]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blame - fs/io_uring.c
io-wq: kill hashed waitqueue before manager exits
[mirror_ubuntu-jammy-kernel.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>
52de1fe1 47#include <net/compat.h>
2b188cc1
JA
48#include <linux/refcount.h>
49#include <linux/uio.h>
6b47ee6e 50#include <linux/bits.h>
2b188cc1
JA
51
52#include <linux/sched/signal.h>
53#include <linux/fs.h>
54#include <linux/file.h>
55#include <linux/fdtable.h>
56#include <linux/mm.h>
57#include <linux/mman.h>
2b188cc1
JA
58#include <linux/percpu.h>
59#include <linux/slab.h>
2b188cc1 60#include <linux/blkdev.h>
edafccee 61#include <linux/bvec.h>
2b188cc1
JA
62#include <linux/net.h>
63#include <net/sock.h>
64#include <net/af_unix.h>
6b06314c 65#include <net/scm.h>
2b188cc1
JA
66#include <linux/anon_inodes.h>
67#include <linux/sched/mm.h>
68#include <linux/uaccess.h>
69#include <linux/nospec.h>
edafccee
JA
70#include <linux/sizes.h>
71#include <linux/hugetlb.h>
aa4c3967 72#include <linux/highmem.h>
15b71abe
JA
73#include <linux/namei.h>
74#include <linux/fsnotify.h>
4840e418 75#include <linux/fadvise.h>
3e4827b0 76#include <linux/eventpoll.h>
7d67af2c 77#include <linux/splice.h>
b41e9852 78#include <linux/task_work.h>
bcf5a063 79#include <linux/pagemap.h>
0f212204 80#include <linux/io_uring.h>
e4b4a13f 81#include <linux/freezer.h>
2b188cc1 82
c826bd7a
DD
83#define CREATE_TRACE_POINTS
84#include <trace/events/io_uring.h>
85
2b188cc1
JA
86#include <uapi/linux/io_uring.h>
87
88#include "internal.h"
561fb04a 89#include "io-wq.h"
2b188cc1 90
5277deaa 91#define IORING_MAX_ENTRIES 32768
33a107f0 92#define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
65e19f54
JA
93
94/*
95 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
96 */
97#define IORING_FILE_TABLE_SHIFT 9
98#define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
99#define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
100#define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
21b55dbc
SG
101#define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
102 IORING_REGISTER_LAST + IORING_OP_LAST)
2b188cc1 103
b16fed66
PB
104#define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
105 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
106 IOSQE_BUFFER_SELECT)
107
2b188cc1
JA
108struct io_uring {
109 u32 head ____cacheline_aligned_in_smp;
110 u32 tail ____cacheline_aligned_in_smp;
111};
112
1e84b97b 113/*
75b28aff
HV
114 * This data is shared with the application through the mmap at offsets
115 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
1e84b97b
SB
116 *
117 * The offsets to the member fields are published through struct
118 * io_sqring_offsets when calling io_uring_setup.
119 */
75b28aff 120struct io_rings {
1e84b97b
SB
121 /*
122 * Head and tail offsets into the ring; the offsets need to be
123 * masked to get valid indices.
124 *
75b28aff
HV
125 * The kernel controls head of the sq ring and the tail of the cq ring,
126 * and the application controls tail of the sq ring and the head of the
127 * cq ring.
1e84b97b 128 */
75b28aff 129 struct io_uring sq, cq;
1e84b97b 130 /*
75b28aff 131 * Bitmasks to apply to head and tail offsets (constant, equals
1e84b97b
SB
132 * ring_entries - 1)
133 */
75b28aff
HV
134 u32 sq_ring_mask, cq_ring_mask;
135 /* Ring sizes (constant, power of 2) */
136 u32 sq_ring_entries, cq_ring_entries;
1e84b97b
SB
137 /*
138 * Number of invalid entries dropped by the kernel due to
139 * invalid index stored in array
140 *
141 * Written by the kernel, shouldn't be modified by the
142 * application (i.e. get number of "new events" by comparing to
143 * cached value).
144 *
145 * After a new SQ head value was read by the application this
146 * counter includes all submissions that were dropped reaching
147 * the new SQ head (and possibly more).
148 */
75b28aff 149 u32 sq_dropped;
1e84b97b 150 /*
0d9b5b3a 151 * Runtime SQ flags
1e84b97b
SB
152 *
153 * Written by the kernel, shouldn't be modified by the
154 * application.
155 *
156 * The application needs a full memory barrier before checking
157 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
158 */
75b28aff 159 u32 sq_flags;
0d9b5b3a
SG
160 /*
161 * Runtime CQ flags
162 *
163 * Written by the application, shouldn't be modified by the
164 * kernel.
165 */
166 u32 cq_flags;
1e84b97b
SB
167 /*
168 * Number of completion events lost because the queue was full;
169 * this should be avoided by the application by making sure
0b4295b5 170 * there are not more requests pending than there is space in
1e84b97b
SB
171 * the completion queue.
172 *
173 * Written by the kernel, shouldn't be modified by the
174 * application (i.e. get number of "new events" by comparing to
175 * cached value).
176 *
177 * As completion events come in out of order this counter is not
178 * ordered with any other data.
179 */
75b28aff 180 u32 cq_overflow;
1e84b97b
SB
181 /*
182 * Ring buffer of completion events.
183 *
184 * The kernel writes completion events fresh every time they are
185 * produced, so the application is allowed to modify pending
186 * entries.
187 */
75b28aff 188 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
2b188cc1
JA
189};
190
45d189c6
PB
191enum io_uring_cmd_flags {
192 IO_URING_F_NONBLOCK = 1,
889fca73 193 IO_URING_F_COMPLETE_DEFER = 2,
45d189c6
PB
194};
195
edafccee
JA
196struct io_mapped_ubuf {
197 u64 ubuf;
198 size_t len;
199 struct bio_vec *bvec;
200 unsigned int nr_bvecs;
de293938 201 unsigned long acct_pages;
edafccee
JA
202};
203
50238531
BM
204struct io_ring_ctx;
205
269bbe5f
BM
206struct io_rsrc_put {
207 struct list_head list;
50238531
BM
208 union {
209 void *rsrc;
210 struct file *file;
211 };
269bbe5f
BM
212};
213
214struct fixed_rsrc_table {
65e19f54 215 struct file **files;
31b51510
JA
216};
217
269bbe5f 218struct fixed_rsrc_ref_node {
05589553
XW
219 struct percpu_ref refs;
220 struct list_head node;
269bbe5f
BM
221 struct list_head rsrc_list;
222 struct fixed_rsrc_data *rsrc_data;
50238531
BM
223 void (*rsrc_put)(struct io_ring_ctx *ctx,
224 struct io_rsrc_put *prsrc);
4a38aed2 225 struct llist_node llist;
e297822b 226 bool done;
05589553
XW
227};
228
269bbe5f
BM
229struct fixed_rsrc_data {
230 struct fixed_rsrc_table *table;
05f3fb3c
JA
231 struct io_ring_ctx *ctx;
232
269bbe5f 233 struct fixed_rsrc_ref_node *node;
05f3fb3c 234 struct percpu_ref refs;
05f3fb3c 235 struct completion done;
8bad28d8 236 bool quiesce;
05f3fb3c
JA
237};
238
5a2e745d
JA
239struct io_buffer {
240 struct list_head list;
241 __u64 addr;
242 __s32 len;
243 __u16 bid;
244};
245
21b55dbc
SG
246struct io_restriction {
247 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
248 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
249 u8 sqe_flags_allowed;
250 u8 sqe_flags_required;
7e84e1c7 251 bool registered;
21b55dbc
SG
252};
253
37d1e2e3
JA
254enum {
255 IO_SQ_THREAD_SHOULD_STOP = 0,
256 IO_SQ_THREAD_SHOULD_PARK,
257};
258
534ca6d6
JA
259struct io_sq_data {
260 refcount_t refs;
69fb2131
JA
261 struct mutex lock;
262
263 /* ctx's that are using this sqd */
264 struct list_head ctx_list;
265 struct list_head ctx_new_list;
266 struct mutex ctx_lock;
267
534ca6d6
JA
268 struct task_struct *thread;
269 struct wait_queue_head wait;
08369246
XW
270
271 unsigned sq_thread_idle;
37d1e2e3
JA
272 int sq_cpu;
273 pid_t task_pid;
274
275 unsigned long state;
276 struct completion startup;
277 struct completion completion;
278 struct completion exited;
534ca6d6
JA
279};
280
258b29a9 281#define IO_IOPOLL_BATCH 8
6dd0be1e 282#define IO_COMPL_BATCH 32
6ff119a6 283#define IO_REQ_CACHE_SIZE 32
bf019da7 284#define IO_REQ_ALLOC_BATCH 8
258b29a9
PB
285
286struct io_comp_state {
6dd0be1e 287 struct io_kiocb *reqs[IO_COMPL_BATCH];
1b4c351f 288 unsigned int nr;
c7dae4ba
JA
289 unsigned int locked_free_nr;
290 /* inline/task_work completion list, under ->uring_lock */
1b4c351f 291 struct list_head free_list;
c7dae4ba
JA
292 /* IRQ completion list, under ->completion_lock */
293 struct list_head locked_free_list;
258b29a9
PB
294};
295
a1ab7b35
PB
296struct io_submit_link {
297 struct io_kiocb *head;
298 struct io_kiocb *last;
299};
300
258b29a9
PB
301struct io_submit_state {
302 struct blk_plug plug;
a1ab7b35 303 struct io_submit_link link;
258b29a9
PB
304
305 /*
306 * io_kiocb alloc cache
307 */
bf019da7 308 void *reqs[IO_REQ_CACHE_SIZE];
258b29a9
PB
309 unsigned int free_reqs;
310
311 bool plug_started;
312
313 /*
314 * Batch completion logic
315 */
316 struct io_comp_state comp;
317
318 /*
319 * File reference cache
320 */
321 struct file *file;
322 unsigned int fd;
323 unsigned int file_refs;
324 unsigned int ios_left;
325};
326
2b188cc1
JA
327struct io_ring_ctx {
328 struct {
329 struct percpu_ref refs;
330 } ____cacheline_aligned_in_smp;
331
332 struct {
333 unsigned int flags;
e1d85334 334 unsigned int compat: 1;
e1d85334
RD
335 unsigned int cq_overflow_flushed: 1;
336 unsigned int drain_next: 1;
337 unsigned int eventfd_async: 1;
21b55dbc 338 unsigned int restricted: 1;
5f3f26f9 339 unsigned int sqo_exec: 1;
2b188cc1 340
75b28aff
HV
341 /*
342 * Ring buffer of indices into array of io_uring_sqe, which is
343 * mmapped by the application using the IORING_OFF_SQES offset.
344 *
345 * This indirection could e.g. be used to assign fixed
346 * io_uring_sqe entries to operations and only submit them to
347 * the queue when needed.
348 *
349 * The kernel modifies neither the indices array nor the entries
350 * array.
351 */
352 u32 *sq_array;
2b188cc1
JA
353 unsigned cached_sq_head;
354 unsigned sq_entries;
355 unsigned sq_mask;
6c271ce2 356 unsigned sq_thread_idle;
498ccd9e 357 unsigned cached_sq_dropped;
2c3bac6d 358 unsigned cached_cq_overflow;
ad3eb2c8 359 unsigned long sq_check_overflow;
de0617e4 360
e941894e
JA
361 /* hashed buffered write serialization */
362 struct io_wq_hash *hash_map;
363
de0617e4 364 struct list_head defer_list;
5262f567 365 struct list_head timeout_list;
1d7bb1d5 366 struct list_head cq_overflow_list;
fcb323cc 367
ad3eb2c8 368 struct io_uring_sqe *sq_sqes;
2b188cc1
JA
369 } ____cacheline_aligned_in_smp;
370
3c1a2ead
JA
371 struct {
372 struct mutex uring_lock;
373 wait_queue_head_t wait;
374 } ____cacheline_aligned_in_smp;
375
376 struct io_submit_state submit_state;
377
206aefde
JA
378 struct io_rings *rings;
379
2aede0e4
JA
380 /* Only used for accounting purposes */
381 struct mm_struct *mm_account;
382
534ca6d6
JA
383 struct io_sq_data *sq_data; /* if using sq thread polling */
384
90554200 385 struct wait_queue_head sqo_sq_wait;
69fb2131 386 struct list_head sqd_list;
75b28aff 387
6b06314c
JA
388 /*
389 * If used, fixed file set. Writers must ensure that ->refs is dead,
390 * readers must ensure that ->refs is alive as long as the file* is
391 * used. Only updated through io_uring_register(2).
392 */
269bbe5f 393 struct fixed_rsrc_data *file_data;
6b06314c
JA
394 unsigned nr_user_files;
395
edafccee
JA
396 /* if used, fixed mapped user buffers */
397 unsigned nr_user_bufs;
398 struct io_mapped_ubuf *user_bufs;
399
2b188cc1
JA
400 struct user_struct *user;
401
0f158b4c
JA
402 struct completion ref_comp;
403 struct completion sq_thread_comp;
206aefde
JA
404
405#if defined(CONFIG_UNIX)
406 struct socket *ring_sock;
407#endif
408
5a2e745d
JA
409 struct idr io_buffer_idr;
410
071698e1
JA
411 struct idr personality_idr;
412
206aefde
JA
413 struct {
414 unsigned cached_cq_tail;
415 unsigned cq_entries;
416 unsigned cq_mask;
417 atomic_t cq_timeouts;
f010505b 418 unsigned cq_last_tm_flush;
ad3eb2c8 419 unsigned long cq_check_overflow;
206aefde
JA
420 struct wait_queue_head cq_wait;
421 struct fasync_struct *cq_fasync;
422 struct eventfd_ctx *cq_ev_fd;
423 } ____cacheline_aligned_in_smp;
2b188cc1 424
2b188cc1
JA
425 struct {
426 spinlock_t completion_lock;
e94f141b 427
def596e9 428 /*
540e32a0 429 * ->iopoll_list is protected by the ctx->uring_lock for
def596e9
JA
430 * io_uring instances that don't use IORING_SETUP_SQPOLL.
431 * For SQPOLL, only the single threaded io_sq_thread() will
432 * manipulate the list, hence no extra locking is needed there.
433 */
540e32a0 434 struct list_head iopoll_list;
78076bb6
JA
435 struct hlist_head *cancel_hash;
436 unsigned cancel_hash_bits;
e94f141b 437 bool poll_multi_file;
31b51510 438
fcb323cc
JA
439 spinlock_t inflight_lock;
440 struct list_head inflight_list;
2b188cc1 441 } ____cacheline_aligned_in_smp;
85faa7b8 442
269bbe5f
BM
443 struct delayed_work rsrc_put_work;
444 struct llist_head rsrc_put_llist;
d67d2263
BM
445 struct list_head rsrc_ref_list;
446 spinlock_t rsrc_ref_lock;
4a38aed2 447
21b55dbc 448 struct io_restriction restrictions;
3c1a2ead 449
7c25c0d1
JA
450 /* exit task_work */
451 struct callback_head *exit_task_work;
452
e941894e
JA
453 struct wait_queue_head hash_wait;
454
3c1a2ead
JA
455 /* Keep this last, we don't need it for the fast path */
456 struct work_struct exit_work;
2b188cc1
JA
457};
458
09bb8394
JA
459/*
460 * First field must be the file pointer in all the
461 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
462 */
221c5eb2
JA
463struct io_poll_iocb {
464 struct file *file;
018043be 465 struct wait_queue_head *head;
221c5eb2 466 __poll_t events;
8c838788 467 bool done;
221c5eb2 468 bool canceled;
392edb45 469 struct wait_queue_entry wait;
221c5eb2
JA
470};
471
018043be
PB
472struct io_poll_remove {
473 struct file *file;
474 u64 addr;
475};
476
b5dba59e
JA
477struct io_close {
478 struct file *file;
b5dba59e
JA
479 int fd;
480};
481
ad8a48ac
JA
482struct io_timeout_data {
483 struct io_kiocb *req;
484 struct hrtimer timer;
485 struct timespec64 ts;
486 enum hrtimer_mode mode;
487};
488
8ed8d3c3
JA
489struct io_accept {
490 struct file *file;
491 struct sockaddr __user *addr;
492 int __user *addr_len;
493 int flags;
09952e3e 494 unsigned long nofile;
8ed8d3c3
JA
495};
496
497struct io_sync {
498 struct file *file;
499 loff_t len;
500 loff_t off;
501 int flags;
d63d1b5e 502 int mode;
8ed8d3c3
JA
503};
504
fbf23849
JA
505struct io_cancel {
506 struct file *file;
507 u64 addr;
508};
509
b29472ee
JA
510struct io_timeout {
511 struct file *file;
bfe68a22
PB
512 u32 off;
513 u32 target_seq;
135fcde8 514 struct list_head list;
90cd7e42
PB
515 /* head of the link, used by linked timeouts only */
516 struct io_kiocb *head;
b29472ee
JA
517};
518
0bdf7a2d
PB
519struct io_timeout_rem {
520 struct file *file;
521 u64 addr;
9c8e11b3
PB
522
523 /* timeout update */
524 struct timespec64 ts;
525 u32 flags;
0bdf7a2d
PB
526};
527
9adbd45d
JA
528struct io_rw {
529 /* NOTE: kiocb has the file as the first member, so don't do it here */
530 struct kiocb kiocb;
531 u64 addr;
532 u64 len;
533};
534
3fbb51c1
JA
535struct io_connect {
536 struct file *file;
537 struct sockaddr __user *addr;
538 int addr_len;
539};
540
e47293fd
JA
541struct io_sr_msg {
542 struct file *file;
fddaface 543 union {
270a5940 544 struct user_msghdr __user *umsg;
fddaface
JA
545 void __user *buf;
546 };
e47293fd 547 int msg_flags;
bcda7baa 548 int bgid;
fddaface 549 size_t len;
bcda7baa 550 struct io_buffer *kbuf;
e47293fd
JA
551};
552
15b71abe
JA
553struct io_open {
554 struct file *file;
555 int dfd;
15b71abe 556 struct filename *filename;
c12cedf2 557 struct open_how how;
4022e7af 558 unsigned long nofile;
15b71abe
JA
559};
560
269bbe5f 561struct io_rsrc_update {
05f3fb3c
JA
562 struct file *file;
563 u64 arg;
564 u32 nr_args;
565 u32 offset;
566};
567
4840e418
JA
568struct io_fadvise {
569 struct file *file;
570 u64 offset;
571 u32 len;
572 u32 advice;
573};
574
c1ca757b
JA
575struct io_madvise {
576 struct file *file;
577 u64 addr;
578 u32 len;
579 u32 advice;
580};
581
3e4827b0
JA
582struct io_epoll {
583 struct file *file;
584 int epfd;
585 int op;
586 int fd;
587 struct epoll_event event;
e47293fd
JA
588};
589
7d67af2c
PB
590struct io_splice {
591 struct file *file_out;
592 struct file *file_in;
593 loff_t off_out;
594 loff_t off_in;
595 u64 len;
596 unsigned int flags;
597};
598
ddf0322d
JA
599struct io_provide_buf {
600 struct file *file;
601 __u64 addr;
602 __s32 len;
603 __u32 bgid;
604 __u16 nbufs;
605 __u16 bid;
606};
607
1d9e1288
BM
608struct io_statx {
609 struct file *file;
610 int dfd;
611 unsigned int mask;
612 unsigned int flags;
e62753e4 613 const char __user *filename;
1d9e1288
BM
614 struct statx __user *buffer;
615};
616
36f4fa68
JA
617struct io_shutdown {
618 struct file *file;
619 int how;
620};
621
80a261fd
JA
622struct io_rename {
623 struct file *file;
624 int old_dfd;
625 int new_dfd;
626 struct filename *oldpath;
627 struct filename *newpath;
628 int flags;
629};
630
14a1143b
JA
631struct io_unlink {
632 struct file *file;
633 int dfd;
634 int flags;
635 struct filename *filename;
636};
637
3ca405eb
PB
638struct io_completion {
639 struct file *file;
640 struct list_head list;
0f7e466b 641 int cflags;
3ca405eb
PB
642};
643
f499a021
JA
644struct io_async_connect {
645 struct sockaddr_storage address;
646};
647
03b1230c
JA
648struct io_async_msghdr {
649 struct iovec fast_iov[UIO_FASTIOV];
257e84a5
PB
650 /* points to an allocated iov, if NULL we use fast_iov instead */
651 struct iovec *free_iov;
03b1230c
JA
652 struct sockaddr __user *uaddr;
653 struct msghdr msg;
b537916c 654 struct sockaddr_storage addr;
03b1230c
JA
655};
656
f67676d1
JA
657struct io_async_rw {
658 struct iovec fast_iov[UIO_FASTIOV];
ff6165b2
JA
659 const struct iovec *free_iovec;
660 struct iov_iter iter;
227c0c96 661 size_t bytes_done;
bcf5a063 662 struct wait_page_queue wpq;
f67676d1
JA
663};
664
6b47ee6e
PB
665enum {
666 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
667 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
668 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
669 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
670 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
bcda7baa 671 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
6b47ee6e 672
6b47ee6e
PB
673 REQ_F_FAIL_LINK_BIT,
674 REQ_F_INFLIGHT_BIT,
675 REQ_F_CUR_POS_BIT,
676 REQ_F_NOWAIT_BIT,
6b47ee6e 677 REQ_F_LINK_TIMEOUT_BIT,
6b47ee6e 678 REQ_F_ISREG_BIT,
99bc4c38 679 REQ_F_NEED_CLEANUP_BIT,
d7718a9d 680 REQ_F_POLLED_BIT,
bcda7baa 681 REQ_F_BUFFER_SELECTED_BIT,
5b0bbee4 682 REQ_F_NO_FILE_TABLE_BIT,
900fad45 683 REQ_F_LTIMEOUT_ACTIVE_BIT,
e342c807 684 REQ_F_COMPLETE_INLINE_BIT,
84557871
JA
685
686 /* not a real bit, just to check we're not overflowing the space */
687 __REQ_F_LAST_BIT,
6b47ee6e
PB
688};
689
690enum {
691 /* ctx owns file */
692 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
693 /* drain existing IO first */
694 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
695 /* linked sqes */
696 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
697 /* doesn't sever on completion < 0 */
698 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
699 /* IOSQE_ASYNC */
700 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
bcda7baa
JA
701 /* IOSQE_BUFFER_SELECT */
702 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
6b47ee6e 703
6b47ee6e
PB
704 /* fail rest of links */
705 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
b05a1bcd 706 /* on inflight list, should be cancelled and waited on exit reliably */
6b47ee6e
PB
707 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
708 /* read/write uses file position */
709 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
710 /* must not punt to workers */
711 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
900fad45 712 /* has or had linked timeout */
6b47ee6e 713 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
6b47ee6e
PB
714 /* regular file */
715 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
99bc4c38
PB
716 /* needs cleanup */
717 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
d7718a9d
JA
718 /* already went through poll handler */
719 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
bcda7baa
JA
720 /* buffer already selected */
721 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
5b0bbee4
JA
722 /* doesn't need file table for this request */
723 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
900fad45
PB
724 /* linked timeout is active, i.e. prepared by link's head */
725 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
e342c807
PB
726 /* completion is deferred through io_comp_state */
727 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
d7718a9d
JA
728};
729
730struct async_poll {
731 struct io_poll_iocb poll;
807abcb0 732 struct io_poll_iocb *double_poll;
6b47ee6e
PB
733};
734
7cbf1722
JA
735struct io_task_work {
736 struct io_wq_work_node node;
737 task_work_func_t func;
738};
739
09bb8394
JA
740/*
741 * NOTE! Each of the iocb union members has the file pointer
742 * as the first entry in their struct definition. So you can
743 * access the file pointer through any of the sub-structs,
744 * or directly as just 'ki_filp' in this struct.
745 */
2b188cc1 746struct io_kiocb {
221c5eb2 747 union {
09bb8394 748 struct file *file;
9adbd45d 749 struct io_rw rw;
221c5eb2 750 struct io_poll_iocb poll;
018043be 751 struct io_poll_remove poll_remove;
8ed8d3c3
JA
752 struct io_accept accept;
753 struct io_sync sync;
fbf23849 754 struct io_cancel cancel;
b29472ee 755 struct io_timeout timeout;
0bdf7a2d 756 struct io_timeout_rem timeout_rem;
3fbb51c1 757 struct io_connect connect;
e47293fd 758 struct io_sr_msg sr_msg;
15b71abe 759 struct io_open open;
b5dba59e 760 struct io_close close;
269bbe5f 761 struct io_rsrc_update rsrc_update;
4840e418 762 struct io_fadvise fadvise;
c1ca757b 763 struct io_madvise madvise;
3e4827b0 764 struct io_epoll epoll;
7d67af2c 765 struct io_splice splice;
ddf0322d 766 struct io_provide_buf pbuf;
1d9e1288 767 struct io_statx statx;
36f4fa68 768 struct io_shutdown shutdown;
80a261fd 769 struct io_rename rename;
14a1143b 770 struct io_unlink unlink;
3ca405eb
PB
771 /* use only after cleaning per-op data, see io_clean_op() */
772 struct io_completion compl;
221c5eb2 773 };
2b188cc1 774
e8c2bc1f
JA
775 /* opcode allocated if it needs to store data for async defer */
776 void *async_data;
d625c6ee 777 u8 opcode;
65a6543d
XW
778 /* polled IO has completed */
779 u8 iopoll_completed;
2b188cc1 780
4f4eeba8 781 u16 buf_index;
9cf7c104 782 u32 result;
4f4eeba8 783
010e8e6b
PB
784 struct io_ring_ctx *ctx;
785 unsigned int flags;
786 refcount_t refs;
787 struct task_struct *task;
788 u64 user_data;
d7718a9d 789
f2f87370 790 struct io_kiocb *link;
269bbe5f 791 struct percpu_ref *fixed_rsrc_refs;
fcb323cc 792
d21ffe7e
PB
793 /*
794 * 1. used with ctx->iopoll_list with reads/writes
795 * 2. to track reqs with ->files (see io_op_def::file_table)
796 */
010e8e6b 797 struct list_head inflight_entry;
7cbf1722
JA
798 union {
799 struct io_task_work io_task_work;
800 struct callback_head task_work;
801 };
010e8e6b
PB
802 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
803 struct hlist_node hash_node;
804 struct async_poll *apoll;
805 struct io_wq_work work;
2b188cc1 806};
05589553 807
27dc8338
PB
808struct io_defer_entry {
809 struct list_head list;
810 struct io_kiocb *req;
9cf7c104 811 u32 seq;
2b188cc1
JA
812};
813
d3656344 814struct io_op_def {
d3656344
JA
815 /* needs req->file assigned */
816 unsigned needs_file : 1;
d3656344
JA
817 /* hash wq insertion if file is a regular file */
818 unsigned hash_reg_file : 1;
819 /* unbound wq insertion if file is a non-regular file */
820 unsigned unbound_nonreg_file : 1;
66f4af93
JA
821 /* opcode is not supported by this kernel */
822 unsigned not_supported : 1;
8a72758c
JA
823 /* set if opcode supports polled "wait" */
824 unsigned pollin : 1;
825 unsigned pollout : 1;
bcda7baa
JA
826 /* op supports buffer selection */
827 unsigned buffer_select : 1;
e8c2bc1f
JA
828 /* must always have async data allocated */
829 unsigned needs_async_data : 1;
27926b68
JA
830 /* should block plug */
831 unsigned plug : 1;
e8c2bc1f
JA
832 /* size of async data needed, if any */
833 unsigned short async_size;
d3656344
JA
834};
835
0918682b 836static const struct io_op_def io_op_defs[] = {
0463b6c5
PB
837 [IORING_OP_NOP] = {},
838 [IORING_OP_READV] = {
d3656344
JA
839 .needs_file = 1,
840 .unbound_nonreg_file = 1,
8a72758c 841 .pollin = 1,
4d954c25 842 .buffer_select = 1,
e8c2bc1f 843 .needs_async_data = 1,
27926b68 844 .plug = 1,
e8c2bc1f 845 .async_size = sizeof(struct io_async_rw),
d3656344 846 },
0463b6c5 847 [IORING_OP_WRITEV] = {
d3656344
JA
848 .needs_file = 1,
849 .hash_reg_file = 1,
850 .unbound_nonreg_file = 1,
8a72758c 851 .pollout = 1,
e8c2bc1f 852 .needs_async_data = 1,
27926b68 853 .plug = 1,
e8c2bc1f 854 .async_size = sizeof(struct io_async_rw),
d3656344 855 },
0463b6c5 856 [IORING_OP_FSYNC] = {
d3656344
JA
857 .needs_file = 1,
858 },
0463b6c5 859 [IORING_OP_READ_FIXED] = {
d3656344
JA
860 .needs_file = 1,
861 .unbound_nonreg_file = 1,
8a72758c 862 .pollin = 1,
27926b68 863 .plug = 1,
e8c2bc1f 864 .async_size = sizeof(struct io_async_rw),
d3656344 865 },
0463b6c5 866 [IORING_OP_WRITE_FIXED] = {
d3656344
JA
867 .needs_file = 1,
868 .hash_reg_file = 1,
869 .unbound_nonreg_file = 1,
8a72758c 870 .pollout = 1,
27926b68 871 .plug = 1,
e8c2bc1f 872 .async_size = sizeof(struct io_async_rw),
d3656344 873 },
0463b6c5 874 [IORING_OP_POLL_ADD] = {
d3656344
JA
875 .needs_file = 1,
876 .unbound_nonreg_file = 1,
877 },
0463b6c5
PB
878 [IORING_OP_POLL_REMOVE] = {},
879 [IORING_OP_SYNC_FILE_RANGE] = {
d3656344
JA
880 .needs_file = 1,
881 },
0463b6c5 882 [IORING_OP_SENDMSG] = {
d3656344
JA
883 .needs_file = 1,
884 .unbound_nonreg_file = 1,
8a72758c 885 .pollout = 1,
e8c2bc1f
JA
886 .needs_async_data = 1,
887 .async_size = sizeof(struct io_async_msghdr),
d3656344 888 },
0463b6c5 889 [IORING_OP_RECVMSG] = {
d3656344
JA
890 .needs_file = 1,
891 .unbound_nonreg_file = 1,
8a72758c 892 .pollin = 1,
52de1fe1 893 .buffer_select = 1,
e8c2bc1f
JA
894 .needs_async_data = 1,
895 .async_size = sizeof(struct io_async_msghdr),
d3656344 896 },
0463b6c5 897 [IORING_OP_TIMEOUT] = {
e8c2bc1f
JA
898 .needs_async_data = 1,
899 .async_size = sizeof(struct io_timeout_data),
d3656344 900 },
9c8e11b3
PB
901 [IORING_OP_TIMEOUT_REMOVE] = {
902 /* used by timeout updates' prep() */
9c8e11b3 903 },
0463b6c5 904 [IORING_OP_ACCEPT] = {
d3656344
JA
905 .needs_file = 1,
906 .unbound_nonreg_file = 1,
8a72758c 907 .pollin = 1,
d3656344 908 },
0463b6c5
PB
909 [IORING_OP_ASYNC_CANCEL] = {},
910 [IORING_OP_LINK_TIMEOUT] = {
e8c2bc1f
JA
911 .needs_async_data = 1,
912 .async_size = sizeof(struct io_timeout_data),
d3656344 913 },
0463b6c5 914 [IORING_OP_CONNECT] = {
d3656344
JA
915 .needs_file = 1,
916 .unbound_nonreg_file = 1,
8a72758c 917 .pollout = 1,
e8c2bc1f
JA
918 .needs_async_data = 1,
919 .async_size = sizeof(struct io_async_connect),
d3656344 920 },
0463b6c5 921 [IORING_OP_FALLOCATE] = {
d3656344 922 .needs_file = 1,
d3656344 923 },
44526bed
JA
924 [IORING_OP_OPENAT] = {},
925 [IORING_OP_CLOSE] = {},
926 [IORING_OP_FILES_UPDATE] = {},
927 [IORING_OP_STATX] = {},
0463b6c5 928 [IORING_OP_READ] = {
3a6820f2
JA
929 .needs_file = 1,
930 .unbound_nonreg_file = 1,
8a72758c 931 .pollin = 1,
bcda7baa 932 .buffer_select = 1,
27926b68 933 .plug = 1,
e8c2bc1f 934 .async_size = sizeof(struct io_async_rw),
3a6820f2 935 },
0463b6c5 936 [IORING_OP_WRITE] = {
3a6820f2
JA
937 .needs_file = 1,
938 .unbound_nonreg_file = 1,
8a72758c 939 .pollout = 1,
27926b68 940 .plug = 1,
e8c2bc1f 941 .async_size = sizeof(struct io_async_rw),
3a6820f2 942 },
0463b6c5 943 [IORING_OP_FADVISE] = {
4840e418 944 .needs_file = 1,
c1ca757b 945 },
44526bed 946 [IORING_OP_MADVISE] = {},
0463b6c5 947 [IORING_OP_SEND] = {
fddaface
JA
948 .needs_file = 1,
949 .unbound_nonreg_file = 1,
8a72758c 950 .pollout = 1,
fddaface 951 },
0463b6c5 952 [IORING_OP_RECV] = {
fddaface
JA
953 .needs_file = 1,
954 .unbound_nonreg_file = 1,
8a72758c 955 .pollin = 1,
bcda7baa 956 .buffer_select = 1,
fddaface 957 },
0463b6c5 958 [IORING_OP_OPENAT2] = {
cebdb986 959 },
3e4827b0
JA
960 [IORING_OP_EPOLL_CTL] = {
961 .unbound_nonreg_file = 1,
3e4827b0 962 },
7d67af2c
PB
963 [IORING_OP_SPLICE] = {
964 .needs_file = 1,
965 .hash_reg_file = 1,
966 .unbound_nonreg_file = 1,
ddf0322d
JA
967 },
968 [IORING_OP_PROVIDE_BUFFERS] = {},
067524e9 969 [IORING_OP_REMOVE_BUFFERS] = {},
f2a8d5c7
PB
970 [IORING_OP_TEE] = {
971 .needs_file = 1,
972 .hash_reg_file = 1,
973 .unbound_nonreg_file = 1,
974 },
36f4fa68
JA
975 [IORING_OP_SHUTDOWN] = {
976 .needs_file = 1,
977 },
44526bed
JA
978 [IORING_OP_RENAMEAT] = {},
979 [IORING_OP_UNLINKAT] = {},
d3656344
JA
980};
981
9936c7c2
PB
982static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
983 struct task_struct *task,
984 struct files_struct *files);
37d1e2e3 985static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx);
269bbe5f 986static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node);
bc9744cd 987static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
1ffc5422 988 struct io_ring_ctx *ctx);
f2303b1f 989static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
1ffc5422 990
23faba36 991static bool io_rw_reissue(struct io_kiocb *req);
78e19bbe 992static void io_cqring_fill_event(struct io_kiocb *req, long res);
ec9c02ad 993static void io_put_req(struct io_kiocb *req);
216578e5 994static void io_put_req_deferred(struct io_kiocb *req, int nr);
c40f6379 995static void io_double_put_req(struct io_kiocb *req);
c7dae4ba
JA
996static void io_dismantle_req(struct io_kiocb *req);
997static void io_put_task(struct task_struct *task, int nr);
998static void io_queue_next(struct io_kiocb *req);
94ae5e77 999static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
7271ef3a 1000static void __io_queue_linked_timeout(struct io_kiocb *req);
94ae5e77 1001static void io_queue_linked_timeout(struct io_kiocb *req);
05f3fb3c 1002static int __io_sqe_files_update(struct io_ring_ctx *ctx,
269bbe5f 1003 struct io_uring_rsrc_update *ip,
05f3fb3c 1004 unsigned nr_args);
3ca405eb 1005static void __io_clean_op(struct io_kiocb *req);
8371adf5
PB
1006static struct file *io_file_get(struct io_submit_state *state,
1007 struct io_kiocb *req, int fd, bool fixed);
c5eef2b9 1008static void __io_queue_sqe(struct io_kiocb *req);
269bbe5f 1009static void io_rsrc_put_work(struct work_struct *work);
de0617e4 1010
847595de
PB
1011static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
1012 struct iov_iter *iter, bool needs_lock);
ff6165b2
JA
1013static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1014 const struct iovec *fast_iov,
227c0c96 1015 struct iov_iter *iter, bool force);
907d1df3 1016static void io_req_task_queue(struct io_kiocb *req);
65453d1e
JA
1017static void io_submit_flush_completions(struct io_comp_state *cs,
1018 struct io_ring_ctx *ctx);
de0617e4 1019
2b188cc1
JA
1020static struct kmem_cache *req_cachep;
1021
0918682b 1022static const struct file_operations io_uring_fops;
2b188cc1
JA
1023
1024struct sock *io_uring_get_socket(struct file *file)
1025{
1026#if defined(CONFIG_UNIX)
1027 if (file->f_op == &io_uring_fops) {
1028 struct io_ring_ctx *ctx = file->private_data;
1029
1030 return ctx->ring_sock->sk;
1031 }
1032#endif
1033 return NULL;
1034}
1035EXPORT_SYMBOL(io_uring_get_socket);
1036
f2f87370
PB
1037#define io_for_each_link(pos, head) \
1038 for (pos = (head); pos; pos = pos->link)
1039
3ca405eb
PB
1040static inline void io_clean_op(struct io_kiocb *req)
1041{
9d5c8190 1042 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
3ca405eb
PB
1043 __io_clean_op(req);
1044}
1045
36f72fe2
PB
1046static inline void io_set_resource_node(struct io_kiocb *req)
1047{
1048 struct io_ring_ctx *ctx = req->ctx;
1049
269bbe5f
BM
1050 if (!req->fixed_rsrc_refs) {
1051 req->fixed_rsrc_refs = &ctx->file_data->node->refs;
1052 percpu_ref_get(req->fixed_rsrc_refs);
36f72fe2
PB
1053 }
1054}
1055
08d23634
PB
1056static bool io_match_task(struct io_kiocb *head,
1057 struct task_struct *task,
1058 struct files_struct *files)
1059{
1060 struct io_kiocb *req;
1061
84965ff8
JA
1062 if (task && head->task != task) {
1063 /* in terms of cancelation, always match if req task is dead */
1064 if (head->task->flags & PF_EXITING)
1065 return true;
08d23634 1066 return false;
84965ff8 1067 }
08d23634
PB
1068 if (!files)
1069 return true;
1070
1071 io_for_each_link(req, head) {
b05a1bcd 1072 if (req->flags & REQ_F_INFLIGHT)
02a13674 1073 return true;
4379bf8b 1074 if (req->task->files == files)
08d23634
PB
1075 return true;
1076 }
1077 return false;
1078}
1079
c40f6379
JA
1080static inline void req_set_fail_links(struct io_kiocb *req)
1081{
1082 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1083 req->flags |= REQ_F_FAIL_LINK;
1084}
4a38aed2 1085
2b188cc1
JA
1086static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1087{
1088 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1089
0f158b4c 1090 complete(&ctx->ref_comp);
2b188cc1
JA
1091}
1092
8eb7e2d0
PB
1093static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1094{
1095 return !req->timeout.off;
1096}
1097
2b188cc1
JA
1098static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1099{
1100 struct io_ring_ctx *ctx;
78076bb6 1101 int hash_bits;
2b188cc1
JA
1102
1103 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1104 if (!ctx)
1105 return NULL;
1106
78076bb6
JA
1107 /*
1108 * Use 5 bits less than the max cq entries, that should give us around
1109 * 32 entries per hash list if totally full and uniformly spread.
1110 */
1111 hash_bits = ilog2(p->cq_entries);
1112 hash_bits -= 5;
1113 if (hash_bits <= 0)
1114 hash_bits = 1;
1115 ctx->cancel_hash_bits = hash_bits;
1116 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1117 GFP_KERNEL);
1118 if (!ctx->cancel_hash)
1119 goto err;
1120 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1121
21482896 1122 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
206aefde
JA
1123 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1124 goto err;
2b188cc1
JA
1125
1126 ctx->flags = p->flags;
90554200 1127 init_waitqueue_head(&ctx->sqo_sq_wait);
69fb2131 1128 INIT_LIST_HEAD(&ctx->sqd_list);
2b188cc1 1129 init_waitqueue_head(&ctx->cq_wait);
1d7bb1d5 1130 INIT_LIST_HEAD(&ctx->cq_overflow_list);
0f158b4c
JA
1131 init_completion(&ctx->ref_comp);
1132 init_completion(&ctx->sq_thread_comp);
5a2e745d 1133 idr_init(&ctx->io_buffer_idr);
071698e1 1134 idr_init(&ctx->personality_idr);
2b188cc1
JA
1135 mutex_init(&ctx->uring_lock);
1136 init_waitqueue_head(&ctx->wait);
1137 spin_lock_init(&ctx->completion_lock);
540e32a0 1138 INIT_LIST_HEAD(&ctx->iopoll_list);
de0617e4 1139 INIT_LIST_HEAD(&ctx->defer_list);
5262f567 1140 INIT_LIST_HEAD(&ctx->timeout_list);
fcb323cc
JA
1141 spin_lock_init(&ctx->inflight_lock);
1142 INIT_LIST_HEAD(&ctx->inflight_list);
d67d2263
BM
1143 spin_lock_init(&ctx->rsrc_ref_lock);
1144 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
269bbe5f
BM
1145 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1146 init_llist_head(&ctx->rsrc_put_llist);
1b4c351f 1147 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
c7dae4ba 1148 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
2b188cc1 1149 return ctx;
206aefde 1150err:
78076bb6 1151 kfree(ctx->cancel_hash);
206aefde
JA
1152 kfree(ctx);
1153 return NULL;
2b188cc1
JA
1154}
1155
9cf7c104 1156static bool req_need_defer(struct io_kiocb *req, u32 seq)
7adf4eaf 1157{
2bc9930e
JA
1158 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1159 struct io_ring_ctx *ctx = req->ctx;
a197f664 1160
9cf7c104 1161 return seq != ctx->cached_cq_tail
2c3bac6d 1162 + READ_ONCE(ctx->cached_cq_overflow);
2bc9930e 1163 }
de0617e4 1164
9d858b21 1165 return false;
de0617e4
JA
1166}
1167
ce3d5aae
PB
1168static void io_req_track_inflight(struct io_kiocb *req)
1169{
1170 struct io_ring_ctx *ctx = req->ctx;
1171
1172 if (!(req->flags & REQ_F_INFLIGHT)) {
ce3d5aae
PB
1173 req->flags |= REQ_F_INFLIGHT;
1174
1175 spin_lock_irq(&ctx->inflight_lock);
1176 list_add(&req->inflight_entry, &ctx->inflight_list);
1177 spin_unlock_irq(&ctx->inflight_lock);
1178 }
1179}
1180
1e6fa521
JA
1181static void io_prep_async_work(struct io_kiocb *req)
1182{
1183 const struct io_op_def *def = &io_op_defs[req->opcode];
1e6fa521
JA
1184 struct io_ring_ctx *ctx = req->ctx;
1185
feaadc4f
PB
1186 if (req->flags & REQ_F_FORCE_ASYNC)
1187 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1188
1e6fa521
JA
1189 if (req->flags & REQ_F_ISREG) {
1190 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1191 io_wq_hash_work(&req->work, file_inode(req->file));
1192 } else {
1193 if (def->unbound_nonreg_file)
1194 req->work.flags |= IO_WQ_WORK_UNBOUND;
1195 }
561fb04a 1196}
cccf0ee8 1197
cbdcb435 1198static void io_prep_async_link(struct io_kiocb *req)
561fb04a 1199{
cbdcb435 1200 struct io_kiocb *cur;
54a91f3b 1201
f2f87370
PB
1202 io_for_each_link(cur, req)
1203 io_prep_async_work(cur);
561fb04a
JA
1204}
1205
ebf93667 1206static void io_queue_async_work(struct io_kiocb *req)
561fb04a 1207{
a197f664 1208 struct io_ring_ctx *ctx = req->ctx;
cbdcb435 1209 struct io_kiocb *link = io_prep_linked_timeout(req);
5aa75ed5 1210 struct io_uring_task *tctx = req->task->io_uring;
561fb04a 1211
3bfe6106
JA
1212 BUG_ON(!tctx);
1213 BUG_ON(!tctx->io_wq);
561fb04a 1214
8766dd51
PB
1215 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1216 &req->work, req->flags);
cbdcb435
PB
1217 /* init ->work of the whole link before punting */
1218 io_prep_async_link(req);
ebf93667 1219 io_wq_enqueue(tctx->io_wq, &req->work);
7271ef3a
JA
1220 if (link)
1221 io_queue_linked_timeout(link);
cbdcb435
PB
1222}
1223
5262f567
JA
1224static void io_kill_timeout(struct io_kiocb *req)
1225{
e8c2bc1f 1226 struct io_timeout_data *io = req->async_data;
5262f567
JA
1227 int ret;
1228
e8c2bc1f 1229 ret = hrtimer_try_to_cancel(&io->timer);
5262f567 1230 if (ret != -1) {
01cec8c1
PB
1231 atomic_set(&req->ctx->cq_timeouts,
1232 atomic_read(&req->ctx->cq_timeouts) + 1);
135fcde8 1233 list_del_init(&req->timeout.list);
78e19bbe 1234 io_cqring_fill_event(req, 0);
216578e5 1235 io_put_req_deferred(req, 1);
5262f567
JA
1236 }
1237}
1238
76e1b642
JA
1239/*
1240 * Returns true if we found and killed one or more timeouts
1241 */
6b81928d
PB
1242static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1243 struct files_struct *files)
5262f567
JA
1244{
1245 struct io_kiocb *req, *tmp;
76e1b642 1246 int canceled = 0;
5262f567
JA
1247
1248 spin_lock_irq(&ctx->completion_lock);
f3606e3a 1249 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
6b81928d 1250 if (io_match_task(req, tsk, files)) {
f3606e3a 1251 io_kill_timeout(req);
76e1b642
JA
1252 canceled++;
1253 }
f3606e3a 1254 }
5262f567 1255 spin_unlock_irq(&ctx->completion_lock);
76e1b642 1256 return canceled != 0;
5262f567
JA
1257}
1258
04518945 1259static void __io_queue_deferred(struct io_ring_ctx *ctx)
de0617e4 1260{
04518945 1261 do {
27dc8338
PB
1262 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1263 struct io_defer_entry, list);
de0617e4 1264
9cf7c104 1265 if (req_need_defer(de->req, de->seq))
04518945 1266 break;
27dc8338 1267 list_del_init(&de->list);
907d1df3 1268 io_req_task_queue(de->req);
27dc8338 1269 kfree(de);
04518945
PB
1270 } while (!list_empty(&ctx->defer_list));
1271}
1272
360428f8 1273static void io_flush_timeouts(struct io_ring_ctx *ctx)
de0617e4 1274{
f010505b
MDG
1275 u32 seq;
1276
1277 if (list_empty(&ctx->timeout_list))
1278 return;
1279
1280 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1281
1282 do {
1283 u32 events_needed, events_got;
360428f8 1284 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
135fcde8 1285 struct io_kiocb, timeout.list);
de0617e4 1286
8eb7e2d0 1287 if (io_is_timeout_noseq(req))
360428f8 1288 break;
f010505b
MDG
1289
1290 /*
1291 * Since seq can easily wrap around over time, subtract
1292 * the last seq at which timeouts were flushed before comparing.
