eventfd_unregister(ctx);
3662 break;
edafccee
JA
3663 default:
3664 ret = -EINVAL;
3665 break;
3666 }
3667
3668 /* bring the ctx back to life */
3669 reinit_completion(&ctx->ctx_done);
3670 percpu_ref_reinit(&ctx->refs);
3671 return ret;
3672}
3673
3674SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
3675 void __user *, arg, unsigned int, nr_args)
3676{
3677 struct io_ring_ctx *ctx;
3678 long ret = -EBADF;
3679 struct fd f;
3680
3681 f = fdget(fd);
3682 if (!f.file)
3683 return -EBADF;
3684
3685 ret = -EOPNOTSUPP;
3686 if (f.file->f_op != &io_uring_fops)
3687 goto out_fput;
3688
3689 ctx = f.file->private_data;
3690
3691 mutex_lock(&ctx->uring_lock);
3692 ret = __io_uring_register(ctx, opcode, arg, nr_args);
3693 mutex_unlock(&ctx->uring_lock);
3694out_fput:
3695 fdput(f);
3696 return ret;
3697}
3698
2b188cc1
JA
3699static int __init io_uring_init(void)
3700{
3701 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
3702 return 0;
3703};
3704__initcall(io_uring_init);