1293 * Assuming not more than 2^31-1 events have happened since,
1294 * these subtractions won't have wrapped, so we can check if
1295 * target is in [last_seq, current_seq] by comparing the two.
1296 */
1297 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1298 events_got = seq - ctx->cq_last_tm_flush;
1299 if (events_got < events_needed)
360428f8 1300 break;
bfe68a22 1301
135fcde8 1302 list_del_init(&req->timeout.list);
5262f567 1303 io_kill_timeout(req);
f010505b
MDG
1304 } while (!list_empty(&ctx->timeout_list));
1305
1306 ctx->cq_last_tm_flush = seq;
360428f8 1307}
5262f567 1308
360428f8
PB
1309static void io_commit_cqring(struct io_ring_ctx *ctx)
1310{
1311 io_flush_timeouts(ctx);
ec30e04b
PB
1312
1313 /* order cqe stores with ring update */
1314 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
de0617e4 1315
04518945
PB
1316 if (unlikely(!list_empty(&ctx->defer_list)))
1317 __io_queue_deferred(ctx);
de0617e4
JA
1318}
1319
90554200
JA
1320static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1321{
1322 struct io_rings *r = ctx->rings;
1323
1324 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1325}
1326
888aae2e
PB
1327static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1328{
1329 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1330}
1331
2b188cc1
JA
1332static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1333{
75b28aff 1334 struct io_rings *rings = ctx->rings;
2b188cc1
JA
1335 unsigned tail;
1336
115e12e5
SB
1337 /*
1338 * writes to the cq entry need to come after reading head; the
1339 * control dependency is enough as we're using WRITE_ONCE to
1340 * fill the cq entry
1341 */
888aae2e 1342 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
2b188cc1
JA
1343 return NULL;
1344
888aae2e 1345 tail = ctx->cached_cq_tail++;
75b28aff 1346 return &rings->cqes[tail & ctx->cq_mask];
2b188cc1
JA
1347}
1348
f2842ab5
JA
1349static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1350{
f0b493e6
JA
1351 if (!ctx->cq_ev_fd)
1352 return false;
7e55a19c
SG
1353 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1354 return false;
f2842ab5
JA
1355 if (!ctx->eventfd_async)
1356 return true;
b41e9852 1357 return io_wq_current_is_worker();
f2842ab5
JA
1358}
1359
b41e9852 1360static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1d7bb1d5 1361{
b1445e59
PB
1362 /* see waitqueue_active() comment */
1363 smp_mb();
1364
1d7bb1d5
JA
1365 if (waitqueue_active(&ctx->wait))
1366 wake_up(&ctx->wait);
534ca6d6
JA
1367 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1368 wake_up(&ctx->sq_data->wait);
b41e9852 1369 if (io_should_trigger_evfd(ctx))
1d7bb1d5 1370 eventfd_signal(ctx->cq_ev_fd, 1);
b1445e59 1371 if (waitqueue_active(&ctx->cq_wait)) {
4aa84f2f
PB
1372 wake_up_interruptible(&ctx->cq_wait);
1373 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1374 }
1d7bb1d5
JA
1375}
1376
80c18e4a
PB
1377static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1378{
b1445e59
PB
1379 /* see waitqueue_active() comment */
1380 smp_mb();
1381
80c18e4a
PB
1382 if (ctx->flags & IORING_SETUP_SQPOLL) {
1383 if (waitqueue_active(&ctx->wait))
1384 wake_up(&ctx->wait);
1385 }
1386 if (io_should_trigger_evfd(ctx))
1387 eventfd_signal(ctx->cq_ev_fd, 1);
b1445e59 1388 if (waitqueue_active(&ctx->cq_wait)) {
4aa84f2f
PB
1389 wake_up_interruptible(&ctx->cq_wait);
1390 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1391 }
80c18e4a
PB
1392}
1393
c4a2ed72 1394/* Returns true if there are no backlogged entries after the flush */
6c503150
PB
1395static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1396 struct task_struct *tsk,
1397 struct files_struct *files)
1d7bb1d5
JA
1398{
1399 struct io_rings *rings = ctx->rings;
e6c8aa9a 1400 struct io_kiocb *req, *tmp;
1d7bb1d5 1401 struct io_uring_cqe *cqe;
1d7bb1d5 1402 unsigned long flags;
b18032bb 1403 bool all_flushed, posted;
1d7bb1d5
JA
1404 LIST_HEAD(list);
1405
e23de15f
PB
1406 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1407 return false;
1d7bb1d5 1408
b18032bb 1409 posted = false;
1d7bb1d5 1410 spin_lock_irqsave(&ctx->completion_lock, flags);
e6c8aa9a 1411 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
08d23634 1412 if (!io_match_task(req, tsk, files))
e6c8aa9a
JA
1413 continue;
1414
1d7bb1d5
JA
1415 cqe = io_get_cqring(ctx);
1416 if (!cqe && !force)
1417 break;
1418
40d8ddd4 1419 list_move(&req->compl.list, &list);
1d7bb1d5
JA
1420 if (cqe) {
1421 WRITE_ONCE(cqe->user_data, req->user_data);
1422 WRITE_ONCE(cqe->res, req->result);
0f7e466b 1423 WRITE_ONCE(cqe->flags, req->compl.cflags);
1d7bb1d5 1424 } else {
2c3bac6d 1425 ctx->cached_cq_overflow++;
1d7bb1d5 1426 WRITE_ONCE(ctx->rings->cq_overflow,
2c3bac6d 1427 ctx->cached_cq_overflow);
1d7bb1d5 1428 }
b18032bb 1429 posted = true;
1d7bb1d5
JA
1430 }
1431
09e88404
PB
1432 all_flushed = list_empty(&ctx->cq_overflow_list);
1433 if (all_flushed) {
1434 clear_bit(0, &ctx->sq_check_overflow);
1435 clear_bit(0, &ctx->cq_check_overflow);
1436 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1437 }
46930143 1438
b18032bb
JA
1439 if (posted)
1440 io_commit_cqring(ctx);
1d7bb1d5 1441 spin_unlock_irqrestore(&ctx->completion_lock, flags);
b18032bb
JA
1442 if (posted)
1443 io_cqring_ev_posted(ctx);
1d7bb1d5
JA
1444
1445 while (!list_empty(&list)) {
40d8ddd4
PB
1446 req = list_first_entry(&list, struct io_kiocb, compl.list);
1447 list_del(&req->compl.list);
ec9c02ad 1448 io_put_req(req);
1d7bb1d5 1449 }
c4a2ed72 1450
09e88404 1451 return all_flushed;
1d7bb1d5
JA
1452}
1453
ca0a2651 1454static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
6c503150
PB
1455 struct task_struct *tsk,
1456 struct files_struct *files)
1457{
ca0a2651
JA
1458 bool ret = true;
1459
6c503150
PB
1460 if (test_bit(0, &ctx->cq_check_overflow)) {
1461 /* iopoll syncs against uring_lock, not completion_lock */
1462 if (ctx->flags & IORING_SETUP_IOPOLL)
1463 mutex_lock(&ctx->uring_lock);
ca0a2651 1464 ret = __io_cqring_overflow_flush(ctx, force, tsk, files);
6c503150
PB
1465 if (ctx->flags & IORING_SETUP_IOPOLL)
1466 mutex_unlock(&ctx->uring_lock);
1467 }
ca0a2651
JA
1468
1469 return ret;
6c503150
PB
1470}
1471
bcda7baa 1472static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
2b188cc1 1473{
78e19bbe 1474 struct io_ring_ctx *ctx = req->ctx;
2b188cc1
JA
1475 struct io_uring_cqe *cqe;
1476
78e19bbe 1477 trace_io_uring_complete(ctx, req->user_data, res);
51c3ff62 1478
2b188cc1
JA
1479 /*
1480 * If we can't get a cq entry, userspace overflowed the
1481 * submission (by quite a lot). Increment the overflow count in
1482 * the ring.
1483 */
1484 cqe = io_get_cqring(ctx);
1d7bb1d5 1485 if (likely(cqe)) {
78e19bbe 1486 WRITE_ONCE(cqe->user_data, req->user_data);
2b188cc1 1487 WRITE_ONCE(cqe->res, res);
bcda7baa 1488 WRITE_ONCE(cqe->flags, cflags);
fdaf083c
JA
1489 } else if (ctx->cq_overflow_flushed ||
1490 atomic_read(&req->task->io_uring->in_idle)) {
0f212204
JA
1491 /*
1492 * If we're in ring overflow flush mode, or in task cancel mode,
1493 * then we cannot store the request for later flushing, we need
1494 * to drop it on the floor.
1495 */
2c3bac6d
PB
1496 ctx->cached_cq_overflow++;
1497 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1d7bb1d5 1498 } else {
ad3eb2c8
JA
1499 if (list_empty(&ctx->cq_overflow_list)) {
1500 set_bit(0, &ctx->sq_check_overflow);
1501 set_bit(0, &ctx->cq_check_overflow);
6d5f9049 1502 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
ad3eb2c8 1503 }
40d8ddd4 1504 io_clean_op(req);
1d7bb1d5 1505 req->result = res;
0f7e466b 1506 req->compl.cflags = cflags;
40d8ddd4
PB
1507 refcount_inc(&req->refs);
1508 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
2b188cc1
JA
1509 }
1510}
1511
bcda7baa
JA
1512static void io_cqring_fill_event(struct io_kiocb *req, long res)
1513{
1514 __io_cqring_fill_event(req, res, 0);
1515}
1516
c7dae4ba
JA
1517static inline void io_req_complete_post(struct io_kiocb *req, long res,
1518 unsigned int cflags)
2b188cc1 1519{
78e19bbe 1520 struct io_ring_ctx *ctx = req->ctx;
2b188cc1
JA
1521 unsigned long flags;
1522
1523 spin_lock_irqsave(&ctx->completion_lock, flags);
bcda7baa 1524 __io_cqring_fill_event(req, res, cflags);
2b188cc1 1525 io_commit_cqring(ctx);
c7dae4ba
JA
1526 /*
1527 * If we're the last reference to this request, add to our locked
1528 * free_list cache.
1529 */
1530 if (refcount_dec_and_test(&req->refs)) {
1531 struct io_comp_state *cs = &ctx->submit_state.comp;
1532
1533 io_dismantle_req(req);
1534 io_put_task(req->task, 1);
1535 list_add(&req->compl.list, &cs->locked_free_list);
1536 cs->locked_free_nr++;
1537 } else
1538 req = NULL;
2b188cc1
JA
1539 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1540
8c838788 1541 io_cqring_ev_posted(ctx);
c7dae4ba
JA
1542 if (req) {
1543 io_queue_next(req);
1544 percpu_ref_put(&ctx->refs);
229a7b63 1545 }
229a7b63
JA
1546}
1547
a38d68db 1548static void io_req_complete_state(struct io_kiocb *req, long res,
889fca73 1549 unsigned int cflags)
229a7b63 1550{
a38d68db
PB
1551 io_clean_op(req);
1552 req->result = res;
1553 req->compl.cflags = cflags;
e342c807 1554 req->flags |= REQ_F_COMPLETE_INLINE;
e1e16097
JA
1555}
1556
889fca73
PB
1557static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1558 long res, unsigned cflags)
bcda7baa 1559{
889fca73
PB
1560 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1561 io_req_complete_state(req, res, cflags);
a38d68db 1562 else
c7dae4ba 1563 io_req_complete_post(req, res, cflags);
bcda7baa
JA
1564}
1565
a38d68db 1566static inline void io_req_complete(struct io_kiocb *req, long res)
0ddf92e8 1567{
889fca73 1568 __io_req_complete(req, 0, res, 0);
0ddf92e8
JA
1569}
1570
c7dae4ba 1571static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
0ddf92e8 1572{
c7dae4ba
JA
1573 struct io_submit_state *state = &ctx->submit_state;
1574 struct io_comp_state *cs = &state->comp;
e5d1bc0a 1575 struct io_kiocb *req = NULL;
0ddf92e8 1576
c7dae4ba
JA
1577 /*
1578 * If we have more than a batch's worth of requests in our IRQ side
1579 * locked cache, grab the lock and move them over to our submission
1580 * side cache.
1581 */
1582 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1583 spin_lock_irq(&ctx->completion_lock);
1584 list_splice_init(&cs->locked_free_list, &cs->free_list);
1585 cs->locked_free_nr = 0;
1586 spin_unlock_irq(&ctx->completion_lock);
1587 }
0ddf92e8 1588
c7dae4ba
JA
1589 while (!list_empty(&cs->free_list)) {
1590 req = list_first_entry(&cs->free_list, struct io_kiocb,
1b4c351f
JA
1591 compl.list);
1592 list_del(&req->compl.list);
e5d1bc0a
PB
1593 state->reqs[state->free_reqs++] = req;
1594 if (state->free_reqs == ARRAY_SIZE(state->reqs))
1595 break;
1b4c351f
JA
1596 }
1597
e5d1bc0a 1598 return req != NULL;
0ddf92e8
JA
1599}
1600
e5d1bc0a 1601static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
2b188cc1 1602{
e5d1bc0a
PB
1603 struct io_submit_state *state = &ctx->submit_state;
1604
1605 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1606
f6b6c7d6 1607 if (!state->free_reqs) {
291b2821 1608 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
2579f913
JA
1609 int ret;
1610
c7dae4ba 1611 if (io_flush_cached_reqs(ctx))
e5d1bc0a
PB
1612 goto got_req;
1613
bf019da7
PB
1614 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1615 state->reqs);
fd6fab2c
JA
1616
1617 /*
1618 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1619 * retry single alloc to be on the safe side.
1620 */
1621 if (unlikely(ret <= 0)) {
1622 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1623 if (!state->reqs[0])
3893f39f 1624 return NULL;
fd6fab2c
JA
1625 ret = 1;
1626 }
291b2821 1627 state->free_reqs = ret;
2b188cc1 1628 }
e5d1bc0a 1629got_req:
291b2821
PB
1630 state->free_reqs--;
1631 return state->reqs[state->free_reqs];
2b188cc1
JA
1632}
1633
8da11c19
PB
1634static inline void io_put_file(struct io_kiocb *req, struct file *file,
1635 bool fixed)
1636{
36f72fe2 1637 if (!fixed)
8da11c19
PB
1638 fput(file);
1639}
1640
4edf20f9 1641static void io_dismantle_req(struct io_kiocb *req)
2b188cc1 1642{
3ca405eb 1643 io_clean_op(req);
929a3af9 1644
e8c2bc1f
JA
1645 if (req->async_data)
1646 kfree(req->async_data);
8da11c19
PB
1647 if (req->file)
1648 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
269bbe5f
BM
1649 if (req->fixed_rsrc_refs)
1650 percpu_ref_put(req->fixed_rsrc_refs);
f85c310a
PB
1651
1652 if (req->flags & REQ_F_INFLIGHT) {
1653 struct io_ring_ctx *ctx = req->ctx;
f85c310a
PB
1654 unsigned long flags;
1655
1656 spin_lock_irqsave(&ctx->inflight_lock, flags);
1657 list_del(&req->inflight_entry);
1658 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1659 req->flags &= ~REQ_F_INFLIGHT;
f85c310a 1660 }
e65ef56d
JA
1661}
1662
b23fcf47 1663/* must to be called somewhat shortly after putting a request */
7c660731
PB
1664static inline void io_put_task(struct task_struct *task, int nr)
1665{
1666 struct io_uring_task *tctx = task->io_uring;
1667
1668 percpu_counter_sub(&tctx->inflight, nr);
1669 if (unlikely(atomic_read(&tctx->in_idle)))
1670 wake_up(&tctx->wait);
1671 put_task_struct_many(task, nr);
1672}
1673
216578e5 1674static void __io_free_req(struct io_kiocb *req)
c6ca97b3 1675{
51a4cc11 1676 struct io_ring_ctx *ctx = req->ctx;
c6ca97b3 1677
216578e5 1678 io_dismantle_req(req);
7c660731 1679 io_put_task(req->task, 1);
c6ca97b3 1680
3893f39f 1681 kmem_cache_free(req_cachep, req);
ecfc5177 1682 percpu_ref_put(&ctx->refs);
e65ef56d
JA
1683}
1684
f2f87370
PB
1685static inline void io_remove_next_linked(struct io_kiocb *req)
1686{
1687 struct io_kiocb *nxt = req->link;
1688
1689 req->link = nxt->link;
1690 nxt->link = NULL;
1691}
1692
c9abd7ad 1693static void io_kill_linked_timeout(struct io_kiocb *req)
2665abfd 1694{
a197f664 1695 struct io_ring_ctx *ctx = req->ctx;
7c86ffee 1696 struct io_kiocb *link;
c9abd7ad
PB
1697 bool cancelled = false;
1698 unsigned long flags;
7c86ffee 1699
c9abd7ad 1700 spin_lock_irqsave(&ctx->completion_lock, flags);
f2f87370
PB
1701 link = req->link;
1702
900fad45
PB
1703 /*
1704 * Can happen if a linked timeout fired and link had been like
1705 * req -> link t-out -> link t-out [-> ...]
1706 */
c9abd7ad
PB
1707 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1708 struct io_timeout_data *io = link->async_data;
1709 int ret;
7c86ffee 1710
f2f87370 1711 io_remove_next_linked(req);
90cd7e42 1712 link->timeout.head = NULL;
c9abd7ad
PB
1713 ret = hrtimer_try_to_cancel(&io->timer);
1714 if (ret != -1) {
1715 io_cqring_fill_event(link, -ECANCELED);
1716 io_commit_cqring(ctx);
1717 cancelled = true;
1718 }
1719 }
7c86ffee 1720 req->flags &= ~REQ_F_LINK_TIMEOUT;
216578e5 1721 spin_unlock_irqrestore(&ctx->completion_lock, flags);
ab0b6451 1722
c9abd7ad 1723 if (cancelled) {
7c86ffee 1724 io_cqring_ev_posted(ctx);
c9abd7ad
PB
1725 io_put_req(link);
1726 }
7c86ffee
PB
1727}
1728
9e645e11 1729
d148ca4b 1730static void io_fail_links(struct io_kiocb *req)
9e645e11 1731{
f2f87370 1732 struct io_kiocb *link, *nxt;
2665abfd 1733 struct io_ring_ctx *ctx = req->ctx;
d148ca4b 1734 unsigned long flags;
9e645e11 1735
d148ca4b 1736 spin_lock_irqsave(&ctx->completion_lock, flags);
f2f87370
PB
1737 link = req->link;
1738 req->link = NULL;
9e645e11 1739
f2f87370
PB
1740 while (link) {
1741 nxt = link->link;
1742 link->link = NULL;
2665abfd 1743
f2f87370 1744 trace_io_uring_fail_link(req, link);
7c86ffee 1745 io_cqring_fill_event(link, -ECANCELED);
216578e5 1746
1575f21a 1747 io_put_req_deferred(link, 2);
f2f87370 1748 link = nxt;
9e645e11 1749 }
2665abfd 1750 io_commit_cqring(ctx);
216578e5 1751 spin_unlock_irqrestore(&ctx->completion_lock, flags);
9e645e11 1752
2665abfd 1753 io_cqring_ev_posted(ctx);
9e645e11
JA
1754}
1755
3fa5e0f3 1756static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
c69f8dbe 1757{
7c86ffee
PB
1758 if (req->flags & REQ_F_LINK_TIMEOUT)
1759 io_kill_linked_timeout(req);
944e58bf 1760
9e645e11
JA
1761 /*
1762 * If LINK is set, we have dependent requests in this chain. If we
1763 * didn't fail this request, queue the first one up, moving any other
1764 * dependencies to the next request. In case of failure, fail the rest
1765 * of the chain.
1766 */
f2f87370
PB
1767 if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
1768 struct io_kiocb *nxt = req->link;
1769
1770 req->link = NULL;
1771 return nxt;
1772 }
9b5f7bd9
PB
1773 io_fail_links(req);
1774 return NULL;
4d7dd462 1775}
9e645e11 1776
f2f87370 1777static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
3fa5e0f3 1778{
cdbff982 1779 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
3fa5e0f3
PB
1780 return NULL;
1781 return __io_req_find_next(req);
1782}
1783
2c32395d
PB
1784static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1785{
1786 if (!ctx)
1787 return;
1788 if (ctx->submit_state.comp.nr) {
1789 mutex_lock(&ctx->uring_lock);
1790 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1791 mutex_unlock(&ctx->uring_lock);
1792 }
1793 percpu_ref_put(&ctx->refs);
1794}
1795
7cbf1722 1796static bool __tctx_task_work(struct io_uring_task *tctx)
c2c4c83c 1797{
65453d1e 1798 struct io_ring_ctx *ctx = NULL;
7cbf1722
JA
1799 struct io_wq_work_list list;
1800 struct io_wq_work_node *node;
c2c4c83c 1801
7cbf1722
JA
1802 if (wq_list_empty(&tctx->task_list))
1803 return false;
6200b0ae 1804
0b81e80c 1805 spin_lock_irq(&tctx->task_lock);
7cbf1722
JA
1806 list = tctx->task_list;
1807 INIT_WQ_LIST(&tctx->task_list);
0b81e80c 1808 spin_unlock_irq(&tctx->task_lock);
c2c4c83c 1809
7cbf1722
JA
1810 node = list.first;
1811 while (node) {
1812 struct io_wq_work_node *next = node->next;
1813 struct io_kiocb *req;
0ba9c9ed 1814
7cbf1722 1815 req = container_of(node, struct io_kiocb, io_task_work.node);
2c32395d
PB
1816 if (req->ctx != ctx) {
1817 ctx_flush_and_put(ctx);
1818 ctx = req->ctx;
1819 percpu_ref_get(&ctx->refs);
65453d1e 1820 }
65453d1e 1821
2c32395d
PB
1822 req->task_work.func(&req->task_work);
1823 node = next;
7cbf1722
JA
1824 }
1825
2c32395d 1826 ctx_flush_and_put(ctx);
7cbf1722 1827 return list.first != NULL;
c2c4c83c
JA
1828}
1829
7cbf1722 1830static void tctx_task_work(struct callback_head *cb)
c40f6379 1831{
7cbf1722 1832 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
c40f6379 1833
1d5f360d
JA
1834 clear_bit(0, &tctx->task_state);
1835
7cbf1722
JA
1836 while (__tctx_task_work(tctx))
1837 cond_resched();
7cbf1722
JA
1838}
1839
1840static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1841 enum task_work_notify_mode notify)
1842{
1843 struct io_uring_task *tctx = tsk->io_uring;
1844 struct io_wq_work_node *node, *prev;
0b81e80c 1845 unsigned long flags;
7cbf1722
JA
1846 int ret;
1847
1848 WARN_ON_ONCE(!tctx);
1849
0b81e80c 1850 spin_lock_irqsave(&tctx->task_lock, flags);
7cbf1722 1851 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
0b81e80c 1852 spin_unlock_irqrestore(&tctx->task_lock, flags);
7cbf1722
JA
1853
1854 /* task_work already pending, we're done */
1855 if (test_bit(0, &tctx->task_state) ||
1856 test_and_set_bit(0, &tctx->task_state))
1857 return 0;
1858
1859 if (!task_work_add(tsk, &tctx->task_work, notify))
1860 return 0;
1861
1862 /*
1863 * Slow path - we failed, find and delete work. if the work is not
1864 * in the list, it got run and we're fine.
1865 */
1866 ret = 0;
0b81e80c 1867 spin_lock_irqsave(&tctx->task_lock, flags);
7cbf1722
JA
1868 wq_list_for_each(node, prev, &tctx->task_list) {
1869 if (&req->io_task_work.node == node) {
1870 wq_list_del(&tctx->task_list, node, prev);
1871 ret = 1;
1872 break;
1873 }
1874 }
0b81e80c 1875 spin_unlock_irqrestore(&tctx->task_lock, flags);
7cbf1722
JA
1876 clear_bit(0, &tctx->task_state);
1877 return ret;
1878}
1879
355fb9e2 1880static int io_req_task_work_add(struct io_kiocb *req)
c2c4c83c
JA
1881{
1882 struct task_struct *tsk = req->task;
1883 struct io_ring_ctx *ctx = req->ctx;
91989c70
JA
1884 enum task_work_notify_mode notify;
1885 int ret;
c2c4c83c 1886
6200b0ae
JA
1887 if (tsk->flags & PF_EXITING)
1888 return -ESRCH;
1889
c2c4c83c 1890 /*
0ba9c9ed
JA
1891 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1892 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1893 * processing task_work. There's no reliable way to tell if TWA_RESUME
1894 * will do the job.
c2c4c83c 1895 */
91989c70 1896 notify = TWA_NONE;
355fb9e2 1897 if (!(ctx->flags & IORING_SETUP_SQPOLL))
c2c4c83c
JA
1898 notify = TWA_SIGNAL;
1899
7cbf1722 1900 ret = io_task_work_add(tsk, req, notify);
c2c4c83c
JA
1901 if (!ret)
1902 wake_up_process(tsk);
0ba9c9ed 1903
c2c4c83c
JA
1904 return ret;
1905}
1906
eab30c4d 1907static void io_req_task_work_add_fallback(struct io_kiocb *req,
7cbf1722 1908 task_work_func_t cb)
eab30c4d 1909{
7c25c0d1
JA
1910 struct io_ring_ctx *ctx = req->ctx;
1911 struct callback_head *head;
eab30c4d
PB
1912
1913 init_task_work(&req->task_work, cb);
7c25c0d1
JA
1914 do {
1915 head = READ_ONCE(ctx->exit_task_work);
1916 req->task_work.next = head;
1917 } while (cmpxchg(&ctx->exit_task_work, head, &req->task_work) != head);
eab30c4d
PB
1918}
1919
c40f6379
JA
1920static void __io_req_task_cancel(struct io_kiocb *req, int error)
1921{
1922 struct io_ring_ctx *ctx = req->ctx;
1923
1924 spin_lock_irq(&ctx->completion_lock);
1925 io_cqring_fill_event(req, error);
1926 io_commit_cqring(ctx);
1927 spin_unlock_irq(&ctx->completion_lock);
1928
1929 io_cqring_ev_posted(ctx);
1930 req_set_fail_links(req);
1931 io_double_put_req(req);
1932}
1933
1934static void io_req_task_cancel(struct callback_head *cb)
1935{
1936 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
87ceb6a6 1937 struct io_ring_ctx *ctx = req->ctx;
c40f6379 1938
792bb6eb 1939 mutex_lock(&ctx->uring_lock);
a3df7698 1940 __io_req_task_cancel(req, req->result);
792bb6eb 1941 mutex_unlock(&ctx->uring_lock);
87ceb6a6 1942 percpu_ref_put(&ctx->refs);
c40f6379
JA
1943}
1944
1945static void __io_req_task_submit(struct io_kiocb *req)
1946{
1947 struct io_ring_ctx *ctx = req->ctx;
1948
04fc6c80 1949 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
81b6d05c 1950 mutex_lock(&ctx->uring_lock);
70aacfe6 1951 if (!(current->flags & PF_EXITING) && !current->in_execve)
c5eef2b9 1952 __io_queue_sqe(req);
81b6d05c 1953 else
c40f6379 1954 __io_req_task_cancel(req, -EFAULT);
81b6d05c 1955 mutex_unlock(&ctx->uring_lock);
c40f6379
JA
1956}
1957
1958static void io_req_task_submit(struct callback_head *cb)
1959{
1960 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1961
1962 __io_req_task_submit(req);
1963}
1964
1965static void io_req_task_queue(struct io_kiocb *req)
1966{
c40f6379
JA
1967 int ret;
1968
7cbf1722 1969 req->task_work.func = io_req_task_submit;
355fb9e2 1970 ret = io_req_task_work_add(req);
c40f6379 1971 if (unlikely(ret)) {
a3df7698 1972 req->result = -ECANCELED;
04fc6c80 1973 percpu_ref_get(&req->ctx->refs);
eab30c4d 1974 io_req_task_work_add_fallback(req, io_req_task_cancel);
c40f6379 1975 }
c40f6379
JA
1976}
1977
a3df7698
PB
1978static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1979{
1980 percpu_ref_get(&req->ctx->refs);
1981 req->result = ret;
1982 req->task_work.func = io_req_task_cancel;
1983
1984 if (unlikely(io_req_task_work_add(req)))
1985 io_req_task_work_add_fallback(req, io_req_task_cancel);
1986}
1987
f2f87370 1988static inline void io_queue_next(struct io_kiocb *req)
c69f8dbe 1989{
9b5f7bd9 1990 struct io_kiocb *nxt = io_req_find_next(req);
944e58bf
PB
1991
1992 if (nxt)
906a8c3f 1993 io_req_task_queue(nxt);
c69f8dbe
JL
1994}
1995
c3524383 1996static void io_free_req(struct io_kiocb *req)
7a743e22 1997{
c3524383
PB
1998 io_queue_next(req);
1999 __io_free_req(req);
2000}
8766dd51 2001
2d6500d4 2002struct req_batch {
5af1d13e
PB
2003 struct task_struct *task;
2004 int task_refs;
1b4c351f 2005 int ctx_refs;
2d6500d4
PB
2006};
2007
5af1d13e
PB
2008static inline void io_init_req_batch(struct req_batch *rb)
2009{
5af1d13e 2010 rb->task_refs = 0;
9ae72463 2011 rb->ctx_refs = 0;
5af1d13e
PB
2012 rb->task = NULL;
2013}
2014
2d6500d4
PB
2015static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2016 struct req_batch *rb)
2017{
6e833d53 2018 if (rb->task)
7c660731 2019 io_put_task(rb->task, rb->task_refs);
9ae72463
PB
2020 if (rb->ctx_refs)
2021 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2d6500d4
PB
2022}
2023
6ff119a6
PB
2024static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2025 struct io_submit_state *state)
2d6500d4 2026{
f2f87370 2027 io_queue_next(req);
2d6500d4 2028
e3bc8e9d 2029 if (req->task != rb->task) {
7c660731
PB
2030 if (rb->task)
2031 io_put_task(rb->task, rb->task_refs);
e3bc8e9d
JA
2032 rb->task = req->task;
2033 rb->task_refs = 0;
5af1d13e 2034 }
e3bc8e9d 2035 rb->task_refs++;
9ae72463 2036 rb->ctx_refs++;
5af1d13e 2037
4edf20f9 2038 io_dismantle_req(req);
bd759045 2039 if (state->free_reqs != ARRAY_SIZE(state->reqs))
6ff119a6 2040 state->reqs[state->free_reqs++] = req;
bd759045
PB
2041 else
2042 list_add(&req->compl.list, &state->comp.free_list);
7a743e22
PB
2043}
2044
905c172f
PB
2045static void io_submit_flush_completions(struct io_comp_state *cs,
2046 struct io_ring_ctx *ctx)
2047{
2048 int i, nr = cs->nr;
2049 struct io_kiocb *req;
2050 struct req_batch rb;
2051
2052 io_init_req_batch(&rb);
2053 spin_lock_irq(&ctx->completion_lock);
2054 for (i = 0; i < nr; i++) {
2055 req = cs->reqs[i];
2056 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2057 }
2058 io_commit_cqring(ctx);
2059 spin_unlock_irq(&ctx->completion_lock);
2060
2061 io_cqring_ev_posted(ctx);
2062 for (i = 0; i < nr; i++) {
2063 req = cs->reqs[i];
2064
2065 /* submission and completion refs */
2066 if (refcount_sub_and_test(2, &req->refs))
6ff119a6 2067 io_req_free_batch(&rb, req, &ctx->submit_state);
905c172f
PB
2068 }
2069
2070 io_req_free_batch_finish(ctx, &rb);
2071 cs->nr = 0;
7a743e22
PB
2072}
2073
ba816ad6
JA
2074/*
2075 * Drop reference to request, return next in chain (if there is one) if this
2076 * was the last reference to this request.
2077 */
9b5f7bd9 2078static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
e65ef56d 2079{
9b5f7bd9
PB
2080 struct io_kiocb *nxt = NULL;
2081
2a44f467 2082 if (refcount_dec_and_test(&req->refs)) {
9b5f7bd9 2083 nxt = io_req_find_next(req);
4d7dd462 2084 __io_free_req(req);
2a44f467 2085 }
9b5f7bd9 2086 return nxt;
2b188cc1
JA
2087}
2088
e65ef56d
JA
2089static void io_put_req(struct io_kiocb *req)
2090{
2091 if (refcount_dec_and_test(&req->refs))
2092 io_free_req(req);
2b188cc1
JA
2093}
2094
216578e5
PB
2095static void io_put_req_deferred_cb(struct callback_head *cb)
2096{
2097 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2098
2099 io_free_req(req);
2100}
2101
2102static void io_free_req_deferred(struct io_kiocb *req)
2103{
2104 int ret;
2105
7cbf1722 2106 req->task_work.func = io_put_req_deferred_cb;
355fb9e2 2107 ret = io_req_task_work_add(req);
eab30c4d
PB
2108 if (unlikely(ret))
2109 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
216578e5
PB
2110}
2111
2112static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2113{
2114 if (refcount_sub_and_test(refs, &req->refs))
2115 io_free_req_deferred(req);
2116}
2117
978db57e
JA
2118static void io_double_put_req(struct io_kiocb *req)
2119{
2120 /* drop both submit and complete references */
2121 if (refcount_sub_and_test(2, &req->refs))
2122 io_free_req(req);
2123}
2124
6c503150 2125static unsigned io_cqring_events(struct io_ring_ctx *ctx)
a3a0e43f
JA
2126{
2127 /* See comment at the top of this file */
2128 smp_rmb();
e23de15f 2129 return __io_cqring_events(ctx);
a3a0e43f
JA
2130}
2131
fb5ccc98
PB
2132static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2133{
2134 struct io_rings *rings = ctx->rings;
2135
2136 /* make sure SQ entry isn't read before tail */
2137 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2138}
2139
8ff069bf 2140static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
e94f141b 2141{
8ff069bf 2142 unsigned int cflags;
e94f141b 2143
bcda7baa
JA
2144 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2145 cflags |= IORING_CQE_F_BUFFER;
0e1b6fe3 2146 req->flags &= ~REQ_F_BUFFER_SELECTED;
bcda7baa
JA
2147 kfree(kbuf);
2148 return cflags;
e94f141b
JA
2149}
2150
8ff069bf 2151static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
bcda7baa 2152{
4d954c25 2153 struct io_buffer *kbuf;
bcda7baa 2154
4d954c25 2155 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
8ff069bf
PB
2156 return io_put_kbuf(req, kbuf);
2157}
2158
4c6e277c
JA
2159static inline bool io_run_task_work(void)
2160{
6200b0ae
JA
2161 /*
2162 * Not safe to run on exiting task, and the task_work handling will
2163 * not add work to such a task.
2164 */
2165 if (unlikely(current->flags & PF_EXITING))
2166 return false;
4c6e277c
JA
2167 if (current->task_works) {
2168 __set_current_state(TASK_RUNNING);
2169 task_work_run();
2170 return true;
2171 }
2172
2173 return false;
bcda7baa
JA
2174}
2175
def596e9
JA
2176/*
2177 * Find and free completed poll iocbs
2178 */
2179static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2180 struct list_head *done)
2181{
8237e045 2182 struct req_batch rb;
def596e9 2183 struct io_kiocb *req;
bbde017a
XW
2184
2185 /* order with ->result store in io_complete_rw_iopoll() */
2186 smp_rmb();
def596e9 2187
5af1d13e 2188 io_init_req_batch(&rb);
def596e9 2189 while (!list_empty(done)) {
bcda7baa
JA
2190 int cflags = 0;
2191
d21ffe7e 2192 req = list_first_entry(done, struct io_kiocb, inflight_entry);
f161340d
PB
2193 list_del(&req->inflight_entry);
2194
bbde017a
XW
2195 if (READ_ONCE(req->result) == -EAGAIN) {
2196 req->iopoll_completed = 0;
23faba36 2197 if (io_rw_reissue(req))
f161340d 2198 continue;
bbde017a 2199 }
def596e9 2200
bcda7baa 2201 if (req->flags & REQ_F_BUFFER_SELECTED)
8ff069bf 2202 cflags = io_put_rw_kbuf(req);
bcda7baa
JA
2203
2204 __io_cqring_fill_event(req, req->result, cflags);
def596e9
JA
2205 (*nr_events)++;
2206
c3524383 2207 if (refcount_dec_and_test(&req->refs))
6ff119a6 2208 io_req_free_batch(&rb, req, &ctx->submit_state);
def596e9 2209 }
def596e9 2210
09bb8394 2211 io_commit_cqring(ctx);
80c18e4a 2212 io_cqring_ev_posted_iopoll(ctx);
2d6500d4 2213 io_req_free_batch_finish(ctx, &rb);
581f9810
BM
2214}
2215
def596e9
JA
2216static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2217 long min)
2218{
2219 struct io_kiocb *req, *tmp;
2220 LIST_HEAD(done);
2221 bool spin;
2222 int ret;
2223
2224 /*
2225 * Only spin for completions if we don't have multiple devices hanging
2226 * off our complete list, and we're under the requested amount.
2227 */
2228 spin = !ctx->poll_multi_file && *nr_events < min;
2229
2230 ret = 0;
d21ffe7e 2231 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
9adbd45d 2232 struct kiocb *kiocb = &req->rw.kiocb;
def596e9
JA
2233
2234 /*
581f9810
BM
2235 * Move completed and retryable entries to our local lists.
2236 * If we find a request that requires polling, break out
2237 * and complete those lists first, if we have entries there.
def596e9 2238 */
65a6543d 2239 if (READ_ONCE(req->iopoll_completed)) {
d21ffe7e 2240 list_move_tail(&req->inflight_entry, &done);
def596e9
JA
2241 continue;
2242 }
2243 if (!list_empty(&done))
2244 break;
2245
2246 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2247 if (ret < 0)
2248 break;
2249
3aadc23e
PB
2250 /* iopoll may have completed current req */
2251 if (READ_ONCE(req->iopoll_completed))
d21ffe7e 2252 list_move_tail(&req->inflight_entry, &done);
3aadc23e 2253
def596e9
JA
2254 if (ret && spin)
2255 spin = false;
2256 ret = 0;
2257 }
2258
2259 if (!list_empty(&done))
2260 io_iopoll_complete(ctx, nr_events, &done);
2261
2262 return ret;
2263}
2264
2265/*
d195a66e 2266 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
def596e9
JA
2267 * non-spinning poll check - we'll still enter the driver poll loop, but only
2268 * as a non-spinning completion check.
2269 */
2270static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2271 long min)
2272{
540e32a0 2273 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
def596e9
JA
2274 int ret;
2275
2276 ret = io_do_iopoll(ctx, nr_events, min);
2277 if (ret < 0)
2278 return ret;
eba0a4dd 2279 if (*nr_events >= min)
def596e9
JA
2280 return 0;
2281 }
2282
2283 return 1;
2284}
2285
2286/*
2287 * We can't just wait for polled events to come to us, we have to actively
2288 * find and complete them.
2289 */
b2edc0a7 2290static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
def596e9
JA
2291{
2292 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2293 return;
2294
2295 mutex_lock(&ctx->uring_lock);
540e32a0 2296 while (!list_empty(&ctx->iopoll_list)) {
def596e9
JA
2297 unsigned int nr_events = 0;
2298
b2edc0a7 2299 io_do_iopoll(ctx, &nr_events, 0);
08f5439f 2300
b2edc0a7
PB
2301 /* let it sleep and repeat later if can't complete a request */
2302 if (nr_events == 0)
2303 break;
08f5439f
JA
2304 /*
2305 * Ensure we allow local-to-the-cpu processing to take place,
2306 * in this case we need to ensure that we reap all events.
3fcee5a6 2307 * Also let task_work, etc. to progress by releasing the mutex
08f5439f 2308 */
3fcee5a6
PB
2309 if (need_resched()) {
2310 mutex_unlock(&ctx->uring_lock);
2311 cond_resched();
2312 mutex_lock(&ctx->uring_lock);
2313 }
def596e9
JA
2314 }
2315 mutex_unlock(&ctx->uring_lock);
2316}
2317
7668b92a 2318static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
def596e9 2319{
7668b92a 2320 unsigned int nr_events = 0;
2b2ed975 2321 int iters = 0, ret = 0;
500f9fba 2322
c7849be9
XW
2323 /*
2324 * We disallow the app entering submit/complete with polling, but we
2325 * still need to lock the ring to prevent racing with polled issue
2326 * that got punted to a workqueue.
2327 */
2328 mutex_lock(&ctx->uring_lock);
def596e9 2329 do {
a3a0e43f
JA
2330 /*
2331 * Don't enter poll loop if we already have events pending.
2332 * If we do, we can potentially be spinning for commands that
2333 * already triggered a CQE (eg in error).
2334 */
6c503150
PB
2335 if (test_bit(0, &ctx->cq_check_overflow))
2336 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2337 if (io_cqring_events(ctx))
a3a0e43f
JA
2338 break;
2339
500f9fba
JA
2340 /*
2341 * If a submit got punted to a workqueue, we can have the
2342 * application entering polling for a command before it gets
2343 * issued. That app will hold the uring_lock for the duration
2344 * of the poll right here, so we need to take a breather every
2345 * now and then to ensure that the issue has a chance to add
2346 * the poll to the issued list. Otherwise we can spin here
2347 * forever, while the workqueue is stuck trying to acquire the
2348 * very same mutex.
2349 */
2350 if (!(++iters & 7)) {
2351 mutex_unlock(&ctx->uring_lock);
4c6e277c 2352 io_run_task_work();
500f9fba
JA
2353 mutex_lock(&ctx->uring_lock);
2354 }
2355
7668b92a 2356 ret = io_iopoll_getevents(ctx, &nr_events, min);
def596e9
JA
2357 if (ret <= 0)
2358 break;
2359 ret = 0;
7668b92a 2360 } while (min && !nr_events && !need_resched());
def596e9 2361
500f9fba 2362 mutex_unlock(&ctx->uring_lock);
def596e9
JA
2363 return ret;
2364}
2365
491381ce 2366static void kiocb_end_write(struct io_kiocb *req)
2b188cc1 2367{
491381ce
JA
2368 /*
2369 * Tell lockdep we inherited freeze protection from submission
2370 * thread.
2371 */
2372 if (req->flags & REQ_F_ISREG) {
2373 struct inode *inode = file_inode(req->file);
2b188cc1 2374
491381ce 2375 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2b188cc1 2376 }
491381ce 2377 file_end_write(req->file);
2b188cc1
JA
2378}
2379
b63534c4 2380#ifdef CONFIG_BLOCK
dc2a6e9a 2381static bool io_resubmit_prep(struct io_kiocb *req)
b63534c4
JA
2382{
2383 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
4a245479 2384 int rw, ret;
b63534c4 2385 struct iov_iter iter;
b63534c4 2386
dc2a6e9a
PB
2387 /* already prepared */
2388 if (req->async_data)
2389 return true;
b63534c4
JA
2390
2391 switch (req->opcode) {
2392 case IORING_OP_READV:
2393 case IORING_OP_READ_FIXED:
2394 case IORING_OP_READ:
2395 rw = READ;
2396 break;
2397 case IORING_OP_WRITEV:
2398 case IORING_OP_WRITE_FIXED:
2399 case IORING_OP_WRITE:
2400 rw = WRITE;
2401 break;
2402 default:
2403 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2404 req->opcode);
dc2a6e9a 2405 return false;
b63534c4
JA
2406 }
2407
dc2a6e9a
PB
2408 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2409 if (ret < 0)
2410 return false;
6bf985dc 2411 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
b63534c4 2412}
b63534c4 2413
3e6a0d3c 2414static bool io_rw_should_reissue(struct io_kiocb *req)
b63534c4 2415{
355afaeb 2416 umode_t mode = file_inode(req->file)->i_mode;
3e6a0d3c 2417 struct io_ring_ctx *ctx = req->ctx;
b63534c4 2418
355afaeb
JA
2419 if (!S_ISBLK(mode) && !S_ISREG(mode))
2420 return false;
3e6a0d3c
JA
2421 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2422 !(ctx->flags & IORING_SETUP_IOPOLL)))
b63534c4 2423 return false;
7c977a58
JA
2424 /*
2425 * If ref is dying, we might be running poll reap from the exit work.
2426 * Don't attempt to reissue from that path, just let it fail with
2427 * -EAGAIN.
2428 */
3e6a0d3c
JA
2429 if (percpu_ref_is_dying(&ctx->refs))
2430 return false;
2431 return true;
2432}
2433#endif
2434
2435static bool io_rw_reissue(struct io_kiocb *req)
2436{
2437#ifdef CONFIG_BLOCK
2438 if (!io_rw_should_reissue(req))
7c977a58 2439 return false;
b63534c4 2440
55e6ac1e
PB
2441 lockdep_assert_held(&req->ctx->uring_lock);
2442
37d1e2e3 2443 if (io_resubmit_prep(req)) {
fdee946d
JA
2444 refcount_inc(&req->refs);
2445 io_queue_async_work(req);
b63534c4 2446 return true;
fdee946d 2447 }
dc2a6e9a 2448 req_set_fail_links(req);
b63534c4
JA
2449#endif
2450 return false;
2451}
2452
a1d7c393 2453static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
889fca73 2454 unsigned int issue_flags)
a1d7c393 2455{
2f8e45f1
PB
2456 int cflags = 0;
2457
23faba36
PB
2458 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2459 return;
2f8e45f1
PB
2460 if (res != req->result)
2461 req_set_fail_links(req);
23faba36 2462
2f8e45f1
PB
2463 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2464 kiocb_end_write(req);
2465 if (req->flags & REQ_F_BUFFER_SELECTED)
2466 cflags = io_put_rw_kbuf(req);
2467 __io_req_complete(req, issue_flags, res, cflags);
ba816ad6
JA
2468}
2469
2470static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2471{
9adbd45d 2472 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
ba816ad6 2473
889fca73 2474 __io_complete_rw(req, res, res2, 0);
2b188cc1
JA
2475}
2476
def596e9
JA
2477static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2478{
9adbd45d 2479 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
def596e9 2480
3e6a0d3c
JA
2481#ifdef CONFIG_BLOCK
2482 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2483 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2484 struct io_async_rw *rw = req->async_data;
2485
2486 if (rw)
2487 iov_iter_revert(&rw->iter,
2488 req->result - iov_iter_count(&rw->iter));
2489 else if (!io_resubmit_prep(req))
2490 res = -EIO;
2491 }
2492#endif
2493
491381ce
JA
2494 if (kiocb->ki_flags & IOCB_WRITE)
2495 kiocb_end_write(req);
def596e9 2496
2d7d6792 2497 if (res != -EAGAIN && res != req->result)
4e88d6e7 2498 req_set_fail_links(req);
bbde017a
XW
2499
2500 WRITE_ONCE(req->result, res);
2501 /* order with io_poll_complete() checking ->result */
cd664b0e
PB
2502 smp_wmb();
2503 WRITE_ONCE(req->iopoll_completed, 1);
def596e9
JA
2504}
2505
2506/*
2507 * After the iocb has been issued, it's safe to be found on the poll list.
2508 * Adding the kiocb to the list AFTER submission ensures that we don't
2509 * find it from a io_iopoll_getevents() thread before the issuer is done
2510 * accessing the kiocb cookie.
2511 */
2e9dbe90 2512static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
def596e9
JA
2513{
2514 struct io_ring_ctx *ctx = req->ctx;
2515
2516 /*
2517 * Track whether we have multiple files in our lists. This will impact
2518 * how we do polling eventually, not spinning if we're on potentially
2519 * different devices.
2520 */
540e32a0 2521 if (list_empty(&ctx->iopoll_list)) {
def596e9
JA
2522 ctx->poll_multi_file = false;
2523 } else if (!ctx->poll_multi_file) {
2524 struct io_kiocb *list_req;
2525
540e32a0 2526 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
d21ffe7e 2527 inflight_entry);
9adbd45d 2528 if (list_req->file != req->file)
def596e9
JA
2529 ctx->poll_multi_file = true;
2530 }
2531
2532 /*
2533 * For fast devices, IO may have already completed. If it has, add
2534 * it to the front so we find it first.
2535 */
65a6543d 2536 if (READ_ONCE(req->iopoll_completed))
d21ffe7e 2537 list_add(&req->inflight_entry, &ctx->iopoll_list);
def596e9 2538 else
d21ffe7e 2539 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
bdcd3eab 2540
2e9dbe90
XW
2541 /*
2542 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2543 * task context or in io worker task context. If current task context is
2544 * sq thread, we don't need to check whether should wake up sq thread.
2545 */
2546 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
534ca6d6
JA
2547 wq_has_sleeper(&ctx->sq_data->wait))
2548 wake_up(&ctx->sq_data->wait);
def596e9
JA
2549}
2550
9f13c35b
PB
2551static inline void io_state_file_put(struct io_submit_state *state)
2552{
02b23a9a
PB
2553 if (state->file_refs) {
2554 fput_many(state->file, state->file_refs);
2555 state->file_refs = 0;
2556 }
9a56a232
JA
2557}
2558
2559/*
2560 * Get as many references to a file as we have IOs left in this submission,
2561 * assuming most submissions are for one file, or at least that each file
2562 * has more than one submission.
2563 */
8da11c19 2564static struct file *__io_file_get(struct io_submit_state *state, int fd)
9a56a232
JA
2565{
2566 if (!state)
2567 return fget(fd);
2568
6e1271e6 2569 if (state->file_refs) {
9a56a232 2570 if (state->fd == fd) {
6e1271e6 2571 state->file_refs--;
9a56a232
JA
2572 return state->file;
2573 }
02b23a9a 2574 io_state_file_put(state);
9a56a232
JA
2575 }
2576 state->file = fget_many(fd, state->ios_left);
6e1271e6 2577 if (unlikely(!state->file))
9a56a232
JA
2578 return NULL;
2579
2580 state->fd = fd;
6e1271e6 2581 state->file_refs = state->ios_left - 1;
9a56a232
JA
2582 return state->file;
2583}
2584
4503b767
JA
2585static bool io_bdev_nowait(struct block_device *bdev)
2586{
9ba0d0c8 2587 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
4503b767
JA
2588}
2589
2b188cc1
JA
2590/*
2591 * If we tracked the file through the SCM inflight mechanism, we could support
2592 * any file. For now, just ensure that anything potentially problematic is done
2593 * inline.
2594 */
af197f50 2595static bool io_file_supports_async(struct file *file, int rw)
2b188cc1
JA
2596{
2597 umode_t mode = file_inode(file)->i_mode;
2598
4503b767 2599 if (S_ISBLK(mode)) {
4e7b5671
CH
2600 if (IS_ENABLED(CONFIG_BLOCK) &&
2601 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
4503b767
JA
2602 return true;
2603 return false;
2604 }
2605 if (S_ISCHR(mode) || S_ISSOCK(mode))
2b188cc1 2606 return true;
4503b767 2607 if (S_ISREG(mode)) {
4e7b5671
CH
2608 if (IS_ENABLED(CONFIG_BLOCK) &&
2609 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
4503b767
JA
2610 file->f_op != &io_uring_fops)
2611 return true;
2612 return false;
2613 }
2b188cc1 2614
c5b85625
JA
2615 /* any ->read/write should understand O_NONBLOCK */
2616 if (file->f_flags & O_NONBLOCK)
2617 return true;
2618
af197f50
JA
2619 if (!(file->f_mode & FMODE_NOWAIT))
2620 return false;
2621
2622 if (rw == READ)
2623 return file->f_op->read_iter != NULL;
2624
2625 return file->f_op->write_iter != NULL;
2b188cc1
JA
2626}
2627
a88fc400 2628static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2b188cc1 2629{
def596e9 2630 struct io_ring_ctx *ctx = req->ctx;
9adbd45d 2631 struct kiocb *kiocb = &req->rw.kiocb;
75c668cd 2632 struct file *file = req->file;
09bb8394
JA
2633 unsigned ioprio;
2634 int ret;
2b188cc1 2635
75c668cd 2636 if (S_ISREG(file_inode(file)->i_mode))
491381ce
JA
2637 req->flags |= REQ_F_ISREG;
2638
2b188cc1 2639 kiocb->ki_pos = READ_ONCE(sqe->off);
75c668cd 2640 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
ba04291e 2641 req->flags |= REQ_F_CUR_POS;
75c668cd 2642 kiocb->ki_pos = file->f_pos;
ba04291e 2643 }
2b188cc1 2644 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
3e577dcd
PB
2645 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2646 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2647 if (unlikely(ret))
2648 return ret;
2b188cc1 2649
75c668cd
PB
2650 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2651 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2652 req->flags |= REQ_F_NOWAIT;
2653
2b188cc1
JA
2654 ioprio = READ_ONCE(sqe->ioprio);
2655 if (ioprio) {
2656 ret = ioprio_check_cap(ioprio);
2657 if (ret)
09bb8394 2658 return ret;
2b188cc1
JA
2659
2660 kiocb->ki_ioprio = ioprio;
2661 } else
2662 kiocb->ki_ioprio = get_current_ioprio();
2663
def596e9 2664 if (ctx->flags & IORING_SETUP_IOPOLL) {
def596e9
JA
2665 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2666 !kiocb->ki_filp->f_op->iopoll)
09bb8394 2667 return -EOPNOTSUPP;
2b188cc1 2668
def596e9
JA
2669 kiocb->ki_flags |= IOCB_HIPRI;
2670 kiocb->ki_complete = io_complete_rw_iopoll;
65a6543d 2671 req->iopoll_completed = 0;
def596e9 2672 } else {
09bb8394
JA
2673 if (kiocb->ki_flags & IOCB_HIPRI)
2674 return -EINVAL;
def596e9
JA
2675 kiocb->ki_complete = io_complete_rw;
2676 }
9adbd45d 2677
3529d8c2
JA
2678 req->rw.addr = READ_ONCE(sqe->addr);
2679 req->rw.len = READ_ONCE(sqe->len);
4f4eeba8 2680 req->buf_index = READ_ONCE(sqe->buf_index);
2b188cc1 2681 return 0;
2b188cc1
JA
2682}
2683
2684static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2685{
2686 switch (ret) {
2687 case -EIOCBQUEUED:
2688 break;
2689 case -ERESTARTSYS:
2690 case -ERESTARTNOINTR:
2691 case -ERESTARTNOHAND:
2692 case -ERESTART_RESTARTBLOCK:
2693 /*
2694 * We can't just restart the syscall, since previously
2695 * submitted sqes may already be in progress. Just fail this
2696 * IO with EINTR.
2697 */
2698 ret = -EINTR;
df561f66 2699 fallthrough;
2b188cc1
JA
2700 default:
2701 kiocb->ki_complete(kiocb, ret, 0);
2702 }
2703}
2704
a1d7c393 2705static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
889fca73 2706 unsigned int issue_flags)
ba816ad6 2707{
ba04291e 2708 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
e8c2bc1f 2709 struct io_async_rw *io = req->async_data;
ba04291e 2710
227c0c96 2711 /* add previously done IO, if any */
e8c2bc1f 2712 if (io && io->bytes_done > 0) {
227c0c96 2713 if (ret < 0)
e8c2bc1f 2714 ret = io->bytes_done;
227c0c96 2715 else
e8c2bc1f 2716 ret += io->bytes_done;
227c0c96
JA
2717 }
2718
ba04291e
JA
2719 if (req->flags & REQ_F_CUR_POS)
2720 req->file->f_pos = kiocb->ki_pos;
bcaec089 2721 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
889fca73 2722 __io_complete_rw(req, ret, 0, issue_flags);
ba816ad6
JA
2723 else
2724 io_rw_done(kiocb, ret);
2725}
2726
847595de 2727static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
edafccee 2728{
9adbd45d
JA
2729 struct io_ring_ctx *ctx = req->ctx;
2730 size_t len = req->rw.len;
edafccee 2731 struct io_mapped_ubuf *imu;
4be1c615 2732 u16 index, buf_index = req->buf_index;
edafccee
JA
2733 size_t offset;
2734 u64 buf_addr;
2735
edafccee
JA
2736 if (unlikely(buf_index >= ctx->nr_user_bufs))
2737 return -EFAULT;
edafccee
JA
2738 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2739 imu = &ctx->user_bufs[index];
9adbd45d 2740 buf_addr = req->rw.addr;
edafccee
JA
2741
2742 /* overflow */
2743 if (buf_addr + len < buf_addr)
2744 return -EFAULT;
2745 /* not inside the mapped region */
2746 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2747 return -EFAULT;
2748
2749 /*
2750 * May not be a start of buffer, set size appropriately
2751 * and advance us to the beginning.
2752 */
2753 offset = buf_addr - imu->ubuf;
2754 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
bd11b3a3
JA
2755
2756 if (offset) {
2757 /*
2758 * Don't use iov_iter_advance() here, as it's really slow for
2759 * using the latter parts of a big fixed buffer - it iterates
2760 * over each segment manually. We can cheat a bit here, because
2761 * we know that:
2762 *
2763 * 1) it's a BVEC iter, we set it up
2764 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2765 * first and last bvec
2766 *
2767 * So just find our index, and adjust the iterator afterwards.
2768 * If the offset is within the first bvec (or the whole first
2769 * bvec, just use iov_iter_advance(). This makes it easier
2770 * since we can just skip the first segment, which may not
2771 * be PAGE_SIZE aligned.
2772 */
2773 const struct bio_vec *bvec = imu->bvec;
2774
2775 if (offset <= bvec->bv_len) {
2776 iov_iter_advance(iter, offset);
2777 } else {
2778 unsigned long seg_skip;
2779
2780 /* skip first vec */
2781 offset -= bvec->bv_len;
2782 seg_skip = 1 + (offset >> PAGE_SHIFT);
2783
2784 iter->bvec = bvec + seg_skip;
2785 iter->nr_segs -= seg_skip;
99c79f66 2786 iter->count -= bvec->bv_len + offset;
bd11b3a3 2787 iter->iov_offset = offset & ~PAGE_MASK;
bd11b3a3
JA
2788 }
2789 }
2790
847595de 2791 return 0;
edafccee
JA
2792}
2793
bcda7baa
JA
2794static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2795{
2796 if (needs_lock)
2797 mutex_unlock(&ctx->uring_lock);
2798}
2799
2800static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2801{
2802 /*
2803 * "Normal" inline submissions always hold the uring_lock, since we
2804 * grab it from the system call. Same is true for the SQPOLL offload.
2805 * The only exception is when we've detached the request and issue it
2806 * from an async worker thread, grab the lock for that case.
2807 */
2808 if (needs_lock)
2809 mutex_lock(&ctx->uring_lock);
2810}
2811
2812static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2813 int bgid, struct io_buffer *kbuf,
2814 bool needs_lock)
2815{
2816 struct io_buffer *head;
2817
2818 if (req->flags & REQ_F_BUFFER_SELECTED)
2819 return kbuf;
2820
2821 io_ring_submit_lock(req->ctx, needs_lock);
2822
2823 lockdep_assert_held(&req->ctx->uring_lock);
2824
2825 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2826 if (head) {
2827 if (!list_empty(&head->list)) {
2828 kbuf = list_last_entry(&head->list, struct io_buffer,
2829 list);
2830 list_del(&kbuf->list);
2831 } else {
2832 kbuf = head;
2833 idr_remove(&req->ctx->io_buffer_idr, bgid);
2834 }
2835 if (*len > kbuf->len)
2836 *len = kbuf->len;
2837 } else {
2838 kbuf = ERR_PTR(-ENOBUFS);
2839 }
2840
2841 io_ring_submit_unlock(req->ctx, needs_lock);
2842
2843 return kbuf;
2844}
2845
4d954c25
JA
2846static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2847 bool needs_lock)
2848{
2849 struct io_buffer *kbuf;
4f4eeba8 2850 u16 bgid;
4d954c25
JA
2851
2852 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
4f4eeba8 2853 bgid = req->buf_index;
4d954c25
JA
2854 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2855 if (IS_ERR(kbuf))
2856 return kbuf;
2857 req->rw.addr = (u64) (unsigned long) kbuf;
2858 req->flags |= REQ_F_BUFFER_SELECTED;
2859 return u64_to_user_ptr(kbuf->addr);
2860}
2861
2862#ifdef CONFIG_COMPAT
2863static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2864 bool needs_lock)
2865{
2866 struct compat_iovec __user *uiov;
2867 compat_ssize_t clen;
2868 void __user *buf;
2869 ssize_t len;
2870
2871 uiov = u64_to_user_ptr(req->rw.addr);
2872 if (!access_ok(uiov, sizeof(*uiov)))
2873 return -EFAULT;
2874 if (__get_user(clen, &uiov->iov_len))
2875 return -EFAULT;
2876 if (clen < 0)
2877 return -EINVAL;
2878
2879 len = clen;
2880 buf = io_rw_buffer_select(req, &len, needs_lock);
2881 if (IS_ERR(buf))
2882 return PTR_ERR(buf);
2883 iov[0].iov_base = buf;
2884 iov[0].iov_len = (compat_size_t) len;
2885 return 0;
2886}
2887#endif
2888
2889static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2890 bool needs_lock)
2891{
2892 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2893 void __user *buf;
2894 ssize_t len;
2895
2896 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2897 return -EFAULT;
2898
2899 len = iov[0].iov_len;
2900 if (len < 0)
2901 return -EINVAL;
2902 buf = io_rw_buffer_select(req, &len, needs_lock);
2903 if (IS_ERR(buf))
2904 return PTR_ERR(buf);
2905 iov[0].iov_base = buf;
2906 iov[0].iov_len = len;
2907 return 0;
2908}
2909
2910static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2911 bool needs_lock)
2912{
dddb3e26
JA
2913 if (req->flags & REQ_F_BUFFER_SELECTED) {
2914 struct io_buffer *kbuf;
2915
2916 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2917 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2918 iov[0].iov_len = kbuf->len;
4d954c25 2919 return 0;
dddb3e26 2920 }
dd201662 2921 if (req->rw.len != 1)
4d954c25
JA
2922 return -EINVAL;
2923
2924#ifdef CONFIG_COMPAT
2925 if (req->ctx->compat)
2926 return io_compat_import(req, iov, needs_lock);
2927#endif
2928
2929 return __io_iov_buffer_select(req, iov, needs_lock);
2930}
2931
847595de
PB
2932static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2933 struct iov_iter *iter, bool needs_lock)
2b188cc1 2934{
9adbd45d
JA
2935 void __user *buf = u64_to_user_ptr(req->rw.addr);
2936 size_t sqe_len = req->rw.len;
847595de 2937 u8 opcode = req->opcode;
4d954c25 2938 ssize_t ret;
edafccee 2939
7d009165 2940 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
edafccee 2941 *iovec = NULL;
9adbd45d 2942 return io_import_fixed(req, rw, iter);
edafccee 2943 }
2b188cc1 2944
bcda7baa 2945 /* buffer index only valid with fixed read/write, or buffer select */
4f4eeba8 2946 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
9adbd45d
JA
2947 return -EINVAL;
2948
3a6820f2 2949 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
bcda7baa 2950 if (req->flags & REQ_F_BUFFER_SELECT) {
4d954c25 2951 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
867a23ea 2952 if (IS_ERR(buf))
4d954c25 2953 return PTR_ERR(buf);
3f9d6441 2954 req->rw.len = sqe_len;
bcda7baa
JA
2955 }
2956
3a6820f2
JA
2957 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2958 *iovec = NULL;
10fc72e4 2959 return ret;
3a6820f2
JA
2960 }
2961
4d954c25
JA
2962 if (req->flags & REQ_F_BUFFER_SELECT) {
2963 ret = io_iov_buffer_select(req, *iovec, needs_lock);
847595de
PB
2964 if (!ret)
2965 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
4d954c25
JA
2966 *iovec = NULL;
2967 return ret;
2968 }
2969
89cd35c5
CH
2970 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
2971 req->ctx->compat);
2b188cc1
JA
2972}
2973
0fef9483
JA
2974static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
2975{
5b09e37e 2976 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
0fef9483
JA
2977}
2978
31b51510 2979/*
32960613
JA
2980 * For files that don't have ->read_iter() and ->write_iter(), handle them
2981 * by looping over ->read() or ->write() manually.
31b51510 2982 */
4017eb91 2983static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
32960613 2984{
4017eb91
JA
2985 struct kiocb *kiocb = &req->rw.kiocb;
2986 struct file *file = req->file;
32960613
JA
2987 ssize_t ret = 0;
2988
2989 /*
2990 * Don't support polled IO through this interface, and we can't
2991 * support non-blocking either. For the latter, this just causes
2992 * the kiocb to be handled from an async context.
2993 */
2994 if (kiocb->ki_flags & IOCB_HIPRI)
2995 return -EOPNOTSUPP;
2996 if (kiocb->ki_flags & IOCB_NOWAIT)
2997 return -EAGAIN;
2998
2999 while (iov_iter_count(iter)) {
311ae9e1 3000 struct iovec iovec;
32960613
JA
3001 ssize_t nr;
3002
311ae9e1
PB
3003 if (!iov_iter_is_bvec(iter)) {
3004 iovec = iov_iter_iovec(iter);
3005 } else {
4017eb91
JA
3006 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3007 iovec.iov_len = req->rw.len;
311ae9e1
PB
3008 }
3009
32960613
JA
3010 if (rw == READ) {
3011 nr = file->f_op->read(file, iovec.iov_base,
0fef9483 3012 iovec.iov_len, io_kiocb_ppos(kiocb));
32960613
JA
3013 } else {
3014 nr = file->f_op->write(file, iovec.iov_base,
0fef9483 3015 iovec.iov_len, io_kiocb_ppos(kiocb));
32960613
JA
3016 }
3017
3018 if (nr < 0) {
3019 if (!ret)
3020 ret = nr;
3021 break;
3022 }
3023 ret += nr;
3024 if (nr != iovec.iov_len)
3025 break;
4017eb91
JA
3026 req->rw.len -= nr;
3027 req->rw.addr += nr;
32960613
JA
3028 iov_iter_advance(iter, nr);
3029 }
3030
3031 return ret;
3032}
3033
ff6165b2
JA
3034static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3035 const struct iovec *fast_iov, struct iov_iter *iter)
f67676d1 3036{
e8c2bc1f 3037 struct io_async_rw *rw = req->async_data;
b64e3444 3038
ff6165b2 3039 memcpy(&rw->iter, iter, sizeof(*iter));
afb87658 3040 rw->free_iovec = iovec;
227c0c96 3041 rw->bytes_done = 0;
ff6165b2 3042 /* can only be fixed buffers, no need to do anything */
9c3a205c 3043 if (iov_iter_is_bvec(iter))
ff6165b2 3044 return;
b64e3444 3045 if (!iovec) {
ff6165b2
JA
3046 unsigned iov_off = 0;
3047
3048 rw->iter.iov = rw->fast_iov;
3049 if (iter->iov != fast_iov) {
3050 iov_off = iter->iov - fast_iov;
3051 rw->iter.iov += iov_off;
3052 }
3053 if (rw->fast_iov != fast_iov)
3054 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
45097dae 3055 sizeof(struct iovec) * iter->nr_segs);
99bc4c38
PB
3056 } else {
3057 req->flags |= REQ_F_NEED_CLEANUP;
f67676d1
JA
3058 }
3059}
3060
e8c2bc1f 3061static inline int __io_alloc_async_data(struct io_kiocb *req)
3d9932a8 3062{
e8c2bc1f
JA
3063 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3064 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3065 return req->async_data == NULL;
3d9932a8
XW
3066}
3067
e8c2bc1f 3068static int io_alloc_async_data(struct io_kiocb *req)
f67676d1 3069{
e8c2bc1f 3070 if (!io_op_defs[req->opcode].needs_async_data)
d3656344 3071 return 0;
3d9932a8 3072
e8c2bc1f 3073 return __io_alloc_async_data(req);
b7bb4f7d
JA
3074}
3075
ff6165b2
JA
3076static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3077 const struct iovec *fast_iov,
227c0c96 3078 struct iov_iter *iter, bool force)
b7bb4f7d 3079{
e8c2bc1f 3080 if (!force && !io_op_defs[req->opcode].needs_async_data)
74566df3 3081 return 0;
e8c2bc1f 3082 if (!req->async_data) {
6bf985dc
PB
3083 if (__io_alloc_async_data(req)) {
3084 kfree(iovec);
5d204bcf 3085 return -ENOMEM;
6bf985dc 3086 }
b7bb4f7d 3087
ff6165b2 3088 io_req_map_rw(req, iovec, fast_iov, iter);
5d204bcf 3089 }
b7bb4f7d 3090 return 0;
f67676d1
JA
3091}
3092
73debe68 3093static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
c3e330a4 3094{
e8c2bc1f 3095 struct io_async_rw *iorw = req->async_data;
f4bff104 3096 struct iovec *iov = iorw->fast_iov;
847595de 3097 int ret;
c3e330a4 3098
2846c481 3099 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
c3e330a4
PB
3100 if (unlikely(ret < 0))
3101 return ret;
3102
ab0b196c
PB
3103 iorw->bytes_done = 0;
3104 iorw->free_iovec = iov;
3105 if (iov)
3106 req->flags |= REQ_F_NEED_CLEANUP;
c3e330a4
PB
3107 return 0;
3108}
3109
73debe68 3110static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
f67676d1 3111{
3529d8c2
JA
3112 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3113 return -EBADF;
93642ef8 3114 return io_prep_rw(req, sqe);
f67676d1
JA
3115}
3116
c1dd91d1
JA
3117/*
3118 * This is our waitqueue callback handler, registered through lock_page_async()
3119 * when we initially tried to do the IO with the iocb armed our waitqueue.
3120 * This gets called when the page is unlocked, and we generally expect that to
3121 * happen when the page IO is completed and the page is now uptodate. This will
3122 * queue a task_work based retry of the operation, attempting to copy the data
3123 * again. If the latter fails because the page was NOT uptodate, then we will
3124 * do a thread based blocking retry of the operation. That's the unexpected
3125 * slow path.
3126 */
bcf5a063
JA
3127static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3128 int sync, void *arg)
3129{
3130 struct wait_page_queue *wpq;
3131 struct io_kiocb *req = wait->private;
bcf5a063 3132 struct wait_page_key *key = arg;
bcf5a063
JA
3133
3134 wpq = container_of(wait, struct wait_page_queue, wait);
3135
cdc8fcb4
LT
3136 if (!wake_page_match(wpq, key))
3137 return 0;
3138
c8d317aa 3139 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
bcf5a063
JA
3140 list_del_init(&wait->entry);
3141
bcf5a063
JA
3142 /* submit ref gets dropped, acquire a new one */
3143 refcount_inc(&req->refs);
921b9054 3144 io_req_task_queue(req);
bcf5a063
JA
3145 return 1;
3146}
3147
c1dd91d1
JA
3148/*
3149 * This controls whether a given IO request should be armed for async page
3150 * based retry. If we return false here, the request is handed to the async
3151 * worker threads for retry. If we're doing buffered reads on a regular file,
3152 * we prepare a private wait_page_queue entry and retry the operation. This
3153 * will either succeed because the page is now uptodate and unlocked, or it
3154 * will register a callback when the page is unlocked at IO completion. Through
3155 * that callback, io_uring uses task_work to setup a retry of the operation.
3156 * That retry will attempt the buffered read again. The retry will generally
3157 * succeed, or in rare cases where it fails, we then fall back to using the
3158 * async worker threads for a blocking retry.
3159 */
227c0c96 3160static bool io_rw_should_retry(struct io_kiocb *req)
f67676d1 3161{
e8c2bc1f
JA
3162 struct io_async_rw *rw = req->async_data;
3163 struct wait_page_queue *wait = &rw->wpq;
bcf5a063 3164 struct kiocb *kiocb = &req->rw.kiocb;
f67676d1 3165
bcf5a063
JA
3166 /* never retry for NOWAIT, we just complete with -EAGAIN */
3167 if (req->flags & REQ_F_NOWAIT)
3168 return false;
f67676d1 3169
227c0c96 3170 /* Only for buffered IO */
3b2a4439 3171 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
bcf5a063 3172 return false;
3b2a4439 3173
bcf5a063
JA
3174 /*
3175 * just use poll if we can, and don't attempt if the fs doesn't
3176 * support callback based unlocks
3177 */
3178 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3179 return false;
f67676d1 3180
3b2a4439
JA
3181 wait->wait.func = io_async_buf_func;
3182 wait->wait.private = req;
3183 wait->wait.flags = 0;
3184 INIT_LIST_HEAD(&wait->wait.entry);
3185 kiocb->ki_flags |= IOCB_WAITQ;
c8d317aa 3186 kiocb->ki_flags &= ~IOCB_NOWAIT;
3b2a4439 3187 kiocb->ki_waitq = wait;
3b2a4439 3188 return true;
bcf5a063
JA
3189}
3190
3191static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3192{
3193 if (req->file->f_op->read_iter)
3194 return call_read_iter(req->file, &req->rw.kiocb, iter);
2dd2111d 3195 else if (req->file->f_op->read)
4017eb91 3196 return loop_rw_iter(READ, req, iter);
2dd2111d
GH
3197 else
3198 return -EINVAL;
f67676d1
JA
3199}
3200
889fca73 3201static int io_read(struct io_kiocb *req, unsigned int issue_flags)
2b188cc1
JA
3202{
3203 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
9adbd45d 3204 struct kiocb *kiocb = &req->rw.kiocb;
ff6165b2 3205 struct iov_iter __iter, *iter = &__iter;
e8c2bc1f 3206 struct io_async_rw *rw = req->async_data;
227c0c96 3207 ssize_t io_size, ret, ret2;
45d189c6 3208 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
ff6165b2 3209
2846c481 3210 if (rw) {
e8c2bc1f 3211 iter = &rw->iter;
2846c481
PB
3212 iovec = NULL;
3213 } else {
3214 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3215 if (ret < 0)
3216 return ret;
3217 }
632546c4 3218 io_size = iov_iter_count(iter);
fa15bafb 3219 req->result = io_size;
2b188cc1 3220
fd6c2e4c
JA
3221 /* Ensure we clear previously set non-block flag */
3222 if (!force_nonblock)
29de5f6a 3223 kiocb->ki_flags &= ~IOCB_NOWAIT;
a88fc400
PB
3224 else
3225 kiocb->ki_flags |= IOCB_NOWAIT;
3226
24c74678 3227 /* If the file doesn't support async, just async punt */
6713e7a6
PB
3228 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3229 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
6bf985dc 3230 return ret ?: -EAGAIN;
6713e7a6 3231 }
9e645e11 3232
632546c4 3233 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
5ea5dd45
PB
3234 if (unlikely(ret)) {
3235 kfree(iovec);
3236 return ret;
3237 }
2b188cc1 3238
227c0c96 3239 ret = io_iter_do_read(req, iter);
32960613 3240
57cd657b 3241 if (ret == -EIOCBQUEUED) {
3e6a0d3c
JA
3242 if (req->async_data)
3243 iov_iter_revert(iter, io_size - iov_iter_count(iter));
fe1cdd55 3244 goto out_free;
227c0c96 3245 } else if (ret == -EAGAIN) {
eefdf30f
JA
3246 /* IOPOLL retry should happen for io-wq threads */
3247 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
f91daf56 3248 goto done;
75c668cd
PB
3249 /* no retry on NONBLOCK nor RWF_NOWAIT */
3250 if (req->flags & REQ_F_NOWAIT)
355afaeb 3251 goto done;
84216315 3252 /* some cases will consume bytes even on error returns */
632546c4 3253 iov_iter_revert(iter, io_size - iov_iter_count(iter));
f38c7e3a 3254 ret = 0;
7335e3bf 3255 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
75c668cd 3256 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
7335e3bf 3257 /* read all, failed, already did sync or don't want to retry */
00d23d51 3258 goto done;
227c0c96
JA
3259 }
3260
227c0c96 3261 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
6bf985dc
PB
3262 if (ret2)
3263 return ret2;
3264
fe1cdd55 3265 iovec = NULL;
e8c2bc1f 3266 rw = req->async_data;
227c0c96 3267 /* now use our persistent iterator, if we aren't already */
e8c2bc1f 3268 iter = &rw->iter;
227c0c96 3269
b23df91b
PB
3270 do {
3271 io_size -= ret;
3272 rw->bytes_done += ret;
3273 /* if we can retry, do so with the callbacks armed */
3274 if (!io_rw_should_retry(req)) {
3275 kiocb->ki_flags &= ~IOCB_WAITQ;
3276 return -EAGAIN;
3277 }
3278
3279 /*
3280 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3281 * we get -EIOCBQUEUED, then we'll get a notification when the
3282 * desired page gets unlocked. We can also get a partial read
3283 * here, and if we do, then just retry at the new offset.
3284 */
3285 ret = io_iter_do_read(req, iter);
3286 if (ret == -EIOCBQUEUED)
3287 return 0;
227c0c96 3288 /* we got some bytes, but not all. retry. */
b5b0ecb7 3289 kiocb->ki_flags &= ~IOCB_WAITQ;
b23df91b 3290 } while (ret > 0 && ret < io_size);
227c0c96 3291done:
889fca73 3292 kiocb_done(kiocb, ret, issue_flags);
fe1cdd55
PB
3293out_free:
3294 /* it's faster to check here then delegate to kfree */
3295 if (iovec)
3296 kfree(iovec);
5ea5dd45 3297 return 0;
2b188cc1
JA
3298}
3299
73debe68 3300static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
f67676d1 3301{
3529d8c2
JA
3302 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3303 return -EBADF;
93642ef8 3304 return io_prep_rw(req, sqe);
f67676d1
JA
3305}
3306
889fca73 3307static int io_write(struct io_kiocb *req, unsigned int issue_flags)
2b188cc1
JA
3308{
3309 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
9adbd45d 3310 struct kiocb *kiocb = &req->rw.kiocb;
ff6165b2 3311 struct iov_iter __iter, *iter = &__iter;
e8c2bc1f 3312 struct io_async_rw *rw = req->async_data;
fa15bafb 3313 ssize_t ret, ret2, io_size;
45d189c6 3314 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
2b188cc1 3315
2846c481 3316 if (rw) {
e8c2bc1f 3317 iter = &rw->iter;
2846c481
PB
3318 iovec = NULL;
3319 } else {
3320 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3321 if (ret < 0)
3322 return ret;
3323 }
632546c4 3324 io_size = iov_iter_count(iter);
fa15bafb 3325 req->result = io_size;
2b188cc1 3326
fd6c2e4c
JA
3327 /* Ensure we clear previously set non-block flag */
3328 if (!force_nonblock)
a88fc400
PB
3329 kiocb->ki_flags &= ~IOCB_NOWAIT;
3330 else
3331 kiocb->ki_flags |= IOCB_NOWAIT;
fd6c2e4c 3332
24c74678 3333 /* If the file doesn't support async, just async punt */
af197f50 3334 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
f67676d1 3335 goto copy_iov;
31b51510 3336
10d59345
JA
3337 /* file path doesn't support NOWAIT for non-direct_IO */
3338 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3339 (req->flags & REQ_F_ISREG))
f67676d1 3340 goto copy_iov;
31b51510 3341
632546c4 3342 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
fa15bafb
PB
3343 if (unlikely(ret))
3344 goto out_free;
4ed734b0 3345
fa15bafb
PB
3346 /*
3347 * Open-code file_start_write here to grab freeze protection,
3348 * which will be released by another thread in
3349 * io_complete_rw(). Fool lockdep by telling it the lock got
3350 * released so that it doesn't complain about the held lock when
3351 * we return to userspace.
3352 */
3353 if (req->flags & REQ_F_ISREG) {
8a3c84b6 3354 sb_start_write(file_inode(req->file)->i_sb);
fa15bafb
PB
3355 __sb_writers_release(file_inode(req->file)->i_sb,
3356 SB_FREEZE_WRITE);
3357 }
3358 kiocb->ki_flags |= IOCB_WRITE;
4ed734b0 3359
fa15bafb 3360 if (req->file->f_op->write_iter)
ff6165b2 3361 ret2 = call_write_iter(req->file, kiocb, iter);
2dd2111d 3362 else if (req->file->f_op->write)
4017eb91 3363 ret2 = loop_rw_iter(WRITE, req, iter);
2dd2111d
GH
3364 else
3365 ret2 = -EINVAL;
4ed734b0 3366
fa15bafb
PB
3367 /*
3368 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3369 * retry them without IOCB_NOWAIT.
3370 */
3371 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3372 ret2 = -EAGAIN;
75c668cd
PB
3373 /* no retry on NONBLOCK nor RWF_NOWAIT */
3374 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
355afaeb 3375 goto done;
3e6a0d3c
JA
3376 if (ret2 == -EIOCBQUEUED && req->async_data)
3377 iov_iter_revert(iter, io_size - iov_iter_count(iter));
fa15bafb 3378 if (!force_nonblock || ret2 != -EAGAIN) {
eefdf30f
JA
3379 /* IOPOLL retry should happen for io-wq threads */
3380 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3381 goto copy_iov;
355afaeb 3382done:
889fca73 3383 kiocb_done(kiocb, ret2, issue_flags);
fa15bafb 3384 } else {
f67676d1 3385copy_iov:
84216315 3386 /* some cases will consume bytes even on error returns */
632546c4 3387 iov_iter_revert(iter, io_size - iov_iter_count(iter));
227c0c96 3388 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
6bf985dc 3389 return ret ?: -EAGAIN;
2b188cc1 3390 }
31b51510 3391out_free:
f261c168 3392 /* it's reportedly faster than delegating the null check to kfree() */
252917c3 3393 if (iovec)
6f2cc166 3394 kfree(iovec);
2b188cc1
JA
3395 return ret;
3396}
3397
80a261fd
JA
3398static int io_renameat_prep(struct io_kiocb *req,
3399 const struct io_uring_sqe *sqe)
3400{
3401 struct io_rename *ren = &req->rename;
3402 const char __user *oldf, *newf;
3403
3404 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3405 return -EBADF;
3406
3407 ren->old_dfd = READ_ONCE(sqe->fd);
3408 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3409 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3410 ren->new_dfd = READ_ONCE(sqe->len);
3411 ren->flags = READ_ONCE(sqe->rename_flags);
3412
3413 ren->oldpath = getname(oldf);
3414 if (IS_ERR(ren->oldpath))
3415 return PTR_ERR(ren->oldpath);
3416
3417 ren->newpath = getname(newf);
3418 if (IS_ERR(ren->newpath)) {
3419 putname(ren->oldpath);
3420 return PTR_ERR(ren->newpath);
3421 }
3422
3423 req->flags |= REQ_F_NEED_CLEANUP;
3424 return 0;
3425}
3426
45d189c6 3427static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
80a261fd
JA
3428{
3429 struct io_rename *ren = &req->rename;
3430 int ret;
3431
45d189c6 3432 if (issue_flags & IO_URING_F_NONBLOCK)
80a261fd
JA
3433 return -EAGAIN;
3434
3435 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3436 ren->newpath, ren->flags);
3437
3438 req->flags &= ~REQ_F_NEED_CLEANUP;
3439 if (ret < 0)
3440 req_set_fail_links(req);
3441 io_req_complete(req, ret);
3442 return 0;
3443}
3444
14a1143b
JA
3445static int io_unlinkat_prep(struct io_kiocb *req,
3446 const struct io_uring_sqe *sqe)
3447{
3448 struct io_unlink *un = &req->unlink;
3449 const char __user *fname;
3450
3451 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3452 return -EBADF;
3453
3454 un->dfd = READ_ONCE(sqe->fd);
3455
3456 un->flags = READ_ONCE(sqe->unlink_flags);
3457 if (un->flags & ~AT_REMOVEDIR)
3458 return -EINVAL;
3459
3460 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3461 un->filename = getname(fname);
3462 if (IS_ERR(un->filename))
3463 return PTR_ERR(un->filename);
3464
3465 req->flags |= REQ_F_NEED_CLEANUP;
3466 return 0;
3467}
3468
45d189c6 3469static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
14a1143b
JA
3470{
3471 struct io_unlink *un = &req->unlink;
3472 int ret;
3473
45d189c6 3474 if (issue_flags & IO_URING_F_NONBLOCK)
14a1143b
JA
3475 return -EAGAIN;
3476
3477 if (un->flags & AT_REMOVEDIR)
3478 ret = do_rmdir(un->dfd, un->filename);
3479 else
3480 ret = do_unlinkat(un->dfd, un->filename);
3481
3482 req->flags &= ~REQ_F_NEED_CLEANUP;
3483 if (ret < 0)
3484 req_set_fail_links(req);
3485 io_req_complete(req, ret);
3486 return 0;
3487}
3488
36f4fa68
JA
3489static int io_shutdown_prep(struct io_kiocb *req,
3490 const struct io_uring_sqe *sqe)
3491{
3492#if defined(CONFIG_NET)
3493 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3494 return -EINVAL;
3495 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3496 sqe->buf_index)
3497 return -EINVAL;
3498
3499 req->shutdown.how = READ_ONCE(sqe->len);
3500 return 0;
3501#else
3502 return -EOPNOTSUPP;
3503#endif
3504}
3505
45d189c6 3506static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
36f4fa68
JA
3507{
3508#if defined(CONFIG_NET)
3509 struct socket *sock;
3510 int ret;
3511
45d189c6 3512 if (issue_flags & IO_URING_F_NONBLOCK)
36f4fa68
JA
3513 return -EAGAIN;
3514
48aba79b 3515 sock = sock_from_file(req->file);
36f4fa68 3516 if (unlikely(!sock))
48aba79b 3517 return -ENOTSOCK;
36f4fa68
JA
3518
3519 ret = __sys_shutdown_sock(sock, req->shutdown.how);
a146468d
JA
3520 if (ret < 0)
3521 req_set_fail_links(req);
36f4fa68
JA
3522 io_req_complete(req, ret);
3523 return 0;
3524#else
3525 return -EOPNOTSUPP;
3526#endif
3527}
3528
f2a8d5c7
PB
3529static int __io_splice_prep(struct io_kiocb *req,
3530 const struct io_uring_sqe *sqe)
7d67af2c
PB
3531{
3532 struct io_splice* sp = &req->splice;
3533 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
7d67af2c 3534
3232dd02
PB
3535 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3536 return -EINVAL;
7d67af2c
PB
3537
3538 sp->file_in = NULL;
7d67af2c
PB
3539 sp->len = READ_ONCE(sqe->len);
3540 sp->flags = READ_ONCE(sqe->splice_flags);
3541
3542 if (unlikely(sp->flags & ~valid_flags))
3543 return -EINVAL;
3544
8371adf5
PB
3545 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3546 (sp->flags & SPLICE_F_FD_IN_FIXED));
3547 if (!sp->file_in)
3548 return -EBADF;
7d67af2c
PB
3549 req->flags |= REQ_F_NEED_CLEANUP;
3550
7cdaf587
XW
3551 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3552 /*
3553 * Splice operation will be punted aync, and here need to
3554 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3555 */
7d67af2c 3556 req->work.flags |= IO_WQ_WORK_UNBOUND;
7cdaf587 3557 }
7d67af2c
PB
3558
3559 return 0;
3560}
3561
f2a8d5c7
PB
3562static int io_tee_prep(struct io_kiocb *req,
3563 const struct io_uring_sqe *sqe)
3564{
3565 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3566 return -EINVAL;
3567 return __io_splice_prep(req, sqe);
3568}
3569
45d189c6 3570static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
f2a8d5c7
PB
3571{
3572 struct io_splice *sp = &req->splice;
3573 struct file *in = sp->file_in;
3574 struct file *out = sp->file_out;
3575 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3576 long ret = 0;
3577
45d189c6 3578 if (issue_flags & IO_URING_F_NONBLOCK)
f2a8d5c7
PB
3579 return -EAGAIN;
3580 if (sp->len)
3581 ret = do_tee(in, out, sp->len, flags);
3582
3583 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3584 req->flags &= ~REQ_F_NEED_CLEANUP;
3585
f2a8d5c7
PB
3586 if (ret != sp->len)
3587 req_set_fail_links(req);
e1e16097 3588 io_req_complete(req, ret);
f2a8d5c7
PB
3589 return 0;
3590}
3591
3592static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3593{
3594 struct io_splice* sp = &req->splice;
3595
3596 sp->off_in = READ_ONCE(sqe->splice_off_in);
3597 sp->off_out = READ_ONCE(sqe->off);
3598 return __io_splice_prep(req, sqe);
3599}
3600
45d189c6 3601static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
7d67af2c
PB
3602{
3603 struct io_splice *sp = &req->splice;
3604 struct file *in = sp->file_in;
3605 struct file *out = sp->file_out;
3606 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3607 loff_t *poff_in, *poff_out;
c9687426 3608 long ret = 0;
7d67af2c 3609
45d189c6 3610 if (issue_flags & IO_URING_F_NONBLOCK)
2fb3e822 3611 return -EAGAIN;
7d67af2c
PB
3612
3613 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3614 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
c9687426 3615
948a7749 3616 if (sp->len)
c9687426 3617 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
7d67af2c
PB
3618
3619 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3620 req->flags &= ~REQ_F_NEED_CLEANUP;
3621
7d67af2c
PB
3622 if (ret != sp->len)
3623 req_set_fail_links(req);
e1e16097 3624 io_req_complete(req, ret);
7d67af2c
PB
3625 return 0;
3626}
3627
2b188cc1
JA
3628/*
3629 * IORING_OP_NOP just posts a completion event, nothing else.
3630 */
889fca73 3631static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
2b188cc1
JA
3632{
3633 struct io_ring_ctx *ctx = req->ctx;
2b188cc1 3634
def596e9
JA
3635 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3636 return -EINVAL;
3637
889fca73 3638 __io_req_complete(req, issue_flags, 0, 0);
2b188cc1
JA
3639 return 0;
3640}
3641
1155c76a 3642static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
c992fe29 3643{
6b06314c 3644 struct io_ring_ctx *ctx = req->ctx;
c992fe29 3645
09bb8394
JA
3646 if (!req->file)
3647 return -EBADF;
c992fe29 3648
6b06314c 3649 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
def596e9 3650 return -EINVAL;
edafccee 3651 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
c992fe29
CH
3652 return -EINVAL;
3653
8ed8d3c3
JA
3654 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3655 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3656 return -EINVAL;
3657
3658 req->sync.off = READ_ONCE(sqe->off);
3659 req->sync.len = READ_ONCE(sqe->len);
c992fe29
CH
3660 return 0;
3661}
3662
45d189c6 3663static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
8ed8d3c3 3664{
8ed8d3c3 3665 loff_t end = req->sync.off + req->sync.len;
8ed8d3c3
JA
3666 int ret;
3667
ac45abc0 3668 /* fsync always requires a blocking context */
45d189c6 3669 if (issue_flags & IO_URING_F_NONBLOCK)
ac45abc0
PB
3670 return -EAGAIN;
3671
9adbd45d 3672 ret = vfs_fsync_range(req->file, req->sync.off,
8ed8d3c3
JA
3673 end > 0 ? end : LLONG_MAX,
3674 req->sync.flags & IORING_FSYNC_DATASYNC);
3675 if (ret < 0)
3676 req_set_fail_links(req);
e1e16097 3677 io_req_complete(req, ret);
c992fe29
CH
3678 return 0;
3679}
3680
d63d1b5e
JA
3681static int io_fallocate_prep(struct io_kiocb *req,
3682 const struct io_uring_sqe *sqe)
3683{
3684 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3685 return -EINVAL;
3232dd02
PB
3686 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3687 return -EINVAL;
d63d1b5e
JA
3688
3689 req->sync.off = READ_ONCE(sqe->off);
3690 req->sync.len = READ_ONCE(sqe->addr);
3691 req->sync.mode = READ_ONCE(sqe->len);
3692 return 0;
3693}
3694
45d189c6 3695static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
5d17b4a4 3696{
ac45abc0
PB
3697 int ret;
3698
d63d1b5e 3699 /* fallocate always requiring blocking context */
45d189c6 3700 if (issue_flags & IO_URING_F_NONBLOCK)
5d17b4a4 3701 return -EAGAIN;
ac45abc0
PB
3702 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3703 req->sync.len);
ac45abc0
PB
3704 if (ret < 0)
3705 req_set_fail_links(req);
e1e16097 3706 io_req_complete(req, ret);
5d17b4a4
JA
3707 return 0;
3708}
3709
ec65fea5 3710static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
b7bb4f7d 3711{
f8748881 3712 const char __user *fname;
15b71abe 3713 int ret;
b7bb4f7d 3714
ec65fea5 3715 if (unlikely(sqe->ioprio || sqe->buf_index))
15b71abe 3716 return -EINVAL;
ec65fea5 3717 if (unlikely(req->flags & REQ_F_FIXED_FILE))
cf3040ca 3718 return -EBADF;
03b1230c 3719
ec65fea5
PB
3720 /* open.how should be already initialised */
3721 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
08a1d26e 3722 req->open.how.flags |= O_LARGEFILE;
3529d8c2 3723
25e72d10
PB
3724 req->open.dfd = READ_ONCE(sqe->fd);
3725 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
f8748881 3726 req->open.filename = getname(fname);
15b71abe
JA
3727 if (IS_ERR(req->open.filename)) {
3728 ret = PTR_ERR(req->open.filename);
3729 req->open.filename = NULL;
3730 return ret;
3731 }
4022e7af 3732 req->open.nofile = rlimit(RLIMIT_NOFILE);
8fef80bf 3733 req->flags |= REQ_F_NEED_CLEANUP;
15b71abe 3734 return 0;
03b1230c
JA
3735}
3736
ec65fea5
PB
3737static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3738{
3739 u64 flags, mode;
3740
14587a46 3741 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4eb8dded 3742 return -EINVAL;
ec65fea5
PB
3743 mode = READ_ONCE(sqe->len);
3744 flags = READ_ONCE(sqe->open_flags);
3745 req->open.how = build_open_how(flags, mode);
3746 return __io_openat_prep(req, sqe);
3747}
3748
cebdb986 3749static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
aa1fa28f 3750{
cebdb986 3751 struct open_how __user *how;
cebdb986 3752 size_t len;
0fa03c62
JA
3753 int ret;
3754
14587a46 3755 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4eb8dded 3756 return -EINVAL;
cebdb986
JA
3757 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3758 len = READ_ONCE(sqe->len);
cebdb986
JA
3759 if (len < OPEN_HOW_SIZE_VER0)
3760 return -EINVAL;
3529d8c2 3761
cebdb986
JA
3762 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3763 len);
3764 if (ret)
3765 return ret;
3529d8c2 3766
ec65fea5 3767 return __io_openat_prep(req, sqe);
cebdb986
JA
3768}
3769
45d189c6 3770static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
15b71abe
JA
3771{
3772 struct open_flags op;
15b71abe 3773 struct file *file;
3a81fd02
JA
3774 bool nonblock_set;
3775 bool resolve_nonblock;
15b71abe
JA
3776 int ret;
3777
cebdb986 3778 ret = build_open_flags(&req->open.how, &op);
15b71abe
JA
3779 if (ret)
3780 goto err;
3a81fd02
JA
3781 nonblock_set = op.open_flag & O_NONBLOCK;
3782 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
45d189c6 3783 if (issue_flags & IO_URING_F_NONBLOCK) {
3a81fd02
JA
3784 /*
3785 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3786 * it'll always -EAGAIN
3787 */
3788 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3789 return -EAGAIN;
3790 op.lookup_flags |= LOOKUP_CACHED;
3791 op.open_flag |= O_NONBLOCK;
3792 }
15b71abe 3793
4022e7af 3794 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
15b71abe
JA
3795 if (ret < 0)
3796 goto err;
3797
3798 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3a81fd02 3799 /* only retry if RESOLVE_CACHED wasn't already set by application */
45d189c6
PB
3800 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3801 file == ERR_PTR(-EAGAIN)) {
944d1444 3802 /*
3a81fd02
JA
3803 * We could hang on to this 'fd', but seems like marginal
3804 * gain for something that is now known to be a slower path.
3805 * So just put it, and we'll get a new one when we retry.
944d1444 3806 */
3a81fd02
JA
3807 put_unused_fd(ret);
3808 return -EAGAIN;
3809 }
3810
15b71abe
JA
3811 if (IS_ERR(file)) {
3812 put_unused_fd(ret);
3813 ret = PTR_ERR(file);
3814 } else {
45d189c6 3815 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3a81fd02 3816 file->f_flags &= ~O_NONBLOCK;
15b71abe
JA
3817 fsnotify_open(file);
3818 fd_install(ret, file);
3819 }
3820err:
3821 putname(req->open.filename);
8fef80bf 3822 req->flags &= ~REQ_F_NEED_CLEANUP;
15b71abe
JA
3823 if (ret < 0)
3824 req_set_fail_links(req);
e1e16097 3825 io_req_complete(req, ret);
15b71abe
JA
3826 return 0;
3827}
3828
45d189c6 3829static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
cebdb986 3830{
45d189c6 3831 return io_openat2(req, issue_flags & IO_URING_F_NONBLOCK);
cebdb986
JA
3832}
3833
067524e9
JA
3834static int io_remove_buffers_prep(struct io_kiocb *req,
3835 const struct io_uring_sqe *sqe)
3836{
3837 struct io_provide_buf *p = &req->pbuf;
3838 u64 tmp;
3839
3840 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3841 return -EINVAL;
3842
3843 tmp = READ_ONCE(sqe->fd);
3844 if (!tmp || tmp > USHRT_MAX)
3845 return -EINVAL;
3846
3847 memset(p, 0, sizeof(*p));
3848 p->nbufs = tmp;
3849 p->bgid = READ_ONCE(sqe->buf_group);
3850 return 0;
3851}
3852
3853static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3854 int bgid, unsigned nbufs)
3855{
3856 unsigned i = 0;
3857
3858 /* shouldn't happen */
3859 if (!nbufs)
3860 return 0;
3861
3862 /* the head kbuf is the list itself */
3863 while (!list_empty(&buf->list)) {
3864 struct io_buffer *nxt;
3865
3866 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3867 list_del(&nxt->list);
3868 kfree(nxt);
3869 if (++i == nbufs)
3870 return i;
3871 }
3872 i++;
3873 kfree(buf);
3874 idr_remove(&ctx->io_buffer_idr, bgid);
3875
3876 return i;
3877}
3878
889fca73 3879static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
067524e9
JA
3880{
3881 struct io_provide_buf *p = &req->pbuf;
3882 struct io_ring_ctx *ctx = req->ctx;
3883 struct io_buffer *head;
3884 int ret = 0;
45d189c6 3885 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
067524e9
JA
3886
3887 io_ring_submit_lock(ctx, !force_nonblock);
3888
3889 lockdep_assert_held(&ctx->uring_lock);
3890
3891 ret = -ENOENT;
3892 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3893 if (head)
3894 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
067524e9
JA
3895 if (ret < 0)
3896 req_set_fail_links(req);
067524e9 3897
31bff9a5
PB
3898 /* need to hold the lock to complete IOPOLL requests */
3899 if (ctx->flags & IORING_SETUP_IOPOLL) {
889fca73 3900 __io_req_complete(req, issue_flags, ret, 0);
31bff9a5
PB
3901 io_ring_submit_unlock(ctx, !force_nonblock);
3902 } else {
3903 io_ring_submit_unlock(ctx, !force_nonblock);
889fca73 3904 __io_req_complete(req, issue_flags, ret, 0);
31bff9a5 3905 }
067524e9
JA
3906 return 0;
3907}
3908
ddf0322d
JA
3909static int io_provide_buffers_prep(struct io_kiocb *req,
3910 const struct io_uring_sqe *sqe)
3911{
3912 struct io_provide_buf *p = &req->pbuf;
3913 u64 tmp;
3914
3915 if (sqe->ioprio || sqe->rw_flags)
3916 return -EINVAL;
3917
3918 tmp = READ_ONCE(sqe->fd);
3919 if (!tmp || tmp > USHRT_MAX)
3920 return -E2BIG;
3921 p->nbufs = tmp;
3922 p->addr = READ_ONCE(sqe->addr);
3923 p->len = READ_ONCE(sqe->len);
3924
efe68c1c 3925 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
ddf0322d
JA
3926 return -EFAULT;
3927
3928 p->bgid = READ_ONCE(sqe->buf_group);
3929 tmp = READ_ONCE(sqe->off);
3930 if (tmp > USHRT_MAX)
3931 return -E2BIG;
3932 p->bid = tmp;
3933 return 0;
3934}
3935
3936static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3937{
3938 struct io_buffer *buf;
3939 u64 addr = pbuf->addr;
3940 int i, bid = pbuf->bid;
3941
3942 for (i = 0; i < pbuf->nbufs; i++) {
3943 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3944 if (!buf)
3945 break;
3946
3947 buf->addr = addr;
3948 buf->len = pbuf->len;
3949 buf->bid = bid;
3950 addr += pbuf->len;
3951 bid++;
3952 if (!*head) {
3953 INIT_LIST_HEAD(&buf->list);
3954 *head = buf;
3955 } else {
3956 list_add_tail(&buf->list, &(*head)->list);
3957 }
3958 }
3959
3960 return i ? i : -ENOMEM;
3961}
3962
889fca73 3963static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
ddf0322d
JA
3964{
3965 struct io_provide_buf *p = &req->pbuf;
3966 struct io_ring_ctx *ctx = req->ctx;
3967 struct io_buffer *head, *list;
3968 int ret = 0;
45d189c6 3969 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
ddf0322d
JA
3970
3971 io_ring_submit_lock(ctx, !force_nonblock);
3972
3973 lockdep_assert_held(&ctx->uring_lock);
3974
3975 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3976
3977 ret = io_add_buffers(p, &head);
3978 if (ret < 0)
3979 goto out;
3980
3981 if (!list) {
3982 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3983 GFP_KERNEL);
3984 if (ret < 0) {
067524e9 3985 __io_remove_buffers(ctx, head, p->bgid, -1U);
ddf0322d
JA
3986 goto out;
3987 }
3988 }
3989out:
ddf0322d
JA
3990 if (ret < 0)
3991 req_set_fail_links(req);
31bff9a5
PB
3992
3993 /* need to hold the lock to complete IOPOLL requests */
3994 if (ctx->flags & IORING_SETUP_IOPOLL) {
889fca73 3995 __io_req_complete(req, issue_flags, ret, 0);
31bff9a5
PB
3996 io_ring_submit_unlock(ctx, !force_nonblock);
3997 } else {
3998 io_ring_submit_unlock(ctx, !force_nonblock);
889fca73 3999 __io_req_complete(req, issue_flags, ret, 0);
31bff9a5 4000 }
ddf0322d 4001 return 0;
cebdb986
JA
4002}
4003
3e4827b0
JA
4004static int io_epoll_ctl_prep(struct io_kiocb *req,
4005 const struct io_uring_sqe *sqe)
4006{
4007#if defined(CONFIG_EPOLL)
4008 if (sqe->ioprio || sqe->buf_index)
4009 return -EINVAL;
6ca56f84 4010 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
3232dd02 4011 return -EINVAL;
3e4827b0
JA
4012
4013 req->epoll.epfd = READ_ONCE(sqe->fd);
4014 req->epoll.op = READ_ONCE(sqe->len);
4015 req->epoll.fd = READ_ONCE(sqe->off);
4016
4017 if (ep_op_has_event(req->epoll.op)) {
4018 struct epoll_event __user *ev;
4019
4020 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4021 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4022 return -EFAULT;
4023 }
4024
4025 return 0;
4026#else
4027 return -EOPNOTSUPP;
4028#endif
4029}
4030
889fca73 4031static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
3e4827b0
JA
4032{
4033#if defined(CONFIG_EPOLL)
4034 struct io_epoll *ie = &req->epoll;
4035 int ret;
45d189c6 4036 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3e4827b0
JA
4037
4038 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4039 if (force_nonblock && ret == -EAGAIN)
4040 return -EAGAIN;
4041
4042 if (ret < 0)
4043 req_set_fail_links(req);
889fca73 4044 __io_req_complete(req, issue_flags, ret, 0);
3e4827b0
JA
4045 return 0;
4046#else
4047 return -EOPNOTSUPP;
4048#endif
4049}
4050
c1ca757b
JA
4051static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4052{
4053#if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4054 if (sqe->ioprio || sqe->buf_index || sqe->off)
4055 return -EINVAL;
3232dd02
PB
4056 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4057 return -EINVAL;
c1ca757b
JA
4058
4059 req->madvise.addr = READ_ONCE(sqe->addr);
4060 req->madvise.len = READ_ONCE(sqe->len);
4061 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4062 return 0;
4063#else
4064 return -EOPNOTSUPP;
4065#endif
4066}
4067
45d189c6 4068static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
c1ca757b
JA
4069{
4070#if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4071 struct io_madvise *ma = &req->madvise;
4072 int ret;
4073
45d189c6 4074 if (issue_flags & IO_URING_F_NONBLOCK)
c1ca757b
JA
4075 return -EAGAIN;
4076
0726b01e 4077 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
c1ca757b
JA
4078 if (ret < 0)
4079 req_set_fail_links(req);
e1e16097 4080 io_req_complete(req, ret);
c1ca757b
JA
4081 return 0;
4082#else
4083 return -EOPNOTSUPP;
4084#endif
4085}
4086
4840e418
JA
4087static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4088{
4089 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4090 return -EINVAL;
3232dd02
PB
4091 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4092 return -EINVAL;
4840e418
JA
4093
4094 req->fadvise.offset = READ_ONCE(sqe->off);
4095 req->fadvise.len = READ_ONCE(sqe->len);
4096 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4097 return 0;
4098}
4099
45d189c6 4100static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4840e418
JA
4101{
4102 struct io_fadvise *fa = &req->fadvise;
4103 int ret;
4104
45d189c6 4105 if (issue_flags & IO_URING_F_NONBLOCK) {
3e69426d
JA
4106 switch (fa->advice) {
4107 case POSIX_FADV_NORMAL:
4108 case POSIX_FADV_RANDOM:
4109 case POSIX_FADV_SEQUENTIAL:
4110 break;
4111 default:
4112 return -EAGAIN;
4113 }
4114 }
4840e418
JA
4115
4116 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4117 if (ret < 0)
4118 req_set_fail_links(req);
e1e16097 4119 io_req_complete(req, ret);
4840e418
JA
4120 return 0;
4121}
4122
eddc7ef5
JA
4123static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4124{
6ca56f84 4125 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
3232dd02 4126 return -EINVAL;
eddc7ef5
JA
4127 if (sqe->ioprio || sqe->buf_index)
4128 return -EINVAL;
9c280f90 4129 if (req->flags & REQ_F_FIXED_FILE)
cf3040ca 4130 return -EBADF;
eddc7ef5 4131
1d9e1288
BM
4132 req->statx.dfd = READ_ONCE(sqe->fd);
4133 req->statx.mask = READ_ONCE(sqe->len);
e62753e4 4134 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
1d9e1288
BM
4135 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4136 req->statx.flags = READ_ONCE(sqe->statx_flags);
eddc7ef5
JA
4137
4138 return 0;
4139}
4140
45d189c6 4141static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
eddc7ef5 4142{
1d9e1288 4143 struct io_statx *ctx = &req->statx;
eddc7ef5
JA
4144 int ret;
4145
45d189c6 4146 if (issue_flags & IO_URING_F_NONBLOCK) {
5b0bbee4
JA
4147 /* only need file table for an actual valid fd */
4148 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4149 req->flags |= REQ_F_NO_FILE_TABLE;
eddc7ef5 4150 return -EAGAIN;
5b0bbee4 4151 }
eddc7ef5 4152
e62753e4
BM
4153 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4154 ctx->buffer);
eddc7ef5 4155
eddc7ef5
JA
4156 if (ret < 0)
4157 req_set_fail_links(req);
e1e16097 4158 io_req_complete(req, ret);
eddc7ef5
JA
4159 return 0;
4160}
4161
b5dba59e
JA
4162static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4163{
14587a46 4164 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3232dd02 4165 return -EINVAL;
b5dba59e
JA
4166 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4167 sqe->rw_flags || sqe->buf_index)
4168 return -EINVAL;
9c280f90 4169 if (req->flags & REQ_F_FIXED_FILE)
cf3040ca 4170 return -EBADF;
b5dba59e
JA
4171
4172 req->close.fd = READ_ONCE(sqe->fd);
b5dba59e 4173 return 0;
b5dba59e
JA
4174}
4175
889fca73 4176static int io_close(struct io_kiocb *req, unsigned int issue_flags)
b5dba59e 4177{
9eac1904 4178 struct files_struct *files = current->files;
3af73b28 4179 struct io_close *close = &req->close;
9eac1904
JA
4180 struct fdtable *fdt;
4181 struct file *file;
b5dba59e
JA
4182 int ret;
4183
9eac1904
JA
4184 file = NULL;
4185 ret = -EBADF;
4186 spin_lock(&files->file_lock);
4187 fdt = files_fdtable(files);
4188 if (close->fd >= fdt->max_fds) {
4189 spin_unlock(&files->file_lock);
4190 goto err;
4191 }
4192 file = fdt->fd[close->fd];
4193 if (!file) {
4194 spin_unlock(&files->file_lock);
4195 goto err;
4196 }
4197
4198 if (file->f_op == &io_uring_fops) {
4199 spin_unlock(&files->file_lock);
4200 file = NULL;
4201 goto err;
3af73b28 4202 }
b5dba59e
JA
4203
4204 /* if the file has a flush method, be safe and punt to async */
45d189c6 4205 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
9eac1904 4206 spin_unlock(&files->file_lock);
0bf0eefd 4207 return -EAGAIN;
a2100672 4208 }
b5dba59e 4209
9eac1904
JA
4210 ret = __close_fd_get_file(close->fd, &file);
4211 spin_unlock(&files->file_lock);
4212 if (ret < 0) {
4213 if (ret == -ENOENT)
4214 ret = -EBADF;
4215 goto err;
4216 }
4217
3af73b28 4218 /* No ->flush() or already async, safely close from here */
9eac1904
JA
4219 ret = filp_close(file, current->files);
4220err:
3af73b28
PB
4221 if (ret < 0)
4222 req_set_fail_links(req);
9eac1904
JA
4223 if (file)
4224 fput(file);
889fca73 4225 __io_req_complete(req, issue_flags, ret, 0);
1a417f4e 4226 return 0;
b5dba59e
JA
4227}
4228
1155c76a 4229static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5d17b4a4
JA
4230{
4231 struct io_ring_ctx *ctx = req->ctx;
5d17b4a4 4232
5d17b4a4
JA
4233 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4234 return -EINVAL;
4235 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4236 return -EINVAL;
4237
8ed8d3c3
JA
4238 req->sync.off = READ_ONCE(sqe->off);
4239 req->sync.len = READ_ONCE(sqe->len);
4240 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
8ed8d3c3
JA
4241 return 0;
4242}
4243
45d189c6 4244static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
8ed8d3c3 4245{
8ed8d3c3
JA
4246 int ret;
4247
ac45abc0 4248 /* sync_file_range always requires a blocking context */
45d189c6 4249 if (issue_flags & IO_URING_F_NONBLOCK)
ac45abc0
PB
4250 return -EAGAIN;
4251
9adbd45d 4252 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
8ed8d3c3
JA
4253 req->sync.flags);
4254 if (ret < 0)
4255 req_set_fail_links(req);
e1e16097 4256 io_req_complete(req, ret);
5d17b4a4
JA
4257 return 0;
4258}
4259
469956e8 4260#if defined(CONFIG_NET)
02d27d89
PB
4261static int io_setup_async_msg(struct io_kiocb *req,
4262 struct io_async_msghdr *kmsg)
4263{
e8c2bc1f
JA
4264 struct io_async_msghdr *async_msg = req->async_data;
4265
4266 if (async_msg)
02d27d89 4267 return -EAGAIN;
e8c2bc1f 4268 if (io_alloc_async_data(req)) {
257e84a5 4269 kfree(kmsg->free_iov);
02d27d89
PB
4270 return -ENOMEM;
4271 }
e8c2bc1f 4272 async_msg = req->async_data;
02d27d89 4273 req->flags |= REQ_F_NEED_CLEANUP;
e8c2bc1f 4274 memcpy(async_msg, kmsg, sizeof(*kmsg));
2a780802 4275 async_msg->msg.msg_name = &async_msg->addr;
257e84a5
PB
4276 /* if were using fast_iov, set it to the new one */
4277 if (!async_msg->free_iov)
4278 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4279
02d27d89
PB
4280 return -EAGAIN;
4281}
4282
2ae523ed
PB
4283static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4284 struct io_async_msghdr *iomsg)
4285{
2ae523ed 4286 iomsg->msg.msg_name = &iomsg->addr;
257e84a5 4287 iomsg->free_iov = iomsg->fast_iov;
2ae523ed 4288 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
257e84a5 4289 req->sr_msg.msg_flags, &iomsg->free_iov);
2ae523ed
PB
4290}
4291
93642ef8
PB
4292static int io_sendmsg_prep_async(struct io_kiocb *req)
4293{
4294 int ret;
4295
4296 if (!io_op_defs[req->opcode].needs_async_data)
4297 return 0;
4298 ret = io_sendmsg_copy_hdr(req, req->async_data);
4299 if (!ret)
4300 req->flags |= REQ_F_NEED_CLEANUP;
4301 return ret;
4302}
4303
3529d8c2 4304static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
03b1230c 4305{
e47293fd 4306 struct io_sr_msg *sr = &req->sr_msg;
03b1230c 4307
d2b6f48b
PB
4308 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4309 return -EINVAL;
4310
e47293fd 4311 sr->msg_flags = READ_ONCE(sqe->msg_flags);
270a5940 4312 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
fddaface 4313 sr->len = READ_ONCE(sqe->len);
3529d8c2 4314
d8768362
JA
4315#ifdef CONFIG_COMPAT
4316 if (req->ctx->compat)
4317 sr->msg_flags |= MSG_CMSG_COMPAT;
4318#endif
93642ef8 4319 return 0;
03b1230c
JA
4320}
4321
889fca73 4322static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
aa1fa28f 4323{
6b754c8b 4324 struct io_async_msghdr iomsg, *kmsg;
0fa03c62 4325 struct socket *sock;
7a7cacba 4326 unsigned flags;
0fa03c62
JA
4327 int ret;
4328
dba4a925 4329 sock = sock_from_file(req->file);
7a7cacba 4330 if (unlikely(!sock))
dba4a925 4331 return -ENOTSOCK;
3529d8c2 4332
257e84a5
PB
4333 kmsg = req->async_data;
4334 if (!kmsg) {
7a7cacba
PB
4335 ret = io_sendmsg_copy_hdr(req, &iomsg);
4336 if (ret)
4337 return ret;
4338 kmsg = &iomsg;
0fa03c62 4339 }
0fa03c62 4340
7a7cacba
PB
4341 flags = req->sr_msg.msg_flags;
4342 if (flags & MSG_DONTWAIT)
4343 req->flags |= REQ_F_NOWAIT;
45d189c6 4344 else if (issue_flags & IO_URING_F_NONBLOCK)
7a7cacba 4345 flags |= MSG_DONTWAIT;
e47293fd 4346
7a7cacba 4347 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
45d189c6 4348 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
7a7cacba
PB
4349 return io_setup_async_msg(req, kmsg);
4350 if (ret == -ERESTARTSYS)
4351 ret = -EINTR;
0fa03c62 4352
257e84a5
PB
4353 /* fast path, check for non-NULL to avoid function call */
4354 if (kmsg->free_iov)
4355 kfree(kmsg->free_iov);
99bc4c38 4356 req->flags &= ~REQ_F_NEED_CLEANUP;
4e88d6e7
JA
4357 if (ret < 0)
4358 req_set_fail_links(req);
889fca73 4359 __io_req_complete(req, issue_flags, ret, 0);
5d17b4a4 4360 return 0;
03b1230c 4361}
aa1fa28f 4362
889fca73 4363static int io_send(struct io_kiocb *req, unsigned int issue_flags)
fddaface 4364{
7a7cacba
PB
4365 struct io_sr_msg *sr = &req->sr_msg;
4366 struct msghdr msg;
4367 struct iovec iov;
fddaface 4368 struct socket *sock;
7a7cacba 4369 unsigned flags;
fddaface
JA
4370 int ret;
4371
dba4a925 4372 sock = sock_from_file(req->file);
7a7cacba 4373 if (unlikely(!sock))
dba4a925 4374 return -ENOTSOCK;
fddaface 4375
7a7cacba
PB
4376 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4377 if (unlikely(ret))
14db8411 4378 return ret;
fddaface 4379
7a7cacba
PB
4380 msg.msg_name = NULL;
4381 msg.msg_control = NULL;
4382 msg.msg_controllen = 0;
4383 msg.msg_namelen = 0;
fddaface 4384
7a7cacba
PB
4385 flags = req->sr_msg.msg_flags;
4386 if (flags & MSG_DONTWAIT)
4387 req->flags |= REQ_F_NOWAIT;
45d189c6 4388 else if (issue_flags & IO_URING_F_NONBLOCK)
7a7cacba 4389 flags |= MSG_DONTWAIT;
fddaface 4390
7a7cacba
PB
4391 msg.msg_flags = flags;
4392 ret = sock_sendmsg(sock, &msg);
45d189c6 4393 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
7a7cacba
PB
4394 return -EAGAIN;
4395 if (ret == -ERESTARTSYS)
4396 ret = -EINTR;
fddaface 4397
fddaface
JA
4398 if (ret < 0)
4399 req_set_fail_links(req);
889fca73 4400 __io_req_complete(req, issue_flags, ret, 0);
fddaface 4401 return 0;
fddaface
JA
4402}
4403
1400e697
PB
4404static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4405 struct io_async_msghdr *iomsg)
52de1fe1
JA
4406{
4407 struct io_sr_msg *sr = &req->sr_msg;
4408 struct iovec __user *uiov;
4409 size_t iov_len;
4410 int ret;
4411
1400e697
PB
4412 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4413 &iomsg->uaddr, &uiov, &iov_len);
52de1fe1
JA
4414 if (ret)
4415 return ret;
4416
4417 if (req->flags & REQ_F_BUFFER_SELECT) {
4418 if (iov_len > 1)
4419 return -EINVAL;
5476dfed 4420 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
52de1fe1 4421 return -EFAULT;
5476dfed 4422 sr->len = iomsg->fast_iov[0].iov_len;
257e84a5 4423 iomsg->free_iov = NULL;
52de1fe1 4424 } else {
257e84a5 4425 iomsg->free_iov = iomsg->fast_iov;
89cd35c5 4426 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
257e84a5 4427 &iomsg->free_iov, &iomsg->msg.msg_iter,
89cd35c5 4428 false);
52de1fe1
JA
4429 if (ret > 0)
4430 ret = 0;
4431 }
4432
4433 return ret;
4434}
4435
4436#ifdef CONFIG_COMPAT
4437static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
1400e697 4438 struct io_async_msghdr *iomsg)
52de1fe1
JA
4439{
4440 struct compat_msghdr __user *msg_compat;
4441 struct io_sr_msg *sr = &req->sr_msg;
4442 struct compat_iovec __user *uiov;
4443 compat_uptr_t ptr;
4444 compat_size_t len;
4445 int ret;
4446
270a5940 4447 msg_compat = (struct compat_msghdr __user *) sr->umsg;
1400e697 4448 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
52de1fe1
JA
4449 &ptr, &len);
4450 if (ret)
4451 return ret;
4452
4453 uiov = compat_ptr(ptr);
4454 if (req->flags & REQ_F_BUFFER_SELECT) {
4455 compat_ssize_t clen;
4456
4457 if (len > 1)
4458 return -EINVAL;
4459 if (!access_ok(uiov, sizeof(*uiov)))
4460 return -EFAULT;
4461 if (__get_user(clen, &uiov->iov_len))
4462 return -EFAULT;
4463 if (clen < 0)
4464 return -EINVAL;
2d280bc8 4465 sr->len = clen;
257e84a5 4466 iomsg->free_iov = NULL;
52de1fe1 4467 } else {
257e84a5 4468 iomsg->free_iov = iomsg->fast_iov;
89cd35c5 4469 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
257e84a5 4470 UIO_FASTIOV, &iomsg->free_iov,
89cd35c5 4471 &iomsg->msg.msg_iter, true);
52de1fe1
JA
4472 if (ret < 0)
4473 return ret;
4474 }
4475
4476 return 0;
4477}
4478#endif
4479
1400e697
PB
4480static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4481 struct io_async_msghdr *iomsg)
52de1fe1 4482{
1400e697 4483 iomsg->msg.msg_name = &iomsg->addr;
52de1fe1
JA
4484
4485#ifdef CONFIG_COMPAT
4486 if (req->ctx->compat)
1400e697 4487 return __io_compat_recvmsg_copy_hdr(req, iomsg);
fddaface 4488#endif
52de1fe1 4489
1400e697 4490 return __io_recvmsg_copy_hdr(req, iomsg);
52de1fe1
JA
4491}
4492
bcda7baa 4493static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
7fbb1b54 4494 bool needs_lock)
bcda7baa
JA
4495{
4496 struct io_sr_msg *sr = &req->sr_msg;
4497 struct io_buffer *kbuf;
4498
bcda7baa
JA
4499 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4500 if (IS_ERR(kbuf))
4501 return kbuf;
4502
4503 sr->kbuf = kbuf;
4504 req->flags |= REQ_F_BUFFER_SELECTED;
bcda7baa 4505 return kbuf;
fddaface
JA
4506}
4507
7fbb1b54
PB
4508static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4509{
4510 return io_put_kbuf(req, req->sr_msg.kbuf);
4511}
4512
93642ef8 4513static int io_recvmsg_prep_async(struct io_kiocb *req)
aa1fa28f 4514{
99bc4c38 4515 int ret;
3529d8c2 4516
93642ef8
PB
4517 if (!io_op_defs[req->opcode].needs_async_data)
4518 return 0;
4519 ret = io_recvmsg_copy_hdr(req, req->async_data);
4520 if (!ret)
4521 req->flags |= REQ_F_NEED_CLEANUP;
4522 return ret;
4523}
4524
4525static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4526{
4527 struct io_sr_msg *sr = &req->sr_msg;
4528
d2b6f48b
PB
4529 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4530 return -EINVAL;
4531
3529d8c2 4532 sr->msg_flags = READ_ONCE(sqe->msg_flags);
270a5940 4533 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
0b7b21e4 4534 sr->len = READ_ONCE(sqe->len);
bcda7baa 4535 sr->bgid = READ_ONCE(sqe->buf_group);
06b76d44 4536
d8768362
JA
4537#ifdef CONFIG_COMPAT
4538 if (req->ctx->compat)
4539 sr->msg_flags |= MSG_CMSG_COMPAT;
4540#endif
93642ef8 4541 return 0;
aa1fa28f
JA
4542}
4543
889fca73 4544static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
aa1fa28f 4545{
6b754c8b 4546 struct io_async_msghdr iomsg, *kmsg;
03b1230c 4547 struct socket *sock;
7fbb1b54 4548 struct io_buffer *kbuf;
7a7cacba 4549 unsigned flags;
52de1fe1 4550 int ret, cflags = 0;
45d189c6 4551 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
03b1230c 4552
dba4a925 4553 sock = sock_from_file(req->file);
7a7cacba 4554 if (unlikely(!sock))
dba4a925 4555 return -ENOTSOCK;
3529d8c2 4556
257e84a5
PB
4557 kmsg = req->async_data;
4558 if (!kmsg) {
7a7cacba
PB
4559 ret = io_recvmsg_copy_hdr(req, &iomsg);
4560 if (ret)
681fda8d 4561 return ret;
7a7cacba
PB
4562 kmsg = &iomsg;
4563 }
03b1230c 4564
bc02ef33 4565 if (req->flags & REQ_F_BUFFER_SELECT) {
7fbb1b54 4566 kbuf = io_recv_buffer_select(req, !force_nonblock);
bc02ef33 4567 if (IS_ERR(kbuf))
52de1fe1 4568 return PTR_ERR(kbuf);
7a7cacba 4569 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
5476dfed
PB
4570 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4571 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
7a7cacba
PB
4572 1, req->sr_msg.len);
4573 }
52de1fe1 4574
7a7cacba
PB
4575 flags = req->sr_msg.msg_flags;
4576 if (flags & MSG_DONTWAIT)
4577 req->flags |= REQ_F_NOWAIT;
4578 else if (force_nonblock)
4579 flags |= MSG_DONTWAIT;
e47293fd 4580
7a7cacba
PB
4581 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4582 kmsg->uaddr, flags);
0e1b6fe3
PB
4583 if (force_nonblock && ret == -EAGAIN)
4584 return io_setup_async_msg(req, kmsg);
7a7cacba
PB
4585 if (ret == -ERESTARTSYS)
4586 ret = -EINTR;
03b1230c 4587
7fbb1b54
PB
4588 if (req->flags & REQ_F_BUFFER_SELECTED)
4589 cflags = io_put_recv_kbuf(req);
257e84a5
PB
4590 /* fast path, check for non-NULL to avoid function call */
4591 if (kmsg->free_iov)
4592 kfree(kmsg->free_iov);
99bc4c38 4593 req->flags &= ~REQ_F_NEED_CLEANUP;
4e88d6e7
JA
4594 if (ret < 0)
4595 req_set_fail_links(req);
889fca73 4596 __io_req_complete(req, issue_flags, ret, cflags);
03b1230c 4597 return 0;
0fa03c62 4598}
5d17b4a4 4599
889fca73 4600static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
fddaface 4601{
6b754c8b 4602 struct io_buffer *kbuf;
7a7cacba
PB
4603 struct io_sr_msg *sr = &req->sr_msg;
4604 struct msghdr msg;
4605 void __user *buf = sr->buf;
fddaface 4606 struct socket *sock;
7a7cacba
PB
4607 struct iovec iov;
4608 unsigned flags;
bcda7baa 4609 int ret, cflags = 0;
45d189c6 4610 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
fddaface 4611
dba4a925 4612 sock = sock_from_file(req->file);
7a7cacba 4613 if (unlikely(!sock))
dba4a925 4614 return -ENOTSOCK;
fddaface 4615
bc02ef33 4616 if (req->flags & REQ_F_BUFFER_SELECT) {
7fbb1b54 4617 kbuf = io_recv_buffer_select(req, !force_nonblock);
bcda7baa
JA
4618 if (IS_ERR(kbuf))
4619 return PTR_ERR(kbuf);
7a7cacba 4620 buf = u64_to_user_ptr(kbuf->addr);
bc02ef33 4621 }
bcda7baa 4622
7a7cacba 4623 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
14c32eee
PB
4624 if (unlikely(ret))
4625 goto out_free;
fddaface 4626
7a7cacba
PB
4627 msg.msg_name = NULL;
4628 msg.msg_control = NULL;
4629 msg.msg_controllen = 0;
4630 msg.msg_namelen = 0;
4631 msg.msg_iocb = NULL;
4632 msg.msg_flags = 0;
fddaface 4633
7a7cacba
PB
4634 flags = req->sr_msg.msg_flags;
4635 if (flags & MSG_DONTWAIT)
4636 req->flags |= REQ_F_NOWAIT;
4637 else if (force_nonblock)
4638 flags |= MSG_DONTWAIT;
4639
4640 ret = sock_recvmsg(sock, &msg, flags);
4641 if (force_nonblock && ret == -EAGAIN)
4642 return -EAGAIN;
4643 if (ret == -ERESTARTSYS)
4644 ret = -EINTR;
14c32eee 4645out_free:
7fbb1b54
PB
4646 if (req->flags & REQ_F_BUFFER_SELECTED)
4647 cflags = io_put_recv_kbuf(req);
fddaface
JA
4648 if (ret < 0)
4649 req_set_fail_links(req);
889fca73 4650 __io_req_complete(req, issue_flags, ret, cflags);
fddaface 4651 return 0;
fddaface
JA
4652}
4653
3529d8c2 4654static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
17f2fe35 4655{
8ed8d3c3
JA
4656 struct io_accept *accept = &req->accept;
4657
14587a46 4658 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
17f2fe35 4659 return -EINVAL;
8042d6ce 4660 if (sqe->ioprio || sqe->len || sqe->buf_index)
17f2fe35
JA
4661 return -EINVAL;
4662
d55e5f5b
JA
4663 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4664 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
8ed8d3c3 4665 accept->flags = READ_ONCE(sqe->accept_flags);
09952e3e 4666 accept->nofile = rlimit(RLIMIT_NOFILE);
8ed8d3c3 4667 return 0;
8ed8d3c3 4668}
17f2fe35 4669
889fca73 4670static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
8ed8d3c3
JA
4671{
4672 struct io_accept *accept = &req->accept;
45d189c6 4673 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
ac45abc0 4674 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
8ed8d3c3
JA
4675 int ret;
4676
e697deed
JX
4677 if (req->file->f_flags & O_NONBLOCK)
4678 req->flags |= REQ_F_NOWAIT;
4679
8ed8d3c3 4680 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
09952e3e
JA
4681 accept->addr_len, accept->flags,
4682 accept->nofile);
8ed8d3c3 4683 if (ret == -EAGAIN && force_nonblock)
17f2fe35 4684 return -EAGAIN;
ac45abc0
PB
4685 if (ret < 0) {
4686 if (ret == -ERESTARTSYS)
4687 ret = -EINTR;
4e88d6e7 4688 req_set_fail_links(req);
ac45abc0 4689 }
889fca73 4690 __io_req_complete(req, issue_flags, ret, 0);
17f2fe35 4691 return 0;
8ed8d3c3
JA
4692}
4693
93642ef8
PB
4694static int io_connect_prep_async(struct io_kiocb *req)
4695{
4696 struct io_async_connect *io = req->async_data;
4697 struct io_connect *conn = &req->connect;
4698
4699 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4700}
4701
3529d8c2 4702static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
f499a021 4703{
3529d8c2 4704 struct io_connect *conn = &req->connect;
f499a021 4705
14587a46 4706 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3fbb51c1
JA
4707 return -EINVAL;
4708 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4709 return -EINVAL;
4710
3529d8c2
JA
4711 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4712 conn->addr_len = READ_ONCE(sqe->addr2);
93642ef8 4713 return 0;
f499a021
JA
4714}
4715
889fca73 4716static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
f8e85cf2 4717{
e8c2bc1f 4718 struct io_async_connect __io, *io;
f8e85cf2 4719 unsigned file_flags;
3fbb51c1 4720 int ret;
45d189c6 4721 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
f8e85cf2 4722
e8c2bc1f
JA
4723 if (req->async_data) {
4724 io = req->async_data;
f499a021 4725 } else {
3529d8c2
JA
4726 ret = move_addr_to_kernel(req->connect.addr,
4727 req->connect.addr_len,
e8c2bc1f 4728 &__io.address);
f499a021
JA
4729 if (ret)
4730 goto out;
4731 io = &__io;
4732 }
4733
3fbb51c1
JA
4734 file_flags = force_nonblock ? O_NONBLOCK : 0;
4735
e8c2bc1f 4736 ret = __sys_connect_file(req->file, &io->address,
3fbb51c1 4737 req->connect.addr_len, file_flags);
87f80d62 4738 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
e8c2bc1f 4739 if (req->async_data)
b7bb4f7d 4740 return -EAGAIN;
e8c2bc1f 4741 if (io_alloc_async_data(req)) {
f499a021
JA
4742 ret = -ENOMEM;
4743 goto out;
4744 }
e8c2bc1f
JA
4745 io = req->async_data;
4746 memcpy(req->async_data, &__io, sizeof(__io));
f8e85cf2 4747 return -EAGAIN;
f499a021 4748 }
f8e85cf2
JA
4749 if (ret == -ERESTARTSYS)
4750 ret = -EINTR;
f499a021 4751out:
4e88d6e7
JA
4752 if (ret < 0)
4753 req_set_fail_links(req);
889fca73 4754 __io_req_complete(req, issue_flags, ret, 0);
f8e85cf2 4755 return 0;
469956e8
Y
4756}
4757#else /* !CONFIG_NET */
99a10081
JA
4758#define IO_NETOP_FN(op) \
4759static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4760{ \
4761 return -EOPNOTSUPP; \
4762}
4763
4764#define IO_NETOP_PREP(op) \
4765IO_NETOP_FN(op) \
4766static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4767{ \
4768 return -EOPNOTSUPP; \
4769} \
4770
4771#define IO_NETOP_PREP_ASYNC(op) \
4772IO_NETOP_PREP(op) \
4773static int io_##op##_prep_async(struct io_kiocb *req) \
4774{ \
4775 return -EOPNOTSUPP; \
4776}
4777
4778IO_NETOP_PREP_ASYNC(sendmsg);
4779IO_NETOP_PREP_ASYNC(recvmsg);
4780IO_NETOP_PREP_ASYNC(connect);
4781IO_NETOP_PREP(accept);
4782IO_NETOP_FN(send);
4783IO_NETOP_FN(recv);
469956e8 4784#endif /* CONFIG_NET */
f8e85cf2 4785
d7718a9d
JA
4786struct io_poll_table {
4787 struct poll_table_struct pt;
4788 struct io_kiocb *req;
4789 int error;
4790};
ce593a6c 4791
d7718a9d
JA
4792static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4793 __poll_t mask, task_work_func_t func)
4794{
aa96bf8a 4795 int ret;
d7718a9d
JA
4796
4797 /* for instances that support it check for an event match first: */
4798 if (mask && !(mask & poll->events))
4799 return 0;
4800
4801 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4802
4803 list_del_init(&poll->wait.entry);
4804
d7718a9d 4805 req->result = mask;
7cbf1722 4806 req->task_work.func = func;
6d816e08
JA
4807 percpu_ref_get(&req->ctx->refs);
4808
d7718a9d 4809 /*
e3aabf95
JA
4810 * If this fails, then the task is exiting. When a task exits, the
4811 * work gets canceled, so just cancel this request as well instead
4812 * of executing it. We can't safely execute it anyway, as we may not
4813 * have the needed state needed for it anyway.
d7718a9d 4814 */
355fb9e2 4815 ret = io_req_task_work_add(req);
aa96bf8a 4816 if (unlikely(ret)) {
e3aabf95 4817 WRITE_ONCE(poll->canceled, true);
eab30c4d 4818 io_req_task_work_add_fallback(req, func);
aa96bf8a 4819 }
d7718a9d
JA
4820 return 1;
4821}
4822
74ce6ce4
JA
4823static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4824 __acquires(&req->ctx->completion_lock)
4825{
4826 struct io_ring_ctx *ctx = req->ctx;
4827
4828 if (!req->result && !READ_ONCE(poll->canceled)) {
4829 struct poll_table_struct pt = { ._key = poll->events };
4830
4831 req->result = vfs_poll(req->file, &pt) & poll->events;
4832 }
4833
4834 spin_lock_irq(&ctx->completion_lock);
4835 if (!req->result && !READ_ONCE(poll->canceled)) {
4836 add_wait_queue(poll->head, &poll->wait);
4837 return true;
4838 }
4839
4840 return false;
4841}
4842
d4e7cd36 4843static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
18bceab1 4844{
e8c2bc1f 4845 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
d4e7cd36 4846 if (req->opcode == IORING_OP_POLL_ADD)
e8c2bc1f 4847 return req->async_data;
d4e7cd36
JA
4848 return req->apoll->double_poll;
4849}
4850
4851static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4852{
4853 if (req->opcode == IORING_OP_POLL_ADD)
4854 return &req->poll;
4855 return &req->apoll->poll;
4856}
4857
4858static void io_poll_remove_double(struct io_kiocb *req)
4859{
4860 struct io_poll_iocb *poll = io_poll_get_double(req);
18bceab1
JA
4861
4862 lockdep_assert_held(&req->ctx->completion_lock);
4863
4864 if (poll && poll->head) {
4865 struct wait_queue_head *head = poll->head;
4866
4867 spin_lock(&head->lock);
4868 list_del_init(&poll->wait.entry);
4869 if (poll->wait.private)
4870 refcount_dec(&req->refs);
4871 poll->head = NULL;
4872 spin_unlock(&head->lock);
4873 }
4874}
4875
4876static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4877{
4878 struct io_ring_ctx *ctx = req->ctx;
4879
d4e7cd36 4880 io_poll_remove_double(req);
18bceab1
JA
4881 req->poll.done = true;
4882 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4883 io_commit_cqring(ctx);
4884}
4885
dd221f46 4886static void io_poll_task_func(struct callback_head *cb)
18bceab1 4887{
dd221f46 4888 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
18bceab1 4889 struct io_ring_ctx *ctx = req->ctx;
dd221f46 4890 struct io_kiocb *nxt;
18bceab1
JA
4891
4892 if (io_poll_rewait(req, &req->poll)) {
4893 spin_unlock_irq(&ctx->completion_lock);
dd221f46
PB
4894 } else {
4895 hash_del(&req->hash_node);
4896 io_poll_complete(req, req->result, 0);
4897 spin_unlock_irq(&ctx->completion_lock);
18bceab1 4898
dd221f46
PB
4899 nxt = io_put_req_find_next(req);
4900 io_cqring_ev_posted(ctx);
4901 if (nxt)
4902 __io_req_task_submit(nxt);
4903 }
18bceab1 4904
6d816e08 4905 percpu_ref_put(&ctx->refs);
18bceab1
JA
4906}
4907
4908static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4909 int sync, void *key)
4910{
4911 struct io_kiocb *req = wait->private;
d4e7cd36 4912 struct io_poll_iocb *poll = io_poll_get_single(req);
18bceab1
JA
4913 __poll_t mask = key_to_poll(key);
4914
4915 /* for instances that support it check for an event match first: */
4916 if (mask && !(mask & poll->events))
4917 return 0;
4918
8706e04e
JA
4919 list_del_init(&wait->entry);
4920
807abcb0 4921 if (poll && poll->head) {
18bceab1
JA
4922 bool done;
4923
807abcb0
JA
4924 spin_lock(&poll->head->lock);
4925 done = list_empty(&poll->wait.entry);
18bceab1 4926 if (!done)
807abcb0 4927 list_del_init(&poll->wait.entry);
d4e7cd36
JA
4928 /* make sure double remove sees this as being gone */
4929 wait->private = NULL;
807abcb0 4930 spin_unlock(&poll->head->lock);
c8b5e260
JA
4931 if (!done) {
4932 /* use wait func handler, so it matches the rq type */
4933 poll->wait.func(&poll->wait, mode, sync, key);
4934 }
18bceab1
JA
4935 }
4936 refcount_dec(&req->refs);
4937 return 1;
4938}
4939
4940static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4941 wait_queue_func_t wake_func)
4942{
4943 poll->head = NULL;
4944 poll->done = false;
4945 poll->canceled = false;
4946 poll->events = events;
4947 INIT_LIST_HEAD(&poll->wait.entry);
4948 init_waitqueue_func_entry(&poll->wait, wake_func);
4949}
4950
4951static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
807abcb0
JA
4952 struct wait_queue_head *head,
4953 struct io_poll_iocb **poll_ptr)
18bceab1
JA
4954{
4955 struct io_kiocb *req = pt->req;
4956
4957 /*
4958 * If poll->head is already set, it's because the file being polled
4959 * uses multiple waitqueues for poll handling (eg one for read, one
4960 * for write). Setup a separate io_poll_iocb if this happens.
4961 */
4962 if (unlikely(poll->head)) {
58852d4d
PB
4963 struct io_poll_iocb *poll_one = poll;
4964
18bceab1 4965 /* already have a 2nd entry, fail a third attempt */
807abcb0 4966 if (*poll_ptr) {
18bceab1
JA
4967 pt->error = -EINVAL;
4968 return;
4969 }
1c3b3e65
JA
4970 /* double add on the same waitqueue head, ignore */
4971 if (poll->head == head)
4972 return;
18bceab1
JA
4973 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4974 if (!poll) {
4975 pt->error = -ENOMEM;
4976 return;
4977 }
58852d4d 4978 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
18bceab1
JA
4979 refcount_inc(&req->refs);
4980 poll->wait.private = req;
807abcb0 4981 *poll_ptr = poll;
18bceab1
JA
4982 }
4983
4984 pt->error = 0;
4985 poll->head = head;
a31eb4a2
JX
4986
4987 if (poll->events & EPOLLEXCLUSIVE)
4988 add_wait_queue_exclusive(head, &poll->wait);
4989 else
4990 add_wait_queue(head, &poll->wait);
18bceab1
JA
4991}
4992
4993static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4994 struct poll_table_struct *p)
4995{
4996 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
807abcb0 4997 struct async_poll *apoll = pt->req->apoll;
18bceab1 4998
807abcb0 4999 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
18bceab1
JA
5000}
5001
d7718a9d
JA
5002static void io_async_task_func(struct callback_head *cb)
5003{
5004 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5005 struct async_poll *apoll = req->apoll;
5006 struct io_ring_ctx *ctx = req->ctx;
5007
5008 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5009
74ce6ce4 5010 if (io_poll_rewait(req, &apoll->poll)) {
d7718a9d 5011 spin_unlock_irq(&ctx->completion_lock);
6d816e08 5012 percpu_ref_put(&ctx->refs);
74ce6ce4 5013 return;
d7718a9d
JA
5014 }
5015
31067255 5016 /* If req is still hashed, it cannot have been canceled. Don't check. */
0be0b0e3 5017 if (hash_hashed(&req->hash_node))
74ce6ce4 5018 hash_del(&req->hash_node);
2bae047e 5019
d4e7cd36 5020 io_poll_remove_double(req);
74ce6ce4
JA
5021 spin_unlock_irq(&ctx->completion_lock);
5022
0be0b0e3
PB
5023 if (!READ_ONCE(apoll->poll.canceled))
5024 __io_req_task_submit(req);
5025 else
5026 __io_req_task_cancel(req, -ECANCELED);
aa340845 5027
6d816e08 5028 percpu_ref_put(&ctx->refs);
807abcb0 5029 kfree(apoll->double_poll);
31067255 5030 kfree(apoll);
d7718a9d
JA
5031}
5032
5033static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5034 void *key)
5035{
5036 struct io_kiocb *req = wait->private;
5037 struct io_poll_iocb *poll = &req->apoll->poll;
5038
5039 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5040 key_to_poll(key));
5041
5042 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5043}
5044
5045static void io_poll_req_insert(struct io_kiocb *req)
5046{
5047 struct io_ring_ctx *ctx = req->ctx;
5048 struct hlist_head *list;
5049
5050 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5051 hlist_add_head(&req->hash_node, list);
5052}
5053
5054static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5055 struct io_poll_iocb *poll,
5056 struct io_poll_table *ipt, __poll_t mask,
5057 wait_queue_func_t wake_func)
5058 __acquires(&ctx->completion_lock)
5059{
5060 struct io_ring_ctx *ctx = req->ctx;
5061 bool cancel = false;
5062
4d52f338 5063 INIT_HLIST_NODE(&req->hash_node);
18bceab1 5064 io_init_poll_iocb(poll, mask, wake_func);
b90cd197 5065 poll->file = req->file;
18bceab1 5066 poll->wait.private = req;
d7718a9d
JA
5067
5068 ipt->pt._key = mask;
5069 ipt->req = req;
5070 ipt->error = -EINVAL;
5071
d7718a9d
JA
5072 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5073
5074 spin_lock_irq(&ctx->completion_lock);
5075 if (likely(poll->head)) {
5076 spin_lock(&poll->head->lock);
5077 if (unlikely(list_empty(&poll->wait.entry))) {
5078 if (ipt->error)
5079 cancel = true;
5080 ipt->error = 0;
5081 mask = 0;
5082 }
5083 if (mask || ipt->error)
5084 list_del_init(&poll->wait.entry);
5085 else if (cancel)
5086 WRITE_ONCE(poll->canceled, true);
5087 else if (!poll->done) /* actually waiting for an event */
5088 io_poll_req_insert(req);
5089 spin_unlock(&poll->head->lock);
5090 }
5091
5092 return mask;
5093}
5094
5095static bool io_arm_poll_handler(struct io_kiocb *req)
5096{
5097 const struct io_op_def *def = &io_op_defs[req->opcode];
5098 struct io_ring_ctx *ctx = req->ctx;
5099 struct async_poll *apoll;
5100 struct io_poll_table ipt;
5101 __poll_t mask, ret;
9dab14b8 5102 int rw;
d7718a9d
JA
5103
5104 if (!req->file || !file_can_poll(req->file))
5105 return false;
24c74678 5106 if (req->flags & REQ_F_POLLED)
d7718a9d 5107 return false;
9dab14b8
JA
5108 if (def->pollin)
5109 rw = READ;
5110 else if (def->pollout)
5111 rw = WRITE;
5112 else
5113 return false;
5114 /* if we can't nonblock try, then no point in arming a poll handler */
5115 if (!io_file_supports_async(req->file, rw))
d7718a9d
JA
5116 return false;
5117
5118 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5119 if (unlikely(!apoll))
5120 return false;
807abcb0 5121 apoll->double_poll = NULL;
d7718a9d
JA
5122
5123 req->flags |= REQ_F_POLLED;
d7718a9d 5124 req->apoll = apoll;
d7718a9d 5125
8755d97a 5126 mask = 0;
d7718a9d 5127 if (def->pollin)
8755d97a 5128 mask |= POLLIN | POLLRDNORM;
d7718a9d
JA
5129 if (def->pollout)
5130 mask |= POLLOUT | POLLWRNORM;
901341bb
LH
5131
5132 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5133 if ((req->opcode == IORING_OP_RECVMSG) &&
5134 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5135 mask &= ~POLLIN;
5136
d7718a9d
JA
5137 mask |= POLLERR | POLLPRI;
5138
5139 ipt.pt._qproc = io_async_queue_proc;
5140
5141 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5142 io_async_wake);
a36da65c 5143 if (ret || ipt.error) {
d4e7cd36 5144 io_poll_remove_double(req);
d7718a9d 5145 spin_unlock_irq(&ctx->completion_lock);
807abcb0 5146 kfree(apoll->double_poll);
d7718a9d
JA
5147 kfree(apoll);
5148 return false;
5149 }
5150 spin_unlock_irq(&ctx->completion_lock);
5151 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5152 apoll->poll.events);
5153 return true;
5154}
5155
5156static bool __io_poll_remove_one(struct io_kiocb *req,
5157 struct io_poll_iocb *poll)
221c5eb2 5158{
b41e9852 5159 bool do_complete = false;
221c5eb2
JA
5160
5161 spin_lock(&poll->head->lock);
5162 WRITE_ONCE(poll->canceled, true);
392edb45
JA
5163 if (!list_empty(&poll->wait.entry)) {
5164 list_del_init(&poll->wait.entry);
b41e9852 5165 do_complete = true;
221c5eb2
JA
5166 }
5167 spin_unlock(&poll->head->lock);
3bfa5bcb 5168 hash_del(&req->hash_node);
d7718a9d
JA
5169 return do_complete;
5170}
5171
5172static bool io_poll_remove_one(struct io_kiocb *req)
5173{
5174 bool do_complete;
5175
d4e7cd36
JA
5176 io_poll_remove_double(req);
5177
d7718a9d
JA
5178 if (req->opcode == IORING_OP_POLL_ADD) {
5179 do_complete = __io_poll_remove_one(req, &req->poll);
5180 } else {
3bfa5bcb
JA
5181 struct async_poll *apoll = req->apoll;
5182
d7718a9d 5183 /* non-poll requests have submit ref still */
3bfa5bcb
JA
5184 do_complete = __io_poll_remove_one(req, &apoll->poll);
5185 if (do_complete) {
d7718a9d 5186 io_put_req(req);
807abcb0 5187 kfree(apoll->double_poll);
3bfa5bcb
JA
5188 kfree(apoll);
5189 }
b1f573bd
XW
5190 }
5191
b41e9852
JA
5192 if (do_complete) {
5193 io_cqring_fill_event(req, -ECANCELED);
5194 io_commit_cqring(req->ctx);
f254ac04 5195 req_set_fail_links(req);
216578e5 5196 io_put_req_deferred(req, 1);
b41e9852
JA
5197 }
5198
5199 return do_complete;
221c5eb2
JA
5200}
5201
76e1b642
JA
5202/*
5203 * Returns true if we found and killed one or more poll requests
5204 */
6b81928d
PB
5205static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5206 struct files_struct *files)
221c5eb2 5207{
78076bb6 5208 struct hlist_node *tmp;
221c5eb2 5209 struct io_kiocb *req;
8e2e1faf 5210 int posted = 0, i;
221c5eb2
JA
5211
5212 spin_lock_irq(&ctx->completion_lock);
78076bb6
JA
5213 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5214 struct hlist_head *list;
5215
5216 list = &ctx->cancel_hash[i];
f3606e3a 5217 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
6b81928d 5218 if (io_match_task(req, tsk, files))
f3606e3a
JA
5219 posted += io_poll_remove_one(req);
5220 }
221c5eb2
JA
5221 }
5222 spin_unlock_irq(&ctx->completion_lock);
b41e9852 5223
8e2e1faf
JA
5224 if (posted)
5225 io_cqring_ev_posted(ctx);
76e1b642
JA
5226
5227 return posted != 0;
221c5eb2
JA
5228}
5229
47f46768
JA
5230static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5231{
78076bb6 5232 struct hlist_head *list;
47f46768
JA
5233 struct io_kiocb *req;
5234
78076bb6
JA
5235 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5236 hlist_for_each_entry(req, list, hash_node) {
b41e9852
JA
5237 if (sqe_addr != req->user_data)
5238 continue;
5239 if (io_poll_remove_one(req))
eac406c6 5240 return 0;
b41e9852 5241 return -EALREADY;
47f46768
JA
5242 }
5243
5244 return -ENOENT;
5245}
5246
3529d8c2
JA
5247static int io_poll_remove_prep(struct io_kiocb *req,
5248 const struct io_uring_sqe *sqe)
0969e783 5249{
0969e783
JA
5250 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5251 return -EINVAL;
5252 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5253 sqe->poll_events)
5254 return -EINVAL;
5255
018043be 5256 req->poll_remove.addr = READ_ONCE(sqe->addr);
0969e783
JA
5257 return 0;
5258}
5259
221c5eb2
JA
5260/*
5261 * Find a running poll command that matches one specified in sqe->addr,
5262 * and remove it if found.
5263 */
61e98203 5264static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
221c5eb2
JA
5265{
5266 struct io_ring_ctx *ctx = req->ctx;
47f46768 5267 int ret;
221c5eb2 5268
221c5eb2 5269 spin_lock_irq(&ctx->completion_lock);
018043be 5270 ret = io_poll_cancel(ctx, req->poll_remove.addr);
221c5eb2
JA
5271 spin_unlock_irq(&ctx->completion_lock);
5272
4e88d6e7
JA
5273 if (ret < 0)
5274 req_set_fail_links(req);
e1e16097 5275 io_req_complete(req, ret);
221c5eb2
JA
5276 return 0;
5277}
5278
221c5eb2
JA
5279static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5280 void *key)
5281{
c2f2eb7d
JA
5282 struct io_kiocb *req = wait->private;
5283 struct io_poll_iocb *poll = &req->poll;
221c5eb2 5284
d7718a9d 5285 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
221c5eb2
JA
5286}
5287
221c5eb2
JA
5288static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5289 struct poll_table_struct *p)
5290{
5291 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5292
e8c2bc1f 5293 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
eac406c6
JA
5294}
5295
3529d8c2 5296static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
221c5eb2
JA
5297{
5298 struct io_poll_iocb *poll = &req->poll;
5769a351 5299 u32 events;
221c5eb2
JA
5300
5301 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5302 return -EINVAL;
5303 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5304 return -EINVAL;
5305
5769a351
JX
5306 events = READ_ONCE(sqe->poll32_events);
5307#ifdef __BIG_ENDIAN
5308 events = swahw32(events);
5309#endif
a31eb4a2
JX
5310 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5311 (events & EPOLLEXCLUSIVE);
0969e783
JA
5312 return 0;
5313}
5314
61e98203 5315static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
0969e783
JA
5316{
5317 struct io_poll_iocb *poll = &req->poll;
5318 struct io_ring_ctx *ctx = req->ctx;
5319 struct io_poll_table ipt;
0969e783 5320 __poll_t mask;
0969e783 5321
d7718a9d 5322 ipt.pt._qproc = io_poll_queue_proc;
36703247 5323
d7718a9d
JA
5324 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5325 io_poll_wake);
221c5eb2 5326
8c838788 5327 if (mask) { /* no async, we'd stolen it */
221c5eb2 5328 ipt.error = 0;
b0dd8a41 5329 io_poll_complete(req, mask, 0);
221c5eb2 5330 }
221c5eb2
JA
5331 spin_unlock_irq(&ctx->completion_lock);
5332
8c838788
JA
5333 if (mask) {
5334 io_cqring_ev_posted(ctx);
014db007 5335 io_put_req(req);
221c5eb2 5336 }
8c838788 5337 return ipt.error;
221c5eb2
JA
5338}
5339
5262f567
JA
5340static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5341{
ad8a48ac
JA
5342 struct io_timeout_data *data = container_of(timer,
5343 struct io_timeout_data, timer);
5344 struct io_kiocb *req = data->req;
5345 struct io_ring_ctx *ctx = req->ctx;
5262f567
JA
5346 unsigned long flags;
5347
5262f567 5348 spin_lock_irqsave(&ctx->completion_lock, flags);
a71976f3 5349 list_del_init(&req->timeout.list);
01cec8c1
PB
5350 atomic_set(&req->ctx->cq_timeouts,
5351 atomic_read(&req->ctx->cq_timeouts) + 1);
5352
78e19bbe 5353 io_cqring_fill_event(req, -ETIME);
5262f567
JA
5354 io_commit_cqring(ctx);
5355 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5356
5357 io_cqring_ev_posted(ctx);
4e88d6e7 5358 req_set_fail_links(req);
5262f567
JA
5359 io_put_req(req);
5360 return HRTIMER_NORESTART;
5361}
5362
fbd15848
PB
5363static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5364 __u64 user_data)
f254ac04 5365{
fbd15848 5366 struct io_timeout_data *io;
47f46768
JA
5367 struct io_kiocb *req;
5368 int ret = -ENOENT;
f254ac04 5369
135fcde8 5370 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
47f46768 5371 if (user_data == req->user_data) {
47f46768
JA
5372 ret = 0;
5373 break;
5374 }
5375 }
5376
5377 if (ret == -ENOENT)
fbd15848
PB
5378 return ERR_PTR(ret);
5379
5380 io = req->async_data;
e8c2bc1f 5381 ret = hrtimer_try_to_cancel(&io->timer);
f254ac04 5382 if (ret == -1)
fbd15848 5383 return ERR_PTR(-EALREADY);
a71976f3 5384 list_del_init(&req->timeout.list);
fbd15848
PB
5385 return req;
5386}
47f46768 5387
fbd15848
PB
5388static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5389{
5390 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5391
5392 if (IS_ERR(req))
5393 return PTR_ERR(req);
f254ac04
JA
5394
5395 req_set_fail_links(req);
f254ac04 5396 io_cqring_fill_event(req, -ECANCELED);
216578e5 5397 io_put_req_deferred(req, 1);
f254ac04
JA
5398 return 0;
5399}
5400
9c8e11b3
PB
5401static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5402 struct timespec64 *ts, enum hrtimer_mode mode)
47f46768 5403{
9c8e11b3
PB
5404 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5405 struct io_timeout_data *data;
47f46768 5406
9c8e11b3
PB
5407 if (IS_ERR(req))
5408 return PTR_ERR(req);
47f46768 5409
9c8e11b3
PB
5410 req->timeout.off = 0; /* noseq */
5411 data = req->async_data;
5412 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5413 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5414 data->timer.function = io_timeout_fn;
5415 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5416 return 0;
47f46768
JA
5417}
5418
3529d8c2
JA
5419static int io_timeout_remove_prep(struct io_kiocb *req,
5420 const struct io_uring_sqe *sqe)
b29472ee 5421{
9c8e11b3
PB
5422 struct io_timeout_rem *tr = &req->timeout_rem;
5423
b29472ee
JA
5424 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5425 return -EINVAL;
61710e43
DA
5426 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5427 return -EINVAL;
9c8e11b3 5428 if (sqe->ioprio || sqe->buf_index || sqe->len)
b29472ee
JA
5429 return -EINVAL;
5430
9c8e11b3
PB
5431 tr->addr = READ_ONCE(sqe->addr);
5432 tr->flags = READ_ONCE(sqe->timeout_flags);
5433 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5434 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5435 return -EINVAL;
5436 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5437 return -EFAULT;
5438 } else if (tr->flags) {
5439 /* timeout removal doesn't support flags */
b29472ee 5440 return -EINVAL;
9c8e11b3 5441 }
b29472ee 5442
b29472ee
JA
5443 return 0;
5444}
5445
8662daec
PB
5446static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5447{
5448 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5449 : HRTIMER_MODE_REL;
5450}
5451
11365043
JA
5452/*
5453 * Remove or update an existing timeout command
5454 */
61e98203 5455static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
11365043 5456{
9c8e11b3 5457 struct io_timeout_rem *tr = &req->timeout_rem;
11365043 5458 struct io_ring_ctx *ctx = req->ctx;
47f46768 5459 int ret;
11365043 5460
11365043 5461 spin_lock_irq(&ctx->completion_lock);
8662daec 5462 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
9c8e11b3 5463 ret = io_timeout_cancel(ctx, tr->addr);
8662daec
PB
5464 else
5465 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5466 io_translate_timeout_mode(tr->flags));
11365043 5467
47f46768 5468 io_cqring_fill_event(req, ret);
11365043
JA
5469 io_commit_cqring(ctx);
5470 spin_unlock_irq(&ctx->completion_lock);
5262f567 5471 io_cqring_ev_posted(ctx);
4e88d6e7
JA
5472 if (ret < 0)
5473 req_set_fail_links(req);
ec9c02ad 5474 io_put_req(req);
11365043 5475 return 0;
5262f567
JA
5476}
5477
3529d8c2 5478static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2d28390a 5479 bool is_timeout_link)
5262f567 5480{
ad8a48ac 5481 struct io_timeout_data *data;
a41525ab 5482 unsigned flags;
56080b02 5483 u32 off = READ_ONCE(sqe->off);
5262f567 5484
ad8a48ac 5485 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5262f567 5486 return -EINVAL;
ad8a48ac 5487 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
a41525ab 5488 return -EINVAL;
56080b02 5489 if (off && is_timeout_link)
2d28390a 5490 return -EINVAL;
a41525ab
JA
5491 flags = READ_ONCE(sqe->timeout_flags);
5492 if (flags & ~IORING_TIMEOUT_ABS)
5262f567 5493 return -EINVAL;
bdf20073 5494
bfe68a22 5495 req->timeout.off = off;
26a61679 5496
e8c2bc1f 5497 if (!req->async_data && io_alloc_async_data(req))
26a61679
JA
5498 return -ENOMEM;
5499
e8c2bc1f 5500 data = req->async_data;
ad8a48ac 5501 data->req = req;
ad8a48ac
JA
5502
5503 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5262f567
JA
5504 return -EFAULT;
5505
8662daec 5506 data->mode = io_translate_timeout_mode(flags);
ad8a48ac 5507 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
dd59a3d5 5508 io_req_track_inflight(req);
ad8a48ac
JA
5509 return 0;
5510}
5511
61e98203 5512static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
ad8a48ac 5513{
ad8a48ac 5514 struct io_ring_ctx *ctx = req->ctx;
e8c2bc1f 5515 struct io_timeout_data *data = req->async_data;
ad8a48ac 5516 struct list_head *entry;
bfe68a22 5517 u32 tail, off = req->timeout.off;
ad8a48ac 5518
733f5c95 5519 spin_lock_irq(&ctx->completion_lock);
93bd25bb 5520
5262f567
JA
5521 /*
5522 * sqe->off holds how many events that need to occur for this
93bd25bb
JA
5523 * timeout event to be satisfied. If it isn't set, then this is
5524 * a pure timeout request, sequence isn't used.
5262f567 5525 */
8eb7e2d0 5526 if (io_is_timeout_noseq(req)) {
93bd25bb
JA
5527 entry = ctx->timeout_list.prev;
5528 goto add;
5529 }
5262f567 5530
bfe68a22
PB
5531 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5532 req->timeout.target_seq = tail + off;
5262f567 5533
f010505b
MDG
5534 /* Update the last seq here in case io_flush_timeouts() hasn't.
5535 * This is safe because ->completion_lock is held, and submissions
5536 * and completions are never mixed in the same ->completion_lock section.
5537 */
5538 ctx->cq_last_tm_flush = tail;
5539
5262f567
JA
5540 /*
5541 * Insertion sort, ensuring the first entry in the list is always
5542 * the one we need first.
5543 */
5262f567 5544 list_for_each_prev(entry, &ctx->timeout_list) {
135fcde8
PB
5545 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5546 timeout.list);
5262f567 5547
8eb7e2d0 5548 if (io_is_timeout_noseq(nxt))
93bd25bb 5549 continue;
bfe68a22
PB
5550 /* nxt.seq is behind @tail, otherwise would've been completed */
5551 if (off >= nxt->timeout.target_seq - tail)
5262f567
JA
5552 break;
5553 }
93bd25bb 5554add:
135fcde8 5555 list_add(&req->timeout.list, entry);
ad8a48ac
JA
5556 data->timer.function = io_timeout_fn;
5557 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5262f567 5558 spin_unlock_irq(&ctx->completion_lock);
5262f567
JA
5559 return 0;
5560}
5262f567 5561
62755e35
JA
5562static bool io_cancel_cb(struct io_wq_work *work, void *data)
5563{
5564 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5565
5566 return req->user_data == (unsigned long) data;
5567}
5568
5aa75ed5 5569static int io_async_cancel_one(struct io_uring_task *tctx, void *sqe_addr)
62755e35 5570{
62755e35 5571 enum io_wq_cancel cancel_ret;
62755e35
JA
5572 int ret = 0;
5573
5aa75ed5
JA
5574 if (!tctx->io_wq)
5575 return -ENOENT;
5576
5577 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, sqe_addr, false);
62755e35
JA
5578 switch (cancel_ret) {
5579 case IO_WQ_CANCEL_OK:
5580 ret = 0;
5581 break;
5582 case IO_WQ_CANCEL_RUNNING:
5583 ret = -EALREADY;
5584 break;
5585 case IO_WQ_CANCEL_NOTFOUND:
5586 ret = -ENOENT;
5587 break;
5588 }
5589
e977d6d3
JA
5590 return ret;
5591}
5592
47f46768
JA
5593static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5594 struct io_kiocb *req, __u64 sqe_addr,
014db007 5595 int success_ret)
47f46768
JA
5596{
5597 unsigned long flags;
5598 int ret;
5599
5aa75ed5
JA
5600 ret = io_async_cancel_one(req->task->io_uring,
5601 (void *) (unsigned long) sqe_addr);
47f46768
JA
5602 if (ret != -ENOENT) {
5603 spin_lock_irqsave(&ctx->completion_lock, flags);
5604 goto done;
5605 }
5606
5607 spin_lock_irqsave(&ctx->completion_lock, flags);
5608 ret = io_timeout_cancel(ctx, sqe_addr);
5609 if (ret != -ENOENT)
5610 goto done;
5611 ret = io_poll_cancel(ctx, sqe_addr);
5612done:
b0dd8a41
JA
5613 if (!ret)
5614 ret = success_ret;
47f46768
JA
5615 io_cqring_fill_event(req, ret);
5616 io_commit_cqring(ctx);
5617 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5618 io_cqring_ev_posted(ctx);
5619
4e88d6e7
JA
5620 if (ret < 0)
5621 req_set_fail_links(req);
014db007 5622 io_put_req(req);
47f46768
JA
5623}
5624
3529d8c2
JA
5625static int io_async_cancel_prep(struct io_kiocb *req,
5626 const struct io_uring_sqe *sqe)
e977d6d3 5627{
fbf23849 5628 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
e977d6d3 5629 return -EINVAL;
61710e43
DA
5630 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5631 return -EINVAL;
5632 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
e977d6d3
JA
5633 return -EINVAL;
5634
fbf23849
JA
5635 req->cancel.addr = READ_ONCE(sqe->addr);
5636 return 0;
5637}
5638
61e98203 5639static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
fbf23849
JA
5640{
5641 struct io_ring_ctx *ctx = req->ctx;
fbf23849 5642
014db007 5643 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5262f567
JA
5644 return 0;
5645}
5646
269bbe5f 5647static int io_rsrc_update_prep(struct io_kiocb *req,
05f3fb3c
JA
5648 const struct io_uring_sqe *sqe)
5649{
6ca56f84
JA
5650 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5651 return -EINVAL;
61710e43
DA
5652 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5653 return -EINVAL;
5654 if (sqe->ioprio || sqe->rw_flags)
05f3fb3c
JA
5655 return -EINVAL;
5656
269bbe5f
BM
5657 req->rsrc_update.offset = READ_ONCE(sqe->off);
5658 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5659 if (!req->rsrc_update.nr_args)
05f3fb3c 5660 return -EINVAL;
269bbe5f 5661 req->rsrc_update.arg = READ_ONCE(sqe->addr);
05f3fb3c
JA
5662 return 0;
5663}
5664
889fca73 5665static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
fbf23849
JA
5666{
5667 struct io_ring_ctx *ctx = req->ctx;
269bbe5f 5668 struct io_uring_rsrc_update up;
05f3fb3c 5669 int ret;
fbf23849 5670
45d189c6 5671 if (issue_flags & IO_URING_F_NONBLOCK)
05f3fb3c 5672 return -EAGAIN;
05f3fb3c 5673
269bbe5f
BM
5674 up.offset = req->rsrc_update.offset;
5675 up.data = req->rsrc_update.arg;
05f3fb3c
JA
5676
5677 mutex_lock(&ctx->uring_lock);
269bbe5f 5678 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
05f3fb3c
JA
5679 mutex_unlock(&ctx->uring_lock);
5680
5681 if (ret < 0)
5682 req_set_fail_links(req);
889fca73 5683 __io_req_complete(req, issue_flags, ret, 0);
5262f567
JA
5684 return 0;
5685}
5686
bfe76559 5687static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
f67676d1 5688{
d625c6ee 5689 switch (req->opcode) {
e781573e 5690 case IORING_OP_NOP:
bfe76559 5691 return 0;
f67676d1
JA
5692 case IORING_OP_READV:
5693 case IORING_OP_READ_FIXED:
3a6820f2 5694 case IORING_OP_READ:
bfe76559 5695 return io_read_prep(req, sqe);
f67676d1
JA
5696 case IORING_OP_WRITEV:
5697 case IORING_OP_WRITE_FIXED:
3a6820f2 5698 case IORING_OP_WRITE:
bfe76559 5699 return io_write_prep(req, sqe);
0969e783 5700 case IORING_OP_POLL_ADD:
bfe76559 5701 return io_poll_add_prep(req, sqe);
0969e783 5702 case IORING_OP_POLL_REMOVE:
bfe76559 5703 return io_poll_remove_prep(req, sqe);
8ed8d3c3 5704 case IORING_OP_FSYNC:
1155c76a 5705 return io_fsync_prep(req, sqe);
8ed8d3c3 5706 case IORING_OP_SYNC_FILE_RANGE:
1155c76a 5707 return io_sfr_prep(req, sqe);
03b1230c 5708 case IORING_OP_SENDMSG:
fddaface 5709 case IORING_OP_SEND:
bfe76559 5710 return io_sendmsg_prep(req, sqe);
03b1230c 5711 case IORING_OP_RECVMSG:
fddaface 5712 case IORING_OP_RECV:
bfe76559 5713 return io_recvmsg_prep(req, sqe);
f499a021 5714 case IORING_OP_CONNECT:
bfe76559 5715 return io_connect_prep(req, sqe);
2d28390a 5716 case IORING_OP_TIMEOUT:
bfe76559 5717 return io_timeout_prep(req, sqe, false);
b29472ee 5718 case IORING_OP_TIMEOUT_REMOVE:
bfe76559 5719 return io_timeout_remove_prep(req, sqe);
fbf23849 5720 case IORING_OP_ASYNC_CANCEL:
bfe76559 5721 return io_async_cancel_prep(req, sqe);
2d28390a 5722 case IORING_OP_LINK_TIMEOUT:
bfe76559 5723 return io_timeout_prep(req, sqe, true);
8ed8d3c3 5724 case IORING_OP_ACCEPT:
bfe76559 5725 return io_accept_prep(req, sqe);
d63d1b5e 5726 case IORING_OP_FALLOCATE:
bfe76559 5727 return io_fallocate_prep(req, sqe);
15b71abe 5728 case IORING_OP_OPENAT:
bfe76559 5729 return io_openat_prep(req, sqe);
b5dba59e 5730 case IORING_OP_CLOSE:
bfe76559 5731 return io_close_prep(req, sqe);
05f3fb3c 5732 case IORING_OP_FILES_UPDATE:
269bbe5f 5733 return io_rsrc_update_prep(req, sqe);
eddc7ef5 5734 case IORING_OP_STATX:
bfe76559 5735 return io_statx_prep(req, sqe);
4840e418 5736 case IORING_OP_FADVISE:
bfe76559 5737 return io_fadvise_prep(req, sqe);
c1ca757b 5738 case IORING_OP_MADVISE:
bfe76559 5739 return io_madvise_prep(req, sqe);
cebdb986 5740 case IORING_OP_OPENAT2:
bfe76559 5741 return io_openat2_prep(req, sqe);
3e4827b0 5742 case IORING_OP_EPOLL_CTL:
bfe76559 5743 return io_epoll_ctl_prep(req, sqe);
7d67af2c 5744 case IORING_OP_SPLICE:
bfe76559 5745 return io_splice_prep(req, sqe);
ddf0322d 5746 case IORING_OP_PROVIDE_BUFFERS:
bfe76559 5747 return io_provide_buffers_prep(req, sqe);
067524e9 5748 case IORING_OP_REMOVE_BUFFERS:
bfe76559 5749 return io_remove_buffers_prep(req, sqe);
f2a8d5c7 5750 case IORING_OP_TEE:
bfe76559 5751 return io_tee_prep(req, sqe);
36f4fa68
JA
5752 case IORING_OP_SHUTDOWN:
5753 return io_shutdown_prep(req, sqe);
80a261fd
JA
5754 case IORING_OP_RENAMEAT:
5755 return io_renameat_prep(req, sqe);
14a1143b
JA
5756 case IORING_OP_UNLINKAT:
5757 return io_unlinkat_prep(req, sqe);
f67676d1
JA
5758 }
5759
bfe76559
PB
5760 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5761 req->opcode);
5762 return-EINVAL;
5763}
5764
93642ef8 5765static int io_req_prep_async(struct io_kiocb *req)
bfe76559 5766{
93642ef8
PB
5767 switch (req->opcode) {
5768 case IORING_OP_READV:
5769 case IORING_OP_READ_FIXED:
5770 case IORING_OP_READ:
5771 return io_rw_prep_async(req, READ);
5772 case IORING_OP_WRITEV:
5773 case IORING_OP_WRITE_FIXED:
5774 case IORING_OP_WRITE:
5775 return io_rw_prep_async(req, WRITE);
5776 case IORING_OP_SENDMSG:
5777 case IORING_OP_SEND:
5778 return io_sendmsg_prep_async(req);
5779 case IORING_OP_RECVMSG:
5780 case IORING_OP_RECV:
5781 return io_recvmsg_prep_async(req);
5782 case IORING_OP_CONNECT:
5783 return io_connect_prep_async(req);
5784 }
5785 return 0;
5786}
5787
be7053b7 5788static int io_req_defer_prep(struct io_kiocb *req)
bfe76559 5789{
be7053b7 5790 if (!io_op_defs[req->opcode].needs_async_data)
bfe76559 5791 return 0;
be7053b7 5792 /* some opcodes init it during the inital prep */
93642ef8 5793 if (req->async_data)
be7053b7
PB
5794 return 0;
5795 if (__io_alloc_async_data(req))
bfe76559 5796 return -EAGAIN;
be7053b7 5797 return io_req_prep_async(req);
f67676d1
JA
5798}
5799
9cf7c104
PB
5800static u32 io_get_sequence(struct io_kiocb *req)
5801{
5802 struct io_kiocb *pos;
5803 struct io_ring_ctx *ctx = req->ctx;
f2f87370 5804 u32 total_submitted, nr_reqs = 0;
9cf7c104 5805
f2f87370
PB
5806 io_for_each_link(pos, req)
5807 nr_reqs++;
9cf7c104
PB
5808
5809 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5810 return total_submitted - nr_reqs;
5811}
5812
be7053b7 5813static int io_req_defer(struct io_kiocb *req)
de0617e4 5814{
a197f664 5815 struct io_ring_ctx *ctx = req->ctx;
27dc8338 5816 struct io_defer_entry *de;
f67676d1 5817 int ret;
9cf7c104 5818 u32 seq;
de0617e4 5819
9d858b21 5820 /* Still need defer if there is pending req in defer list. */
9cf7c104
PB
5821 if (likely(list_empty_careful(&ctx->defer_list) &&
5822 !(req->flags & REQ_F_IO_DRAIN)))
5823 return 0;
5824
5825 seq = io_get_sequence(req);
5826 /* Still a chance to pass the sequence check */
5827 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
de0617e4
JA
5828 return 0;
5829
be7053b7
PB
5830 ret = io_req_defer_prep(req);
5831 if (ret)
5832 return ret;
cbdcb435 5833 io_prep_async_link(req);
27dc8338
PB
5834 de = kmalloc(sizeof(*de), GFP_KERNEL);
5835 if (!de)
5836 return -ENOMEM;
2d28390a 5837
de0617e4 5838 spin_lock_irq(&ctx->completion_lock);
9cf7c104 5839 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
de0617e4 5840 spin_unlock_irq(&ctx->completion_lock);
27dc8338 5841 kfree(de);
ae34817b
PB
5842 io_queue_async_work(req);
5843 return -EIOCBQUEUED;
de0617e4
JA
5844 }
5845
915967f6 5846 trace_io_uring_defer(ctx, req, req->user_data);
27dc8338 5847 de->req = req;
9cf7c104 5848 de->seq = seq;
27dc8338 5849 list_add_tail(&de->list, &ctx->defer_list);
de0617e4
JA
5850 spin_unlock_irq(&ctx->completion_lock);
5851 return -EIOCBQUEUED;
5852}
5853
3ca405eb 5854static void __io_clean_op(struct io_kiocb *req)
99bc4c38 5855{
0e1b6fe3
PB
5856 if (req->flags & REQ_F_BUFFER_SELECTED) {
5857 switch (req->opcode) {
5858 case IORING_OP_READV:
5859 case IORING_OP_READ_FIXED:
5860 case IORING_OP_READ:
bcda7baa 5861 kfree((void *)(unsigned long)req->rw.addr);
0e1b6fe3
PB
5862 break;
5863 case IORING_OP_RECVMSG:
5864 case IORING_OP_RECV:
bcda7baa 5865 kfree(req->sr_msg.kbuf);
0e1b6fe3
PB
5866 break;
5867 }
5868 req->flags &= ~REQ_F_BUFFER_SELECTED;
99bc4c38
PB
5869 }
5870
0e1b6fe3
PB
5871 if (req->flags & REQ_F_NEED_CLEANUP) {
5872 switch (req->opcode) {
5873 case IORING_OP_READV:
5874 case IORING_OP_READ_FIXED:
5875 case IORING_OP_READ:
5876 case IORING_OP_WRITEV:
5877 case IORING_OP_WRITE_FIXED:
e8c2bc1f
JA
5878 case IORING_OP_WRITE: {
5879 struct io_async_rw *io = req->async_data;
5880 if (io->free_iovec)
5881 kfree(io->free_iovec);
0e1b6fe3 5882 break;
e8c2bc1f 5883 }
0e1b6fe3 5884 case IORING_OP_RECVMSG:
e8c2bc1f
JA
5885 case IORING_OP_SENDMSG: {
5886 struct io_async_msghdr *io = req->async_data;
257e84a5
PB
5887
5888 kfree(io->free_iov);
0e1b6fe3 5889 break;
e8c2bc1f 5890 }
0e1b6fe3
PB
5891 case IORING_OP_SPLICE:
5892 case IORING_OP_TEE:
5893 io_put_file(req, req->splice.file_in,
5894 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5895 break;
f3cd4850
JA
5896 case IORING_OP_OPENAT:
5897 case IORING_OP_OPENAT2:
5898 if (req->open.filename)
5899 putname(req->open.filename);
5900 break;
80a261fd
JA
5901 case IORING_OP_RENAMEAT:
5902 putname(req->rename.oldpath);
5903 putname(req->rename.newpath);
5904 break;
14a1143b
JA
5905 case IORING_OP_UNLINKAT:
5906 putname(req->unlink.filename);
5907 break;
0e1b6fe3
PB
5908 }
5909 req->flags &= ~REQ_F_NEED_CLEANUP;
99bc4c38 5910 }
99bc4c38
PB
5911}
5912
889fca73 5913static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
2b188cc1 5914{
a197f664 5915 struct io_ring_ctx *ctx = req->ctx;
5730b27e 5916 const struct cred *creds = NULL;
d625c6ee 5917 int ret;
2b188cc1 5918
5730b27e
JA
5919 if (req->work.personality) {
5920 const struct cred *new_creds;
5921
5922 if (!(issue_flags & IO_URING_F_NONBLOCK))
5923 mutex_lock(&ctx->uring_lock);
5924 new_creds = idr_find(&ctx->personality_idr, req->work.personality);
5925 if (!(issue_flags & IO_URING_F_NONBLOCK))
5926 mutex_unlock(&ctx->uring_lock);
5927 if (!new_creds)
5928 return -EINVAL;
5929 creds = override_creds(new_creds);
5930 }
5931
d625c6ee 5932 switch (req->opcode) {
2b188cc1 5933 case IORING_OP_NOP:
889fca73 5934 ret = io_nop(req, issue_flags);
2b188cc1
JA
5935 break;
5936 case IORING_OP_READV:
edafccee 5937 case IORING_OP_READ_FIXED:
3a6820f2 5938 case IORING_OP_READ:
889fca73 5939 ret = io_read(req, issue_flags);
edafccee 5940 break;
3529d8c2 5941 case IORING_OP_WRITEV:
edafccee 5942 case IORING_OP_WRITE_FIXED:
3a6820f2 5943 case IORING_OP_WRITE:
889fca73 5944 ret = io_write(req, issue_flags);
2b188cc1 5945 break;
c992fe29 5946 case IORING_OP_FSYNC:
45d189c6 5947 ret = io_fsync(req, issue_flags);
c992fe29 5948 break;
221c5eb2 5949 case IORING_OP_POLL_ADD:
61e98203 5950 ret = io_poll_add(req, issue_flags);
221c5eb2
JA
5951 break;
5952 case IORING_OP_POLL_REMOVE:
61e98203 5953 ret = io_poll_remove(req, issue_flags);
221c5eb2 5954 break;
5d17b4a4 5955 case IORING_OP_SYNC_FILE_RANGE:
45d189c6 5956 ret = io_sync_file_range(req, issue_flags);
5d17b4a4 5957 break;
0fa03c62 5958 case IORING_OP_SENDMSG:
889fca73 5959 ret = io_sendmsg(req, issue_flags);
062d04d7 5960 break;
fddaface 5961 case IORING_OP_SEND:
889fca73 5962 ret = io_send(req, issue_flags);
0fa03c62 5963 break;
aa1fa28f 5964 case IORING_OP_RECVMSG:
889fca73 5965 ret = io_recvmsg(req, issue_flags);
062d04d7 5966 break;
fddaface 5967 case IORING_OP_RECV:
889fca73 5968 ret = io_recv(req, issue_flags);
aa1fa28f 5969 break;
5262f567 5970 case IORING_OP_TIMEOUT:
61e98203 5971 ret = io_timeout(req, issue_flags);
5262f567 5972 break;
11365043 5973 case IORING_OP_TIMEOUT_REMOVE:
61e98203 5974 ret = io_timeout_remove(req, issue_flags);
11365043 5975 break;
17f2fe35 5976 case IORING_OP_ACCEPT:
889fca73 5977 ret = io_accept(req, issue_flags);
17f2fe35 5978 break;
f8e85cf2 5979 case IORING_OP_CONNECT:
889fca73 5980 ret = io_connect(req, issue_flags);
f8e85cf2 5981 break;
62755e35 5982 case IORING_OP_ASYNC_CANCEL:
61e98203 5983 ret = io_async_cancel(req, issue_flags);
62755e35 5984 break;
d63d1b5e 5985 case IORING_OP_FALLOCATE:
45d189c6 5986 ret = io_fallocate(req, issue_flags);
d63d1b5e 5987 break;
15b71abe 5988 case IORING_OP_OPENAT:
45d189c6 5989 ret = io_openat(req, issue_flags);
15b71abe 5990 break;
b5dba59e 5991 case IORING_OP_CLOSE:
889fca73 5992 ret = io_close(req, issue_flags);
b5dba59e 5993 break;
05f3fb3c 5994 case IORING_OP_FILES_UPDATE:
889fca73 5995 ret = io_files_update(req, issue_flags);
05f3fb3c 5996 break;
eddc7ef5 5997 case IORING_OP_STATX:
45d189c6 5998 ret = io_statx(req, issue_flags);
eddc7ef5 5999 break;
4840e418 6000 case IORING_OP_FADVISE:
45d189c6 6001 ret = io_fadvise(req, issue_flags);
4840e418 6002 break;
c1ca757b 6003 case IORING_OP_MADVISE:
45d189c6 6004 ret = io_madvise(req, issue_flags);
c1ca757b 6005 break;
cebdb986 6006 case IORING_OP_OPENAT2:
45d189c6 6007 ret = io_openat2(req, issue_flags);
cebdb986 6008 break;
3e4827b0 6009 case IORING_OP_EPOLL_CTL:
889fca73 6010 ret = io_epoll_ctl(req, issue_flags);
3e4827b0 6011 break;
7d67af2c 6012 case IORING_OP_SPLICE:
45d189c6 6013 ret = io_splice(req, issue_flags);
7d67af2c 6014 break;
ddf0322d 6015 case IORING_OP_PROVIDE_BUFFERS:
889fca73 6016 ret = io_provide_buffers(req, issue_flags);
ddf0322d 6017 break;
067524e9 6018 case IORING_OP_REMOVE_BUFFERS:
889fca73 6019 ret = io_remove_buffers(req, issue_flags);
3e4827b0 6020 break;
f2a8d5c7 6021 case IORING_OP_TEE:
45d189c6 6022 ret = io_tee(req, issue_flags);
f2a8d5c7 6023 break;
36f4fa68 6024 case IORING_OP_SHUTDOWN:
45d189c6 6025 ret = io_shutdown(req, issue_flags);
36f4fa68 6026 break;
80a261fd 6027 case IORING_OP_RENAMEAT:
45d189c6 6028 ret = io_renameat(req, issue_flags);
80a261fd 6029 break;
14a1143b 6030 case IORING_OP_UNLINKAT:
45d189c6 6031 ret = io_unlinkat(req, issue_flags);
14a1143b 6032 break;
2b188cc1
JA
6033 default:
6034 ret = -EINVAL;
6035 break;
6036 }
6037
5730b27e
JA
6038 if (creds)
6039 revert_creds(creds);
6040
def596e9
JA
6041 if (ret)
6042 return ret;
6043
b532576e
JA
6044 /* If the op doesn't have a file, we're not polling for it */
6045 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
11ba820b
JA
6046 const bool in_async = io_wq_current_is_worker();
6047
11ba820b
JA
6048 /* workqueue context doesn't hold uring_lock, grab it now */
6049 if (in_async)
6050 mutex_lock(&ctx->uring_lock);
6051
2e9dbe90 6052 io_iopoll_req_issued(req, in_async);
11ba820b
JA
6053
6054 if (in_async)
6055 mutex_unlock(&ctx->uring_lock);
def596e9
JA
6056 }
6057
6058 return 0;
2b188cc1
JA
6059}
6060
5280f7e5 6061static void io_wq_submit_work(struct io_wq_work *work)
2b188cc1
JA
6062{
6063 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6df1db6b 6064 struct io_kiocb *timeout;
561fb04a 6065 int ret = 0;
2b188cc1 6066
6df1db6b
PB
6067 timeout = io_prep_linked_timeout(req);
6068 if (timeout)
6069 io_queue_linked_timeout(timeout);
d4c81f38 6070
4014d943 6071 if (work->flags & IO_WQ_WORK_CANCEL)
561fb04a 6072 ret = -ECANCELED;
31b51510 6073
561fb04a 6074 if (!ret) {
561fb04a 6075 do {
889fca73 6076 ret = io_issue_sqe(req, 0);
561fb04a
JA
6077 /*
6078 * We can get EAGAIN for polled IO even though we're
6079 * forcing a sync submission from here, since we can't
6080 * wait for request slots on the block side.
6081 */
6082 if (ret != -EAGAIN)
6083 break;
6084 cond_resched();
6085 } while (1);
6086 }
31b51510 6087
a3df7698 6088 /* avoid locking problems by failing it from a clean context */
561fb04a 6089 if (ret) {
a3df7698
PB
6090 /* io-wq is going to take one down */
6091 refcount_inc(&req->refs);
6092 io_req_task_queue_fail(req, ret);
edafccee 6093 }
2b188cc1
JA
6094}
6095
65e19f54
JA
6096static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6097 int index)
6098{
269bbe5f 6099 struct fixed_rsrc_table *table;
65e19f54 6100
05f3fb3c 6101 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
84695089 6102 return table->files[index & IORING_FILE_TABLE_MASK];
65e19f54
JA
6103}
6104
8371adf5
PB
6105static struct file *io_file_get(struct io_submit_state *state,
6106 struct io_kiocb *req, int fd, bool fixed)
09bb8394 6107{
a197f664 6108 struct io_ring_ctx *ctx = req->ctx;
8da11c19 6109 struct file *file;
09bb8394 6110
8da11c19 6111 if (fixed) {
479f517b 6112 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
8371adf5 6113 return NULL;
b7620121 6114 fd = array_index_nospec(fd, ctx->nr_user_files);
8da11c19 6115 file = io_file_from_index(ctx, fd);
36f72fe2 6116 io_set_resource_node(req);
09bb8394 6117 } else {
c826bd7a 6118 trace_io_uring_file_get(ctx, fd);
8da11c19 6119 file = __io_file_get(state, fd);
09bb8394
JA
6120 }
6121
ce3d5aae
PB
6122 if (file && unlikely(file->f_op == &io_uring_fops))
6123 io_req_track_inflight(req);
8371adf5 6124 return file;
09bb8394
JA
6125}
6126
2665abfd 6127static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
2b188cc1 6128{
ad8a48ac
JA
6129 struct io_timeout_data *data = container_of(timer,
6130 struct io_timeout_data, timer);
90cd7e42 6131 struct io_kiocb *prev, *req = data->req;
2665abfd 6132 struct io_ring_ctx *ctx = req->ctx;
2665abfd 6133 unsigned long flags;
2665abfd
JA
6134
6135 spin_lock_irqsave(&ctx->completion_lock, flags);
90cd7e42
PB
6136 prev = req->timeout.head;
6137 req->timeout.head = NULL;
2665abfd
JA
6138
6139 /*
6140 * We don't expect the list to be empty, that will only happen if we
6141 * race with the completion of the linked work.
6142 */
90cd7e42 6143 if (prev && refcount_inc_not_zero(&prev->refs))
f2f87370 6144 io_remove_next_linked(prev);
90cd7e42
PB
6145 else
6146 prev = NULL;
2665abfd
JA
6147 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6148
6149 if (prev) {
4e88d6e7 6150 req_set_fail_links(prev);
014db007 6151 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
9ae1f8dd 6152 io_put_req_deferred(prev, 1);
47f46768 6153 } else {
9ae1f8dd
PB
6154 io_req_complete_post(req, -ETIME, 0);
6155 io_put_req_deferred(req, 1);
2665abfd 6156 }
2665abfd
JA
6157 return HRTIMER_NORESTART;
6158}
6159
7271ef3a 6160static void __io_queue_linked_timeout(struct io_kiocb *req)
2665abfd 6161{
76a46e06 6162 /*
f2f87370
PB
6163 * If the back reference is NULL, then our linked request finished
6164 * before we got a chance to setup the timer
76a46e06 6165 */
90cd7e42 6166 if (req->timeout.head) {
e8c2bc1f 6167 struct io_timeout_data *data = req->async_data;
94ae5e77 6168
ad8a48ac
JA
6169 data->timer.function = io_link_timeout_fn;
6170 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6171 data->mode);
2665abfd 6172 }
7271ef3a
JA
6173}
6174
6175static void io_queue_linked_timeout(struct io_kiocb *req)
6176{
6177 struct io_ring_ctx *ctx = req->ctx;
6178
6179 spin_lock_irq(&ctx->completion_lock);
6180 __io_queue_linked_timeout(req);
76a46e06 6181 spin_unlock_irq(&ctx->completion_lock);
2665abfd 6182
2665abfd 6183 /* drop submission reference */
76a46e06
JA
6184 io_put_req(req);
6185}
2665abfd 6186
ad8a48ac 6187static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
2665abfd 6188{
f2f87370 6189 struct io_kiocb *nxt = req->link;
2665abfd 6190
f2f87370
PB
6191 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6192 nxt->opcode != IORING_OP_LINK_TIMEOUT)
76a46e06 6193 return NULL;
2665abfd 6194
90cd7e42 6195 nxt->timeout.head = req;
900fad45 6196 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
76a46e06 6197 req->flags |= REQ_F_LINK_TIMEOUT;
76a46e06 6198 return nxt;
2665abfd
JA
6199}
6200
c5eef2b9 6201static void __io_queue_sqe(struct io_kiocb *req)
2b188cc1 6202{
d3d7298d 6203 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
e0c5c576 6204 int ret;
2b188cc1 6205
c5eef2b9 6206 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
193155c8 6207
491381ce
JA
6208 /*
6209 * We async punt it if the file wasn't marked NOWAIT, or if the file
6210 * doesn't support non-blocking read/write attempts
6211 */
24c74678 6212 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
f063c547 6213 if (!io_arm_poll_handler(req)) {
f063c547
PB
6214 /*
6215 * Queued up for async execution, worker will release
6216 * submit reference when the iocb is actually submitted.
6217 */
6218 io_queue_async_work(req);
2b188cc1 6219 }
0d63c148
PB
6220 } else if (likely(!ret)) {
6221 /* drop submission reference */
e342c807 6222 if (req->flags & REQ_F_COMPLETE_INLINE) {
c5eef2b9
PB
6223 struct io_ring_ctx *ctx = req->ctx;
6224 struct io_comp_state *cs = &ctx->submit_state.comp;
e65ef56d 6225
6dd0be1e 6226 cs->reqs[cs->nr++] = req;
d3d7298d 6227 if (cs->nr == ARRAY_SIZE(cs->reqs))
c5eef2b9 6228 io_submit_flush_completions(cs, ctx);
9affd664 6229 } else {
d3d7298d 6230 io_put_req(req);
0d63c148
PB
6231 }
6232 } else {
4e88d6e7 6233 req_set_fail_links(req);
e65ef56d 6234 io_put_req(req);
e1e16097 6235 io_req_complete(req, ret);
9e645e11 6236 }
d3d7298d
PB
6237 if (linked_timeout)
6238 io_queue_linked_timeout(linked_timeout);
2b188cc1
JA
6239}
6240
be7053b7 6241static void io_queue_sqe(struct io_kiocb *req)
4fe2c963
JL
6242{
6243 int ret;
6244
be7053b7 6245 ret = io_req_defer(req);
4fe2c963
JL
6246 if (ret) {
6247 if (ret != -EIOCBQUEUED) {
1118591a 6248fail_req:
4e88d6e7 6249 req_set_fail_links(req);
e1e16097
JA
6250 io_put_req(req);
6251 io_req_complete(req, ret);
4fe2c963 6252 }
2550878f 6253 } else if (req->flags & REQ_F_FORCE_ASYNC) {
be7053b7
PB
6254 ret = io_req_defer_prep(req);
6255 if (unlikely(ret))
6256 goto fail_req;
ce35a47a
JA
6257 io_queue_async_work(req);
6258 } else {
c5eef2b9 6259 __io_queue_sqe(req);
ce35a47a 6260 }
4fe2c963
JL
6261}
6262
b16fed66
PB
6263/*
6264 * Check SQE restrictions (opcode and flags).
6265 *
6266 * Returns 'true' if SQE is allowed, 'false' otherwise.
6267 */
6268static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6269 struct io_kiocb *req,
6270 unsigned int sqe_flags)
4fe2c963 6271{
b16fed66
PB
6272 if (!ctx->restricted)
6273 return true;
6274
6275 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6276 return false;
6277
6278 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6279 ctx->restrictions.sqe_flags_required)
6280 return false;
6281
6282 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6283 ctx->restrictions.sqe_flags_required))
6284 return false;
6285
6286 return true;
4fe2c963
JL
6287}
6288
b16fed66
PB
6289static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6290 const struct io_uring_sqe *sqe)
6291{
6292 struct io_submit_state *state;
6293 unsigned int sqe_flags;
5730b27e 6294 int ret = 0;
b16fed66
PB
6295
6296 req->opcode = READ_ONCE(sqe->opcode);
6297 /* same numerical values with corresponding REQ_F_*, safe to copy */
6298 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6299 req->user_data = READ_ONCE(sqe->user_data);
6300 req->async_data = NULL;
6301 req->file = NULL;
6302 req->ctx = ctx;
6303 req->link = NULL;
6304 req->fixed_rsrc_refs = NULL;
6305 /* one is dropped after submission, the other at completion */
6306 refcount_set(&req->refs, 2);
6307 req->task = current;
6308 req->result = 0;
6309
6310 /* enforce forwards compatibility on users */
ebf4a5db
PB
6311 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6312 req->flags = 0;
b16fed66 6313 return -EINVAL;
ebf4a5db 6314 }
b16fed66
PB
6315
6316 if (unlikely(req->opcode >= IORING_OP_LAST))
6317 return -EINVAL;
6318
b16fed66
PB
6319 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6320 return -EACCES;
6321
6322 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6323 !io_op_defs[req->opcode].buffer_select)
6324 return -EOPNOTSUPP;
863e0560 6325
5730b27e
JA
6326 req->work.list.next = NULL;
6327 req->work.flags = 0;
6328 req->work.personality = READ_ONCE(sqe->personality);
b16fed66
PB
6329 state = &ctx->submit_state;
6330
6331 /*
6332 * Plug now if we have more than 1 IO left after this, and the target
6333 * is potentially a read/write to block based storage.
6334 */
6335 if (!state->plug_started && state->ios_left > 1 &&
6336 io_op_defs[req->opcode].plug) {
6337 blk_start_plug(&state->plug);
6338 state->plug_started = true;
6339 }
6340
6341 if (io_op_defs[req->opcode].needs_file) {
6342 bool fixed = req->flags & REQ_F_FIXED_FILE;
6343
6344 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6345 if (unlikely(!req->file))
6346 ret = -EBADF;
6347 }
6348
6349 state->ios_left--;
6350 return ret;
6351}
6352
a6b8cadc 6353static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
a1ab7b35 6354 const struct io_uring_sqe *sqe)
9e645e11 6355{
a1ab7b35 6356 struct io_submit_link *link = &ctx->submit_state.link;
ef4ff581 6357 int ret;
9e645e11 6358
a6b8cadc
PB
6359 ret = io_init_req(ctx, req, sqe);
6360 if (unlikely(ret)) {
6361fail_req:
6362 io_put_req(req);
6363 io_req_complete(req, ret);
de59bc10
PB
6364 if (link->head) {
6365 /* fail even hard links since we don't submit */
cf109604 6366 link->head->flags |= REQ_F_FAIL_LINK;
de59bc10
PB
6367 io_put_req(link->head);
6368 io_req_complete(link->head, -ECANCELED);
6369 link->head = NULL;
6370 }
a6b8cadc
PB
6371 return ret;
6372 }
be7053b7
PB
6373 ret = io_req_prep(req, sqe);
6374 if (unlikely(ret))
6375 goto fail_req;
a6b8cadc 6376
be7053b7 6377 /* don't need @sqe from now on */
a6b8cadc
PB
6378 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6379 true, ctx->flags & IORING_SETUP_SQPOLL);
6380
9e645e11
JA
6381 /*
6382 * If we already have a head request, queue this one for async
6383 * submittal once the head completes. If we don't have a head but
6384 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6385 * submitted sync once the chain is complete. If none of those
6386 * conditions are true (normal request), then just queue it.
6387 */
863e0560
PB
6388 if (link->head) {
6389 struct io_kiocb *head = link->head;
4e88d6e7 6390
8cdf2193
PB
6391 /*
6392 * Taking sequential execution of a link, draining both sides
6393 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6394 * requests in the link. So, it drains the head and the
6395 * next after the link request. The last one is done via
6396 * drain_next flag to persist the effect across calls.
6397 */
ef4ff581 6398 if (req->flags & REQ_F_IO_DRAIN) {
711be031
PB
6399 head->flags |= REQ_F_IO_DRAIN;
6400 ctx->drain_next = 1;
6401 }
be7053b7 6402 ret = io_req_defer_prep(req);
cf109604 6403 if (unlikely(ret))
a6b8cadc 6404 goto fail_req;
9d76377f 6405 trace_io_uring_link(ctx, req, head);
f2f87370 6406 link->last->link = req;
863e0560 6407 link->last = req;
32fe525b
PB
6408
6409 /* last request of a link, enqueue the link */
ef4ff581 6410 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
de59bc10 6411 io_queue_sqe(head);
863e0560 6412 link->head = NULL;
32fe525b 6413 }
9e645e11 6414 } else {
711be031
PB
6415 if (unlikely(ctx->drain_next)) {
6416 req->flags |= REQ_F_IO_DRAIN;
ef4ff581 6417 ctx->drain_next = 0;
711be031 6418 }
ef4ff581 6419 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
863e0560
PB
6420 link->head = req;
6421 link->last = req;
711be031 6422 } else {
be7053b7 6423 io_queue_sqe(req);
711be031 6424 }
9e645e11 6425 }
2e6e1fde 6426
1d4240cc 6427 return 0;
9e645e11
JA
6428}
6429
9a56a232
JA
6430/*
6431 * Batched submission is done, ensure local IO is flushed out.
6432 */
ba88ff11
PB
6433static void io_submit_state_end(struct io_submit_state *state,
6434 struct io_ring_ctx *ctx)
9a56a232 6435{
a1ab7b35 6436 if (state->link.head)
de59bc10 6437 io_queue_sqe(state->link.head);
6dd0be1e 6438 if (state->comp.nr)
ba88ff11 6439 io_submit_flush_completions(&state->comp, ctx);
27926b68
JA
6440 if (state->plug_started)
6441 blk_finish_plug(&state->plug);
9f13c35b 6442 io_state_file_put(state);
9a56a232
JA
6443}
6444
6445/*
6446 * Start submission side cache.
6447 */
6448static void io_submit_state_start(struct io_submit_state *state,
ba88ff11 6449 unsigned int max_ios)
9a56a232 6450{
27926b68 6451 state->plug_started = false;
9a56a232 6452 state->ios_left = max_ios;
a1ab7b35
PB
6453 /* set only head, no need to init link_last in advance */
6454 state->link.head = NULL;
9a56a232
JA
6455}
6456
2b188cc1
JA
6457static void io_commit_sqring(struct io_ring_ctx *ctx)
6458{
75b28aff 6459 struct io_rings *rings = ctx->rings;
2b188cc1 6460
caf582c6
PB
6461 /*
6462 * Ensure any loads from the SQEs are done at this point,
6463 * since once we write the new head, the application could
6464 * write new data to them.
6465 */
6466 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2b188cc1
JA
6467}
6468
2b188cc1 6469/*
3529d8c2 6470 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
2b188cc1
JA
6471 * that is mapped by userspace. This means that care needs to be taken to
6472 * ensure that reads are stable, as we cannot rely on userspace always
6473 * being a good citizen. If members of the sqe are validated and then later
6474 * used, it's important that those reads are done through READ_ONCE() to
6475 * prevent a re-load down the line.
6476 */
709b302f 6477static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2b188cc1 6478{
75b28aff 6479 u32 *sq_array = ctx->sq_array;
2b188cc1
JA
6480 unsigned head;
6481
6482 /*
6483 * The cached sq head (or cq tail) serves two purposes:
6484 *
6485 * 1) allows us to batch the cost of updating the user visible
6486 * head updates.
6487 * 2) allows the kernel side to track the head on its own, even
6488 * though the application is the one updating it.
6489 */
4fccfcbb 6490 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
709b302f
PB
6491 if (likely(head < ctx->sq_entries))
6492 return &ctx->sq_sqes[head];
2b188cc1
JA
6493
6494 /* drop invalid entries */
498ccd9e 6495 ctx->cached_sq_dropped++;
ee7d46d9 6496 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
709b302f
PB
6497 return NULL;
6498}
6499
0f212204 6500static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6c271ce2 6501{
46c4e16a 6502 int submitted = 0;
6c271ce2 6503
c4a2ed72 6504 /* if we have a backlog and couldn't flush it all, return BUSY */
ad3eb2c8 6505 if (test_bit(0, &ctx->sq_check_overflow)) {
6c503150 6506 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
ad3eb2c8
JA
6507 return -EBUSY;
6508 }
6c271ce2 6509
ee7d46d9
PB
6510 /* make sure SQ entry isn't read before tail */
6511 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
9ef4f124 6512
2b85edfc
PB
6513 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6514 return -EAGAIN;
6c271ce2 6515
d8a6df10 6516 percpu_counter_add(&current->io_uring->inflight, nr);
faf7b51c 6517 refcount_add(nr, &current->usage);
ba88ff11 6518 io_submit_state_start(&ctx->submit_state, nr);
b14cca0c 6519
46c4e16a 6520 while (submitted < nr) {
3529d8c2 6521 const struct io_uring_sqe *sqe;
196be95c 6522 struct io_kiocb *req;
fb5ccc98 6523
258b29a9 6524 req = io_alloc_req(ctx);
196be95c
PB
6525 if (unlikely(!req)) {
6526 if (!submitted)
6527 submitted = -EAGAIN;
fb5ccc98 6528 break;
196be95c 6529 }
4fccfcbb
PB
6530 sqe = io_get_sqe(ctx);
6531 if (unlikely(!sqe)) {
6532 kmem_cache_free(req_cachep, req);
6533 break;
6534 }
d3656344
JA
6535 /* will complete beyond this point, count as submitted */
6536 submitted++;
a1ab7b35 6537 if (io_submit_sqe(ctx, req, sqe))
196be95c 6538 break;
6c271ce2
JA
6539 }
6540
9466f437
PB
6541 if (unlikely(submitted != nr)) {
6542 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
d8a6df10
JA
6543 struct io_uring_task *tctx = current->io_uring;
6544 int unused = nr - ref_used;
9466f437 6545
d8a6df10
JA
6546 percpu_ref_put_many(&ctx->refs, unused);
6547 percpu_counter_sub(&tctx->inflight, unused);
6548 put_task_struct_many(current, unused);
9466f437 6549 }
6c271ce2 6550
a1ab7b35 6551 io_submit_state_end(&ctx->submit_state, ctx);
ae9428ca
PB
6552 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6553 io_commit_sqring(ctx);
6554
6c271ce2
JA
6555 return submitted;
6556}
6557
23b3628e
XW
6558static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6559{
6560 /* Tell userspace we may need a wakeup call */
6561 spin_lock_irq(&ctx->completion_lock);
6562 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6563 spin_unlock_irq(&ctx->completion_lock);
6564}
6565
6566static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6567{
6568 spin_lock_irq(&ctx->completion_lock);
6569 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6570 spin_unlock_irq(&ctx->completion_lock);
6571}
6572
08369246 6573static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6c271ce2 6574{
c8d1ba58 6575 unsigned int to_submit;
bdcd3eab 6576 int ret = 0;
6c271ce2 6577
c8d1ba58 6578 to_submit = io_sqring_entries(ctx);
e95eee2d
JA
6579 /* if we're handling multiple rings, cap submit size for fairness */
6580 if (cap_entries && to_submit > 8)
6581 to_submit = 8;
6582
906a3c6f 6583 if (!list_empty(&ctx->iopoll_list) || to_submit) {
c8d1ba58 6584 unsigned nr_events = 0;
a4c0b3de 6585
c8d1ba58 6586 mutex_lock(&ctx->uring_lock);
906a3c6f 6587 if (!list_empty(&ctx->iopoll_list))
c8d1ba58 6588 io_do_iopoll(ctx, &nr_events, 0);
906a3c6f 6589
70aacfe6 6590 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)))
08369246 6591 ret = io_submit_sqes(ctx, to_submit);
c8d1ba58
JA
6592 mutex_unlock(&ctx->uring_lock);
6593 }
6c271ce2 6594
90554200
JA
6595 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6596 wake_up(&ctx->sqo_sq_wait);
6c271ce2 6597
08369246
XW
6598 return ret;
6599}
6c271ce2 6600
08369246
XW
6601static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6602{
6603 struct io_ring_ctx *ctx;
6604 unsigned sq_thread_idle = 0;
6c271ce2 6605
08369246
XW
6606 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6607 if (sq_thread_idle < ctx->sq_thread_idle)
6608 sq_thread_idle = ctx->sq_thread_idle;
c8d1ba58 6609 }
c1edbf5f 6610
08369246 6611 sqd->sq_thread_idle = sq_thread_idle;
c8d1ba58 6612}
6c271ce2 6613
69fb2131
JA
6614static void io_sqd_init_new(struct io_sq_data *sqd)
6615{
6616 struct io_ring_ctx *ctx;
6617
6618 while (!list_empty(&sqd->ctx_new_list)) {
6619 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
69fb2131
JA
6620 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6621 complete(&ctx->sq_thread_comp);
6622 }
08369246
XW
6623
6624 io_sqd_update_thread_idle(sqd);
69fb2131
JA
6625}
6626
37d1e2e3
JA
6627static bool io_sq_thread_should_stop(struct io_sq_data *sqd)
6628{
6629 return test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6630}
6631
6632static bool io_sq_thread_should_park(struct io_sq_data *sqd)
6633{
6634 return test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
6635}
6636
6637static void io_sq_thread_parkme(struct io_sq_data *sqd)
6638{
6639 for (;;) {
6640 /*
6641 * TASK_PARKED is a special state; we must serialize against
6642 * possible pending wakeups to avoid store-store collisions on
6643 * task->state.
6644 *
6645 * Such a collision might possibly result in the task state
6646 * changin from TASK_PARKED and us failing the
6647 * wait_task_inactive() in kthread_park().
6648 */
6649 set_special_state(TASK_PARKED);
6650 if (!test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state))
6651 break;
6652
6653 /*
6654 * Thread is going to call schedule(), do not preempt it,
6655 * or the caller of kthread_park() may spend more time in
6656 * wait_task_inactive().
6657 */
6658 preempt_disable();
6659 complete(&sqd->completion);
6660 schedule_preempt_disabled();
6661 preempt_enable();
6662 }
6663 __set_current_state(TASK_RUNNING);
6664}
6665
c8d1ba58
JA
6666static int io_sq_thread(void *data)
6667{
69fb2131
JA
6668 struct io_sq_data *sqd = data;
6669 struct io_ring_ctx *ctx;
a0d9205f 6670 unsigned long timeout = 0;
37d1e2e3 6671 char buf[TASK_COMM_LEN];
08369246 6672 DEFINE_WAIT(wait);
6c271ce2 6673
37d1e2e3
JA
6674 sprintf(buf, "iou-sqp-%d", sqd->task_pid);
6675 set_task_comm(current, buf);
37d1e2e3
JA
6676 current->pf_io_worker = NULL;
6677
6678 if (sqd->sq_cpu != -1)
6679 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6680 else
6681 set_cpus_allowed_ptr(current, cpu_online_mask);
6682 current->flags |= PF_NO_SETAFFINITY;
6683
37d1e2e3
JA
6684 wait_for_completion(&sqd->startup);
6685
6686 while (!io_sq_thread_should_stop(sqd)) {
08369246
XW
6687 int ret;
6688 bool cap_entries, sqt_spin, needs_sched;
c1edbf5f
JA
6689
6690 /*
69fb2131 6691 * Any changes to the sqd lists are synchronized through the
37d1e2e3 6692 * thread parking. This synchronizes the thread vs users,
69fb2131 6693 * the users are synchronized on the sqd->ctx_lock.
c1edbf5f 6694 */
37d1e2e3
JA
6695 if (io_sq_thread_should_park(sqd)) {
6696 io_sq_thread_parkme(sqd);
6697 continue;
65b2b213 6698 }
08369246 6699 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
69fb2131 6700 io_sqd_init_new(sqd);
08369246
XW
6701 timeout = jiffies + sqd->sq_thread_idle;
6702 }
37d1e2e3
JA
6703 if (fatal_signal_pending(current))
6704 break;
08369246 6705 sqt_spin = false;
e95eee2d 6706 cap_entries = !list_is_singular(&sqd->ctx_list);
69fb2131 6707 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
08369246
XW
6708 ret = __io_sq_thread(ctx, cap_entries);
6709 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6710 sqt_spin = true;
69fb2131 6711 }
6c271ce2 6712
08369246 6713 if (sqt_spin || !time_after(jiffies, timeout)) {
c8d1ba58
JA
6714 io_run_task_work();
6715 cond_resched();
08369246
XW
6716 if (sqt_spin)
6717 timeout = jiffies + sqd->sq_thread_idle;
6718 continue;
6719 }
6720
08369246
XW
6721 needs_sched = true;
6722 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6723 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6724 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6725 !list_empty_careful(&ctx->iopoll_list)) {
6726 needs_sched = false;
6727 break;
6728 }
6729 if (io_sqring_entries(ctx)) {
6730 needs_sched = false;
6731 break;
6732 }
6733 }
6734
37d1e2e3 6735 if (needs_sched && !io_sq_thread_should_park(sqd)) {
69fb2131
JA
6736 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6737 io_ring_set_wakeup_flag(ctx);
08369246 6738
69fb2131 6739 schedule();
e4b4a13f 6740 try_to_freeze();
69fb2131
JA
6741 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6742 io_ring_clear_wakeup_flag(ctx);
6c271ce2 6743 }
08369246
XW
6744
6745 finish_wait(&sqd->wait, &wait);
6746 timeout = jiffies + sqd->sq_thread_idle;
6c271ce2
JA
6747 }
6748
37d1e2e3
JA
6749 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6750 io_uring_cancel_sqpoll(ctx);
06058632 6751
37d1e2e3 6752 io_run_task_work();
28cea78a 6753
8629397e
JA
6754 if (io_sq_thread_should_park(sqd))
6755 io_sq_thread_parkme(sqd);
6756
37d1e2e3
JA
6757 /*
6758 * Clear thread under lock so that concurrent parks work correctly
6759 */
8629397e 6760 complete(&sqd->completion);
37d1e2e3
JA
6761 mutex_lock(&sqd->lock);
6762 sqd->thread = NULL;
5f3f26f9
JA
6763 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6764 ctx->sqo_exec = 1;
6765 io_ring_set_wakeup_flag(ctx);
6766 }
06058632 6767
37d1e2e3 6768 complete(&sqd->exited);
e54945ae 6769 mutex_unlock(&sqd->lock);
37d1e2e3 6770 do_exit(0);
6c271ce2
JA
6771}
6772
bda52162
JA
6773struct io_wait_queue {
6774 struct wait_queue_entry wq;
6775 struct io_ring_ctx *ctx;
6776 unsigned to_wait;
6777 unsigned nr_timeouts;
6778};
6779
6c503150 6780static inline bool io_should_wake(struct io_wait_queue *iowq)
bda52162
JA
6781{
6782 struct io_ring_ctx *ctx = iowq->ctx;
6783
6784 /*
d195a66e 6785 * Wake up if we have enough events, or if a timeout occurred since we
bda52162
JA
6786 * started waiting. For timeouts, we always want to return to userspace,
6787 * regardless of event count.
6788 */
6c503150 6789 return io_cqring_events(ctx) >= iowq->to_wait ||
bda52162
JA
6790 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6791}
6792
6793static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6794 int wake_flags, void *key)
6795{
6796 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6797 wq);
6798
6c503150
PB
6799 /*
6800 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6801 * the task, and the next invocation will do it.
6802 */
6803 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6804 return autoremove_wake_function(curr, mode, wake_flags, key);
6805 return -1;
bda52162
JA
6806}
6807
af9c1a44
JA
6808static int io_run_task_work_sig(void)
6809{
6810 if (io_run_task_work())
6811 return 1;
6812 if (!signal_pending(current))
6813 return 0;
792ee0f6
JA
6814 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
6815 return -ERESTARTSYS;
af9c1a44
JA
6816 return -EINTR;
6817}
6818
eeb60b9a
PB
6819/* when returns >0, the caller should retry */
6820static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6821 struct io_wait_queue *iowq,
6822 signed long *timeout)
6823{
6824 int ret;
6825
6826 /* make sure we run task_work before checking for signals */
6827 ret = io_run_task_work_sig();
6828 if (ret || io_should_wake(iowq))
6829 return ret;
6830 /* let the caller flush overflows, retry */
6831 if (test_bit(0, &ctx->cq_check_overflow))
6832 return 1;
6833
6834 *timeout = schedule_timeout(*timeout);
6835 return !*timeout ? -ETIME : 1;
6836}
6837
2b188cc1
JA
6838/*
6839 * Wait until events become available, if we don't already have some. The
6840 * application must reap them itself, as they reside on the shared cq ring.
6841 */
6842static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
c73ebb68
HX
6843 const sigset_t __user *sig, size_t sigsz,
6844 struct __kernel_timespec __user *uts)
2b188cc1 6845{
bda52162
JA
6846 struct io_wait_queue iowq = {
6847 .wq = {
6848 .private = current,
6849 .func = io_wake_function,
6850 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6851 },
6852 .ctx = ctx,
6853 .to_wait = min_events,
6854 };
75b28aff 6855 struct io_rings *rings = ctx->rings;
c1d5a224
PB
6856 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6857 int ret;
2b188cc1 6858
b41e9852 6859 do {
6c503150
PB
6860 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6861 if (io_cqring_events(ctx) >= min_events)
b41e9852 6862 return 0;
4c6e277c 6863 if (!io_run_task_work())
b41e9852 6864 break;
b41e9852 6865 } while (1);
2b188cc1
JA
6866
6867 if (sig) {
9e75ad5d
AB
6868#ifdef CONFIG_COMPAT
6869 if (in_compat_syscall())
6870 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
b772434b 6871 sigsz);
9e75ad5d
AB
6872 else
6873#endif
b772434b 6874 ret = set_user_sigmask(sig, sigsz);
9e75ad5d 6875
2b188cc1
JA
6876 if (ret)
6877 return ret;
6878 }
6879
c73ebb68 6880 if (uts) {
c1d5a224
PB
6881 struct timespec64 ts;
6882
c73ebb68
HX
6883 if (get_timespec64(&ts, uts))
6884 return -EFAULT;
6885 timeout = timespec64_to_jiffies(&ts);
6886 }
6887
bda52162 6888 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
c826bd7a 6889 trace_io_uring_cqring_wait(ctx, min_events);
bda52162 6890 do {
ca0a2651
JA
6891 /* if we can't even flush overflow, don't wait for more */
6892 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL)) {
6893 ret = -EBUSY;
6894 break;
6895 }
bda52162
JA
6896 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6897 TASK_INTERRUPTIBLE);
eeb60b9a
PB
6898 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6899 finish_wait(&ctx->wait, &iowq.wq);
ca0a2651 6900 cond_resched();
eeb60b9a 6901 } while (ret > 0);
bda52162 6902
b7db41c9 6903 restore_saved_sigmask_unless(ret == -EINTR);
2b188cc1 6904
75b28aff 6905 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2b188cc1
JA
6906}
6907
6b06314c
JA
6908static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6909{
6910#if defined(CONFIG_UNIX)
6911 if (ctx->ring_sock) {
6912 struct sock *sock = ctx->ring_sock->sk;
6913 struct sk_buff *skb;
6914
6915 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6916 kfree_skb(skb);
6917 }
6918#else
6919 int i;
6920
65e19f54
JA
6921 for (i = 0; i < ctx->nr_user_files; i++) {
6922 struct file *file;
6923
6924 file = io_file_from_index(ctx, i);
6925 if (file)
6926 fput(file);
6927 }
6b06314c
JA
6928#endif
6929}
6930
00835dce 6931static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
05f3fb3c 6932{
269bbe5f 6933 struct fixed_rsrc_data *data;
05f3fb3c 6934
269bbe5f 6935 data = container_of(ref, struct fixed_rsrc_data, refs);
05f3fb3c
JA
6936 complete(&data->done);
6937}
6938
2a63b2d9 6939static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
1642b445 6940{
2a63b2d9 6941 spin_lock_bh(&ctx->rsrc_ref_lock);
1642b445
PB
6942}
6943
2a63b2d9 6944static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
6b06314c 6945{
2a63b2d9
BM
6946 spin_unlock_bh(&ctx->rsrc_ref_lock);
6947}
65e19f54 6948
d67d2263
BM
6949static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
6950 struct fixed_rsrc_data *rsrc_data,
269bbe5f 6951 struct fixed_rsrc_ref_node *ref_node)
1642b445 6952{
2a63b2d9 6953 io_rsrc_ref_lock(ctx);
269bbe5f 6954 rsrc_data->node = ref_node;
d67d2263 6955 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
2a63b2d9 6956 io_rsrc_ref_unlock(ctx);
269bbe5f 6957 percpu_ref_get(&rsrc_data->refs);
1642b445
PB
6958}
6959
8bad28d8 6960static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
6b06314c 6961{
8bad28d8 6962 struct fixed_rsrc_ref_node *ref_node = NULL;
6b06314c 6963
2a63b2d9 6964 io_rsrc_ref_lock(ctx);
1e5d770b 6965 ref_node = data->node;
e6cb007c 6966 data->node = NULL;
2a63b2d9 6967 io_rsrc_ref_unlock(ctx);
05589553
XW
6968 if (ref_node)
6969 percpu_ref_kill(&ref_node->refs);
8bad28d8
HX
6970}
6971
6972static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
6973 struct io_ring_ctx *ctx,
f2303b1f
PB
6974 void (*rsrc_put)(struct io_ring_ctx *ctx,
6975 struct io_rsrc_put *prsrc))
8bad28d8 6976{
f2303b1f 6977 struct fixed_rsrc_ref_node *backup_node;
8bad28d8 6978 int ret;
05589553 6979
8bad28d8
HX
6980 if (data->quiesce)
6981 return -ENXIO;
05589553 6982
8bad28d8 6983 data->quiesce = true;
1ffc5422 6984 do {
f2303b1f
PB
6985 ret = -ENOMEM;
6986 backup_node = alloc_fixed_rsrc_ref_node(ctx);
6987 if (!backup_node)
6988 break;
6989 backup_node->rsrc_data = data;
6990 backup_node->rsrc_put = rsrc_put;
6991
8bad28d8
HX
6992 io_sqe_rsrc_kill_node(ctx, data);
6993 percpu_ref_kill(&data->refs);
6994 flush_delayed_work(&ctx->rsrc_put_work);
6995
1ffc5422
PB
6996 ret = wait_for_completion_interruptible(&data->done);
6997 if (!ret)
6998 break;
8bad28d8 6999
cb5e1b81 7000 percpu_ref_resurrect(&data->refs);
8bad28d8
HX
7001 io_sqe_rsrc_set_node(ctx, data, backup_node);
7002 backup_node = NULL;
cb5e1b81 7003 reinit_completion(&data->done);
8bad28d8 7004 mutex_unlock(&ctx->uring_lock);
1ffc5422 7005 ret = io_run_task_work_sig();
8bad28d8 7006 mutex_lock(&ctx->uring_lock);
f2303b1f 7007 } while (ret >= 0);
8bad28d8 7008 data->quiesce = false;
05f3fb3c 7009
8bad28d8
HX
7010 if (backup_node)
7011 destroy_fixed_rsrc_ref_node(backup_node);
7012 return ret;
d7954b2b
BM
7013}
7014
1ad555c6
BM
7015static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7016{
7017 struct fixed_rsrc_data *data;
7018
7019 data = kzalloc(sizeof(*data), GFP_KERNEL);
7020 if (!data)
7021 return NULL;
7022
00835dce 7023 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
1ad555c6
BM
7024 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7025 kfree(data);
7026 return NULL;
7027 }
7028 data->ctx = ctx;
7029 init_completion(&data->done);
7030 return data;
7031}
7032
7033static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7034{
7035 percpu_ref_exit(&data->refs);
7036 kfree(data->table);
7037 kfree(data);
7038}
7039
d7954b2b
BM
7040static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7041{
7042 struct fixed_rsrc_data *data = ctx->file_data;
d7954b2b
BM
7043 unsigned nr_tables, i;
7044 int ret;
7045
8bad28d8
HX
7046 /*
7047 * percpu_ref_is_dying() is to stop parallel files unregister
7048 * Since we possibly drop uring lock later in this function to
7049 * run task work.
7050 */
7051 if (!data || percpu_ref_is_dying(&data->refs))
d7954b2b 7052 return -ENXIO;
f2303b1f 7053 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
d7954b2b
BM
7054 if (ret)
7055 return ret;
7056
6b06314c 7057 __io_sqe_files_unregister(ctx);
65e19f54
JA
7058 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7059 for (i = 0; i < nr_tables; i++)
05f3fb3c 7060 kfree(data->table[i].files);
1ad555c6 7061 free_fixed_rsrc_data(data);
05f3fb3c 7062 ctx->file_data = NULL;
6b06314c
JA
7063 ctx->nr_user_files = 0;
7064 return 0;
7065}
7066
37d1e2e3
JA
7067static void io_sq_thread_unpark(struct io_sq_data *sqd)
7068 __releases(&sqd->lock)
7069{
7070 if (!sqd->thread)
7071 return;
7072 if (sqd->thread == current)
7073 return;
7074 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7075 wake_up_state(sqd->thread, TASK_PARKED);
7076 mutex_unlock(&sqd->lock);
7077}
7078
7079static bool io_sq_thread_park(struct io_sq_data *sqd)
7080 __acquires(&sqd->lock)
7081{
7082 if (sqd->thread == current)
7083 return true;
7084 mutex_lock(&sqd->lock);
7085 if (!sqd->thread) {
7086 mutex_unlock(&sqd->lock);
7087 return false;
7088 }
7089 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7090 wake_up_process(sqd->thread);
7091 wait_for_completion(&sqd->completion);
7092 return true;
7093}
7094
7095static void io_sq_thread_stop(struct io_sq_data *sqd)
7096{
e54945ae 7097 if (test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state))
37d1e2e3 7098 return;
e54945ae
JA
7099 mutex_lock(&sqd->lock);
7100 if (sqd->thread) {
7101 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7102 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state));
7103 wake_up_process(sqd->thread);
7104 mutex_unlock(&sqd->lock);
7105 wait_for_completion(&sqd->exited);
7106 WARN_ON_ONCE(sqd->thread);
7107 } else {
7108 mutex_unlock(&sqd->lock);
7109 }
37d1e2e3
JA
7110}
7111
534ca6d6 7112static void io_put_sq_data(struct io_sq_data *sqd)
6c271ce2 7113{
534ca6d6 7114 if (refcount_dec_and_test(&sqd->refs)) {
37d1e2e3
JA
7115 io_sq_thread_stop(sqd);
7116 kfree(sqd);
7117 }
7118}
7119
7120static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7121{
7122 struct io_sq_data *sqd = ctx->sq_data;
7123
7124 if (sqd) {
eb85890b 7125 complete(&sqd->startup);
534ca6d6 7126 if (sqd->thread) {
37d1e2e3
JA
7127 wait_for_completion(&ctx->sq_thread_comp);
7128 io_sq_thread_park(sqd);
534ca6d6
JA
7129 }
7130
37d1e2e3
JA
7131 mutex_lock(&sqd->ctx_lock);
7132 list_del(&ctx->sqd_list);
7133 io_sqd_update_thread_idle(sqd);
7134 mutex_unlock(&sqd->ctx_lock);
7135
7136 if (sqd->thread)
7137 io_sq_thread_unpark(sqd);
7138
7139 io_put_sq_data(sqd);
7140 ctx->sq_data = NULL;
534ca6d6
JA
7141 }
7142}
7143
aa06165d
JA
7144static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7145{
7146 struct io_ring_ctx *ctx_attach;
7147 struct io_sq_data *sqd;
7148 struct fd f;
7149
7150 f = fdget(p->wq_fd);
7151 if (!f.file)
7152 return ERR_PTR(-ENXIO);
7153 if (f.file->f_op != &io_uring_fops) {
7154 fdput(f);
7155 return ERR_PTR(-EINVAL);
7156 }
7157
7158 ctx_attach = f.file->private_data;
7159 sqd = ctx_attach->sq_data;
7160 if (!sqd) {
7161 fdput(f);
7162 return ERR_PTR(-EINVAL);
7163 }
7164
7165 refcount_inc(&sqd->refs);
7166 fdput(f);
7167 return sqd;
7168}
7169
534ca6d6
JA
7170static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7171{
7172 struct io_sq_data *sqd;
7173
aa06165d
JA
7174 if (p->flags & IORING_SETUP_ATTACH_WQ)
7175 return io_attach_sq_data(p);
7176
534ca6d6
JA
7177 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7178 if (!sqd)
7179 return ERR_PTR(-ENOMEM);
7180
7181 refcount_set(&sqd->refs, 1);
69fb2131
JA
7182 INIT_LIST_HEAD(&sqd->ctx_list);
7183 INIT_LIST_HEAD(&sqd->ctx_new_list);
7184 mutex_init(&sqd->ctx_lock);
7185 mutex_init(&sqd->lock);
534ca6d6 7186 init_waitqueue_head(&sqd->wait);
37d1e2e3
JA
7187 init_completion(&sqd->startup);
7188 init_completion(&sqd->completion);
7189 init_completion(&sqd->exited);
534ca6d6
JA
7190 return sqd;
7191}
7192
6b06314c 7193#if defined(CONFIG_UNIX)
6b06314c
JA
7194/*
7195 * Ensure the UNIX gc is aware of our file set, so we are certain that
7196 * the io_uring can be safely unregistered on process exit, even if we have
7197 * loops in the file referencing.
7198 */
7199static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7200{
7201 struct sock *sk = ctx->ring_sock->sk;
7202 struct scm_fp_list *fpl;
7203 struct sk_buff *skb;
08a45173 7204 int i, nr_files;
6b06314c 7205
6b06314c
JA
7206 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7207 if (!fpl)
7208 return -ENOMEM;
7209
7210 skb = alloc_skb(0, GFP_KERNEL);
7211 if (!skb) {
7212 kfree(fpl);
7213 return -ENOMEM;
7214 }
7215
7216 skb->sk = sk;
6b06314c 7217
08a45173 7218 nr_files = 0;
62e398be 7219 fpl->user = get_uid(current_user());
6b06314c 7220 for (i = 0; i < nr; i++) {
65e19f54
JA
7221 struct file *file = io_file_from_index(ctx, i + offset);
7222
7223 if (!file)
08a45173 7224 continue;
65e19f54 7225 fpl->fp[nr_files] = get_file(file);
08a45173
JA
7226 unix_inflight(fpl->user, fpl->fp[nr_files]);
7227 nr_files++;
6b06314c
JA
7228 }
7229
08a45173
JA
7230 if (nr_files) {
7231 fpl->max = SCM_MAX_FD;
7232 fpl->count = nr_files;
7233 UNIXCB(skb).fp = fpl;
05f3fb3c 7234 skb->destructor = unix_destruct_scm;
08a45173
JA
7235 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7236 skb_queue_head(&sk->sk_receive_queue, skb);
6b06314c 7237
08a45173
JA
7238 for (i = 0; i < nr_files; i++)
7239 fput(fpl->fp[i]);
7240 } else {
7241 kfree_skb(skb);
7242 kfree(fpl);
7243 }
6b06314c
JA
7244
7245 return 0;
7246}
7247
7248/*
7249 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7250 * causes regular reference counting to break down. We rely on the UNIX
7251 * garbage collection to take care of this problem for us.
7252 */
7253static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7254{
7255 unsigned left, total;
7256 int ret = 0;
7257
7258 total = 0;
7259 left = ctx->nr_user_files;
7260 while (left) {
7261 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6b06314c
JA
7262
7263 ret = __io_sqe_files_scm(ctx, this_files, total);
7264 if (ret)
7265 break;
7266 left -= this_files;
7267 total += this_files;
7268 }
7269
7270 if (!ret)
7271 return 0;
7272
7273 while (total < ctx->nr_user_files) {
65e19f54
JA
7274 struct file *file = io_file_from_index(ctx, total);
7275
7276 if (file)
7277 fput(file);
6b06314c
JA
7278 total++;
7279 }
7280
7281 return ret;
7282}
7283#else
7284static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7285{
7286 return 0;
7287}
7288#endif
7289
269bbe5f 7290static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
5398ae69 7291 unsigned nr_tables, unsigned nr_files)
65e19f54
JA
7292{
7293 int i;
7294
7295 for (i = 0; i < nr_tables; i++) {
269bbe5f 7296 struct fixed_rsrc_table *table = &file_data->table[i];
65e19f54
JA
7297 unsigned this_files;
7298
7299 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7300 table->files = kcalloc(this_files, sizeof(struct file *),
7301 GFP_KERNEL);
7302 if (!table->files)
7303 break;
7304 nr_files -= this_files;
7305 }
7306
7307 if (i == nr_tables)
7308 return 0;
7309
7310 for (i = 0; i < nr_tables; i++) {
269bbe5f 7311 struct fixed_rsrc_table *table = &file_data->table[i];
65e19f54
JA
7312 kfree(table->files);
7313 }
7314 return 1;
7315}
7316
50238531 7317static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
05f3fb3c 7318{
50238531 7319 struct file *file = prsrc->file;
05f3fb3c
JA
7320#if defined(CONFIG_UNIX)
7321 struct sock *sock = ctx->ring_sock->sk;
7322 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7323 struct sk_buff *skb;
7324 int i;
7325
7326 __skb_queue_head_init(&list);
7327
7328 /*
7329 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7330 * remove this entry and rearrange the file array.
7331 */
7332 skb = skb_dequeue(head);
7333 while (skb) {
7334 struct scm_fp_list *fp;
7335
7336 fp = UNIXCB(skb).fp;
7337 for (i = 0; i < fp->count; i++) {
7338 int left;
7339
7340 if (fp->fp[i] != file)
7341 continue;
7342
7343 unix_notinflight(fp->user, fp->fp[i]);
7344 left = fp->count - 1 - i;
7345 if (left) {
7346 memmove(&fp->fp[i], &fp->fp[i + 1],
7347 left * sizeof(struct file *));
7348 }
7349 fp->count--;
7350 if (!fp->count) {
7351 kfree_skb(skb);
7352 skb = NULL;
7353 } else {
7354 __skb_queue_tail(&list, skb);
7355 }
7356 fput(file);
7357 file = NULL;
7358 break;
7359 }
7360
7361 if (!file)
7362 break;
7363
7364 __skb_queue_tail(&list, skb);
7365
7366 skb = skb_dequeue(head);
7367 }
7368
7369 if (skb_peek(&list)) {
7370 spin_lock_irq(&head->lock);
7371 while ((skb = __skb_dequeue(&list)) != NULL)
7372 __skb_queue_tail(head, skb);
7373 spin_unlock_irq(&head->lock);
7374 }
7375#else
7376 fput(file);
7377#endif
7378}
7379
269bbe5f 7380static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
65e19f54 7381{
269bbe5f
BM
7382 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7383 struct io_ring_ctx *ctx = rsrc_data->ctx;
7384 struct io_rsrc_put *prsrc, *tmp;
05589553 7385
269bbe5f
BM
7386 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7387 list_del(&prsrc->list);
50238531 7388 ref_node->rsrc_put(ctx, prsrc);
269bbe5f 7389 kfree(prsrc);
65e19f54 7390 }
05589553 7391
05589553
XW
7392 percpu_ref_exit(&ref_node->refs);
7393 kfree(ref_node);
269bbe5f 7394 percpu_ref_put(&rsrc_data->refs);
2faf852d 7395}
65e19f54 7396
269bbe5f 7397static void io_rsrc_put_work(struct work_struct *work)
4a38aed2
JA
7398{
7399 struct io_ring_ctx *ctx;
7400 struct llist_node *node;
7401
269bbe5f
BM
7402 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7403 node = llist_del_all(&ctx->rsrc_put_llist);
4a38aed2
JA
7404
7405 while (node) {
269bbe5f 7406 struct fixed_rsrc_ref_node *ref_node;
4a38aed2
JA
7407 struct llist_node *next = node->next;
7408
269bbe5f
BM
7409 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7410 __io_rsrc_put_work(ref_node);
4a38aed2
JA
7411 node = next;
7412 }
7413}
7414
ea64ec02
PB
7415static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7416 unsigned i)
2faf852d 7417{
ea64ec02
PB
7418 struct fixed_rsrc_table *table;
7419
7420 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7421 return &table->files[i & IORING_FILE_TABLE_MASK];
7422}
7423
00835dce 7424static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
2faf852d 7425{
269bbe5f
BM
7426 struct fixed_rsrc_ref_node *ref_node;
7427 struct fixed_rsrc_data *data;
4a38aed2 7428 struct io_ring_ctx *ctx;
e297822b 7429 bool first_add = false;
4a38aed2 7430 int delay = HZ;
65e19f54 7431
269bbe5f
BM
7432 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7433 data = ref_node->rsrc_data;
e297822b
PB
7434 ctx = data->ctx;
7435
2a63b2d9 7436 io_rsrc_ref_lock(ctx);
e297822b
PB
7437 ref_node->done = true;
7438
d67d2263
BM
7439 while (!list_empty(&ctx->rsrc_ref_list)) {
7440 ref_node = list_first_entry(&ctx->rsrc_ref_list,
269bbe5f 7441 struct fixed_rsrc_ref_node, node);
e297822b
PB
7442 /* recycle ref nodes in order */
7443 if (!ref_node->done)
7444 break;
7445 list_del(&ref_node->node);
269bbe5f 7446 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
e297822b 7447 }
2a63b2d9 7448 io_rsrc_ref_unlock(ctx);
05589553 7449
e297822b 7450 if (percpu_ref_is_dying(&data->refs))
4a38aed2 7451 delay = 0;
05589553 7452
4a38aed2 7453 if (!delay)
269bbe5f 7454 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
4a38aed2 7455 else if (first_add)
269bbe5f 7456 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
05f3fb3c 7457}
65e19f54 7458
6802535d 7459static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
05589553 7460 struct io_ring_ctx *ctx)
05f3fb3c 7461{
269bbe5f 7462 struct fixed_rsrc_ref_node *ref_node;
05f3fb3c 7463
05589553
XW
7464 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7465 if (!ref_node)
3e2224c5 7466 return NULL;
05f3fb3c 7467
00835dce 7468 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
05589553
XW
7469 0, GFP_KERNEL)) {
7470 kfree(ref_node);
3e2224c5 7471 return NULL;
05589553
XW
7472 }
7473 INIT_LIST_HEAD(&ref_node->node);
269bbe5f 7474 INIT_LIST_HEAD(&ref_node->rsrc_list);
e297822b 7475 ref_node->done = false;
05589553 7476 return ref_node;
05589553
XW
7477}
7478
bc9744cd
PB
7479static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7480 struct fixed_rsrc_ref_node *ref_node)
6802535d 7481{
269bbe5f 7482 ref_node->rsrc_data = ctx->file_data;
50238531 7483 ref_node->rsrc_put = io_ring_file_put;
05589553
XW
7484}
7485
269bbe5f 7486static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
05589553
XW
7487{
7488 percpu_ref_exit(&ref_node->refs);
7489 kfree(ref_node);
65e19f54
JA
7490}
7491
ea64ec02 7492
6b06314c
JA
7493static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7494 unsigned nr_args)
7495{
7496 __s32 __user *fds = (__s32 __user *) arg;
600cf3f8 7497 unsigned nr_tables, i;
05f3fb3c 7498 struct file *file;
600cf3f8 7499 int fd, ret = -ENOMEM;
269bbe5f
BM
7500 struct fixed_rsrc_ref_node *ref_node;
7501 struct fixed_rsrc_data *file_data;
6b06314c 7502
05f3fb3c 7503 if (ctx->file_data)
6b06314c
JA
7504 return -EBUSY;
7505 if (!nr_args)
7506 return -EINVAL;
7507 if (nr_args > IORING_MAX_FIXED_FILES)
7508 return -EMFILE;
7509
1ad555c6 7510 file_data = alloc_fixed_rsrc_data(ctx);
5398ae69 7511 if (!file_data)
05f3fb3c 7512 return -ENOMEM;
13770a71 7513 ctx->file_data = file_data;
05f3fb3c 7514
65e19f54 7515 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
035fbafc 7516 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
5398ae69 7517 GFP_KERNEL);
600cf3f8
PB
7518 if (!file_data->table)
7519 goto out_free;
05f3fb3c 7520
600cf3f8 7521 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
1ad555c6 7522 goto out_free;
65e19f54 7523
08a45173 7524 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
600cf3f8
PB
7525 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7526 ret = -EFAULT;
7527 goto out_fput;
7528 }
08a45173 7529 /* allow sparse sets */
600cf3f8 7530 if (fd == -1)
08a45173 7531 continue;
6b06314c 7532
05f3fb3c 7533 file = fget(fd);
6b06314c 7534 ret = -EBADF;
05f3fb3c 7535 if (!file)
600cf3f8 7536 goto out_fput;
05f3fb3c 7537
6b06314c
JA
7538 /*
7539 * Don't allow io_uring instances to be registered. If UNIX
7540 * isn't enabled, then this causes a reference cycle and this
7541 * instance can never get freed. If UNIX is enabled we'll
7542 * handle it just fine, but there's still no point in allowing
7543 * a ring fd as it doesn't support regular read/write anyway.
7544 */
05f3fb3c
JA
7545 if (file->f_op == &io_uring_fops) {
7546 fput(file);
600cf3f8 7547 goto out_fput;
6b06314c 7548 }
ea64ec02 7549 *io_fixed_file_slot(file_data, i) = file;
6b06314c
JA
7550 }
7551
6b06314c 7552 ret = io_sqe_files_scm(ctx);
05589553 7553 if (ret) {
6b06314c 7554 io_sqe_files_unregister(ctx);
05589553
XW
7555 return ret;
7556 }
6b06314c 7557
bc9744cd 7558 ref_node = alloc_fixed_rsrc_ref_node(ctx);
3e2224c5 7559 if (!ref_node) {
05589553 7560 io_sqe_files_unregister(ctx);
3e2224c5 7561 return -ENOMEM;
05589553 7562 }
bc9744cd 7563 init_fixed_file_ref_node(ctx, ref_node);
05589553 7564
d67d2263 7565 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
6b06314c 7566 return ret;
600cf3f8
PB
7567out_fput:
7568 for (i = 0; i < ctx->nr_user_files; i++) {
7569 file = io_file_from_index(ctx, i);
7570 if (file)
7571 fput(file);
7572 }
7573 for (i = 0; i < nr_tables; i++)
7574 kfree(file_data->table[i].files);
7575 ctx->nr_user_files = 0;
600cf3f8 7576out_free:
1ad555c6 7577 free_fixed_rsrc_data(ctx->file_data);
55cbc256 7578 ctx->file_data = NULL;
6b06314c
JA
7579 return ret;
7580}
7581
c3a31e60
JA
7582static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7583 int index)
7584{
7585#if defined(CONFIG_UNIX)
7586 struct sock *sock = ctx->ring_sock->sk;
7587 struct sk_buff_head *head = &sock->sk_receive_queue;
7588 struct sk_buff *skb;
7589
7590 /*
7591 * See if we can merge this file into an existing skb SCM_RIGHTS
7592 * file set. If there's no room, fall back to allocating a new skb
7593 * and filling it in.
7594 */
7595 spin_lock_irq(&head->lock);
7596 skb = skb_peek(head);
7597 if (skb) {
7598 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7599
7600 if (fpl->count < SCM_MAX_FD) {
7601 __skb_unlink(skb, head);
7602 spin_unlock_irq(&head->lock);
7603 fpl->fp[fpl->count] = get_file(file);
7604 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7605 fpl->count++;
7606 spin_lock_irq(&head->lock);
7607 __skb_queue_head(head, skb);
7608 } else {
7609 skb = NULL;
7610 }
7611 }
7612 spin_unlock_irq(&head->lock);
7613
7614 if (skb) {
7615 fput(file);
7616 return 0;
7617 }
7618
7619 return __io_sqe_files_scm(ctx, 1, index);
7620#else
7621 return 0;
7622#endif
7623}
7624
50238531 7625static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
05f3fb3c 7626{
269bbe5f
BM
7627 struct io_rsrc_put *prsrc;
7628 struct fixed_rsrc_ref_node *ref_node = data->node;
05f3fb3c 7629
269bbe5f
BM
7630 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7631 if (!prsrc)
a5318d3c 7632 return -ENOMEM;
05f3fb3c 7633
50238531 7634 prsrc->rsrc = rsrc;
269bbe5f 7635 list_add(&prsrc->list, &ref_node->rsrc_list);
05589553 7636
a5318d3c 7637 return 0;
05f3fb3c
JA
7638}
7639
269bbe5f
BM
7640static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7641 struct file *file)
7642{
50238531 7643 return io_queue_rsrc_removal(data, (void *)file);
269bbe5f
BM
7644}
7645
05f3fb3c 7646static int __io_sqe_files_update(struct io_ring_ctx *ctx,
269bbe5f 7647 struct io_uring_rsrc_update *up,
05f3fb3c
JA
7648 unsigned nr_args)
7649{
269bbe5f
BM
7650 struct fixed_rsrc_data *data = ctx->file_data;
7651 struct fixed_rsrc_ref_node *ref_node;
ea64ec02 7652 struct file *file, **file_slot;
c3a31e60
JA
7653 __s32 __user *fds;
7654 int fd, i, err;
7655 __u32 done;
05589553 7656 bool needs_switch = false;
c3a31e60 7657
05f3fb3c 7658 if (check_add_overflow(up->offset, nr_args, &done))
c3a31e60
JA
7659 return -EOVERFLOW;
7660 if (done > ctx->nr_user_files)
7661 return -EINVAL;
7662
bc9744cd 7663 ref_node = alloc_fixed_rsrc_ref_node(ctx);
3e2224c5
MWO
7664 if (!ref_node)
7665 return -ENOMEM;
bc9744cd 7666 init_fixed_file_ref_node(ctx, ref_node);
05589553 7667
269bbe5f 7668 fds = u64_to_user_ptr(up->data);
67973b93 7669 for (done = 0; done < nr_args; done++) {
c3a31e60
JA
7670 err = 0;
7671 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7672 err = -EFAULT;
7673 break;
7674 }
4e0377a1 7675 if (fd == IORING_REGISTER_FILES_SKIP)
7676 continue;
7677
67973b93 7678 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
ea64ec02
PB
7679 file_slot = io_fixed_file_slot(ctx->file_data, i);
7680
7681 if (*file_slot) {
7682 err = io_queue_file_removal(data, *file_slot);
a5318d3c
HD
7683 if (err)
7684 break;
ea64ec02 7685 *file_slot = NULL;
05589553 7686 needs_switch = true;
c3a31e60
JA
7687 }
7688 if (fd != -1) {
c3a31e60
JA
7689 file = fget(fd);
7690 if (!file) {
7691 err = -EBADF;
7692 break;
7693 }
7694 /*
7695 * Don't allow io_uring instances to be registered. If
7696 * UNIX isn't enabled, then this causes a reference
7697 * cycle and this instance can never get freed. If UNIX
7698 * is enabled we'll handle it just fine, but there's
7699 * still no point in allowing a ring fd as it doesn't
7700 * support regular read/write anyway.
7701 */
7702 if (file->f_op == &io_uring_fops) {
7703 fput(file);
7704 err = -EBADF;
7705 break;
7706 }
e68a3ff8 7707 *file_slot = file;
c3a31e60 7708 err = io_sqe_file_register(ctx, file, i);
f3bd9dae 7709 if (err) {
e68a3ff8 7710 *file_slot = NULL;
f3bd9dae 7711 fput(file);
c3a31e60 7712 break;
f3bd9dae 7713 }
c3a31e60 7714 }
05f3fb3c
JA
7715 }
7716
05589553 7717 if (needs_switch) {
b2e96852 7718 percpu_ref_kill(&data->node->refs);
d67d2263 7719 io_sqe_rsrc_set_node(ctx, data, ref_node);
05589553 7720 } else
269bbe5f 7721 destroy_fixed_rsrc_ref_node(ref_node);
c3a31e60
JA
7722
7723 return done ? done : err;
7724}
05589553 7725
05f3fb3c
JA
7726static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7727 unsigned nr_args)
7728{
269bbe5f 7729 struct io_uring_rsrc_update up;
05f3fb3c
JA
7730
7731 if (!ctx->file_data)
7732 return -ENXIO;
7733 if (!nr_args)
7734 return -EINVAL;
7735 if (copy_from_user(&up, arg, sizeof(up)))
7736 return -EFAULT;
7737 if (up.resv)
7738 return -EINVAL;
7739
7740 return __io_sqe_files_update(ctx, &up, nr_args);
7741}
c3a31e60 7742
5280f7e5 7743static struct io_wq_work *io_free_work(struct io_wq_work *work)
7d723065
JA
7744{
7745 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7746
5280f7e5
PB
7747 req = io_put_req_find_next(req);
7748 return req ? &req->work : NULL;
7d723065
JA
7749}
7750
5aa75ed5 7751static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
24369c2e 7752{
e941894e 7753 struct io_wq_hash *hash;
24369c2e 7754 struct io_wq_data data;
24369c2e 7755 unsigned int concurrency;
24369c2e 7756
e941894e
JA
7757 hash = ctx->hash_map;
7758 if (!hash) {
7759 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7760 if (!hash)
7761 return ERR_PTR(-ENOMEM);
7762 refcount_set(&hash->refs, 1);
7763 init_waitqueue_head(&hash->wait);
7764 ctx->hash_map = hash;
24369c2e
PB
7765 }
7766
e941894e 7767 data.hash = hash;
e9fd9396 7768 data.free_work = io_free_work;
f5fa38c5 7769 data.do_work = io_wq_submit_work;
24369c2e 7770
d25e3a3d
JA
7771 /* Do QD, or 4 * CPUS, whatever is smallest */
7772 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
24369c2e 7773
5aa75ed5 7774 return io_wq_create(concurrency, &data);
24369c2e
PB
7775}
7776
5aa75ed5
JA
7777static int io_uring_alloc_task_context(struct task_struct *task,
7778 struct io_ring_ctx *ctx)
0f212204
JA
7779{
7780 struct io_uring_task *tctx;
d8a6df10 7781 int ret;
0f212204
JA
7782
7783 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7784 if (unlikely(!tctx))
7785 return -ENOMEM;
7786
d8a6df10
JA
7787 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7788 if (unlikely(ret)) {
7789 kfree(tctx);
7790 return ret;
7791 }
7792
5aa75ed5
JA
7793 tctx->io_wq = io_init_wq_offload(ctx);
7794 if (IS_ERR(tctx->io_wq)) {
7795 ret = PTR_ERR(tctx->io_wq);
7796 percpu_counter_destroy(&tctx->inflight);
7797 kfree(tctx);
7798 return ret;
7799 }
7800
0f212204
JA
7801 xa_init(&tctx->xa);
7802 init_waitqueue_head(&tctx->wait);
7803 tctx->last = NULL;
fdaf083c
JA
7804 atomic_set(&tctx->in_idle, 0);
7805 tctx->sqpoll = false;
0f212204 7806 task->io_uring = tctx;
7cbf1722
JA
7807 spin_lock_init(&tctx->task_lock);
7808 INIT_WQ_LIST(&tctx->task_list);
7809 tctx->task_state = 0;
7810 init_task_work(&tctx->task_work, tctx_task_work);
0f212204
JA
7811 return 0;
7812}
7813
7814void __io_uring_free(struct task_struct *tsk)
7815{
7816 struct io_uring_task *tctx = tsk->io_uring;
7817
7818 WARN_ON_ONCE(!xa_empty(&tctx->xa));
ef8eaa4e
PB
7819 WARN_ON_ONCE(tctx->io_wq);
7820
d8a6df10 7821 percpu_counter_destroy(&tctx->inflight);
0f212204
JA
7822 kfree(tctx);
7823 tsk->io_uring = NULL;
7824}
7825
5f3f26f9
JA
7826static int io_sq_thread_fork(struct io_sq_data *sqd, struct io_ring_ctx *ctx)
7827{
46fe18b1 7828 struct task_struct *tsk;
5f3f26f9
JA
7829 int ret;
7830
7831 clear_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7832 reinit_completion(&sqd->completion);
70aacfe6 7833 ctx->sqo_exec = 0;
5f3f26f9 7834 sqd->task_pid = current->pid;
46fe18b1
JA
7835 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7836 if (IS_ERR(tsk))
7837 return PTR_ERR(tsk);
7838 ret = io_uring_alloc_task_context(tsk, ctx);
7839 if (ret)
7840 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7841 sqd->thread = tsk;
7842 wake_up_new_task(tsk);
7843 return ret;
5f3f26f9
JA
7844}
7845
7e84e1c7
SG
7846static int io_sq_offload_create(struct io_ring_ctx *ctx,
7847 struct io_uring_params *p)
2b188cc1
JA
7848{
7849 int ret;
7850
d25e3a3d
JA
7851 /* Retain compatibility with failing for an invalid attach attempt */
7852 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7853 IORING_SETUP_ATTACH_WQ) {
7854 struct fd f;
7855
7856 f = fdget(p->wq_fd);
7857 if (!f.file)
7858 return -ENXIO;
7859 if (f.file->f_op != &io_uring_fops) {
7860 fdput(f);
7861 return -EINVAL;
7862 }
7863 fdput(f);
7864 }
6c271ce2 7865 if (ctx->flags & IORING_SETUP_SQPOLL) {
46fe18b1 7866 struct task_struct *tsk;
534ca6d6
JA
7867 struct io_sq_data *sqd;
7868
3ec482d1 7869 ret = -EPERM;
ce59fc69 7870 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
3ec482d1
JA
7871 goto err;
7872
534ca6d6
JA
7873 sqd = io_get_sq_data(p);
7874 if (IS_ERR(sqd)) {
7875 ret = PTR_ERR(sqd);
7876 goto err;
7877 }
69fb2131 7878
534ca6d6 7879 ctx->sq_data = sqd;
69fb2131
JA
7880 io_sq_thread_park(sqd);
7881 mutex_lock(&sqd->ctx_lock);
7882 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7883 mutex_unlock(&sqd->ctx_lock);
7884 io_sq_thread_unpark(sqd);
534ca6d6 7885
917257da
JA
7886 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7887 if (!ctx->sq_thread_idle)
7888 ctx->sq_thread_idle = HZ;
7889
aa06165d 7890 if (sqd->thread)
5aa75ed5 7891 return 0;
aa06165d 7892
6c271ce2 7893 if (p->flags & IORING_SETUP_SQ_AFF) {
44a9bd18 7894 int cpu = p->sq_thread_cpu;
6c271ce2 7895
917257da 7896 ret = -EINVAL;
44a9bd18
JA
7897 if (cpu >= nr_cpu_ids)
7898 goto err;
7889f44d 7899 if (!cpu_online(cpu))
917257da
JA
7900 goto err;
7901
37d1e2e3 7902 sqd->sq_cpu = cpu;
6c271ce2 7903 } else {
37d1e2e3 7904 sqd->sq_cpu = -1;
6c271ce2 7905 }
37d1e2e3
JA
7906
7907 sqd->task_pid = current->pid;
46fe18b1
JA
7908 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7909 if (IS_ERR(tsk)) {
7910 ret = PTR_ERR(tsk);
6c271ce2
JA
7911 goto err;
7912 }
46fe18b1
JA
7913 ret = io_uring_alloc_task_context(tsk, ctx);
7914 if (ret)
7915 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7916 sqd->thread = tsk;
7917 wake_up_new_task(tsk);
0f212204
JA
7918 if (ret)
7919 goto err;
6c271ce2
JA
7920 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7921 /* Can't have SQ_AFF without SQPOLL */
7922 ret = -EINVAL;
7923 goto err;
7924 }
7925
2b188cc1
JA
7926 return 0;
7927err:
37d1e2e3 7928 io_sq_thread_finish(ctx);
2b188cc1
JA
7929 return ret;
7930}
7931
7e84e1c7
SG
7932static void io_sq_offload_start(struct io_ring_ctx *ctx)
7933{
534ca6d6
JA
7934 struct io_sq_data *sqd = ctx->sq_data;
7935
3ebba796 7936 ctx->flags &= ~IORING_SETUP_R_DISABLED;
eb85890b 7937 if (ctx->flags & IORING_SETUP_SQPOLL)
37d1e2e3 7938 complete(&sqd->startup);
7e84e1c7
SG
7939}
7940
a087e2b5
BM
7941static inline void __io_unaccount_mem(struct user_struct *user,
7942 unsigned long nr_pages)
2b188cc1
JA
7943{
7944 atomic_long_sub(nr_pages, &user->locked_vm);
7945}
7946
a087e2b5
BM
7947static inline int __io_account_mem(struct user_struct *user,
7948 unsigned long nr_pages)
2b188cc1
JA
7949{
7950 unsigned long page_limit, cur_pages, new_pages;
7951
7952 /* Don't allow more pages than we can safely lock */
7953 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7954
7955 do {
7956 cur_pages = atomic_long_read(&user->locked_vm);
7957 new_pages = cur_pages + nr_pages;
7958 if (new_pages > page_limit)
7959 return -ENOMEM;
7960 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7961 new_pages) != cur_pages);
7962
7963 return 0;
7964}
7965
26bfa89e 7966static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
a087e2b5 7967{
62e398be 7968 if (ctx->user)
a087e2b5 7969 __io_unaccount_mem(ctx->user, nr_pages);
30975825 7970
26bfa89e
JA
7971 if (ctx->mm_account)
7972 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
a087e2b5
BM
7973}
7974
26bfa89e 7975static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
a087e2b5 7976{
30975825
BM
7977 int ret;
7978
62e398be 7979 if (ctx->user) {
30975825
BM
7980 ret = __io_account_mem(ctx->user, nr_pages);
7981 if (ret)
7982 return ret;
7983 }
7984
26bfa89e
JA
7985 if (ctx->mm_account)
7986 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
a087e2b5
BM
7987
7988 return 0;
7989}
7990
2b188cc1
JA
7991static void io_mem_free(void *ptr)
7992{
52e04ef4
MR
7993 struct page *page;
7994
7995 if (!ptr)
7996 return;
2b188cc1 7997
52e04ef4 7998 page = virt_to_head_page(ptr);
2b188cc1
JA
7999 if (put_page_testzero(page))
8000 free_compound_page(page);
8001}
8002
8003static void *io_mem_alloc(size_t size)
8004{
8005 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
26bfa89e 8006 __GFP_NORETRY | __GFP_ACCOUNT;
2b188cc1
JA
8007
8008 return (void *) __get_free_pages(gfp_flags, get_order(size));
8009}
8010
75b28aff
HV
8011static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8012 size_t *sq_offset)
8013{
8014 struct io_rings *rings;
8015 size_t off, sq_array_size;
8016
8017 off = struct_size(rings, cqes, cq_entries);
8018 if (off == SIZE_MAX)
8019 return SIZE_MAX;
8020
8021#ifdef CONFIG_SMP
8022 off = ALIGN(off, SMP_CACHE_BYTES);
8023 if (off == 0)
8024 return SIZE_MAX;
8025#endif
8026
b36200f5
DV
8027 if (sq_offset)
8028 *sq_offset = off;
8029
75b28aff
HV
8030 sq_array_size = array_size(sizeof(u32), sq_entries);
8031 if (sq_array_size == SIZE_MAX)
8032 return SIZE_MAX;
8033
8034 if (check_add_overflow(off, sq_array_size, &off))
8035 return SIZE_MAX;
8036
75b28aff
HV
8037 return off;
8038}
8039
0a96bbe4 8040static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
edafccee
JA
8041{
8042 int i, j;
8043
8044 if (!ctx->user_bufs)
8045 return -ENXIO;
8046
8047 for (i = 0; i < ctx->nr_user_bufs; i++) {
8048 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8049
8050 for (j = 0; j < imu->nr_bvecs; j++)
f1f6a7dd 8051 unpin_user_page(imu->bvec[j].bv_page);
edafccee 8052
de293938 8053 if (imu->acct_pages)
26bfa89e 8054 io_unaccount_mem(ctx, imu->acct_pages);
d4ef6475 8055 kvfree(imu->bvec);
edafccee
JA
8056 imu->nr_bvecs = 0;
8057 }
8058
8059 kfree(ctx->user_bufs);
8060 ctx->user_bufs = NULL;
8061 ctx->nr_user_bufs = 0;
8062 return 0;
8063}
8064
8065static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8066 void __user *arg, unsigned index)
8067{
8068 struct iovec __user *src;
8069
8070#ifdef CONFIG_COMPAT
8071 if (ctx->compat) {
8072 struct compat_iovec __user *ciovs;
8073 struct compat_iovec ciov;
8074
8075 ciovs = (struct compat_iovec __user *) arg;
8076 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8077 return -EFAULT;
8078
d55e5f5b 8079 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
edafccee
JA
8080 dst->iov_len = ciov.iov_len;
8081 return 0;
8082 }
8083#endif
8084 src = (struct iovec __user *) arg;
8085 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8086 return -EFAULT;
8087 return 0;
8088}
8089
de293938
JA
8090/*
8091 * Not super efficient, but this is just a registration time. And we do cache
8092 * the last compound head, so generally we'll only do a full search if we don't
8093 * match that one.
8094 *
8095 * We check if the given compound head page has already been accounted, to
8096 * avoid double accounting it. This allows us to account the full size of the
8097 * page, not just the constituent pages of a huge page.
8098 */
8099static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8100 int nr_pages, struct page *hpage)
8101{
8102 int i, j;
8103
8104 /* check current page array */
8105 for (i = 0; i < nr_pages; i++) {
8106 if (!PageCompound(pages[i]))
8107 continue;
8108 if (compound_head(pages[i]) == hpage)
8109 return true;
8110 }
8111
8112 /* check previously registered pages */
8113 for (i = 0; i < ctx->nr_user_bufs; i++) {
8114 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8115
8116 for (j = 0; j < imu->nr_bvecs; j++) {
8117 if (!PageCompound(imu->bvec[j].bv_page))
8118 continue;
8119 if (compound_head(imu->bvec[j].bv_page) == hpage)
8120 return true;
8121 }
8122 }
8123
8124 return false;
8125}
8126
8127static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8128 int nr_pages, struct io_mapped_ubuf *imu,
8129 struct page **last_hpage)
8130{
8131 int i, ret;
8132
8133 for (i = 0; i < nr_pages; i++) {
8134 if (!PageCompound(pages[i])) {
8135 imu->acct_pages++;
8136 } else {
8137 struct page *hpage;
8138
8139 hpage = compound_head(pages[i]);
8140 if (hpage == *last_hpage)
8141 continue;
8142 *last_hpage = hpage;
8143 if (headpage_already_acct(ctx, pages, i, hpage))
8144 continue;
8145 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8146 }
8147 }
8148
8149 if (!imu->acct_pages)
8150 return 0;
8151
26bfa89e 8152 ret = io_account_mem(ctx, imu->acct_pages);
de293938
JA
8153 if (ret)
8154 imu->acct_pages = 0;
8155 return ret;
8156}
8157
0a96bbe4
BM
8158static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8159 struct io_mapped_ubuf *imu,
8160 struct page **last_hpage)
edafccee
JA
8161{
8162 struct vm_area_struct **vmas = NULL;
8163 struct page **pages = NULL;
0a96bbe4
BM
8164 unsigned long off, start, end, ubuf;
8165 size_t size;
8166 int ret, pret, nr_pages, i;
8167
8168 ubuf = (unsigned long) iov->iov_base;
8169 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8170 start = ubuf >> PAGE_SHIFT;
8171 nr_pages = end - start;
8172
8173 ret = -ENOMEM;
8174
8175 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8176 if (!pages)
8177 goto done;
8178
8179 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8180 GFP_KERNEL);
8181 if (!vmas)
8182 goto done;
edafccee 8183
0a96bbe4
BM
8184 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8185 GFP_KERNEL);
8186 if (!imu->bvec)
8187 goto done;
8188
8189 ret = 0;
8190 mmap_read_lock(current->mm);
8191 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8192 pages, vmas);
8193 if (pret == nr_pages) {
8194 /* don't support file backed memory */
8195 for (i = 0; i < nr_pages; i++) {
8196 struct vm_area_struct *vma = vmas[i];
8197
8198 if (vma->vm_file &&
8199 !is_file_hugepages(vma->vm_file)) {
8200 ret = -EOPNOTSUPP;
8201 break;
8202 }
8203 }
8204 } else {
8205 ret = pret < 0 ? pret : -EFAULT;
8206 }
8207 mmap_read_unlock(current->mm);
8208 if (ret) {
8209 /*
8210 * if we did partial map, or found file backed vmas,
8211 * release any pages we did get
8212 */
8213 if (pret > 0)
8214 unpin_user_pages(pages, pret);
8215 kvfree(imu->bvec);
8216 goto done;
8217 }
8218
8219 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8220 if (ret) {
8221 unpin_user_pages(pages, pret);
8222 kvfree(imu->bvec);
8223 goto done;
8224 }
8225
8226 off = ubuf & ~PAGE_MASK;
8227 size = iov->iov_len;
8228 for (i = 0; i < nr_pages; i++) {
8229 size_t vec_len;
8230
8231 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8232 imu->bvec[i].bv_page = pages[i];
8233 imu->bvec[i].bv_len = vec_len;
8234 imu->bvec[i].bv_offset = off;
8235 off = 0;
8236 size -= vec_len;
8237 }
8238 /* store original address for later verification */
8239 imu->ubuf = ubuf;
8240 imu->len = iov->iov_len;
8241 imu->nr_bvecs = nr_pages;
8242 ret = 0;
8243done:
8244 kvfree(pages);
8245 kvfree(vmas);
8246 return ret;
8247}
8248
2b358604 8249static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
0a96bbe4 8250{
edafccee
JA
8251 if (ctx->user_bufs)
8252 return -EBUSY;
8253 if (!nr_args || nr_args > UIO_MAXIOV)
8254 return -EINVAL;
8255
8256 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8257 GFP_KERNEL);
8258 if (!ctx->user_bufs)
8259 return -ENOMEM;
8260
2b358604
BM
8261 return 0;
8262}
edafccee 8263
2b358604
BM
8264static int io_buffer_validate(struct iovec *iov)
8265{
8266 /*
8267 * Don't impose further limits on the size and buffer
8268 * constraints here, we'll -EINVAL later when IO is
8269 * submitted if they are wrong.
8270 */
8271 if (!iov->iov_base || !iov->iov_len)
8272 return -EFAULT;
edafccee 8273
2b358604
BM
8274 /* arbitrary limit, but we need something */
8275 if (iov->iov_len > SZ_1G)
8276 return -EFAULT;
edafccee 8277
2b358604
BM
8278 return 0;
8279}
edafccee 8280
2b358604
BM
8281static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8282 unsigned int nr_args)
8283{
8284 int i, ret;
8285 struct iovec iov;
8286 struct page *last_hpage = NULL;
edafccee 8287
2b358604
BM
8288 ret = io_buffers_map_alloc(ctx, nr_args);
8289 if (ret)
8290 return ret;
edafccee 8291
edafccee
JA
8292 for (i = 0; i < nr_args; i++) {
8293 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
edafccee 8294
edafccee
JA
8295 ret = io_copy_iov(ctx, &iov, arg, i);
8296 if (ret)
0a96bbe4 8297 break;
de293938 8298
2b358604
BM
8299 ret = io_buffer_validate(&iov);
8300 if (ret)
0a96bbe4 8301 break;
edafccee 8302
0a96bbe4
BM
8303 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8304 if (ret)
8305 break;
edafccee
JA
8306
8307 ctx->nr_user_bufs++;
8308 }
0a96bbe4
BM
8309
8310 if (ret)
8311 io_sqe_buffers_unregister(ctx);
8312
edafccee
JA
8313 return ret;
8314}
8315
9b402849
JA
8316static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8317{
8318 __s32 __user *fds = arg;
8319 int fd;
8320
8321 if (ctx->cq_ev_fd)
8322 return -EBUSY;
8323
8324 if (copy_from_user(&fd, fds, sizeof(*fds)))
8325 return -EFAULT;
8326
8327 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8328 if (IS_ERR(ctx->cq_ev_fd)) {
8329 int ret = PTR_ERR(ctx->cq_ev_fd);
8330 ctx->cq_ev_fd = NULL;
8331 return ret;
8332 }
8333
8334 return 0;
8335}
8336
8337static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8338{
8339 if (ctx->cq_ev_fd) {
8340 eventfd_ctx_put(ctx->cq_ev_fd);
8341 ctx->cq_ev_fd = NULL;
8342 return 0;
8343 }
8344
8345 return -ENXIO;
8346}
8347
5a2e745d
JA
8348static int __io_destroy_buffers(int id, void *p, void *data)
8349{
8350 struct io_ring_ctx *ctx = data;
8351 struct io_buffer *buf = p;
8352
067524e9 8353 __io_remove_buffers(ctx, buf, id, -1U);
5a2e745d
JA
8354 return 0;
8355}
8356
8357static void io_destroy_buffers(struct io_ring_ctx *ctx)
8358{
8359 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8360 idr_destroy(&ctx->io_buffer_idr);
8361}
8362
68e68ee6 8363static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
1b4c351f 8364{
68e68ee6 8365 struct io_kiocb *req, *nxt;
1b4c351f 8366
68e68ee6
JA
8367 list_for_each_entry_safe(req, nxt, list, compl.list) {
8368 if (tsk && req->task != tsk)
8369 continue;
1b4c351f
JA
8370 list_del(&req->compl.list);
8371 kmem_cache_free(req_cachep, req);
8372 }
8373}
8374
4010fec4 8375static void io_req_caches_free(struct io_ring_ctx *ctx)
2b188cc1 8376{
bf019da7 8377 struct io_submit_state *submit_state = &ctx->submit_state;
e5547d2c 8378 struct io_comp_state *cs = &ctx->submit_state.comp;
bf019da7 8379
9a4fdbd8
JA
8380 mutex_lock(&ctx->uring_lock);
8381
8e5c66c4 8382 if (submit_state->free_reqs) {
9a4fdbd8
JA
8383 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8384 submit_state->reqs);
8e5c66c4
PB
8385 submit_state->free_reqs = 0;
8386 }
9a4fdbd8
JA
8387
8388 spin_lock_irq(&ctx->completion_lock);
e5547d2c
PB
8389 list_splice_init(&cs->locked_free_list, &cs->free_list);
8390 cs->locked_free_nr = 0;
9a4fdbd8
JA
8391 spin_unlock_irq(&ctx->completion_lock);
8392
e5547d2c
PB
8393 io_req_cache_free(&cs->free_list, NULL);
8394
9a4fdbd8
JA
8395 mutex_unlock(&ctx->uring_lock);
8396}
8397
2b188cc1
JA
8398static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8399{
04fc6c80
PB
8400 /*
8401 * Some may use context even when all refs and requests have been put,
8402 * and they are free to do so while still holding uring_lock, see
8403 * __io_req_task_submit(). Wait for them to finish.
8404 */
8405 mutex_lock(&ctx->uring_lock);
8406 mutex_unlock(&ctx->uring_lock);
8407
37d1e2e3 8408 io_sq_thread_finish(ctx);
0a96bbe4 8409 io_sqe_buffers_unregister(ctx);
2aede0e4 8410
37d1e2e3 8411 if (ctx->mm_account) {
2aede0e4
JA
8412 mmdrop(ctx->mm_account);
8413 ctx->mm_account = NULL;
30975825 8414 }
def596e9 8415
8bad28d8 8416 mutex_lock(&ctx->uring_lock);
6b06314c 8417 io_sqe_files_unregister(ctx);
8bad28d8 8418 mutex_unlock(&ctx->uring_lock);
9b402849 8419 io_eventfd_unregister(ctx);
5a2e745d 8420 io_destroy_buffers(ctx);
41726c9a 8421 idr_destroy(&ctx->personality_idr);
def596e9 8422
2b188cc1 8423#if defined(CONFIG_UNIX)
355e8d26
EB
8424 if (ctx->ring_sock) {
8425 ctx->ring_sock->file = NULL; /* so that iput() is called */
2b188cc1 8426 sock_release(ctx->ring_sock);
355e8d26 8427 }
2b188cc1
JA
8428#endif
8429
75b28aff 8430 io_mem_free(ctx->rings);
2b188cc1 8431 io_mem_free(ctx->sq_sqes);
2b188cc1
JA
8432
8433 percpu_ref_exit(&ctx->refs);
2b188cc1 8434 free_uid(ctx->user);
4010fec4 8435 io_req_caches_free(ctx);
e941894e
JA
8436 if (ctx->hash_map)
8437 io_wq_put_hash(ctx->hash_map);
78076bb6 8438 kfree(ctx->cancel_hash);
2b188cc1
JA
8439 kfree(ctx);
8440}
8441
8442static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8443{
8444 struct io_ring_ctx *ctx = file->private_data;
8445 __poll_t mask = 0;
8446
8447 poll_wait(file, &ctx->cq_wait, wait);
4f7067c3
SB
8448 /*
8449 * synchronizes with barrier from wq_has_sleeper call in
8450 * io_commit_cqring
8451 */
2b188cc1 8452 smp_rmb();
90554200 8453 if (!io_sqring_full(ctx))
2b188cc1 8454 mask |= EPOLLOUT | EPOLLWRNORM;
ed670c3f
HX
8455
8456 /*
8457 * Don't flush cqring overflow list here, just do a simple check.
8458 * Otherwise there could possible be ABBA deadlock:
8459 * CPU0 CPU1
8460 * ---- ----
8461 * lock(&ctx->uring_lock);
8462 * lock(&ep->mtx);
8463 * lock(&ctx->uring_lock);
8464 * lock(&ep->mtx);
8465 *
8466 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8467 * pushs them to do the flush.
8468 */
8469 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
2b188cc1
JA
8470 mask |= EPOLLIN | EPOLLRDNORM;
8471
8472 return mask;
8473}
8474
8475static int io_uring_fasync(int fd, struct file *file, int on)
8476{
8477 struct io_ring_ctx *ctx = file->private_data;
8478
8479 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8480}
8481
0bead8cd 8482static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
071698e1 8483{
4379bf8b 8484 const struct cred *creds;
071698e1 8485
4379bf8b
JA
8486 creds = idr_remove(&ctx->personality_idr, id);
8487 if (creds) {
8488 put_cred(creds);
0bead8cd 8489 return 0;
1e6fa521 8490 }
0bead8cd
YD
8491
8492 return -EINVAL;
8493}
8494
8495static int io_remove_personalities(int id, void *p, void *data)
8496{
8497 struct io_ring_ctx *ctx = data;
8498
8499 io_unregister_personality(ctx, id);
071698e1
JA
8500 return 0;
8501}
8502
ba50a036 8503static bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
7c25c0d1 8504{
28c4721b 8505 struct callback_head *work, *next;
ba50a036 8506 bool executed = false;
7c25c0d1
JA
8507
8508 do {
28c4721b 8509 work = xchg(&ctx->exit_task_work, NULL);
7c25c0d1
JA
8510 if (!work)
8511 break;
8512
8513 do {
8514 next = work->next;
8515 work->func(work);
8516 work = next;
8517 cond_resched();
8518 } while (work);
ba50a036 8519 executed = true;
7c25c0d1 8520 } while (1);
ba50a036
PB
8521
8522 return executed;
7c25c0d1
JA
8523}
8524
85faa7b8
JA
8525static void io_ring_exit_work(struct work_struct *work)
8526{
b2edc0a7
PB
8527 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8528 exit_work);
85faa7b8 8529
56952e91
JA
8530 /*
8531 * If we're doing polled IO and end up having requests being
8532 * submitted async (out-of-line), then completions can come in while
8533 * we're waiting for refs to drop. We need to reap these manually,
8534 * as nobody else will be looking for them.
8535 */
b2edc0a7 8536 do {
9936c7c2 8537 io_uring_try_cancel_requests(ctx, NULL, NULL);
b2edc0a7 8538 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
85faa7b8
JA
8539 io_ring_ctx_free(ctx);
8540}
8541
2b188cc1
JA
8542static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8543{
8544 mutex_lock(&ctx->uring_lock);
8545 percpu_ref_kill(&ctx->refs);
cda286f0
PB
8546 /* if force is set, the ring is going away. always drop after that */
8547 ctx->cq_overflow_flushed = 1;
634578f8 8548 if (ctx->rings)
6c503150 8549 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
5c766a90 8550 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
2b188cc1
JA
8551 mutex_unlock(&ctx->uring_lock);
8552
6b81928d
PB
8553 io_kill_timeouts(ctx, NULL, NULL);
8554 io_poll_remove_all(ctx, NULL, NULL);
561fb04a 8555
15dff286 8556 /* if we failed setting up the ctx, we might not have any rings */
b2edc0a7 8557 io_iopoll_try_reap_events(ctx);
309fc03a 8558
85faa7b8 8559 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
fc666777
JA
8560 /*
8561 * Use system_unbound_wq to avoid spawning tons of event kworkers
8562 * if we're exiting a ton of rings at the same time. It just adds
8563 * noise and overhead, there's no discernable change in runtime
8564 * over using system_wq.
8565 */
8566 queue_work(system_unbound_wq, &ctx->exit_work);
2b188cc1
JA
8567}
8568
8569static int io_uring_release(struct inode *inode, struct file *file)
8570{
8571 struct io_ring_ctx *ctx = file->private_data;
8572
8573 file->private_data = NULL;
8574 io_ring_ctx_wait_and_kill(ctx);
8575 return 0;
8576}
8577
f6edbabb
PB
8578struct io_task_cancel {
8579 struct task_struct *task;
8580 struct files_struct *files;
8581};
f254ac04 8582
f6edbabb 8583static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
b711d4ea 8584{
9a472ef7 8585 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
f6edbabb 8586 struct io_task_cancel *cancel = data;
9a472ef7
PB
8587 bool ret;
8588
f6edbabb 8589 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
9a472ef7
PB
8590 unsigned long flags;
8591 struct io_ring_ctx *ctx = req->ctx;
8592
8593 /* protect against races with linked timeouts */
8594 spin_lock_irqsave(&ctx->completion_lock, flags);
f6edbabb 8595 ret = io_match_task(req, cancel->task, cancel->files);
9a472ef7
PB
8596 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8597 } else {
f6edbabb 8598 ret = io_match_task(req, cancel->task, cancel->files);
9a472ef7
PB
8599 }
8600 return ret;
b711d4ea
JA
8601}
8602
b7ddce3c 8603static void io_cancel_defer_files(struct io_ring_ctx *ctx,
ef9865a4 8604 struct task_struct *task,
b7ddce3c
PB
8605 struct files_struct *files)
8606{
8607 struct io_defer_entry *de = NULL;
8608 LIST_HEAD(list);
8609
8610 spin_lock_irq(&ctx->completion_lock);
8611 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
08d23634 8612 if (io_match_task(de->req, task, files)) {
b7ddce3c
PB
8613 list_cut_position(&list, &ctx->defer_list, &de->list);
8614 break;
8615 }
8616 }
8617 spin_unlock_irq(&ctx->completion_lock);
8618
8619 while (!list_empty(&list)) {
8620 de = list_first_entry(&list, struct io_defer_entry, list);
8621 list_del_init(&de->list);
8622 req_set_fail_links(de->req);
8623 io_put_req(de->req);
8624 io_req_complete(de->req, -ECANCELED);
8625 kfree(de);
8626 }
8627}
8628
9936c7c2
PB
8629static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8630 struct task_struct *task,
8631 struct files_struct *files)
8632{
8633 struct io_task_cancel cancel = { .task = task, .files = files, };
64c72123
PB
8634 struct task_struct *tctx_task = task ?: current;
8635 struct io_uring_task *tctx = tctx_task->io_uring;
9936c7c2
PB
8636
8637 while (1) {
8638 enum io_wq_cancel cret;
8639 bool ret = false;
8640
5aa75ed5
JA
8641 if (tctx && tctx->io_wq) {
8642 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9936c7c2
PB
8643 &cancel, true);
8644 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8645 }
8646
8647 /* SQPOLL thread does its own polling */
8648 if (!(ctx->flags & IORING_SETUP_SQPOLL) && !files) {
8649 while (!list_empty_careful(&ctx->iopoll_list)) {
8650 io_iopoll_try_reap_events(ctx);
8651 ret = true;
8652 }
8653 }
8654
8655 ret |= io_poll_remove_all(ctx, task, files);
8656 ret |= io_kill_timeouts(ctx, task, files);
8657 ret |= io_run_task_work();
ba50a036 8658 ret |= io_run_ctx_fallback(ctx);
9936c7c2
PB
8659 io_cqring_overflow_flush(ctx, true, task, files);
8660 if (!ret)
8661 break;
8662 cond_resched();
8663 }
8664}
8665
ca70f00b
PB
8666static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8667 struct task_struct *task,
8668 struct files_struct *files)
8669{
8670 struct io_kiocb *req;
8671 int cnt = 0;
8672
8673 spin_lock_irq(&ctx->inflight_lock);
8674 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8675 cnt += io_match_task(req, task, files);
8676 spin_unlock_irq(&ctx->inflight_lock);
8677 return cnt;
8678}
8679
b52fda00 8680static void io_uring_cancel_files(struct io_ring_ctx *ctx,
df9923f9 8681 struct task_struct *task,
fcb323cc
JA
8682 struct files_struct *files)
8683{
fcb323cc 8684 while (!list_empty_careful(&ctx->inflight_list)) {
d8f1b971 8685 DEFINE_WAIT(wait);
ca70f00b 8686 int inflight;
fcb323cc 8687
ca70f00b
PB
8688 inflight = io_uring_count_inflight(ctx, task, files);
8689 if (!inflight)
fcb323cc 8690 break;
f6edbabb 8691
9936c7c2 8692 io_uring_try_cancel_requests(ctx, task, files);
ca70f00b 8693
34343786
PB
8694 if (ctx->sq_data)
8695 io_sq_thread_unpark(ctx->sq_data);
ca70f00b
PB
8696 prepare_to_wait(&task->io_uring->wait, &wait,
8697 TASK_UNINTERRUPTIBLE);
8698 if (inflight == io_uring_count_inflight(ctx, task, files))
8699 schedule();
c98de08c 8700 finish_wait(&task->io_uring->wait, &wait);
34343786
PB
8701 if (ctx->sq_data)
8702 io_sq_thread_park(ctx->sq_data);
0f212204 8703 }
0f212204
JA
8704}
8705
8706/*
8707 * We need to iteratively cancel requests, in case a request has dependent
8708 * hard links. These persist even for failure of cancelations, hence keep
8709 * looping until none are found.
8710 */
8711static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8712 struct files_struct *files)
8713{
8714 struct task_struct *task = current;
37d1e2e3 8715 bool did_park = false;
0f212204 8716
fdaf083c 8717 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
70aacfe6
PB
8718 /* never started, nothing to cancel */
8719 if (ctx->flags & IORING_SETUP_R_DISABLED) {
8720 io_sq_offload_start(ctx);
8721 return;
8722 }
37d1e2e3
JA
8723 did_park = io_sq_thread_park(ctx->sq_data);
8724 if (did_park) {
8725 task = ctx->sq_data->thread;
8726 atomic_inc(&task->io_uring->in_idle);
8727 }
fdaf083c 8728 }
0f212204 8729
df9923f9 8730 io_cancel_defer_files(ctx, task, files);
0f212204 8731
3a7efd1a 8732 io_uring_cancel_files(ctx, task, files);
b52fda00 8733 if (!files)
9936c7c2 8734 io_uring_try_cancel_requests(ctx, task, NULL);
fdaf083c 8735
37d1e2e3 8736 if (did_park) {
fdaf083c 8737 atomic_dec(&task->io_uring->in_idle);
fdaf083c
JA
8738 io_sq_thread_unpark(ctx->sq_data);
8739 }
0f212204
JA
8740}
8741
8742/*
8743 * Note that this task has used io_uring. We use it for cancelation purposes.
8744 */
fdaf083c 8745static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
0f212204 8746{
236434c3 8747 struct io_uring_task *tctx = current->io_uring;
a528b04e 8748 int ret;
236434c3
MWO
8749
8750 if (unlikely(!tctx)) {
5aa75ed5 8751 ret = io_uring_alloc_task_context(current, ctx);
0f212204
JA
8752 if (unlikely(ret))
8753 return ret;
236434c3 8754 tctx = current->io_uring;
0f212204 8755 }
236434c3
MWO
8756 if (tctx->last != file) {
8757 void *old = xa_load(&tctx->xa, (unsigned long)file);
0f212204 8758
236434c3 8759 if (!old) {
0f212204 8760 get_file(file);
a528b04e
PB
8761 ret = xa_err(xa_store(&tctx->xa, (unsigned long)file,
8762 file, GFP_KERNEL));
8763 if (ret) {
8764 fput(file);
8765 return ret;
8766 }
0f212204 8767 }
236434c3 8768 tctx->last = file;
0f212204
JA
8769 }
8770
fdaf083c
JA
8771 /*
8772 * This is race safe in that the task itself is doing this, hence it
8773 * cannot be going through the exit/cancel paths at the same time.
8774 * This cannot be modified while exit/cancel is running.
8775 */
8776 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8777 tctx->sqpoll = true;
8778
0f212204
JA
8779 return 0;
8780}
8781
8782/*
8783 * Remove this io_uring_file -> task mapping.
8784 */
8785static void io_uring_del_task_file(struct file *file)
8786{
8787 struct io_uring_task *tctx = current->io_uring;
0f212204
JA
8788
8789 if (tctx->last == file)
8790 tctx->last = NULL;
5e2ed8c4 8791 file = xa_erase(&tctx->xa, (unsigned long)file);
0f212204
JA
8792 if (file)
8793 fput(file);
8794}
8795
8452d4a6 8796static void io_uring_clean_tctx(struct io_uring_task *tctx)
de7f1d9e
PB
8797{
8798 struct file *file;
8799 unsigned long index;
8800
8801 xa_for_each(&tctx->xa, index, file)
8802 io_uring_del_task_file(file);
8452d4a6
PB
8803 if (tctx->io_wq) {
8804 io_wq_put_and_exit(tctx->io_wq);
8805 tctx->io_wq = NULL;
8806 }
de7f1d9e
PB
8807}
8808
0f212204
JA
8809void __io_uring_files_cancel(struct files_struct *files)
8810{
8811 struct io_uring_task *tctx = current->io_uring;
ce765372
MWO
8812 struct file *file;
8813 unsigned long index;
0f212204
JA
8814
8815 /* make sure overflow events are dropped */
fdaf083c 8816 atomic_inc(&tctx->in_idle);
de7f1d9e
PB
8817 xa_for_each(&tctx->xa, index, file)
8818 io_uring_cancel_task_requests(file->private_data, files);
fdaf083c 8819 atomic_dec(&tctx->in_idle);
de7f1d9e 8820
8452d4a6
PB
8821 if (files)
8822 io_uring_clean_tctx(tctx);
fdaf083c
JA
8823}
8824
8825static s64 tctx_inflight(struct io_uring_task *tctx)
8826{
0e9ddb39
PB
8827 return percpu_counter_sum(&tctx->inflight);
8828}
fdaf083c 8829
0e9ddb39
PB
8830static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8831{
37d1e2e3 8832 struct io_sq_data *sqd = ctx->sq_data;
0e9ddb39
PB
8833 struct io_uring_task *tctx;
8834 s64 inflight;
8835 DEFINE_WAIT(wait);
fdaf083c 8836
37d1e2e3 8837 if (!sqd)
0e9ddb39 8838 return;
37d1e2e3
JA
8839 if (!io_sq_thread_park(sqd))
8840 return;
8841 tctx = ctx->sq_data->thread->io_uring;
e54945ae
JA
8842 /* can happen on fork/alloc failure, just ignore that state */
8843 if (!tctx) {
8844 io_sq_thread_unpark(sqd);
8845 return;
8846 }
fdaf083c 8847
0e9ddb39
PB
8848 atomic_inc(&tctx->in_idle);
8849 do {
8850 /* read completions before cancelations */
8851 inflight = tctx_inflight(tctx);
8852 if (!inflight)
8853 break;
8854 io_uring_cancel_task_requests(ctx, NULL);
fdaf083c 8855
0e9ddb39
PB
8856 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8857 /*
8858 * If we've seen completions, retry without waiting. This
8859 * avoids a race where a completion comes in before we did
8860 * prepare_to_wait().
8861 */
8862 if (inflight == tctx_inflight(tctx))
8863 schedule();
8864 finish_wait(&tctx->wait, &wait);
8865 } while (1);
8866 atomic_dec(&tctx->in_idle);
37d1e2e3 8867 io_sq_thread_unpark(sqd);
0f212204
JA
8868}
8869
0f212204
JA
8870/*
8871 * Find any io_uring fd that this task has registered or done IO on, and cancel
8872 * requests.
8873 */
8874void __io_uring_task_cancel(void)
8875{
8876 struct io_uring_task *tctx = current->io_uring;
8877 DEFINE_WAIT(wait);
d8a6df10 8878 s64 inflight;
0f212204
JA
8879
8880 /* make sure overflow events are dropped */
fdaf083c 8881 atomic_inc(&tctx->in_idle);
0f212204 8882
0e9ddb39
PB
8883 if (tctx->sqpoll) {
8884 struct file *file;
8885 unsigned long index;
8886
8887 xa_for_each(&tctx->xa, index, file)
8888 io_uring_cancel_sqpoll(file->private_data);
8889 }
0b5cd6c3 8890
d8a6df10 8891 do {
0f212204 8892 /* read completions before cancelations */
fdaf083c 8893 inflight = tctx_inflight(tctx);
d8a6df10
JA
8894 if (!inflight)
8895 break;
0f212204
JA
8896 __io_uring_files_cancel(NULL);
8897
8898 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8899
8900 /*
a1bb3cd5
PB
8901 * If we've seen completions, retry without waiting. This
8902 * avoids a race where a completion comes in before we did
8903 * prepare_to_wait().
0f212204 8904 */
a1bb3cd5
PB
8905 if (inflight == tctx_inflight(tctx))
8906 schedule();
f57555ed 8907 finish_wait(&tctx->wait, &wait);
d8a6df10 8908 } while (1);
0f212204 8909
fdaf083c 8910 atomic_dec(&tctx->in_idle);
de7f1d9e 8911
8452d4a6
PB
8912 io_uring_clean_tctx(tctx);
8913 /* all current's requests should be gone, we can kill tctx */
8914 __io_uring_free(current);
44e728b8
PB
8915}
8916
6c5c240e
RP
8917static void *io_uring_validate_mmap_request(struct file *file,
8918 loff_t pgoff, size_t sz)
2b188cc1 8919{
2b188cc1 8920 struct io_ring_ctx *ctx = file->private_data;
6c5c240e 8921 loff_t offset = pgoff << PAGE_SHIFT;
2b188cc1
JA
8922 struct page *page;
8923 void *ptr;
8924
8925 switch (offset) {
8926 case IORING_OFF_SQ_RING:
75b28aff
HV
8927 case IORING_OFF_CQ_RING:
8928 ptr = ctx->rings;
2b188cc1
JA
8929 break;
8930 case IORING_OFF_SQES:
8931 ptr = ctx->sq_sqes;
8932 break;
2b188cc1 8933 default:
6c5c240e 8934 return ERR_PTR(-EINVAL);
2b188cc1
JA
8935 }
8936
8937 page = virt_to_head_page(ptr);
a50b854e 8938 if (sz > page_size(page))
6c5c240e
RP
8939 return ERR_PTR(-EINVAL);
8940
8941 return ptr;
8942}
8943
8944#ifdef CONFIG_MMU
8945
8946static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8947{
8948 size_t sz = vma->vm_end - vma->vm_start;
8949 unsigned long pfn;
8950 void *ptr;
8951
8952 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8953 if (IS_ERR(ptr))
8954 return PTR_ERR(ptr);
2b188cc1
JA
8955
8956 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8957 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8958}
8959
6c5c240e
RP
8960#else /* !CONFIG_MMU */
8961
8962static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8963{
8964 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8965}
8966
8967static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8968{
8969 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8970}
8971
8972static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8973 unsigned long addr, unsigned long len,
8974 unsigned long pgoff, unsigned long flags)
8975{
8976 void *ptr;
8977
8978 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8979 if (IS_ERR(ptr))
8980 return PTR_ERR(ptr);
8981
8982 return (unsigned long) ptr;
8983}
8984
8985#endif /* !CONFIG_MMU */
8986
d9d05217 8987static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
90554200 8988{
d9d05217 8989 int ret = 0;
90554200
JA
8990 DEFINE_WAIT(wait);
8991
8992 do {
8993 if (!io_sqring_full(ctx))
8994 break;
90554200
JA
8995 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8996
8997 if (!io_sqring_full(ctx))
8998 break;
90554200
JA
8999 schedule();
9000 } while (!signal_pending(current));
9001
9002 finish_wait(&ctx->sqo_sq_wait, &wait);
d9d05217 9003 return ret;
90554200
JA
9004}
9005
c73ebb68
HX
9006static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9007 struct __kernel_timespec __user **ts,
9008 const sigset_t __user **sig)
9009{
9010 struct io_uring_getevents_arg arg;
9011
9012 /*
9013 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9014 * is just a pointer to the sigset_t.
9015 */
9016 if (!(flags & IORING_ENTER_EXT_ARG)) {
9017 *sig = (const sigset_t __user *) argp;
9018 *ts = NULL;
9019 return 0;
9020 }
9021
9022 /*
9023 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9024 * timespec and sigset_t pointers if good.
9025 */
9026 if (*argsz != sizeof(arg))
9027 return -EINVAL;
9028 if (copy_from_user(&arg, argp, sizeof(arg)))
9029 return -EFAULT;
9030 *sig = u64_to_user_ptr(arg.sigmask);
9031 *argsz = arg.sigmask_sz;
9032 *ts = u64_to_user_ptr(arg.ts);
9033 return 0;
9034}
9035
2b188cc1 9036SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
c73ebb68
HX
9037 u32, min_complete, u32, flags, const void __user *, argp,
9038 size_t, argsz)
2b188cc1
JA
9039{
9040 struct io_ring_ctx *ctx;
9041 long ret = -EBADF;
9042 int submitted = 0;
9043 struct fd f;
9044
4c6e277c 9045 io_run_task_work();
b41e9852 9046
90554200 9047 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
c73ebb68 9048 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
2b188cc1
JA
9049 return -EINVAL;
9050
9051 f = fdget(fd);
9052 if (!f.file)
9053 return -EBADF;
9054
9055 ret = -EOPNOTSUPP;
9056 if (f.file->f_op != &io_uring_fops)
9057 goto out_fput;
9058
9059 ret = -ENXIO;
9060 ctx = f.file->private_data;
9061 if (!percpu_ref_tryget(&ctx->refs))
9062 goto out_fput;
9063
7e84e1c7
SG
9064 ret = -EBADFD;
9065 if (ctx->flags & IORING_SETUP_R_DISABLED)
9066 goto out;
9067
6c271ce2
JA
9068 /*
9069 * For SQ polling, the thread will do all submissions and completions.
9070 * Just return the requested submit count, and wake the thread if
9071 * we were asked to.
9072 */
b2a9eada 9073 ret = 0;
6c271ce2 9074 if (ctx->flags & IORING_SETUP_SQPOLL) {
6c503150 9075 io_cqring_overflow_flush(ctx, false, NULL, NULL);
89448c47 9076
5f3f26f9
JA
9077 if (unlikely(ctx->sqo_exec)) {
9078 ret = io_sq_thread_fork(ctx->sq_data, ctx);
9079 if (ret)
9080 goto out;
9081 ctx->sqo_exec = 0;
9082 }
d9d05217 9083 ret = -EOWNERDEAD;
6c271ce2 9084 if (flags & IORING_ENTER_SQ_WAKEUP)
534ca6d6 9085 wake_up(&ctx->sq_data->wait);
d9d05217
PB
9086 if (flags & IORING_ENTER_SQ_WAIT) {
9087 ret = io_sqpoll_wait_sq(ctx);
9088 if (ret)
9089 goto out;
9090 }
6c271ce2 9091 submitted = to_submit;
b2a9eada 9092 } else if (to_submit) {
fdaf083c 9093 ret = io_uring_add_task_file(ctx, f.file);
0f212204
JA
9094 if (unlikely(ret))
9095 goto out;
2b188cc1 9096 mutex_lock(&ctx->uring_lock);
0f212204 9097 submitted = io_submit_sqes(ctx, to_submit);
2b188cc1 9098 mutex_unlock(&ctx->uring_lock);
7c504e65
PB
9099
9100 if (submitted != to_submit)
9101 goto out;
2b188cc1
JA
9102 }
9103 if (flags & IORING_ENTER_GETEVENTS) {
c73ebb68
HX
9104 const sigset_t __user *sig;
9105 struct __kernel_timespec __user *ts;
9106
9107 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9108 if (unlikely(ret))
9109 goto out;
9110
2b188cc1
JA
9111 min_complete = min(min_complete, ctx->cq_entries);
9112
32b2244a
XW
9113 /*
9114 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9115 * space applications don't need to do io completion events
9116 * polling again, they can rely on io_sq_thread to do polling
9117 * work, which can reduce cpu usage and uring_lock contention.
9118 */
9119 if (ctx->flags & IORING_SETUP_IOPOLL &&
9120 !(ctx->flags & IORING_SETUP_SQPOLL)) {
7668b92a 9121 ret = io_iopoll_check(ctx, min_complete);
def596e9 9122 } else {
c73ebb68 9123 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
def596e9 9124 }
2b188cc1
JA
9125 }
9126
7c504e65 9127out:
6805b32e 9128 percpu_ref_put(&ctx->refs);
2b188cc1
JA
9129out_fput:
9130 fdput(f);
9131 return submitted ? submitted : ret;
9132}
9133
bebdb65e 9134#ifdef CONFIG_PROC_FS
87ce955b
JA
9135static int io_uring_show_cred(int id, void *p, void *data)
9136{
4379bf8b 9137 const struct cred *cred = p;
87ce955b
JA
9138 struct seq_file *m = data;
9139 struct user_namespace *uns = seq_user_ns(m);
9140 struct group_info *gi;
9141 kernel_cap_t cap;
9142 unsigned __capi;
9143 int g;
9144
9145 seq_printf(m, "%5d\n", id);
9146 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9147 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9148 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9149 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9150 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9151 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9152 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9153 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9154 seq_puts(m, "\n\tGroups:\t");
9155 gi = cred->group_info;
9156 for (g = 0; g < gi->ngroups; g++) {
9157 seq_put_decimal_ull(m, g ? " " : "",
9158 from_kgid_munged(uns, gi->gid[g]));
9159 }
9160 seq_puts(m, "\n\tCapEff:\t");
9161 cap = cred->cap_effective;
9162 CAP_FOR_EACH_U32(__capi)
9163 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9164 seq_putc(m, '\n');
9165 return 0;
9166}
9167
9168static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9169{
dbbe9c64 9170 struct io_sq_data *sq = NULL;
fad8e0de 9171 bool has_lock;
87ce955b
JA
9172 int i;
9173
fad8e0de
JA
9174 /*
9175 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9176 * since fdinfo case grabs it in the opposite direction of normal use
9177 * cases. If we fail to get the lock, we just don't iterate any
9178 * structures that could be going away outside the io_uring mutex.
9179 */
9180 has_lock = mutex_trylock(&ctx->uring_lock);
9181
5f3f26f9 9182 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
dbbe9c64 9183 sq = ctx->sq_data;
5f3f26f9
JA
9184 if (!sq->thread)
9185 sq = NULL;
9186 }
dbbe9c64
JQ
9187
9188 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9189 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
87ce955b 9190 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
fad8e0de 9191 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
ea64ec02 9192 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
87ce955b 9193
87ce955b
JA
9194 if (f)
9195 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9196 else
9197 seq_printf(m, "%5u: <none>\n", i);
9198 }
9199 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
fad8e0de 9200 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
87ce955b
JA
9201 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9202
9203 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9204 (unsigned int) buf->len);
9205 }
fad8e0de 9206 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
87ce955b
JA
9207 seq_printf(m, "Personalities:\n");
9208 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9209 }
d7718a9d
JA
9210 seq_printf(m, "PollList:\n");
9211 spin_lock_irq(&ctx->completion_lock);
9212 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9213 struct hlist_head *list = &ctx->cancel_hash[i];
9214 struct io_kiocb *req;
9215
9216 hlist_for_each_entry(req, list, hash_node)
9217 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9218 req->task->task_works != NULL);
9219 }
9220 spin_unlock_irq(&ctx->completion_lock);
fad8e0de
JA
9221 if (has_lock)
9222 mutex_unlock(&ctx->uring_lock);
87ce955b
JA
9223}
9224
9225static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9226{
9227 struct io_ring_ctx *ctx = f->private_data;
9228
9229 if (percpu_ref_tryget(&ctx->refs)) {
9230 __io_uring_show_fdinfo(ctx, m);
9231 percpu_ref_put(&ctx->refs);
9232 }
9233}
bebdb65e 9234#endif
87ce955b 9235
2b188cc1
JA
9236static const struct file_operations io_uring_fops = {
9237 .release = io_uring_release,
9238 .mmap = io_uring_mmap,
6c5c240e
RP
9239#ifndef CONFIG_MMU
9240 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9241 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9242#endif
2b188cc1
JA
9243 .poll = io_uring_poll,
9244 .fasync = io_uring_fasync,
bebdb65e 9245#ifdef CONFIG_PROC_FS
87ce955b 9246 .show_fdinfo = io_uring_show_fdinfo,
bebdb65e 9247#endif
2b188cc1
JA
9248};
9249
9250static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9251 struct io_uring_params *p)
9252{
75b28aff
HV
9253 struct io_rings *rings;
9254 size_t size, sq_array_offset;
2b188cc1 9255
bd740481
JA
9256 /* make sure these are sane, as we already accounted them */
9257 ctx->sq_entries = p->sq_entries;
9258 ctx->cq_entries = p->cq_entries;
9259
75b28aff
HV
9260 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9261 if (size == SIZE_MAX)
9262 return -EOVERFLOW;
9263
9264 rings = io_mem_alloc(size);
9265 if (!rings)
2b188cc1
JA
9266 return -ENOMEM;
9267
75b28aff
HV
9268 ctx->rings = rings;
9269 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9270 rings->sq_ring_mask = p->sq_entries - 1;
9271 rings->cq_ring_mask = p->cq_entries - 1;
9272 rings->sq_ring_entries = p->sq_entries;
9273 rings->cq_ring_entries = p->cq_entries;
9274 ctx->sq_mask = rings->sq_ring_mask;
9275 ctx->cq_mask = rings->cq_ring_mask;
2b188cc1
JA
9276
9277 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
eb065d30
JA
9278 if (size == SIZE_MAX) {
9279 io_mem_free(ctx->rings);
9280 ctx->rings = NULL;
2b188cc1 9281 return -EOVERFLOW;
eb065d30 9282 }
2b188cc1
JA
9283
9284 ctx->sq_sqes = io_mem_alloc(size);
eb065d30
JA
9285 if (!ctx->sq_sqes) {
9286 io_mem_free(ctx->rings);
9287 ctx->rings = NULL;
2b188cc1 9288 return -ENOMEM;
eb065d30 9289 }
2b188cc1 9290
2b188cc1
JA
9291 return 0;
9292}
9293
9faadcc8
PB
9294static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9295{
9296 int ret, fd;
9297
9298 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9299 if (fd < 0)
9300 return fd;
9301
9302 ret = io_uring_add_task_file(ctx, file);
9303 if (ret) {
9304 put_unused_fd(fd);
9305 return ret;
9306 }
9307 fd_install(fd, file);
9308 return fd;
9309}
9310
2b188cc1
JA
9311/*
9312 * Allocate an anonymous fd, this is what constitutes the application
9313 * visible backing of an io_uring instance. The application mmaps this
9314 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9315 * we have to tie this fd to a socket for file garbage collection purposes.
9316 */
9faadcc8 9317static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
2b188cc1
JA
9318{
9319 struct file *file;
9faadcc8 9320#if defined(CONFIG_UNIX)
2b188cc1
JA
9321 int ret;
9322
2b188cc1
JA
9323 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9324 &ctx->ring_sock);
9325 if (ret)
9faadcc8 9326 return ERR_PTR(ret);
2b188cc1
JA
9327#endif
9328
2b188cc1
JA
9329 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9330 O_RDWR | O_CLOEXEC);
2b188cc1 9331#if defined(CONFIG_UNIX)
9faadcc8
PB
9332 if (IS_ERR(file)) {
9333 sock_release(ctx->ring_sock);
9334 ctx->ring_sock = NULL;
9335 } else {
9336 ctx->ring_sock->file = file;
0f212204 9337 }
2b188cc1 9338#endif
9faadcc8 9339 return file;
2b188cc1
JA
9340}
9341
7f13657d
XW
9342static int io_uring_create(unsigned entries, struct io_uring_params *p,
9343 struct io_uring_params __user *params)
2b188cc1 9344{
2b188cc1 9345 struct io_ring_ctx *ctx;
9faadcc8 9346 struct file *file;
2b188cc1
JA
9347 int ret;
9348
8110c1a6 9349 if (!entries)
2b188cc1 9350 return -EINVAL;
8110c1a6
JA
9351 if (entries > IORING_MAX_ENTRIES) {
9352 if (!(p->flags & IORING_SETUP_CLAMP))
9353 return -EINVAL;
9354 entries = IORING_MAX_ENTRIES;
9355 }
2b188cc1
JA
9356
9357 /*
9358 * Use twice as many entries for the CQ ring. It's possible for the
9359 * application to drive a higher depth than the size of the SQ ring,
9360 * since the sqes are only used at submission time. This allows for
33a107f0
JA
9361 * some flexibility in overcommitting a bit. If the application has
9362 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9363 * of CQ ring entries manually.
2b188cc1
JA
9364 */
9365 p->sq_entries = roundup_pow_of_two(entries);
33a107f0
JA
9366 if (p->flags & IORING_SETUP_CQSIZE) {
9367 /*
9368 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9369 * to a power-of-two, if it isn't already. We do NOT impose
9370 * any cq vs sq ring sizing.
9371 */
eb2667b3 9372 if (!p->cq_entries)
33a107f0 9373 return -EINVAL;
8110c1a6
JA
9374 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9375 if (!(p->flags & IORING_SETUP_CLAMP))
9376 return -EINVAL;
9377 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9378 }
eb2667b3
JQ
9379 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9380 if (p->cq_entries < p->sq_entries)
9381 return -EINVAL;
33a107f0
JA
9382 } else {
9383 p->cq_entries = 2 * p->sq_entries;
9384 }
2b188cc1 9385
2b188cc1 9386 ctx = io_ring_ctx_alloc(p);
62e398be 9387 if (!ctx)
2b188cc1 9388 return -ENOMEM;
2b188cc1 9389 ctx->compat = in_compat_syscall();
62e398be
JA
9390 if (!capable(CAP_IPC_LOCK))
9391 ctx->user = get_uid(current_user());
2aede0e4
JA
9392
9393 /*
9394 * This is just grabbed for accounting purposes. When a process exits,
9395 * the mm is exited and dropped before the files, hence we need to hang
9396 * on to this mm purely for the purposes of being able to unaccount
9397 * memory (locked/pinned vm). It's not used for anything else.
9398 */
6b7898eb 9399 mmgrab(current->mm);
2aede0e4 9400 ctx->mm_account = current->mm;
6b7898eb 9401
2b188cc1
JA
9402 ret = io_allocate_scq_urings(ctx, p);
9403 if (ret)
9404 goto err;
9405
7e84e1c7 9406 ret = io_sq_offload_create(ctx, p);
2b188cc1
JA
9407 if (ret)
9408 goto err;
9409
7e84e1c7
SG
9410 if (!(p->flags & IORING_SETUP_R_DISABLED))
9411 io_sq_offload_start(ctx);
9412
2b188cc1 9413 memset(&p->sq_off, 0, sizeof(p->sq_off));
75b28aff
HV
9414 p->sq_off.head = offsetof(struct io_rings, sq.head);
9415 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9416 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9417 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9418 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9419 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9420 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
2b188cc1
JA
9421
9422 memset(&p->cq_off, 0, sizeof(p->cq_off));
75b28aff
HV
9423 p->cq_off.head = offsetof(struct io_rings, cq.head);
9424 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9425 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9426 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9427 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9428 p->cq_off.cqes = offsetof(struct io_rings, cqes);
0d9b5b3a 9429 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
ac90f249 9430
7f13657d
XW
9431 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9432 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
5769a351 9433 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
c73ebb68 9434 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
1c0aa1fa 9435 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
7f13657d
XW
9436
9437 if (copy_to_user(params, p, sizeof(*p))) {
9438 ret = -EFAULT;
9439 goto err;
9440 }
d1719f70 9441
9faadcc8
PB
9442 file = io_uring_get_file(ctx);
9443 if (IS_ERR(file)) {
9444 ret = PTR_ERR(file);
9445 goto err;
9446 }
9447
044c1ab3
JA
9448 /*
9449 * Install ring fd as the very last thing, so we don't risk someone
9450 * having closed it before we finish setup
9451 */
9faadcc8
PB
9452 ret = io_uring_install_fd(ctx, file);
9453 if (ret < 0) {
9454 /* fput will clean it up */
9455 fput(file);
9456 return ret;
9457 }
044c1ab3 9458
c826bd7a 9459 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
2b188cc1
JA
9460 return ret;
9461err:
9462 io_ring_ctx_wait_and_kill(ctx);
9463 return ret;
9464}
9465
9466/*
9467 * Sets up an aio uring context, and returns the fd. Applications asks for a
9468 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9469 * params structure passed in.
9470 */
9471static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9472{
9473 struct io_uring_params p;
2b188cc1
JA
9474 int i;
9475
9476 if (copy_from_user(&p, params, sizeof(p)))
9477 return -EFAULT;
9478 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9479 if (p.resv[i])
9480 return -EINVAL;
9481 }
9482
6c271ce2 9483 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8110c1a6 9484 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
7e84e1c7
SG
9485 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9486 IORING_SETUP_R_DISABLED))
2b188cc1
JA
9487 return -EINVAL;
9488
7f13657d 9489 return io_uring_create(entries, &p, params);
2b188cc1
JA
9490}
9491
9492SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9493 struct io_uring_params __user *, params)
9494{
9495 return io_uring_setup(entries, params);
9496}
9497
66f4af93
JA
9498static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9499{
9500 struct io_uring_probe *p;
9501 size_t size;
9502 int i, ret;
9503
9504 size = struct_size(p, ops, nr_args);
9505 if (size == SIZE_MAX)
9506 return -EOVERFLOW;
9507 p = kzalloc(size, GFP_KERNEL);
9508 if (!p)
9509 return -ENOMEM;
9510
9511 ret = -EFAULT;
9512 if (copy_from_user(p, arg, size))
9513 goto out;
9514 ret = -EINVAL;
9515 if (memchr_inv(p, 0, size))
9516 goto out;
9517
9518 p->last_op = IORING_OP_LAST - 1;
9519 if (nr_args > IORING_OP_LAST)
9520 nr_args = IORING_OP_LAST;
9521
9522 for (i = 0; i < nr_args; i++) {
9523 p->ops[i].op = i;
9524 if (!io_op_defs[i].not_supported)
9525 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9526 }
9527 p->ops_len = i;
9528
9529 ret = 0;
9530 if (copy_to_user(arg, p, size))
9531 ret = -EFAULT;
9532out:
9533 kfree(p);
9534 return ret;
9535}
9536
071698e1
JA
9537static int io_register_personality(struct io_ring_ctx *ctx)
9538{
4379bf8b 9539 const struct cred *creds;
1e6fa521 9540 int ret;
071698e1 9541
4379bf8b 9542 creds = get_current_cred();
1e6fa521 9543
4379bf8b
JA
9544 ret = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
9545 USHRT_MAX, GFP_KERNEL);
9546 if (ret < 0)
9547 put_cred(creds);
1e6fa521 9548 return ret;
071698e1
JA
9549}
9550
21b55dbc
SG
9551static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9552 unsigned int nr_args)
9553{
9554 struct io_uring_restriction *res;
9555 size_t size;
9556 int i, ret;
9557
7e84e1c7
SG
9558 /* Restrictions allowed only if rings started disabled */
9559 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9560 return -EBADFD;
9561
21b55dbc 9562 /* We allow only a single restrictions registration */
7e84e1c7 9563 if (ctx->restrictions.registered)
21b55dbc
SG
9564 return -EBUSY;
9565
9566 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9567 return -EINVAL;
9568
9569 size = array_size(nr_args, sizeof(*res));
9570 if (size == SIZE_MAX)
9571 return -EOVERFLOW;
9572
9573 res = memdup_user(arg, size);
9574 if (IS_ERR(res))
9575 return PTR_ERR(res);
9576
9577 ret = 0;
9578
9579 for (i = 0; i < nr_args; i++) {
9580 switch (res[i].opcode) {
9581 case IORING_RESTRICTION_REGISTER_OP:
9582 if (res[i].register_op >= IORING_REGISTER_LAST) {
9583 ret = -EINVAL;
9584 goto out;
9585 }
9586
9587 __set_bit(res[i].register_op,
9588 ctx->restrictions.register_op);
9589 break;
9590 case IORING_RESTRICTION_SQE_OP:
9591 if (res[i].sqe_op >= IORING_OP_LAST) {
9592 ret = -EINVAL;
9593 goto out;
9594 }
9595
9596 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9597 break;
9598 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9599 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9600 break;
9601 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9602 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9603 break;
9604 default:
9605 ret = -EINVAL;
9606 goto out;
9607 }
9608 }
9609
9610out:
9611 /* Reset all restrictions if an error happened */
9612 if (ret != 0)
9613 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9614 else
7e84e1c7 9615 ctx->restrictions.registered = true;
21b55dbc
SG
9616
9617 kfree(res);
9618 return ret;
9619}
9620
7e84e1c7
SG
9621static int io_register_enable_rings(struct io_ring_ctx *ctx)
9622{
9623 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9624 return -EBADFD;
9625
9626 if (ctx->restrictions.registered)
9627 ctx->restricted = 1;
9628
7e84e1c7 9629 io_sq_offload_start(ctx);
7e84e1c7
SG
9630 return 0;
9631}
9632
071698e1
JA
9633static bool io_register_op_must_quiesce(int op)
9634{
9635 switch (op) {
9636 case IORING_UNREGISTER_FILES:
9637 case IORING_REGISTER_FILES_UPDATE:
9638 case IORING_REGISTER_PROBE:
9639 case IORING_REGISTER_PERSONALITY:
9640 case IORING_UNREGISTER_PERSONALITY:
9641 return false;
9642 default:
9643 return true;
9644 }
9645}
9646
edafccee
JA
9647static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9648 void __user *arg, unsigned nr_args)
b19062a5
JA
9649 __releases(ctx->uring_lock)
9650 __acquires(ctx->uring_lock)
edafccee
JA
9651{
9652 int ret;
9653
35fa71a0
JA
9654 /*
9655 * We're inside the ring mutex, if the ref is already dying, then
9656 * someone else killed the ctx or is already going through
9657 * io_uring_register().
9658 */
9659 if (percpu_ref_is_dying(&ctx->refs))
9660 return -ENXIO;
9661
071698e1 9662 if (io_register_op_must_quiesce(opcode)) {
05f3fb3c 9663 percpu_ref_kill(&ctx->refs);
b19062a5 9664
05f3fb3c
JA
9665 /*
9666 * Drop uring mutex before waiting for references to exit. If
9667 * another thread is currently inside io_uring_enter() it might
9668 * need to grab the uring_lock to make progress. If we hold it
9669 * here across the drain wait, then we can deadlock. It's safe
9670 * to drop the mutex here, since no new references will come in
9671 * after we've killed the percpu ref.
9672 */
9673 mutex_unlock(&ctx->uring_lock);
af9c1a44
JA
9674 do {
9675 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9676 if (!ret)
9677 break;
ed6930c9
JA
9678 ret = io_run_task_work_sig();
9679 if (ret < 0)
9680 break;
af9c1a44
JA
9681 } while (1);
9682
05f3fb3c 9683 mutex_lock(&ctx->uring_lock);
af9c1a44 9684
c150368b
JA
9685 if (ret) {
9686 percpu_ref_resurrect(&ctx->refs);
21b55dbc
SG
9687 goto out_quiesce;
9688 }
9689 }
9690
9691 if (ctx->restricted) {
9692 if (opcode >= IORING_REGISTER_LAST) {
9693 ret = -EINVAL;
9694 goto out;
9695 }
9696
9697 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9698 ret = -EACCES;
c150368b
JA
9699 goto out;
9700 }
05f3fb3c 9701 }
edafccee
JA
9702
9703 switch (opcode) {
9704 case IORING_REGISTER_BUFFERS:
0a96bbe4 9705 ret = io_sqe_buffers_register(ctx, arg, nr_args);
edafccee
JA
9706 break;
9707 case IORING_UNREGISTER_BUFFERS:
9708 ret = -EINVAL;
9709 if (arg || nr_args)
9710 break;
0a96bbe4 9711 ret = io_sqe_buffers_unregister(ctx);
edafccee 9712 break;
6b06314c
JA
9713 case IORING_REGISTER_FILES:
9714 ret = io_sqe_files_register(ctx, arg, nr_args);
9715 break;
9716 case IORING_UNREGISTER_FILES:
9717 ret = -EINVAL;
9718 if (arg || nr_args)
9719 break;
9720 ret = io_sqe_files_unregister(ctx);
9721 break;
c3a31e60
JA
9722 case IORING_REGISTER_FILES_UPDATE:
9723 ret = io_sqe_files_update(ctx, arg, nr_args);
9724 break;
9b402849 9725 case IORING_REGISTER_EVENTFD:
f2842ab5 9726 case IORING_REGISTER_EVENTFD_ASYNC:
9b402849
JA
9727 ret = -EINVAL;
9728 if (nr_args != 1)
9729 break;
9730 ret = io_eventfd_register(ctx, arg);
f2842ab5
JA
9731 if (ret)
9732 break;
9733 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9734 ctx->eventfd_async = 1;
9735 else
9736 ctx->eventfd_async = 0;
9b402849
JA
9737 break;
9738 case IORING_UNREGISTER_EVENTFD:
9739 ret = -EINVAL;
9740 if (arg || nr_args)
9741 break;
9742 ret = io_eventfd_unregister(ctx);
9743 break;
66f4af93
JA
9744 case IORING_REGISTER_PROBE:
9745 ret = -EINVAL;
9746 if (!arg || nr_args > 256)
9747 break;
9748 ret = io_probe(ctx, arg, nr_args);
9749 break;
071698e1
JA
9750 case IORING_REGISTER_PERSONALITY:
9751 ret = -EINVAL;
9752 if (arg || nr_args)
9753 break;
9754 ret = io_register_personality(ctx);
9755 break;
9756 case IORING_UNREGISTER_PERSONALITY:
9757 ret = -EINVAL;
9758 if (arg)
9759 break;
9760 ret = io_unregister_personality(ctx, nr_args);
9761 break;
7e84e1c7
SG
9762 case IORING_REGISTER_ENABLE_RINGS:
9763 ret = -EINVAL;
9764 if (arg || nr_args)
9765 break;
9766 ret = io_register_enable_rings(ctx);
9767 break;
21b55dbc
SG
9768 case IORING_REGISTER_RESTRICTIONS:
9769 ret = io_register_restrictions(ctx, arg, nr_args);
9770 break;
edafccee
JA
9771 default:
9772 ret = -EINVAL;
9773 break;
9774 }
9775
21b55dbc 9776out:
071698e1 9777 if (io_register_op_must_quiesce(opcode)) {
05f3fb3c 9778 /* bring the ctx back to life */
05f3fb3c 9779 percpu_ref_reinit(&ctx->refs);
21b55dbc 9780out_quiesce:
0f158b4c 9781 reinit_completion(&ctx->ref_comp);
05f3fb3c 9782 }
edafccee
JA
9783 return ret;
9784}
9785
9786SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9787 void __user *, arg, unsigned int, nr_args)
9788{
9789 struct io_ring_ctx *ctx;
9790 long ret = -EBADF;
9791 struct fd f;
9792
9793 f = fdget(fd);
9794 if (!f.file)
9795 return -EBADF;
9796
9797 ret = -EOPNOTSUPP;
9798 if (f.file->f_op != &io_uring_fops)
9799 goto out_fput;
9800
9801 ctx = f.file->private_data;
9802
b6c23dd5
PB
9803 io_run_task_work();
9804
edafccee
JA
9805 mutex_lock(&ctx->uring_lock);
9806 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9807 mutex_unlock(&ctx->uring_lock);
c826bd7a
DD
9808 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9809 ctx->cq_ev_fd != NULL, ret);
edafccee
JA
9810out_fput:
9811 fdput(f);
9812 return ret;
9813}
9814
2b188cc1
JA
9815static int __init io_uring_init(void)
9816{
d7f62e82
SM
9817#define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9818 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9819 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9820} while (0)
9821
9822#define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9823 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9824 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9825 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9826 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9827 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9828 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9829 BUILD_BUG_SQE_ELEM(8, __u64, off);
9830 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9831 BUILD_BUG_SQE_ELEM(16, __u64, addr);
7d67af2c 9832 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
d7f62e82
SM
9833 BUILD_BUG_SQE_ELEM(24, __u32, len);
9834 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9835 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9836 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9837 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
5769a351
JX
9838 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9839 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
d7f62e82
SM
9840 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9841 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9842 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9843 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9844 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9845 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9846 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9847 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
7d67af2c 9848 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
d7f62e82
SM
9849 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9850 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9851 BUILD_BUG_SQE_ELEM(42, __u16, personality);
7d67af2c 9852 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
d7f62e82 9853
d3656344 9854 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
84557871 9855 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
91f245d5
JA
9856 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9857 SLAB_ACCOUNT);
2b188cc1
JA
9858 return 0;
9859};
9860__initcall(io_uring_init);